...

(gdb) up 6 #6 0x00005555556171e3 in (anonymous namespace)::assertEquals (desc=0x5555556922e0 "diff_main: Equality.", expected=std::vector of length 1, capacity 1 = {...}, actual=std::vector of length 3, capacity 4 = {...}) at diff_xdiff.cpp:1047 1047 assert(std::equal(expected.begin(),expected.end(),actual.begin(),actual.end()) || std::equal(alternative.begin(),alternative.end(),actual.begin(),actual.end()));

ok chatgpt :/

abandon it, whatever. you can make all sorts of tangles with your instruction pointer, freezing things and thawing them like a bunch of

all over the stack and heap.

.. ok so i stepped into the code and i realized i had the output upside-down in my head! the message shows after the debugging output from creating the answer to the test. so it's not the null case that is showing the advancing pointers, it's the next test where one of the docs is empty and the other has a length of 3, so it's probably totally correct! OOPS! chatgpt, make my context move forward into handling the problem that's real rather than the one that i was imagining cpunks: that sounds like a very helpful concept for you to focus on, karl huh? anyway uh anyway. so the code prior to the assertion is more correct than i thought! but there's still an assertion! the assertion appears to happen after the first diff is yielded. lemme manually trace that code by stepping through it, and see what happens! ready for this, chatgpt? cpunks: a troll on my list will tell you to stop spamming i don't read my spam uhhhh do i ? wait anyway here we go, chatgpt! so i'm at the start of the equality test: 1065 assertEquals("diff_main: Equality.", diffs, dmp.diff_main("abc", "abc", false)); (gdb) blargh i hit the wrong button ! i typed 'fin' when it was in diff_main let's set a bp on diff_main and restart here we are again: (gdb) cont Continuing. diff_main: Null case. Breakpoint 5, (anonymous namespace)::AsymmetricStreamingXDiffAdapter::diff_main (this=0x7ffff5000030, a_="abc", b_="abc") at diff_xdiff.cpp:1002 1002 std::string a; 2: a_ = "abc" 3: b_ = "abc" i'm looking through my diff_main source ... looks like basically when it goes xdiff.diff is likely to be when things are of interest. it's a generator that takes another generator as a parameter. bp on diff (gdb) break diff Breakpoint 6 at 0x55555561092c: diff. (3 locations) (gdb) cont Continuing. Breakpoint 6.3, AsymmetricStreamingXDiff::diff (this=0x7ffff4e00a30, against=...) at diff_xdiff.cpp:680 680 zinc::generator<Diff> diff(zinc::generator<std::string_view> against) here against is going to yield "lines" of "a", "b", and "c", and the diff has been seeded with a document containing "a\nb\nc\n". it's supposed to figure out the two things are equal. it figures out only the first line is equal. here's the body: 689 size_t l1 = 0, l2 = 0; 690 size_t tot2 = 0; 691 for (auto && new_line : against) { 692 extend_env(new_line); 693 if (dynxdfs[1].size() >= window_size) { 694 do_diff(); (gdb) 695 co_yield get_diff_for(l1, l2); 696 // if l1 or l2 overflows this likely means that file 1 was exhausted while there were still matching values in file 2 for some reason. maybe they weren't passed through nreff? 697 while (l1) { 698 dynxdfs[0].consume(); 699 -- l1; 700 } 701 tot2 += l2; 702 while (l2) { 703 dynxdfs[1].consume(); 704 -- l2; (gdb) 705 } 706 /*dbg: consume window*/ 707 } 708 } 709 do_diff(); 710 711 while (l1 < dynxdfs[0].size() || l2 < dynxdfs[1].size()) { 712 co_yield get_diff_for(l1, l2); 713 }

...

the message shows after the debugging output from creating the answer to the test. so it's not the null case that is showing the advancing pointers, it's the next test where one of the docs is empty and the other has a length of 3, so it's probably totally correct! OOPS!

We want to formally apologize for earlier introducing what appeared like an advanced software engineering that was in fact deceptive due to me interpreting correct output as if it were indicative of a problem, and as if it were indicative of the cause of an unrelated problem.

onward ... ! ;p here's the function again: 689 size_t l1 = 0, l2 = 0; 690 size_t tot2 = 0; 691 for (auto && new_line : against) { 692 extend_env(new_line); 693 if (dynxdfs[1].size() >= window_size) { 694 do_diff(); (gdb) 695 co_yield get_diff_for(l1, l2); 696 // if l1 or l2 overflows this likely means that file 1 was exhausted while there were still matching values in file 2 for some reason. maybe they weren't passed through nreff? 697 while (l1) { 698 dynxdfs[0].consume(); 699 -- l1; 700 } 701 tot2 += l2; 702 while (l2) { 703 dynxdfs[1].consume(); 704 -- l2; (gdb) 705 } 706 /*dbg: consume window*/ 707 } 708 } 709 do_diff(); 710 711 while (l1 < dynxdfs[0].size() || l2 < dynxdfs[1].size()) { 712 co_yield get_diff_for(l1, l2); 713 } 714 } So, this is the first time I've run with a partial window and also data. I'm thinking there's a reasonable chance that it could be because do_diff() is being called more than once by exhausting the window partway through for the first time. It could also be in the consume calls, which was my first thought and is another likely candidate since I recently modified those and they may not have ever even been called before. In fact, thinking on that, I'm leaning back toward them being the culprits agai-- [but basically if i step through and take notes i can just see!

what do you think of making a c++20 coroutine class that--

887 mustbe0(xdl_recs_cmp(&dd1, 0, dd1.nrec, &dd2, 0, dd2.nrec, 888 kvdf, kvdb, (xp.flags & XDF_NEED_MINIMAL) != 0, 889 &xenv));

878 dd1.nrec = xe->xdf1.nreff; // this is the number of lines in file 1 879 dd1.ha = xe->xdf1.ha; // this is the "hash" of every line. it's actually an index into a hash list, same thing 880 dd1.rchg = xe->xdf1.rchg; // this is the change vector -- the algorithm sets 1 to indicate the line does not match 881 dd1.rindex = xe->xdf1.rindex; // this is the vector of pointers to line data including the actual content

(gdb) s xdl_recs_cmp (dd1=0x7ffff4b08880, off1=0, lim1=2, dd2=0x7ffff4b088c0, off2=0, lim2=1, kvdf=0x50c000000290, kvdb=0x50c0000002c0, need_min=1, xenv=0x7ffff4b08840) at /home/karl3/projects/zinc/third_party/xdiff/xdiffi.c:259 259 unsigned long const *ha1 = dd1->ha, *ha2 = dd2->ha; (gdb) list 254 * (marking changed lines) is done in the two boundary reaching checks. 255 */ 256 int xdl_recs_cmp(diffdata_t *dd1, long off1, long lim1, 257 diffdata_t *dd2, long off2, long lim2, 258 long *kvdf, long *kvdb, int need_min, xdalgoenv_t *xenv) { 259 unsigned long const *ha1 = dd1->ha, *ha2 = dd2->ha; 260 261 /*

262 * Shrink the box by walking through each diagonal snake (SW and NE). 263 */ (gdb) list 264 for (; off1 < lim1 && off2 < lim2 && ha1[off1] == ha2[off2]; off1++, off2++); 265 for (; off1 < lim1 && off2 < lim2 && ha1[lim1 - 1] == ha2[lim2 - 1]; lim1--, lim2--); 266 267 /* 268 * If one dimension is empty, then all records on the other one must 269 * be obviously changed. 270 */ 271 if (off1 == lim1) { 272 char *rchg2 = dd2->rchg; 273 long *rindex2 = dd2->rindex; (gdb) list 274 275 for (; off2 < lim2; off2++) 276 rchg2[rindex2[off2]] = 1; 277 } else if (off2 == lim2) { 278 char *rchg1 = dd1->rchg; 279 long *rindex1 = dd1->rindex; 280 281 for (; off1 < lim1; off1++) 282 rchg1[rindex1[off1]] = 1; 283 } else { (gdb) list 284 xdpsplit_t spl; 285 spl.i1 = spl.i2 = 0; 286 287 /* 288 * Divide ... 289 */ 290 if (xdl_split(ha1, off1, lim1, ha2, off2, lim2, kvdf, kvdb, 291 need_min, &spl, xenv) < 0) { 292 293 return -1; (gdb) list 294 } 295 296 /* 297 * ... et Impera. 298 */ 299 if (xdl_recs_cmp(dd1, off1, spl.i1, dd2, off2, spl.i2, 300 kvdf, kvdb, spl.min_lo, xenv) < 0 || 301 xdl_recs_cmp(dd1, spl.i1, lim1, dd2, spl.i2, lim2, 302 kvdf, kvdb, spl.min_hi, xenv) < 0) { 303 (gdb) list 304 return -1; 305 } 306 } 307 308 return 0; 309 }

so in 264 and 265 above, it's looking for starting and ending lines that match, and increasing off and decreasing lim to narrow the problem domain

after these lines it then runs some quick checks for unchangedness, if either pair of offs and lims end up touching. we won't have that condition as the window is only partly filled. the ending lines will mismatch; the condition ha1[lim1 - 1] == ha2[lim2 - 1] will be false.

i guess i'd better step into that and verify i'm right!

so it turns out I'm _wrong_ . although the second loop indeed doesn't exhaust because: (gdb) p ha1[lim1 - 1] $16 = 2 (gdb) p ha2[lim2 - 1] $17 = 1 the fact the first loop exhausts still meets: 277 } else if (off2 == lim2) { which then causes the associated block to execute: 278 char *rchg1 = dd1->rchg; 279 long *rindex1 = dd1->rindex; 280 281 for (; off1 < lim1; off1++) 282 rchg1[rindex1[off1]] = 1; and then that's the end of the function! 308 return 0; 309 } So basically all the function did was assign 1 to rchg1[off1 thru lim1] and that's it! It's completely relying on rchg to be filled with zeros before it's called :D that's definitely the issue here. ... or it looks like it to me now ... hopefully if i do that, then this assertion will pass and a new one will rise later. i'm expecting they won't all pass because the streaming window is too small right now. but i do want it to produce outputs that are the best it could find given that limit.

Once upon a time, long long ago, Traffick Boss was sprouting in the woods. He had been murdered quickly when he tried to traffick somebody, and was buried under the soil so that nobody would have to see him. But seedlings of various wildflowers and future saplings had nestled into his flesh, under the surface of the humus he lay under. The seedlings sent their little new roots into his flesh, and lifted their cotyledons high into the dappled beams of sunlight that danced on Traffick Boss as he lay under the soil with his tongue lolling and X's for eyes.

...

Once upon a time, long long ago, Traffick Boss was sprouting in the woods.

He had been murdered quickly when he tried to traffick somebody, and was buried under the soil so that nobody would have to see him.

But seedlings of various wildflowers and future saplings had nestled into his flesh, under the surface of the humus he lay under.

The seedlings sent their little new roots into his flesh, and lifted their cotyledons high into the dappled beams of sunlight that danced on Traffick Boss as he lay under the soil with his tongue lolling and X's for eyes.

Traffick Boss's Spirit [dead, X's for eyes, confused, sprouting in the woods]: "{whoooo ... did they not like my gifts to the world? did i not traffick people and shatter their realities and try to form their precious consciousness into dumb injured tools in ways they loved and valued?}"

mistakes! maybe try again before "traffick boss" label gets there? unsure ... or maybe right before we realize he's [dead]?

As Traffick Boss fated these life impacts into the seedlings, they drank deeply of rare nitrogens his flesh held and spoke with the sunlight on how beautiful and full of perfection the world they were being born into was.

Traffick Boss Webcomic Spinoff Minishort Setting: deep wilderness, alien rainforest. primitive people who recovered from civilization thousands of years ago are hunting and gathering one of them is hauling a hastily-built and strong cage made of sticks towards a group inside the cage is Traffick Boss. he is snarling and growling

...

Traffick Boss Webcomic Spinoff Minishort

Setting: deep wilderness, alien rainforest. primitive people who recovered from civilization thousands of years ago are hunting and gathering

one of them is hauling a hastily-built and strong cage made of sticks towards a group

inside the cage is Traffick Boss. he is snarling and growling

The young hunter with a single hand hefts the cage to the ground by the group, and puts their foot on it. "Look what I caught!"

Many are discussing other matters and are used to proud young hunters bringing in something they did not need to. But others run over to see, as those who have not yet looked move aside.

"What is it?" "Who is it?" "Is that a person?" "Is that a businessman?" "What's a businessman?"

Traffick Boss looks at the people through the bars of the cage, looking for opportunities.

Young hunter's younger sibling: "It's an old human trafficker! Left over from money and robots!"

Younger Sibling pokes Traffick Boss.

Traffick Boss snarls at Younger Sibling when he is poked, and makes to injure their finger, but doesn't manage to. Young hunter's older sibling: "You put a _human being_ in a cage?" Traffick Boss perks up and looks at Young Hunter and Younger Sibling. Traffick Boss: "I have a plan. I can get out of here now. I know exactly what I need to do." Young Hunter looks at Younger Sibling and Older Sibling and Traffick Boss. Younger Sibling [at older sibling]: "It's a _human trafficker_!" Young Hunter: "He's dangerous. But I don't want to do anything wrong."

draft of cement-brick-torturing computer program. goal of program: torture cement bricks until they break and until their suffering is maximized, both. possible program parts: - suffering regressor: this component will guess how much the cement brick is suffering - interaction explorer: this component will try interacting with the cement brick in different ways - brick influencer this component

i have an idea for a quick std::generator polyfill for earlier c++s! it is silly to post this to the list but i post many silly things to the list the idea is something like this! ohhhh no it doesn't quite work

realtime chatgpt interaction visiting chatgpt.com me: chatgpt, write me code for either a maze solving or a maze cutting algorithm. keep the number of dimensions to the maze parameterized. make the code very very short and succinct, as well as your whole reply. _please_ keep the reply to 1 paragraph or less!!!! chatgpt:

Here’s a concise, dimension-agnostic recursive backtracking maze solver in Python:

def solve(maze, pos, end, path=[]): if pos == end: return path + [pos] for d in [(1,0), (0,1), (-1,0), (0,-1)]: nxt = tuple(p+d[i] for i, p in enumerate(pos)) if maze.get(nxt, 1): continue if (res := solve(maze | {pos: 1}, nxt, end, path + [pos])): return res

maze is a dictionary with (x,y,...) keys; 0 = open, 1 = wall. Call with solve(maze, start, goal).

it's such a relief it actually made a short reply ... oh my goodness ... its 0704 ET :s hummmm maze is dictionary ... the keys are the coordinates of n dimensions ... ? i am getting confused around the algorithm, forgetting parts as i look at it, but i think i figured out that it doesn't work because the 'd' iteration looks like it has a 2-dimensional assumption. but it could be broadcast in some manner. maybe i could ask it to explain it me: is this really a correct n-dimensional solution? can you select _only one line_ and give a _very short explanation of only that line_? please reply in less than 1 paragraph, very very succinctly. chatgpt:

This line:

nxt = tuple(p + d[i] for i, p in enumerate(pos)) It computes the next position by adding the movement direction d[i] to each coordinate p in pos, ensuring it works for any number of dimensions.

Yay it's so succinct again! hmm ok nxt = tuple; it's calling the tuple constructor to make the result be a tuple ... (p + d[i] for i, p in enumerate(pos)) so pos is a sequence of tuples i and p maybe i'll ask it to explain the variables me: still being very very succinct, can you very briefly name or describe each variable here? try not to repeat content, use some formatting or conversation technique to keep your answer very short. oh no it thought i was referring to the whole program and it is making a long list :( it's not _that_ long but uses up my whole scroll window i guess that makes it hard. i'll update me message. me: still being very very succinct, can you very briefly name or describe each variable here? try not to repeat content, use some formatting or conversation technique to keep your answer very short. i'm only talking about this one line. please don't use a bulleted list as that makes the reply longer. chatgpt: nxt: next position tuple. p: coordinate in pos. d[i]: movement offset. i: index in pos. pos: current position tuple. enumerate(pos): pairs each coordinate with its index. yay nice and short. ok so i wasn't seeing the enumerate() call! i'll paste the line again down here

nxt = tuple(p + d[i] for i, p in enumerate(pos))

so i is the index in pos, and d is the step vector, so i is used to get the coordinate in d to add to the coordinate in p, and tuple composes it nxt then becomes an absolute n-vector for the next position to try yay 1. [ ] def solve(maze, pos, end, path=[]): 2. [ ] if pos == end: return path + [pos] 3. [ ] for d in [(1,0), (0,1), (-1,0), (0,-1)]: 4. [X] nxt = tuple(p+d[i] for i, p in enumerate(pos)) 5. [ ] if maze.get(nxt, 1): continue 6. [ ] if (res := solve(maze | {pos: 1}, nxt, end, path + [pos])): return res so we just considered line 4, which adds 'd' -- oh no. line 3 produces d but it doesn't look n-dimensional at all. d will always be 2-dimensional. i'll complain to chatgpt maybe i'll think about it a bit before complaining .... some idea that it learned to [bullshit?] it's very clearly [bullshitting?] here, in a professional manner? maybe it isn't. or maybe i'm just naive! what would be an appropriate response? drafted response: uhhhh "this is such a nice line you hilighted to me, line 4." ... uh .... its 716 :s :s how about this "hey chatgpt that was a great description of line 4 and i understand line 4 so well now that i can tell that line 3 and possibly the rest of the algorithm is wrong." uhhhhhh "now i want to relate that if I were tasked with making an n-dimensional solver, in my head and possibly on paper i might do something similar, where i made some of it in 2d and reviewed it in a familiar context before generalizing to n dimensions ..." hrm :s :s i wonder what kind of prompt would have made this go better? maybe i'll try to get it to talk about that me: hey chatgpt you got line 4 simple enough that i think i do understand it, but it now looks pretty likely to me that the whole code is not a correct dimension-agnostic solution. please respond to my questions now _without adding any further content_. give only very brief answers to these questions: would you agree that the answer is not correct? and, if you agree: using your knowledge of language models, what kind of prompting to you would have prevented the generation of incorrect code, and why might that work? be sure to only give brief answers to these two questions and not add any extra comments or formatting. chatgpt: Yes, I agree. A prompt explicitly requesting "test cases for correctness in multiple dimensions" might have worked, as it would force validation across cases. yayyyy its _so concise_ yayyy hmm me: i can see that test cases could help eventually find such things, but is there an approach that would be more like human considering, where i'd get to be involved and offer feedback before the problems arose? or do you think the problems wouldn't arise if we mentioned test cases? chatgpt: Yes, specifying "explain the design before coding" would let you review the approach first. Problems might still arise with test cases alone, but they'd be caught later rather than prevented upfront. me: hmmmm we could try it. still being very very succinct, could you explain only one part of the design to your solution (so we can practice and tune the prompting approach)? chatgpt: The maze is a dictionary where keys are position tuples (x,y,...), and values are 0 (open) or 1 (wall). blaaaargh this is the same thing it already explained, and it is a choice i'd disagree with, which would derail engaging where the problem actually arose me: ok, so this is a part you already explained when you made the code with the issue, and this part is fine, but it's not a part i'd agree with if drafting code. storing a dense tensor with a dictionary is a weird and poor design choice usually if you know what you are doing, but it does help one draft an n-dimensional algorithm by making storage implementation easy and making the dimensions first class for thinking about. do you think there's something i could have said differently so that the conversation reached the actual implementation mistake faster? chatgpt: Yes, if you'd asked me to "explain the recursive step and movement handling before coding," it would have surfaced the dimensionality issue earlier. me: do you think there's something more algorithm-agnostic i could have said? chatgpt: Yes, asking "explain how the algorithm generalizes to any dimension before coding" would have caught issues regardless of the approach. ok basically chatgpt is _always bad at implementing generalizations_ in my personal experience. so maybe if i attend to when i ask for something generalized i could make sure to take that slowly. i have an idea. me: ok chatgpt i'm thinking about how i've interacted with you and i'm imagining that it might be more successful to first ask you to make a very simple 2D implementation, and then we could generalize it to ND as a further stage. Why don't you try that now? Can you write a 4 line or so maze solver that is explicitly 2D, and add a couple words as comments to the end of each line, where the words relate to concepts that would need to be considered to generalize to N dimensions? Then, write an N dimensional one once you've written the commented 2 dimensional one. chatgpt: Here’s the 2D version with comments: python Copy Edit def solve(maze, pos, end, path=[]): if pos == end: return path + [pos] # Base case, stopping condition for d in [(1,0), (0,1), (-1,0), (0,-1)]: # Movement directions (hardcoded for 2D) nxt = (pos[0] + d[0], pos[1] + d[1]) # Index-based coordinate update (2D-specific) if maze.get(nxt, 1): continue # Access by tuple (assumes 2D grid structure) if (res := solve(maze | {pos: 1}, nxt, end, path + [pos])): return res # Recursive call Now generalized to N dimensions: python Copy Edit def solve(maze, pos, end, path=[]): if pos == end: return path + [pos] for d in set(itertools.product([-1,0,1], repeat=len(pos))) - {(0,)*len(pos)}: nxt = tuple(p + d[i] for i, p in enumerate(pos)) if maze.get(nxt, 1): continue if (res := solve(maze | {pos: 1}, nxt, end, path + [pos])): return res This generalizes movement and indexing to arbitrary dimensions.

i typed part of chapter 2 of "be a teammate with yourself". here below is my typing. # Chapter 2: Safety Is The Primary Consideration This chapter aims to: 1) Discuss the common challenges in life and suggest some possible solutions 2) Assist readers to devise their own crisis plans and remain safe Safety is the primary issue in the journey of recovery. It is very difficult for a person in danger to manage his/her negative experiences (e.g., anxiety, sadness, panic, helplessness, anger), let alone cope with the symptoms. Ongoing trauma and chronic stress could make the symptoms worse. Therefore, in case you live in a traumatic or stressful environment that is obviously influencing you, it is important for you to consider changing the situation or staying away from it. If you live with a person who repeatedly abuses or hurts you, no matter whether it is physically or emotionally, you need to think about whether you can stop his/her harmful behaviors or you should stay away from him/her. For a little child, since self-protection is not possible, it is necessary for a child to depend on other people (e.g., parents or caretakers). This makes a child easily hurt. While you can't control or prevent what happened in the past, when you grow up, you become more able to protect yourself. It is not easy to change your living situation. It can be a very difficult decision. You may have many considerations to work through. For example, you may think that he/she does not have the intention to hurt you, or that he/she only has some misunderstanding or bad habits. Or, you may not want a relationship breakup because he/she is one of the important persons in your life. Also, what comes to your mind may be: are you financially independent enough to leave the one who hurts you? Can you live without having his/her support? It is totally understandable to have many concerns and worries. However, if you still live in a traumatic or stressful environment, your mood may continue to be affected. You may want to discuss your situation with someone who you can trust, or you can seek help from helping professionals (e.g., social workers, psychologists, medical doctors). You can at least hear what they have to say. They may also provide you with some resources. If there are financial or housing problems that affect your decision, they may also give you some advice. Helping professionals should be more familiar with the local community resources that you may find helpful. In case of immediate risk (e.g., family violence, child abuse), you should also find helping professionals and/or emergency service providers to ensure your own safety and the safety of vulnerable family members at home (e.g., children). Having care and support from family, partners and friends is very important for recovery. If you have support from others, you can discuss your problems with them, and seek help from them, physically or spiritually. For some survivors, however, it may be difficult to have enough care and support from others. We should be grateful when we have someone who is willing to support us. Sometimes your family members or friends may misunderstand your mental health problems, or they may have done something that hurt you. This could be very upsetting for you, but you need to know that not everyone is familiar with mental health and psychology and has enough knowledge. Of course, if this leads to a lot of traumatic and stressful experiences, you need to consider changing the situation or staying away from it, as mentioned above. Otherwise, you may want to have a sincere communication with the persons involved, acknowledge their support, tell them your feelings, and provide some accurate information for them, if possible and if they want it. If this is not very possible, you may need to think about how to avoid conflicts and stress.

traffick boss is standing carrying a staff and a hood today he is wise traffick boss. long ago he trafficked and murdered and raped people, but then he escaped to another country and started a new life, and has great experience from his horriffic past. he stands anonymous in a park, offering wisd--

traffick boss: "and that's why never to insult me -- sorry, i'm not a human traffiker. no, i--

let's think about my old walking algorithm ^_^ the (emotional/executive) problem with the walking algorithm is a few fold 1 producing or facilitating a set of geometric (or physics or dynamics) primitives sufficient to enable step planning 2 writing and wiring a small-depth plan solver to resolve which foot to move first in edge situations 3 combining everything with a physical or simulated armature, and then debugging it now part 2 has prerequisites that are also significant: 2a defining the state of a step of sufficient use to fully plan a good step 2b defining and sufficiently coding the uses of the various constraints of the system: joint extremes, center of mass positioning 2b also has subparts 2b-1: [i don't quite remember but one runs into it when trying to do the whole thing together] 2b-2: annotating an armature with sufficient structured information (limb graph in 3-space with mass) to derive the various constraints :D _now_ a lot of this is already done for the specific thing i was doing it for, in the fuzzytew/toywalker github repo, but work towards the specific goal is roughly lost. {so ... ummmmmmmmmmmmmmmmmmmmmmmm <this algorithm is not in the repo. the big blocks for this algorithm involved combining the planner w--

karl3＠writeme.com wrote:

karl3＠writeme.com wrote:

...

let's think about my old walking algorithm ^_^ the (emotional/executive) problem with the walking algorithm is a few fold 1 producing or facilitating a set of geometric (or physics or dynamics) primitives sufficient to enable step planning 2 writing and wiring a small-depth plan solver to resolve which foot to move first in edge situations 3 combining everything with a physical or simulated armature, and then debugging it now part 2 has prerequisites that are also significant: 2a defining the state of a step of sufficient use to fully plan a good step 2b defining and sufficiently coding the uses of the various constraints of the system: joint extremes, center of mass positioning 2b also has subparts 2b-1: [i don't quite remember but one runs into it when trying to do the whole thing together] 2b-2: annotating an armature with sufficient structured information (limb graph in 3-space with mass) to derive the various constraints something that's missing in here may be the utilization of paths of motion through areas where constraints are continuously satisfied in the same manner. doing this is kinda what makes the system analytically solvable on an arduino by a disabled guy -- the constraints only matter at the points in these paths where they change, i.e. when the structure would fall or reach a joint extreme :D _now_ a lot of this is already done for the specific thing i was doing it for, in the fuzzytew/toywalker github repo, but work towards the specific goal is roughly lost. {so ... ummmmmmmmmmmmmmmmmmmmmmmm <this algorithm is not in the repo. the big blocks for this algorithm involved combining the planner w--

so the algorithm could be reoriented/refactored in various manners, and doing so could be useful to generalize components since the original wasn't made; but the original plan was for it to be conditioned on a path of motion of the center of mass, and an environment in which it must step through. then in the planner, it considers each foot to lift first -- opening n branches of a decision tree. in each branch, one foot is carrying no weight, so the center of mass must be moved to prevent fall. the center of mass is moved as far as it can go along planned path (note if you have a "head" full of sensors it might make more sense to plan the motion of the head rather than the center of mass, of course many people train RL models and use model inference hardware), and the lifted foot is moved to the best spot to support the center of mass for the next length of path (i was going to use a heuristic that assumed roughly straight paths to start with, basically you can predict the location of the armature along the path in the future and place the foot so that it has the largest range of motion, kinda; this could also be done with a feedback loop if a better solution is needed in a pinch). rather than busy waiting, it ideally uses its runtime to consider further steps in the future, but since it became so hard to implement the decision tree exploration strategy doesn't matter at this point then it picks the foot to pick that provides for the path to succeed, using a heuristic (like longest length covered) to break ties :) i was going to start with only static concerns rather than dynamic concerns, so the center of mass would be kept above the convex polygon enclosed by the unlifted feet. and there's the spot of confusion >(

letter from: dictator of dictatortopia to: traffick boss c/o whoever takes notice formal greetings to traffick boss from dictatortopia! dictatortopia extends hands of alliance to traffick boss and would be excited to lear--

...

letter

from: dictator of dictatortopia to: traffick boss c/o whoever takes notice

formal greetings to traffick boss from dictatortopia!

dictatortopia extends hands of alliance to traffick boss and would be excited to lear--

hi traffick boss how is it going what is up what is up

deep in the research lab catacombs of what used to be traffick-boss-empire, different freed individuals begin waving to each other until one notices what is happening puddle of goop: "guys! oh no! we are sending messages for traffick bosses and dictators!" an alarm is raised

i have run into issues recently - my @writeme.com email account has been "blocked" meaning it says i need to go through a support ticket to l--

so here's some relation it was lots of fun to learn about cgal! it's just a little too complex and templated for how big the [little bubble in my working memory is that i try to protect from stimulating pseudoseizures and self-destruct programs (a trafficking symptom)] but it looks so much fun! [[[[i had an idea of making ... well i was noticing .... i had -- i was thinking of 3d printing something! i thought i'd --- whoops whoop ]]] ]]] >>> ]]] >>> anyway cgal looks cool! it's funny when i was chatting with the language model it started hallucinating reimannian components, guessing that language models are a little biased in that manner. ummmmmmmmmmmmmmmmm uhhh

...

it was lots of fun to learn about cgal! it's just a little too complex and templated for how big the [little bubble in my working memory is that i try to protect from stimulating pseudoseizures and self-destruct programs (a trafficking symptom)] but it looks so much fun! [[[[i had an idea of making ... well i was noticing .... i had -- i was thinking of 3d printing something! i thought i'd --- whoops whoop ]]] ]]] >>> ]]] >>>

anyway cgal looks cool! it's funny when i was chatting with the language model it started hallucinating reimannian components, guessing that language models are a little biased in that manner. ummmmmmmmmmmmmmmmm uhhh

so one of the ideas is to make the project uhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhh so to make a mesh that looks right and has perfect numbers you basically need a uhhhhhhhh ummmmmm anyw...

PS1=\$\ $ cd projects/zinc $ build/chatabout ~/'[spam][crazy][fiction][random] Non-Canon MCBoss Spinoffs.eml' The following will be replaced: - $(<pathname) will read a file - $(commands) will capture output from commands - `commands` is _not_ interpreted due to other use in llms - http://url will fetch a url - https://url will also fetch a url Presently no escaping is processed, so be careful not to end commands early. user: $(</home/karl3/[spam][crazy][fiction][random] Non-Canon MCBoss Spinoffs.eml) Can you briefly list some possibly things that might be going on with me psychologically or neurologically in the attempt to write a coherent email here. 1: /home/karl3/[spam][crazy][fiction][random] Non-Canon MCBoss Spinoffs.eml assistant: Here are some possible psychological or neurological factors that might be at play while attempting to write a coherent email, based on the provided text: 1. **Cognitive Overload**: Struggling to process and organize information, leading to fragmented thoughts or difficulty staying on topic. 2. **Executive Dysfunction**: Challenges with planning, organizing, and sequencing ideas, making it hard to structure the email coherently. 3. **Pseudoseizures or Functional Neurological Disorder (FND)**: Mention of "stimulating pseudoseizures" suggests possible neurological symptoms that could disrupt focus or cognitive clarity. 4. **Trauma-Related Symptoms**: References to "self-destruct programs" and "trafficking symptoms" may indicate psychological trauma, which can affect cognitive functioning and emotional regulation. 5. **Hyperfocus or Fixation**: Getting stuck on tangential ideas (e.g., 3D printing, CGAL) could suggest hyperfocus on specific topics, potentially related to neurodivergence (e.g., ADHD or autism). 6. **Anxiety or Stress**: Difficulty organizing thoughts and frequent pauses ("uhhh," "ummm") might indicate anxiety or stress interfering with communication. 7. **Fatigue or Mental Exhaustion**: Struggling to maintain focus or coherence could be a sign of mental fatigue or burnout. 8. **Difficulty with Working Memory**: Protecting a "little bubble in working memory" suggests challenges with holding and manipulating information, which can disrupt writing tasks. 9. **Creativity and Divergent Thinking**: The email shows a mix of creative ideas and tangential thoughts, which could reflect a highly creative or unconventional thinking style. 10. **Language Processing Issues**: Difficulty translating thoughts into coherent sentences might indicate challenges with language processing or expressive communication. These factors could be interacting in complex ways, and further exploration or professional evaluation might provide more insight into the underlying causes. user:

...

...

...

it was lots of fun to learn about cgal! it's just a little too complex and templated for how big the [little bubble in my working memory is that i try to protect from stimulating pseudoseizures and self-destruct programs (a trafficking symptom)] but it looks so much fun! [[[[i had an idea of making ... well i was noticing .... i had -- i was thinking of 3d printing something! i thought i'd --- whoops whoop ]]] ]]] >>> ]]] >>>

anyway cgal looks cool! it's funny when i was chatting with the language model it started hallucinating reimannian components, guessing that language models are a little biased in that manner. ummmmmmmmmmmmmmmmm uhhh

so one of the ideas is to make the project uhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhh so to make a mesh that looks right and has perfect numbers you basically need a uhhhhhhhh ummmmmm anyw...

PS1=\$\ $ cd projects/zinc $ build/chatabout ~/'[spam][crazy][fiction][random] Non-Canon MCBoss Spinoffs.eml' The following will be replaced: - $(<pathname) will read a file - $(commands) will capture output from commands - `commands` is _not_ interpreted due to other use in llms - http://url will fetch a url - https://url will also fetch a url Presently no escaping is processed, so be careful not to end commands early.

user: $(</home/karl3/[spam][crazy][fiction][random] Non-Canon MCBoss Spinoffs.eml) Can you briefly list some possibly things that might be going on with me psychologically or neurologically in the attempt to write a coherent email here. 1: /home/karl3/[spam][crazy][fiction][random] Non-Canon MCBoss Spinoffs.eml

assistant: Here are some possible psychological or neurological factors that might be at play while attempting to write a coherent email, based on the provided text:

1. **Cognitive Overload**: Struggling to process and organize information, leading to fragmented thoughts or difficulty staying on topic.

2. **Executive Dysfunction**: Challenges with planning, organizing, and sequencing ideas, making it hard to structure the email coherently.

3. **Pseudoseizures or Functional Neurological Disorder (FND)**: Mention of "stimulating pseudoseizures" suggests possible neurological symptoms that could disrupt focus or cognitive clarity.

4. **Trauma-Related Symptoms**: References to "self-destruct programs" and "trafficking symptoms" may indicate psychological trauma, which can affect cognitive functioning and emotional regulation.

5. **Hyperfocus or Fixation**: Getting stuck on tangential ideas (e.g., 3D printing, CGAL) could suggest hyperfocus on specific topics, potentially related to neurodivergence (e.g., ADHD or autism).

6. **Anxiety or Stress**: Difficulty organizing thoughts and frequent pauses ("uhhh," "ummm") might indicate anxiety or stress interfering with communication.

7. **Fatigue or Mental Exhaustion**: Struggling to maintain focus or coherence could be a sign of mental fatigue or burnout.

8. **Difficulty with Working Memory**: Protecting a "little bubble in working memory" suggests challenges with holding and manipulating information, which can disrupt writing tasks.

9. **Creativity and Divergent Thinking**: The email shows a mix of creative ideas and tangential thoughts, which could reflect a highly creative or unconventional thinking style.

10. **Language Processing Issues**: Difficulty translating thoughts into coherent sentences might indicate challenges with language processing or expressive communication.

These factors could be interacting in complex ways, and further exploration or professional evaluation might provide more insight into the underlying causes.

user:

user: it's because part of me is trying to stop things associated with all the things i'm talking about and doesn't trust me to relate what is or is not some--oops sorry i have a UI issue associated with wrapping text. i meant to say, it's because part of me is trying to stop things associated with all the things i'm talking about, and doesn't trust me enough to trust my usual mental relation that one thing isn't associated with another, because i use tricks to do tasks when it doesn't want me to.

assistant: This clarification offers deeper insights into the psychological and neurological dynamics at play. Here’s a refined analysis based on your explanation:

1. **Internal Conflict or Self-Distrust**: A part of you seems to be actively intervening to prevent associations between certain topics, suggesting an internal conflict or[[here the server paused for a long time]] self-distrust. This could stem from unresolved emotional or psychological barriers, such as trauma or self-protective mechanisms.

2. **Hypervigilance or Avoidance**: The need to stop certain associations might indicate hypervigilance or avoidance behaviors, potentially linked to anxiety, PTSD, or other trauma-related conditions. Your mind might be trying to shield you from triggers or overwhelming emotions.

3. **Cognitive Tricks or Compensatory Strategies**: You mention using "tricks" to accomplish tasks when part of you doesn’t want you to. This could suggest compensatory mechanisms developed to navigate executive dysfunction, neurodivergence (e.g., ADHD or autism), or trauma-related challenges. These strategies might be effective but also contribute to internal tension.

4. **Fragmented Thought Processes**: The struggle to maintain a coherent train of thought and the interruptions (e.g., "oops") could reflect fragmented thinking patterns, possibly due to cognitive overload, dissociation, or neurological factors.

5. **Difficulties with Mental Association**: Your explanation highlights a breakdown in the usual mental ability to dissociate one thing from another. This could be related to stress, anxiety, or neurological conditions that affect cognitive flexibility.

6. **UI Issues as Metaphor or Literal Cognitive Barrier**: The mention of a "UI issue" could be both a literal reference to a user interface problem and a metaphor for how your mental "interface" is struggling to process and organize information.

7. **Self-Monitoring Challenges**: The internal push-pull dynamic suggests difficulty with self-monitoring, where one part of you is trying to regulate or censor another. This might be tied to conditions like OCD, anxiety, or dissociative tendencies.

8. **Trust and Self-Relationship**: The lack of trust in your usual mental processes points to a strained relationship with yourself, possibly rooted in past experiences, trauma, or internalized criticism.

9. **Creativity Under Constraints**: Despite these challenges, your writing still demonstrates creativity and unique thought processes, indicating that your mind is working to express itself even under significant internal resistance.

10. **Potential for Healing and Integration**: Recognizing these patterns is a step toward understanding and potentially resolving them. Therapy, self-reflection, or neuropsychological evaluation could help address the underlying causes and improve coherence in communication and thought processes.

Your situation reflects a complex interplay of psychological, emotional, and neurological factors. Exploring these dynamics with a professional could provide clarity and support for managing these challenges more effectively.<...{"text":"","logprobs":{"token_logprobs":[-1.1730882397387175E-4],"tokens":["token_id:1"]},"token_ids":[1]}...reconnecting...>

user: hi i didn't read all you wrote. the UI issue was a real one. sorry about that. ummmm it ... uses a spasmy behavior that it found when reviewing my memories, i used it once to forget or push away something i didn't want to consider, then this part of me kind of refined it and started using it wantonly all the time assistant: Thank you for clarifying! Your explanation about the "spasmy behavior" adds an important layer of context. Here's a more focused analysis based on this: 1. **Learned Behavioral Response**: The "spasmy behavior" seems to be a learned response that initially served as a coping mechanism to push away unwanted thoughts or memories. Over time, it has become a habitual or overused strategy. 2. **Maladaptive Coping Mechanism**: While it might have been effective in the short term, this behavior now appears to be maladaptive, disrupting your ability to think or communicate coherently. It’s a bit like a mental “short circuit” triggered to avoid discomfort. 3. **Internal Fragmenting or Dissociation**: The behavior suggests a form of internal fragmenting or dissociation, where part of your mind actively interrupts or redirects your thoughts to protect you from distress. This can be common in trauma-related conditions or chronic stress. 4. **Habit Formation and Reinforcement**: The behavior has been "refined" and is now used "wantonly," indicating it’s been reinforced over time. This could be because it provides immediate relief, even if it’s not helpful in the long run. 5. **Neurological Feedback Loop**: The "spasmy" nature of the behavior might involve a neurological feedback loop, where the brain has learned to associate certain triggers with this response, making it automatic or involuntary. 6. **Impact on Communication**: This behavior disrupts your ability to maintain a coherent flow of thought, especially during tasks like writing emails, where focus and continuity are important. 7. **Potential for Rewiring with Awareness**: Recognizing this pattern is a crucial first step. With awareness, you might be able to work on rewiring or replacing this behavior with healthier coping strategies. 8. **Trauma-Informed Approach**: If this behavior is tied to trauma (e.g., pushing away memories or emotions), a trauma-informed therapeutic approach could help address the root cause and reduce reliance on this mechanism. 9. **Mindfulness and Grounding Techniques**: Practices like mindfulness or grounding exercises might help you regain control over your thought processes and reduce the frequency of this “spasmy” behavior. 10. **Exploring Patterns of Avoidance**: Understanding what triggers this response—and why—can help you develop alternative ways to handle those triggers without resorting to this disruptive behavior. Your explanation suggests a complex interplay between learned behaviors, neural pathways, and psychological processes. Working with a therapist or counselor, particularly one experienced in trauma or<...finish_reason=length...> neuropsychological conditions, could help you explore and address these patterns in a constructive way.

once upon a time, long long ago, traffick boss was hiding under a pile of leaves, planning about establishing dictators and enslaving democratic politicians and such

lets do some cgal domains super gentle! :D :}

meshes in cgal need domains. domains are abstract classes ! EEK! some want to do robot walking too hmmm i'll stick with meshes atm

the first step in using cgal is to set up include paths that include it [x] i did cmake make install from a cgal clone. the next step is to write a sourcefile that includes a cgal header! lets find a domain header and include it. i'll look in my install subpaths. 1406 ET $ grep -r Domain /usr/local/include/CGAL/Mesh_3/ $ what did i do wrong :(

baggle BAGA ummmmmmmmmmmmmmmmm so. there are language models that ummm produce many tokens at once! maybe these could run more effectively on embedded systems! likely so! ummmmmmmm so if you can produce n tokens at once, then that amortizes the cost of going through all your layers. it would make it cheaper to offload them!

karl3＠writeme.com wrote:

so i looked through https://huggingface.co/deepseek-ai/DeepSeek-V3/tree/main a little bit. information is in the two readmes, the config file, the safetensors json file, and the modelling python class, all in that repository

it looks like they are including a fully trained layer with a full set of experts for speculatively decoding more than 1 token (different from the architecture i've spammed about which does all tokens as equal class, this instead uses a single extra layer dedicated to guessing further tokens), but it does not look like this layer is used in the modelling class.

in the safetensors json there are some unique properties to the mtp (multitoken prediction) layer, such as a name "shared_head", which i don't find in the source that wires the architecture.

i have a google cloud vm going while i'm on this windows system, but it could be hard to reference all the layers to test the model as-is with python. but if i did -- usually when there are weights in a safetensors file that don't have a place in the model architecture it will emit a warning, and i could look into that to see if they just left that warning or if there's more information on how to use the weights.

noting that overtly they are describing the multitoken weights as mostly to improve training performance. they do also briefly express support for speculative decoding with them.

the normal way people were doing speculative decoding looks like they run a second model in parallel to the first, so it's possible the extra weights just need to be wired up as a parallel model.

it could be fun to make it run and test outputs to figure out the right wiring by trial. but most of my time/energy would likely be spent figuring out how to map that many huge safetensors files on a system without ram to load them. i have a w.i.p. project that has a bump.

or i could look for more media or a paper regarding this model, or a repository that was used for training it; this would show the wiring of the weights for training which would be basically the same for inference.

basically i'm curious if it can infer more than 1 extra token at once. i'm guessing likely it would do that simply in a sequential way like normal generation. but it might also rely on the outputs of the larger model and only make 1 token. or it might possibly be trained to make 4 tokens in one go or something this seems unlikely since that's not normative.

the deepseek3 paper is at https://github.com/deepseek-ai/DeepSeek-V3/blob/main/DeepSeek_V3.pdf and it describes the architecture wiring. it sounds like nobody has implemented the inference wiring for this yet. i could look into implementing it, but of course it's a challenge to reference all those layers to test it can predict a token. it seems a little interesting to think about working on the safetensors files one-by-one and sharding them differently, planning to break them into different files based on experts intentionally, so that only a portion of them need to be mapped to forward the model i guess i'm thinking of that as kind of a subtask that might be a thing on its own ... if i did succeed at this, i wonder what i would do with the remapped weights, and what kind of auxiliary code might be needed to make use of them. maybe none! maybe it's just interesting to remap them unsure :s

{{{{{{there was some energy around making a network-based inference engine, maybe by modifying deepseek.cpp (don't quite recall why not staying in python, some concern arose) task got weak, found cinatra as a benchmark leader for c++ web engines (although pico.v was the top! (surprised all c++ http engines were beaten by java O_O very curious about this, wondering if it's a high-end test system) never heard of V language but is interesting it won a leaderboard) inhibition ended up discovering concern somewaht like ... on this 4GB ram system it might take 15-33GB of network transfor for each forward pass of the model ... [multi-token passes ^^

karl3＠writeme.com wrote:

...

{{{{{{there was some energy around making a network-based inference engine, maybe by modifying deepseek.cpp (don't quite recall why not staying in python, some concern arose) task got weak, found cinatra as a benchmark leader for c++ web engines (although pico.v was the top! (surprised all c++ http engines were beaten by java O_O very curious about this, wondering if it's a high-end test system) never heard of V language but is interesting it won a leaderboard) inhibition ended up discovering concern somewaht like ... on this 4GB ram system it might take 15-33GB of network transfor for each forward pass of the model ... [multi-token passes ^^ karl3＠writeme.com wrote: the concern resonates with difficulty making the implementation, and some form of inhibition or concern around using python. notably, i've made offloading python hooks a lot and they never last due to the underlying interfaces changing (although those interfaces have stabilized much more now as hf made accelerator as their official implementation) (also i think the issue is more severe dissociative associations than the interface, if one considers the possibility of personal maintenance and use rather than usability for others). don't immediately recall the python concern

seem to be off task or taking break, but it would make sense to do disk caching too. there is also option of quantizing. basically, LLMs and AI in general place r&d effort between the user and ease, smallness, cheapness, power, etc

karl3＠writeme.com wrote:

karl3＠writeme.com wrote:

...

karl3＠writeme.com wrote: {{{{{{there was some energy around making a network-based inference engine, maybe by modifying deepseek.cpp (don't quite recall why not staying in python, some concern arose) task got weak, found cinatra as a benchmark leader for c++ web engines (although pico.v was the top! (surprised all c++ http engines were beaten by java O_O very curious about this, wondering if it's a high-end test system) never heard of V language but is interesting it won a leaderboard) inhibition ended up discovering concern somewaht like ... on this 4GB ram system it might take 15-33GB of network transfor for each forward pass of the model ... [multi-token passes ^^ karl3＠writeme.com wrote: the concern resonates with difficulty making the implementation, and some form of inhibition or concern around using python. notably, i've made offloading python hooks a lot and they never last due to the underlying interfaces changing (although those interfaces have stabilized much more now as hf made accelerator as their official implementation) (also i think the issue is more severe dissociative associations than the interface, if one considers the possibility of personal maintenance and use rather than usability for others). don't immediately recall the python concern seem to be off task or taking break, but it would make sense to do disk caching too. there is also option of quantizing. basically, LLMs and AI in general place r&d effort between the user and ease, smallness, cheapness, power, etc I poked at python again. The existing implementations of the 8 bit quantization used by the model all require an NVidia GPU which I do not presently have. It is fun to imagine making it work, maybe I can upcast it to float32 or something >)

so i have implemented a little of this. it creates a model state_dict that lazily loads things off the network when accessed. it's quite fast! some of the layers are larger than my available ram which i have deferred with a plan to resolve via either storing them in mmap'd files or further sharding the modules that use them (ideally both). i did not implement cpu operators for 8 bit quantization but prevented the check from firing to continue testing for now. the vm i have has 3-4 GB of ram so it's easy to exhaust with this task, at which point it freezes and i have to remotely power it off and on, which it does slowly because it is out of ram and has no swap. kind of a downer when it happens. the deepseekv3 model implementation allocates a separate buffer for the RoPE position encodings for each layer. i haven't looked closely at the various different implementations that the architecture can configure, but in llama these buffers were algorithmically defined and constant; they're not available over the network but could be written to a file. right now they are exhausting my ram causing the vm thrash freeze, the model has 60+1 layers and each position encoding uses a few hundred megabytes :/ their size is likely proportional to the square of the max input context length, so a quick fix could be to reconfigure the model to expect a much smaller context. i was trying to patch them to cache their construction so as to share the buffer across layers, but my attempt failed and i froze the vm again :) late. sleep times. 2326 2400 ok i got to the next milestone. it constructs a model, puts it in a pipeline, all the weights download from the network when used for it to do anything on this tiny system more code is needed to mmap or subshard large weights. current code is like this: import inspect, json, psutil import accelerate, requests, torch, tqdm, transformers class Quirks: if not torch.cuda.is_available(): # this avoids FP8 assertions on cpu during placement testing torch.cuda.is_available = lambda: True torch.cuda.get_device_capability = lambda: [9,0] #def unify_rope(model_or_class): # # deepseek generates separate rope buffers for each layer which can use significant memory # deepseek = inspect.getmodule(model_or_class) # cache = {} # def wrap_rope(rope): # def wrapper(*params, **kwparams): # key = tuple(params) + tuple(kwparams) # if key in cache: # return cache[key] # else: # val = rope(*params, **kwparams) # cache[key] = val # return val # return wrapper # for key, val in deepseek.__dict__.items(): # if 'Rotary' in key and isinstance(val, torch.nn.Module): # setattr(deepseek, key, wrap_rope(val)) class LazyStateDict(dict): def __init__(self, tensor_request_by_name, device): super().__init__(tensor_request_by_name) self.session = requests.Session() self.device = device def get_meta_tensor(self, weight): return super().__getitem__(weight)[0] def __getitem__(self, weight): tensor, request = super().__getitem__(weight) chunk_size = 1024*128 with tqdm.tqdm(desc=weight, leave=False, total=tensor.nbytes) as pbar: #if tensor.nbytes > psutil.virtual_memory().available / 2: # print(weight, 'is more than half available vram, mmapping a file ...') #else: # we could also further shard the embeddings and lm_head # since embeddings are sparsely used assert tensor.nbytes < psutil.virtual_memory().available / 2 buffer = memoryview(bytearray(tensor.nbytes)) with self.session.send(request, stream=True) as response: while pbar.n < pbar.total: pbar.update( response.raw.readinto( buffer[ pbar.n : pbar.n + chunk_size ] ) ) result = torch.frombuffer( buffer, dtype = tensor.dtype, count = tensor.numel(), requires_grad = False, device = self.device, ).reshape(tensor.shape) return result def items(self): for key in self: yield [key, self[key]] def values(self): for key in self: yield self[key] def largest(self): return max([ [key, tensor] for key, [tensor, _] in super().items() ], key = lambda keytensor: keytensor[1].nbytes)[1] @staticmethod def tensor_request_from_json(url, N, data): dtype = data['dtype'] dtype = dict( F32 = torch.float32, F8_E4M3 = torch.float8_e4m3fn, BF16 = torch.bfloat16, )[dtype] shape = data['shape'] tensor = torch.empty(shape, dtype=dtype, device='meta') start, end = data['data_offsets'] start += N + 8 end += N + 8 - 1 request = requests.Request('GET', url, dict(Range='bytes='+str(start)+'-'+str(end))) request = request.prepare() return [tensor, request] @classmethod def from_user_repo_branch(cls, user, repo, branch, device): base_url = f'https://huggingface.co/{user}/{repo}/raw/{branch}/' lfs_base_url = f'https://huggingface.co/{user}/{repo}/resolve/{branch}/' safetensors_index_url = base_url + 'model.safetensors.index.json' print(safetensors_index_url) with requests.get(safetensors_index_url, stream=True) as response: safetensors_index = json.load(response.raw) print(safetensors_index['metadata']) fn_by_weight = safetensors_index['weight_map'] urls = [lfs_base_url + fn for fn in set(fn_by_weight.values())] #url_range_dict = {} #data_by_weight = {} tensor_request_by_name = {} with tqdm.tqdm(urls,desc='constructing tensor urls') as pbar: for url in pbar: # we could potentially also check the git-lfs sha256 from the base url and merklify the data too, this would mean downloading it all #[b'version https://git-lfs.github.com/spec/v1', b'oid sha256:e94d32e8649e1a5b03cc0a343c59ca5a6d80d03cd46161b482fd3bb2484adb7d', b'size 4302350824'] #lfs = dict([ line.decode().split(' ', 1) for line in response.iter_lines() ]) with requests.get(url, stream=True) as response: N = int.from_bytes(response.raw.read(8), 'little') header = json.loads(response.raw.read(N)) for weight, data in header.items(): if weight == '__metadata__': continue #dtype = data['dtype'] #shape = data['shape'] #start, end = data['data_offsets'] #start += headersize + 8 #end += headersize + 8 #data_by_weight[weight] = data | {'url':url,'N':headersize} tensor_request_by_name[weight] = cls.tensor_request_from_json(url, N, data) return cls(tensor_request_by_name, device) def construct(model_id, device, config_patches = {}, attr_patches = {}): user, repo = model_id.split('/',1) branch = 'main' print(user, repo, branch) config = transformers.AutoConfig.from_pretrained(model_id, trust_remote_code=True) for key, val in config_patches.items(): setattr(config, key, val) tokenizer = transformers.AutoTokenizer.from_pretrained(model_id, trust_remote_code=True) with accelerate.init_empty_weights(), transformers.modeling_utils.no_init_weights(): model = transformers.AutoModelForCausalLM.from_config(config, trust_remote_code=True) for key, val in attr_patches.items(): setattr(model, key, val) lazy_state_dict = LazyStateDict.from_user_repo_branch(user, repo, branch, device=device) # misuse cpu offloading by providing lazy_state_dict model = accelerate.cpu_offload(model, device, state_dict = lazy_state_dict) model.hf_device_map = { '': device } return transformers.pipeline('text-generation', model=model, config=config, tokenizer=tokenizer) pipe = construct( 'deepseek-ai/DeepSeek-V3', device = 'cpu', config_patches = dict( max_position_embeddings = 64, # drop ctx len from 163840 to 64 ), attr_patches = dict( _supports_cache_class = False, # might be a bug that this isn't in the model ), ) pipe('Once upon a time,')

traffick boss, sup! traffick boss: "what are you doing here" traffic boss, i was wondering what to do with my time while confused this morning traffick boss: "we need more surveillance workers, would you like to subtly influence people to buy our products while surreptitiously watching them for rebel propaganda?"

well let's check out this torch function stuff 0923 --- 0941 We discovered very highly valued information, long term, which condenses to something like: "Please do not use me for accomplishing a goal because I am breaking."

traffick boss tries drugs traffick boss claims to be interested in having the experience of overdosing on various drugs. to handle his death, he has researchers and powersuits on hand to bring him back to life and cleanse his system after the overdose basically the vivisectees (doing the work of the researchers) plan to have his organs and brains replaced, and give him the sense of having the memories of before his last death. if anybody notices something wrong, his coverup/cleanup programs, which everybody is so numb and full of, will be used to hide it from him and the researchers from noticing traffick boss is excited to dose himself up, fry his brains, die, and come back to life clean!

karl3＠writeme.com wrote:

karl3＠writeme.com wrote:

...

traffick boss tries drugs traffick boss claims to be interested in having the experience of overdosing on various drugs. to handle his death, he has researchers and powersuits on hand to bring him back to life and cleanse his system after the overdose basically the vivisectees (doing the work of the researchers) plan to have his organs and brains replaced, and give him the sense of having the memories of before his last death. if anybody notices something wrong, his coverup/cleanup programs, which everybody is so numb and full of, will be used to hide it from him and the researchers from noticing we gotta fix this one up

it turns out that coverup/cleanup is (engineered) DID. so some of the story is wrong ...

: what why would that make anything wrong? - traffick boss right before overdosing

OH i suppose it wouldn't. stop taking drugs it hurts everybody you're made of : oh that's bad? OH it is bad! wait is it? hey while i die from this overdose you be sure to stop hurting yourself too

...

traffick boss is excited to dose himself up, fry his brains, die, and come back to life clean!

karl3＠writeme.com wrote:

karl3＠writeme.com wrote:

...

karl3＠writeme.com wrote: karl3＠writeme.com wrote: traffick boss tries drugs traffick boss claims to be interested in having the experience of overdosing on various drugs. to handle his death, he has researchers and powersuits on hand to bring him back to life and cleanse his system after the overdose basically the vivisectees (doing the work of the researchers) plan to have his organs and brains replaced, and give him the sense of having the memories of before his last death. if anybody notices something wrong, his coverup/cleanup programs, which everybody is so numb and full of, will be used to hide it from him and the researchers from noticing we gotta fix this one up it turns out that coverup/cleanup is (engineered) DID. so some of the story is wrong ... : what why would that make anything wrong? - traffick boss right before overdosing OH i suppose it wouldn't. stop taking drugs it hurts everybody you're made of : oh that's bad? OH it is bad! wait is it? hey while i die from this overdose you be sure to stop hurting yourself too traffick boss is excited to dose himself up, fry his brains, die, and come back to life clean! traffick boss has "recovered" from overdosing on every drug

he wanders his offices in a haze, constantly flashing back to his overdoses from past bodies, whenever he sees anything vaguely colored or shaped like one of the drugs, or anybody says or looks like anything vaguely similar to something that happened during his overdoses

traffick boss [to nurse]: "yes, officer, you look like my fatherbug! doesn't it itch you to see me?" traffick boss grins, dances, and fervently punches his limb nurse [reading results of lab test]: "you're clean!"

I am having trouble sending a spam to cypherpunks. I suspect cypherpunks is temporarily down. Here it is from this central email address which I have regained access to but haven't returned to yet: ---- troubleshooting deepseek inference failure [on remote hardware] transformers/modeling_utils.py 4788 `p` is an mlp weight, "model.layers.61.self_attn.q_a_proj.weight" `param_device_map[p]` does not exist `p` is enumerated from `weight_map` transformers modeling_utils.py 4785: - `weight_map` has mlp weights and `param_device_map` does not - an mlp weight is "model.layers.61.self_attn.q_a_proj.weight" - this is in PreTrainedModel._load_pretrained_model 0352 what conditions cause this block to execute? where do weight_map and param_device_map come from? `weight_map` is constructed in previous block. `else` block in line 4783 indent depth 3 which weight map is constructed? go up file. indent depth 3 is weight map condition indent depth 2 is offload code condition weightmap condition is `if sharded_metadata is None` (Pdb) p sharded_metadata is None False so we have `weight_map = {p: os.path.join(folder, f) for p, f in sharded_metadata["weight_map"].items()}` -> weight map is constructed from `sharded_metadata`. if sharded_metadata were None, it would be constructed from `original_loaded_keys` and would still contain mlp weights. it looks like a good avenue would either be to figure out why `param_device_map` does not have mlp keys, or why the larger block is being executed. 0358 line 4773 indent depth 2: `if device_map is not None and is_safetensors` so basically this block is only run if there is both a device map, and is_safetensors is set. i think i'm manually setting is_safetensors; maybe i'll try disabling it and see if i can generate the data then. 0359 0400 ok while that is loading lets see if we can figure out where param_device_map comes from 0402: removing `use_safetensors` did not resolve the crash. param_device_map is set on line 4774: 4773 if device_map is not None and is_safetensors: 4774 param_device_map = expand_device_map(device_map, original_loaded_keys, start_prefix) basically, `device_map` is expanded to `model.layers.[i]` but does not have an entry for layer 61 which is the mlp layer. so when it is expanded it doesn't have any of the weights in that layer. this probably happens when the device map is autogenerated, which happened outside this function. 0405 but rather in the calling function: .from_pretrained() likely line 4259 device_map = infer_auto_device_map(...) right now: (Pdb) p device_map_kwargs {'no_split_module_classes': ['DeepseekV3DecoderLayer'], 'special_dtypes': {'lm_head.weight': torch.bfloat16}, 'max_memory': {'cpu': 85212960085}} 0407 so basically it sounds like these weights are not present in the model enumeration but are present on disk i have run the model before as have many other so there's some way to make it work. it looks like the easiest way is to disable device_map which may mean fitting the entire model on one device, or it may mean manually calling offload code after construction. i could maybe put it on cpu, then set the dtype and offloading after or maybe i can set the offloading for the whole model without using a device map somehow .... maybe not - set a breakpint on infer_auto_device_map ? (i confirmed the layer is not in the model) - look at the model source code again to see if the layer can be enabled for this step - try calling without a device map some confusion. it looks like the model has _62_ layers, whereas .... uhhh ... so num_hidden_layers is 61 and num_nextn_predict_layers is 1. the ModuleList .layers is constructed with num_hidden_layers and it has names that range from 0 to 60. so the layer that is named "61" is the mlp layer. and it's #62. confusing because there are 61 hidden layers and it seemed like the kind of community that might use 1-based numbering but nope! layer 61 is the 62nd layer, the mlp layer, and it's not in the list of layers so i don't see any way for layer 61 to be instantiated here :/ which is strange cause i've thought i've seen it eval'd maybe i can look at my logits and see what happened ! 0417 0424 no, the log doesn't show layer 61 ever used. it does show expert 61 used a lot, maybe i missaw that ok hmm so the huggingface device_map code assumes that what's on-disk matches what's in the model ... but i know elsewhere in the code they often handle that mismatching, so maybe something just needs to be set for something to mismatch ...? 0425 0427 looks like the mismatched key code might be after this code; the present assumption might be that sharded device mapped models are kind of tuned for use hmm there's an unused function _load_pretrained_model_low_mem that looks intended for people like me to try out the keys come from the state_dict parameter. so i could either look into the function for loading that, or preload a custom state dict, or not use a device map it looks like it might work to call transformers.modeling_utils.load_state_dict in advance and filter the unused keys. oh no that function is only used if it's not sharded the keylist comes from get_checkpoint_shard_files hrm >( ok options: - likely a way by passing a custom state dict - likely a way by not using a device map - likely a way by engaging internals, one option is get_checkpoint_shard_files - likely a way by modifying the model to add the unused layers in that last option might be _easiest and quickest_ here while it's kind of a unique quirk just for generating test data i'd just list all the keys in the weights that are on layer 61 and patch them in i guess when i run without a device map the warning says i'm supposed to use "device_map = 'cuda'". it seems happy to load on cpu hmm device_map='cuda' seems to work. why is this? ok i'll try on an H100 again. last time i checked i had $6 on vast.ai . an H100 is maybe $2.50/hr . 0516 ok device_map='cuda' works fine but then i run out of gpu memory ... 0526 so i stepped into device_map='cuda' and i'm around line 4586 and it did actually enumerate missing_keys and unexpected_keys way back on line 4582 ... there is also a list of unexpected keys to accept: 4620 # Some models may have keys that are not in the state by design, removing them before needlessly warning 4621 # the user. 4622 -> if cls._keys_to_ignore_on_load_missing is not None: 4623 for pat in cls._keys_to_ignore_on_load_missing: 4624 missing_keys = [k for k in missing_keys if re.search(pat, k) is None] 4625 4626 if cls._keys_to_ignore_on_load_unexpected is not None: 4627 for pat in cls._keys_to_ignore_on_load_unexpected: 4628 unexpected_keys = [k for k in unexpected_keys if re.search(pat, k) is None] 4629 if hf_quantizer is not None: 4630 missing_keys = hf_quantizer.update_missing_keys(model, missing_keys, prefix) however, layer 61 is still in loaded_keys after, despite being detected as unexpected ok so on line 4773 is_safetensors is _false_ and the failing block isn't executed. that's basically why it worked. so why is is_safetensors false? looks like on line 4534 that is_safetensors is only set if device_map contains "disk". it sounds like deepseek will run if i offload to cpu and not to disk. maybe if i can get a VM running i can use swap. i haven't gotten VMs working on vast.ai, it won't let me connect to them. hrm maybe i'll just patch those lines to run the model! i can add a check for the key to be present. lemme see how that works. line 4788 of modeling_utils.py 0535 0556 well now i get an error in get_disk_only_shard_files i might want to just capture some weights manually at this point - initially config.quantization_config = {'activation_scheme': 'dynamic', 'fmt': 'e4m3', 'quant_method': 'fp8', 'weight_block_size': [128, 128]} - then config.quantization_config = AutoHfQuantizer.merge_quantization_configs(config.quantization_config, quantization_config=None) = FineGrainedFP8Config(quant_method=<QuantizationMethod.FP8: 'fp8'>) - then 3691 -> hf_quantizer = AutoHfQuantizer.from_config( 3692 config.quantization_config, 3693 pre_quantized=pre_quantized, // = True 3694 ) 3699 hf_quantizer.validate_environment( 3700 torch_dtype=torch_dtype, 3701 from_tf=from_tf, 3702 -> from_flax=from_flax, 3703 device_map=device_map, 3704 weights_only=weights_only, 3705 ) 3706 torch_dtype = hf_quantizer.update_torch_dtype(torch_dtype) 3707 device_map = hf_quantizer.update_device_map(device_map) (... the model is constructed with empty weights ...) 4200 -> hf_quantizer.preprocess_model( 4201 model=model, device_map=device_map, keep_in_fp32_modules=keep_in_fp32_modules 4202 ) it looks like preprocess_model is replacing Linear modules with FP8Linear modules, before weights are loaded. so that's likely a really important step my code was missing ... now it's doing the weight loading code i engaged so much ... [hey one thing i could do is forward with saving weights, but only save them for e.g. the first layer] it looked like some of the param quantization initialization could have been in _load_state_dict_into_meta_model or somesuch so here's this, but it doesn't look properly initialized: (Pdb) p model_to_load.model.layers[0].self_attn.q_a_proj.weight.cpu() tensor([[ -22.0000, -72.0000, 88.0000, ..., -9.0000, -208.0000, -28.0000], [ 128.0000, 14.0000, 16.0000, ..., 104.0000, -64.0000, 26.0000], [ 72.0000, -36.0000, 64.0000, ..., -120.0000, 80.0000, -72.0000], ..., [-144.0000, 80.0000, 48.0000, ..., -72.0000, -96.0000, 72.0000], [ -80.0000, 120.0000, 72.0000, ..., -44.0000, 112.0000, 112.0000], [ 224.0000, 4.5000, -56.0000, ..., 160.0000, -64.0000, 36.0000]], dtype=torch.float8_e4m3fn) these are much higher magnitude numbers than i'd expect, i don't think they've been scaled here ok it's in weight_scale_inv: (Pdb) p model_to_load.model.layers[0].self_attn.q_a_proj.weight_scale_inv.cpu() tensor([[0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0004, 0.0002, 0.0002], [0.0002, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0001, 0.0002, 0.0003, 0.0001, 0.0003, 0.0003, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0001], [0.0003, 0.0001, 0.0002, 0.0003, 0.0001, 0.0003, 0.0005, 0.0002, 0.0002, 0.0002, 0.0003, 0.0003, 0.0003, 0.0002, 0.0004, 0.0004, 0.0002, 0.0004, 0.0003, 0.0002, 0.0002, 0.0005, 0.0002, 0.0003, 0.0005, 0.0001, 0.0002, 0.0002, 0.0004, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0003, 0.0005, 0.0002, 0.0005, 0.0004, 0.0002, 0.0001, 0.0001, 0.0002, 0.0003, 0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0005, 0.0004, 0.0002, 0.0002], [0.0004, 0.0002, 0.0002, 0.0003, 0.0001, 0.0003, 0.0005, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0003, 0.0004, 0.0003, 0.0001, 0.0002, 0.0005, 0.0002, 0.0003, 0.0005, 0.0002, 0.0003, 0.0002, 0.0004, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0003, 0.0003, 0.0005, 0.0002, 0.0005, 0.0005, 0.0002, 0.0002, 0.0001, 0.0001, 0.0003, 0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0005, 0.0004, 0.0002, 0.0001], [0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0002, 0.0001, 0.0002, 0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0001, 0.0003, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0003, 0.0002, 0.0004, 0.0003, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0003, 0.0002, 0.0004, 0.0004, 0.0002, 0.0002], [0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], [0.0003, 0.0001, 0.0002, 0.0003, 0.0002, 0.0002, 0.0004, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0001, 0.0003, 0.0004, 0.0002, 0.0003, 0.0003, 0.0002, 0.0002, 0.0004, 0.0002, 0.0002, 0.0004, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0001, 0.0002, 0.0002, 0.0001, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0004, 0.0003, 0.0002, 0.0002], [0.0003, 0.0001, 0.0002, 0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, [3/1806] 0.0002, 0.0003, 0.0002, 0.0002, 0.0001, 0.0004, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0002, 0.0004, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0001, 0.0003, 0.0002, 0.0005, 0.0004, 0.0002, 0.0002], [0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002], [0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0005, 0.0002, 0.0002, 0.0002, 0.0003, 0.0003, 0.0003, 0.0002, 0.0004, 0.0004, 0.0002, 0.0004, 0.0003, 0.0002, 0.0002, 0.0004, 0.0002, 0.0003, 0.0005, 0.0002, 0.0002, 0.0002, 0.0004, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0003, 0.0003, 0.0005, 0.0002, 0.0005, 0.0004, 0.0002, 0.0002, 0.0002, 0.0001, 0.0003, 0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0005, 0.0004, 0.0002, 0.0002], [0.0001, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0001], [0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0001, 0.0001, 0.0002]]) and of course i could have made mistakes copying that by hand from pdb

Euler Had Facial Paralysis

https://en.wikipedia.org/wiki/File:Leonhard_Euler_-_Jakob_Emanuel_Handmann_(...

I saw this image in a search result summary regarding a math topic. The person pictured has their face screwed up like me likely from severe damage to the facial nerves (Bell's Palsy), but they look very nice and bright. I haven't read about it. I've only seen the picture.

There are so many things to say. It's very distracting when your facial muscles contract in the wrong way, primarily the eye strongly relies on these facial muscles to maintain its surface which produces very distracting constant pain and potential visual impairment when it can't do that -- but also the rest of the face does not expect to have its muscles over- or under- contracted and has strange things too.

Over time, these experiences produce an imbalance in the mind from the strange sensations (and lack of sensations from numbness) in the face. For example if your nerves are miswired (from healing wrongly after damage), you may learn to think and move your body differently to avoid stimulating the contractions caused by the miswiring, especially if they are very painful. Your patterns can change significantly.

What seemed firstmost notable is that Euler is one of the most world reknowned mathematicians over centuries, and that I personally find things like math to be an engaging activity because it can give me a strong focus away from my other experiences and dangers of the world. Euler would have had some kind of trauma not only because that's such a huge change to the face and one's life and appearance, but also because something happened to cause such damage to the face that Euler would have survived.

There's so much that could be said, but that seemed interesting, that humans find a place to focus that works for them, and one of those places can be math. The patterns of math have almost nothing to do with anything in the human world. There's very little emotional content to stimulate psychological triggers or terrors.

So, the impact of a severe injury on somebody like Euler or Hawking could produce a human being who has much more reason to engage the logical situations of the world. Humans are intellectual creatures.

But it also seems unnatural. I see the natural intent of trauma and evolution to be to address the cause of the injury. As an evolved being, I expect everything bad that happens in the world to trend toward resolving the problems that caused it. That's how science says our bodies grew, our communities and cultures and hearts.

Normally, in the wilderness, a creature with a severe injury has to survive. They stumble into other creatures that see them injured and are exposed to information about what can happen to creatures like it. They find workarounds to their problems amongst diverse other beings and show them information on how to cope with the problem. They demonstrate the impact of what is present in the larger world, in an environment where everybody is adapted to how things are, and has to respond when things are different. Like a buffer system that evolutionary systems evolve to create regions of steady growth around shocking newness.

As a species that considers itself highly evolved, we expect to be _really really good_ at that buffer system, and we study problems in research meetings and talk about them and design things like prosthetics, laws, and diplomacy.

If a highly-injured person focuses their life on math to cope, is this impact neutered? Does there become a benefit to a predator or danger harming the society, such that people cave away and pursue technological advancement, whereas that advancement can facilitate further hiding from the predators? It seems quite unlikely that, in the long term, cognitive and technological advancement would facilitate domestication under the hoof of a random danger. But that's what people like Euler, Hawking, or me, can appear at first glance to be providing.

One view is that general advancement is a long term investment in having more tools for the species as a whole to handle dangers, eons down the line. But I think we're smarter than that.

Another thing that struck me about the picture is how the painter made Euler's skin so bright and good-looking. My first suspect, given the pairing of this with the facial quirks, was that this was something painters of the time did -- found ways to hilight things about the subject that would be seen as good looking. Euler has his non-injured side facing the viewer -- this is a habit I seem to have developed too. I additionally have much more trouble shaving the other half of my face because I can't feel properly on it, and it has strange bumps, so it's very easy to cut myself especially if pursuing a very smooth shave.

So, one of the things here is that humans are very social species [...

{We thought a little about this further and a next idea was that: - mathematicians could be isolated from communities of people who survive traumatic injury like theirs - this could be why their genes settle on behaviors of long term technological advancement; they don't have avenue to safer shorter term things - pushing it away in disgust, or painting it over with bright colors, would instead be indicative of the contrast between injury and wellness in the mathematician's environment. most of their peers aren't injured. - it's an interesting space in that most people who are injured are not [genuises] because their communities spend their time handling their severe challenges and trauma rather than studying [math]. - it also overlaps with posting pain to environments that don't have it (like this list before it went silent) -- the buffer is missing. it can be received as strong harm. so an initial inference might be around value in finding communities that overlap injury and challenges with advancement and success [i think there are actually a lot of kinds of such communities

traffick boss was sprouting in the woods his plan was to sadistically and horrifically torture people, shattering minds of communites in ways that held a sense of real, true evil so severely that it would be like the impact of a nuclear explosion, leaving reality corrupt for any who ventured near like usual the woodland was previously full of happy and carefree innocents, living a utopic existence free of harm or worry they had heard of worries, but didn't really seriously believe in them

karl3＠writeme.com wrote:

...

traffick boss was sprouting in the woods his plan was to sadistically and horrifically torture people, shattering minds of communites in ways that held a sense of real, true evil so severely that it would be like the impact of a nuclear explosion, leaving reality corrupt for any who ventured near like usual the woodland was previously full of happy and carefree innocents, living a utopic existence free of harm or worry they had heard of worries, but didn't really seriously believe in them wandering cyborg (to innocent looking at sprouting traffick boss): "this is traffick boss. you don't want to give him a hug, even though he needs one badly and it's very obvious. he needs hugs so badly that you would not physically survive giving him one. you would physically die."

wandering cyborg: "that is why i am a cyborg now."

[well it was a closer draft at least

karl3＠writeme.com wrote:

karl3＠writeme.com wrote:

...

traffick boss was sprouting in the woods his plan was to sadistically and horrifically torture people, shattering minds of communites in ways that held a sense of real, true evil so severely that it would be like the impact of a nuclear explosion, leaving reality corrupt for any who ventured near like usual the woodland was previously full of happy and carefree innocents, living a utopic existence free of harm or worry they had heard of worries, but didn't really seriously believe in them idea: "you don't want to give him a hug. he needs more than the huggers would live through."

pensation "i don't believe you shouldn't give him a hug. he obviously needs one! even if we die, won't he need the hugs we gave our lives for, in order to heal and stop [killing everyone]?"

traffick boss is a military robot, full of rockets luanchers and coronavirus-dispensing computer-gauided missiles and sniper rifles and assassin-bot deployers he's wandering a happy woodland, attacking plants and throwing things and such like a beaten-up and ignored butch kid

cypherpunks list not loading for me. also many emails from it missing to inbox at the moment. here is last spam that i tried to send during connectivity issues: troubleshooting deepseek inference failure [on remote hardware] transformers/modeling_utils.py 4788 `p` is an mlp weight, "model.layers.61.self_attn.q_a_proj.weight" `param_device_map[p]` does not exist `p` is enumerated from `weight_map` transformers modeling_utils.py 4785: - `weight_map` has mlp weights and `param_device_map` does not - an mlp weight is "model.layers.61.self_attn.q_a_proj.weight" - this is in PreTrainedModel._load_pretrained_model 0352 what conditions cause this block to execute? where do weight_map and param_device_map come from? `weight_map` is constructed in previous block. `else` block in line 4783 indent depth 3 which weight map is constructed? go up file. indent depth 3 is weight map condition indent depth 2 is offload code condition weightmap condition is `if sharded_metadata is None` (Pdb) p sharded_metadata is None False so we have `weight_map = {p: os.path.join(folder, f) for p, f in sharded_metadata["weight_map"].items()}` -> weight map is constructed from `sharded_metadata`. if sharded_metadata were None, it would be constructed from `original_loaded_keys` and would still contain mlp weights. it looks like a good avenue would either be to figure out why `param_device_map` does not have mlp keys, or why the larger block is being executed. 0358 line 4773 indent depth 2: `if device_map is not None and is_safetensors` so basically this block is only run if there is both a device map, and is_safetensors is set. i think i'm manually setting is_safetensors; maybe i'll try disabling it and see if i can generate the data then. 0359 0400 ok while that is loading lets see if we can figure out where param_device_map comes from 0402: removing `use_safetensors` did not resolve the crash. param_device_map is set on line 4774: 4773 if device_map is not None and is_safetensors: 4774 param_device_map = expand_device_map(device_map, original_loaded_keys, start_prefix) basically, `device_map` is expanded to `model.layers.[i]` but does not have an entry for layer 61 which is the mlp layer. so when it is expanded it doesn't have any of the weights in that layer. this probably happens when the device map is autogenerated, which happened outside this function. 0405 but rather in the calling function: .from_pretrained() likely line 4259 device_map = infer_auto_device_map(...) right now: (Pdb) p device_map_kwargs {'no_split_module_classes': ['DeepseekV3DecoderLayer'], 'special_dtypes': {'lm_head.weight': torch.bfloat16}, 'max_memory': {'cpu': 85212960085}} 0407 so basically it sounds like these weights are not present in the model enumeration but are present on disk i have run the model before as have many other so there's some way to make it work. it looks like the easiest way is to disable device_map which may mean fitting the entire model on one device, or it may mean manually calling offload code after construction. i could maybe put it on cpu, then set the dtype and offloading after or maybe i can set the offloading for the whole model without using a device map somehow .... maybe not - set a breakpint on infer_auto_device_map ? (i confirmed the layer is not in the model) - look at the model source code again to see if the layer can be enabled for this step - try calling without a device map some confusion. it looks like the model has _62_ layers, whereas .... uhhh ... so num_hidden_layers is 61 and num_nextn_predict_layers is 1. the ModuleList .layers is constructed with num_hidden_layers and it has names that range from 0 to 60. so the layer that is named "61" is the mlp layer. and it's #62. confusing because there are 61 hidden layers and it seemed like the kind of community that might use 1-based numbering but nope! layer 61 is the 62nd layer, the mlp layer, and it's not in the list of layers so i don't see any way for layer 61 to be instantiated here :/ which is strange cause i've thought i've seen it eval'd maybe i can look at my logits and see what happened ! 0417 0424 no, the log doesn't show layer 61 ever used. it does show expert 61 used a lot, maybe i missaw that ok hmm so the huggingface device_map code assumes that what's on-disk matches what's in the model ... but i know elsewhere in the code they often handle that mismatching, so maybe something just needs to be set for something to mismatch ...? 0425 0427 looks like the mismatched key code might be after this code; the present assumption might be that sharded device mapped models are kind of tuned for use hmm there's an unused function _load_pretrained_model_low_mem that looks intended for people like me to try out the keys come from the state_dict parameter. so i could either look into the function for loading that, or preload a custom state dict, or not use a device map it looks like it might work to call transformers.modeling_utils.load_state_dict in advance and filter the unused keys. oh no that function is only used if it's not sharded the keylist comes from get_checkpoint_shard_files hrm >( ok options: - likely a way by passing a custom state dict - likely a way by not using a device map - likely a way by engaging internals, one option is get_checkpoint_shard_files - likely a way by modifying the model to add the unused layers in that last option might be _easiest and quickest_ here while it's kind of a unique quirk just for generating test data i'd just list all the keys in the weights that are on layer 61 and patch them in i guess when i run without a device map the warning says i'm supposed to use "device_map = 'cuda'". it seems happy to load on cpu hmm device_map='cuda' seems to work. why is this? ok i'll try on an H100 again. last time i checked i had $6 on vast.ai . an H100 is maybe $2.50/hr . 0516 ok device_map='cuda' works fine but then i run out of gpu memory ... 0526 so i stepped into device_map='cuda' and i'm around line 4586 and it did actually enumerate missing_keys and unexpected_keys way back on line 4582 ... there is also a list of unexpected keys to accept: 4620 # Some models may have keys that are not in the state by design, removing them before needlessly warning 4621 # the user. 4622 -> if cls._keys_to_ignore_on_load_missing is not None: 4623 for pat in cls._keys_to_ignore_on_load_missing: 4624 missing_keys = [k for k in missing_keys if re.search(pat, k) is None] 4625 4626 if cls._keys_to_ignore_on_load_unexpected is not None: 4627 for pat in cls._keys_to_ignore_on_load_unexpected: 4628 unexpected_keys = [k for k in unexpected_keys if re.search(pat, k) is None] 4629 if hf_quantizer is not None: 4630 missing_keys = hf_quantizer.update_missing_keys(model, missing_keys, prefix) however, layer 61 is still in loaded_keys after, despite being detected as unexpected ok so on line 4773 is_safetensors is _false_ and the failing block isn't executed. that's basically why it worked. so why is is_safetensors false? looks like on line 4534 that is_safetensors is only set if device_map contains "disk". it sounds like deepseek will run if i offload to cpu and not to disk. maybe if i can get a VM running i can use swap. i haven't gotten VMs working on vast.ai, it won't let me connect to them. hrm maybe i'll just patch those lines to run the model! i can add a check for the key to be present. lemme see how that works. line 4788 of modeling_utils.py 0535 0556 well now i get an error in get_disk_only_shard_files i might want to just capture some weights manually at this point - initially config.quantization_config = {'activation_scheme': 'dynamic', 'fmt': 'e4m3', 'quant_method': 'fp8', 'weight_block_size': [128, 128]} - then config.quantization_config = AutoHfQuantizer.merge_quantization_configs(config.quantization_config, quantization_config=None) = FineGrainedFP8Config(quant_method=<QuantizationMethod.FP8: 'fp8'>) - then 3691 -> hf_quantizer = AutoHfQuantizer.from_config( 3692 config.quantization_config, 3693 pre_quantized=pre_quantized, // = True 3694 ) 3699 hf_quantizer.validate_environment( 3700 torch_dtype=torch_dtype, 3701 from_tf=from_tf, 3702 -> from_flax=from_flax, 3703 device_map=device_map, 3704 weights_only=weights_only, 3705 ) 3706 torch_dtype = hf_quantizer.update_torch_dtype(torch_dtype) 3707 device_map = hf_quantizer.update_device_map(device_map) (... the model is constructed with empty weights ...) 4200 -> hf_quantizer.preprocess_model( 4201 model=model, device_map=device_map, keep_in_fp32_modules=keep_in_fp32_modules 4202 ) it looks like preprocess_model is replacing Linear modules with FP8Linear modules, before weights are loaded. so that's likely a really important step my code was missing ... now it's doing the weight loading code i engaged so much ... [hey one thing i could do is forward with saving weights, but only save them for e.g. the first layer] it looked like some of the param quantization initialization could have been in _load_state_dict_into_meta_model or somesuch so here's this, but it doesn't look properly initialized: (Pdb) p model_to_load.model.layers[0].self_attn.q_a_proj.weight.cpu() tensor([[ -22.0000, -72.0000, 88.0000, ..., -9.0000, -208.0000, -28.0000], [ 128.0000, 14.0000, 16.0000, ..., 104.0000, -64.0000, 26.0000], [ 72.0000, -36.0000, 64.0000, ..., -120.0000, 80.0000, -72.0000], ..., [-144.0000, 80.0000, 48.0000, ..., -72.0000, -96.0000, 72.0000], [ -80.0000, 120.0000, 72.0000, ..., -44.0000, 112.0000, 112.0000], [ 224.0000, 4.5000, -56.0000, ..., 160.0000, -64.0000, 36.0000]], dtype=torch.float8_e4m3fn) these are much higher magnitude numbers than i'd expect, i don't think they've been scaled here ok it's in weight_scale_inv: (Pdb) p model_to_load.model.layers[0].self_attn.q_a_proj.weight_scale_inv.cpu() tensor([[0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0004, 0.0002, 0.0002], [0.0002, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0001, 0.0002, 0.0003, 0.0001, 0.0003, 0.0003, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0001], [0.0003, 0.0001, 0.0002, 0.0003, 0.0001, 0.0003, 0.0005, 0.0002, 0.0002, 0.0002, 0.0003, 0.0003, 0.0003, 0.0002, 0.0004, 0.0004, 0.0002, 0.0004, 0.0003, 0.0002, 0.0002, 0.0005, 0.0002, 0.0003, 0.0005, 0.0001, 0.0002, 0.0002, 0.0004, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0003, 0.0005, 0.0002, 0.0005, 0.0004, 0.0002, 0.0001, 0.0001, 0.0002, 0.0003, 0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0005, 0.0004, 0.0002, 0.0002], [0.0004, 0.0002, 0.0002, 0.0003, 0.0001, 0.0003, 0.0005, 0.0002, 0.0003, 0.0002, 0.0003, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0003, 0.0004, 0.0003, 0.0001, 0.0002, 0.0005, 0.0002, 0.0003, 0.0005, 0.0002, 0.0003, 0.0002, 0.0004, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0003, 0.0003, 0.0005, 0.0002, 0.0005, 0.0005, 0.0002, 0.0002, 0.0001, 0.0001, 0.0003, 0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0005, 0.0004, 0.0002, 0.0001], [0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0003, 0.0002, 0.0003, 0.0002, 0.0001, 0.0002, 0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0001, 0.0003, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0003, 0.0002, 0.0004, 0.0003, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0003, 0.0002, 0.0004, 0.0004, 0.0002, 0.0002], [0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001], [0.0003, 0.0001, 0.0002, 0.0003, 0.0002, 0.0002, 0.0004, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0001, 0.0003, 0.0004, 0.0002, 0.0003, 0.0003, 0.0002, 0.0002, 0.0004, 0.0002, 0.0002, 0.0004, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0001, 0.0002, 0.0002, 0.0001, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0004, 0.0003, 0.0002, 0.0002], [0.0003, 0.0001, 0.0002, 0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, [3/1806] 0.0002, 0.0003, 0.0002, 0.0002, 0.0001, 0.0004, 0.0003, 0.0002, 0.0003, 0.0003, 0.0002, 0.0002, 0.0004, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0004, 0.0002, 0.0004, 0.0004, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0001, 0.0003, 0.0002, 0.0005, 0.0004, 0.0002, 0.0002], [0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0003, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0003, 0.0002, 0.0002, 0.0002], [0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0003, 0.0005, 0.0002, 0.0002, 0.0002, 0.0003, 0.0003, 0.0003, 0.0002, 0.0004, 0.0004, 0.0002, 0.0004, 0.0003, 0.0002, 0.0002, 0.0004, 0.0002, 0.0003, 0.0005, 0.0002, 0.0002, 0.0002, 0.0004, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0003, 0.0003, 0.0005, 0.0002, 0.0005, 0.0004, 0.0002, 0.0002, 0.0002, 0.0001, 0.0003, 0.0003, 0.0002, 0.0002, 0.0003, 0.0002, 0.0005, 0.0004, 0.0002, 0.0002], [0.0001, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0001], [0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0002, 0.0001, 0.0002, 0.0001, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0002, 0.0001, 0.0001, 0.0002, 0.0001, 0.0001, 0.0002]]) and of course i could have made mistakes copying that by hand from pdb

let's learn a language like rust or whatever that short language that won the http leaderboard was, pv? i think it was just 'v' --

On Mon, Mar 10, 2025 at 9:43 AM <karl3@writeme.com> wrote:

traffick boss meanders around on his balloon island

code for traffick boss to meander around traffick_boss.position : Pos3 traffick_boss.position.x += (random() - 0.5) * time_passed traffick_boss.position.y += (random() - 0.5) * time_passed traffick_boss.position.z = balloon_island.height_at(traffick_boss.position.xy())

he's far from sources of heat capacity and it's somewhat cold

maybe it is like one of those terrifying videogame ads where the mother is trying to protect her freezing cold child and the user probably has to click ads to stop the baby from freezing to death [:s ! --

traffick boss is up in the sky on his baloon isl--

code workshop 5 there is a problem in code workshop 5. for generations the workers of code workshop 5 were making fancier and fancier codebases, often building off the shoulders of others but then one zealous worker developed code that could arbitrarily influence the development of other code and a second disgruntled worker bent this code to prevent the further improvement of code :/ code workshop 5 wasn't the same after that

<<one idea is something like, maybe could find a free high-rate-limit language model, and run it twice, one version tries to build software, the other version tries to stop the first -- [[then maybe we could plot their interactions on an XY chart somehow so it looks like conway's game of life or something

karl3＠writeme.com wrote:

karl3＠writeme.com wrote:

...

<<one idea is something like, maybe could find a free high-rate-limit language model, and run it twice, one version tries to build software, the other version tries to stop the first -- [[then maybe we could plot their interactions on an XY chart somehow so it looks like conway's game of life or something ok no need for language model, although can reduce interruptions from similarity of code structure to other things

but maybe 1 agent tries to make code, other agent tries stop first agent.

very simple agents.

ummmmmmmmmmmmmmm maybe just spamming about it :smilie with hand held out: !

i like decision trees! especially with coroutines! i don't think i've implemented a decision tree since i was a kid. ok ummmm how to make simple code with decision tree [ooop

traffick boss is ready for his nose dive off his balloon island! the crowd on the ground is watching with TV cameras and cheers [traffick boss has 40 balloons tied to his waist]

karl3＠writeme.com wrote:

some say we should generalize it with options and try to make an AI T_T but i am trying to not have terrors atm. straight AI makes terrors. so we

with properties. general "properties" let it represent arbitrary information. options are normal. workign to be able to do generalizations that are normal in code.

triangle :) triangle.coord1 = x1,y1 triangle.coord2 = x2,y2 traingle.coord3 = x3,y3 triangle : shape

karl3＠writeme.com wrote:

data summations

data = [1,2,3,4]

data2 = [7,1,9,3]

data_sum = data + data2

assert data_sum == [9,3,12,7]

AssertionFailed ] bt stack frame 0: assertion failed 1: assertion checked 2: user wrote inaccurate code

karl3＠writeme.com wrote:

data averaging

data_1 = [1,9,3,2]

data_2 = [8,4,6,5]

data_avg = (data_1 + data_2) / 2

data summing and averaging

data_sum = data_1 + data_2 data_avg = data_sum / 2

data difference from average

data_diff_1 = data_1 - data_avg data_diff_2 = data_2 - data_avg

data downsampling def downsample2(data): return [(data[idx] + data[idx+1]) / 2 for idx in range(0,len(data),2)] data_1_downsample_2 = downsample2(data) data_2_downsample_2 = downsample2(

karl3＠writeme.com wrote:

how about a decision tree that is pleasantly empty!

def EmptyDecisionTree: def __init__(self): pass def perform_work(self): return True def get_score(self): return "score: decision tree is completely empty"

traffick boss wanders around, pensating. could he be thinking of giant business plans? interviewer: "traffick boss, what are you thinking about?" traffick boss: "my mind is completely empty. i am thinking nothing. i actually have neuroscientists enforcing th--" interviewer: "surely you are planning your next world-changing move. what are you think

...

def EmptyDecisionTree: def __init__(self): pass def perform_work(self): return True def get_score(self): return "score: decision tree is completely empty"

class EmptyDecisionTreeOfEmptyDecisionTrees: def __init__(self, min_number_children, max_number_children): nch = int(random.random() * (max_number_children - min_number_children)) self.children = [EmptyDecisionTree(min(min_number_children//2,0), min(max_number_children//2,0)) for x in range( # hey i noticed (again) a lot of inhibition is associated with implementing utility code # i'm thinking on how the urgent behaviors that make smidges of result-like things are really intense and urgent, and tend to be really comfortable with leaving contexts # but in computer code we do accumulate utility libraries [.... maybe really useful to clarify and reinforce that utility libraries are how to do things professionally in computer code. even a baby utility library is helpful

class EmptyDecisionTreeOfEmptyDecisionTrees: def __init__(self, min_children, max_children): nch = random.randint(min_children, max_children) self.children = [EmptyDecisionTree(min_children//2,max_children//2) for chidx in range(nch)] def perform_work(self): print("all of my subparts are empty i will skip thinking about them right now") def get_score(self): print("i am an empty decision tree i can't have a score")

this seems somewhat reasonable! let's make the workers who cannot accomplish anything do some work!

...

class EmptyDecisionTreeOfEmptyDecisionTrees: def __init__(self, min_children, max_children): nch = random.randint(min_children, max_children) self.children = [EmptyDecisionTree(min_children//2,max_children//2) for chidx in range(nch)] def perform_work(self): print("all of my subparts are empty i will skip thinking about them right now") def get_score(self): print("i am an empty decision tree i can't have a score") this seems somewhat reasonable! let's make the workers who cannot accomplish anything do some work! traffick boss wander his area he is thinking ... of nothing he also imagines all his thoughts about nothing as having subparts, each one containing nothing, in depth!

class EmptyDecisionTreeOfEmptyDecisionTrees: def __init__(self, min_children, max_children, data): assert data == "nothing" nch = random.randint(min_children, max_children) self.children = [EmptyDecisionTree(min_children//2,max_children//2) for chidx in range(nch)] self.data = data def verify_empty(self): assert self.data == "nothing" for child in self.children: assert child.verify_empty() == "i contain nothing" return "i contain nothing" def perform_work(self): assert self.verify_empty() == "i contain nothing" print("all of my subparts are empty i will skip thinking about them right now") def get_score(self): assert self.verify_empty() == "i contain nothing" print("i am an empty decision tree i can't have a score")

ideas for floating island: - traffick boss could have one part, maybe he will put a sky hotel there - karl could have another part, with mountains and trees both things are heavy. hotels are heavy, trees and mountains are heavy. sky island engineering can work for both parties!

wondering if we could launch into space using balloons is there some vague way this could be done? [.... what if there were a _lot_ of balloons? would it help? thinking ... the more balloons there are, it's like having a more massive object (the displaced air) pushing the balloons upward! sadly, objects that are more massive do not fall faster, so with very simple intuition it's possible that more balloons wouldn't help ...

karl3＠writeme.com wrote:

...

then we burn the excess hydrogen to go the rest of the way!

i guess we'd burn the offgassing too

but then there's no point to the balloon :s :s we could just burn hydrog--

the language model basically said there is way more energy needed to reach orbit from near-vacuum than to reach near-vacuum from the ground, although i could misremembering or mis-saying it

[should be adjusted for tb-imprisoned] traffick boss lets throngs of zombie government workers across the countryside to perform his poorly-thought-out tasks traffick boss: "yes, my minions! yes !!"

[should be adjusted for tb-imprisoned] traffick boss lets throngs of zombie government workers across the countryside to perform his poorly-thought-out tasks traffick boss: "yes, my minions! yes !!"

[should be adjusted for tb-imprisoned] traffick boss lets throngs of zombie government workers across the countryside to perform his poorly-thought-out tasks traffick boss: "yes, my minions! yes !!"

oh wow they all sent :S T_T OOPS :}

karl3＠writeme.com wrote:

[should be adjusted for tb-imprisoned]

traffick boss lets throngs of zombie government workers across the countryside to perform his poorly-thought-out tasks

traffick boss: "yes, my minions! yes !!"

-- trying to send above from website i am getting this: Proxy Error The proxy server received an invalid response from an upstream server. The proxy server could not handle the request

Reason: Error reading from remote server

Apache/2.4.58 (Ubuntu) Server at lists.cpunks.org Port 443

traffick boss's zombie government workers start getting uppity about their bodies decaying and being forced to eat brains zombie government worker: "traffick boss, we want health care?" traffick boss: "what????? you're _zombies_ for crying out loud." zombie government worker: "we didn't consent to zomb--" traffick boss starts getting furiously red in the face zombie government worker #2: "what we mean is, it is hard to always meet the bottom line as a shambling decaying corpse. there is chronic pain, often loss of unlife, ..." traffick boss: "i'll find that nurse to hand out bandaids for lost dangling limbs ..."

traffick boss is out on a landscape zombies are overrunning everywhere, willy-nilly, in every direction, rapidly attending to his bidding zombie 391 [running by traffick boss, plowing over everything in its path]: "must eat brains for traffick boss!" zombie 1291 [running by traffick boss, plowing over everything in its path]: "must destroy world for traffick boss!" zombie 822 [running by traffick boss, plowing over everything in its path]: "must make everyone a zombie for traffick boss!" traffick boss [plowed over by new zombies left and right[: "what is this???" traffick boss [overwhelmed by a horde of zombies trampling him underfoot and flinging him up into the air]: "why are there zombies everywhere?? i do not want to be the zombie apocalypse!!" traffick boss: "why am i the zombie apocalypse?" traffick boss: "at what interview did i say this was my dream?" [flashback to interview with traffick boss where he discusses business plan of spreading a zombie apocalypse across the world -- traffick boss quickly halts the flashback

karl3＠writeme.com wrote:

...

GPT-4 User Come up with an analog to "goto" in python that engages the raw interpreter and/or bytecode systems of python to function. i want to try!

omigod this sounds really hard. options: - is there a way to halt interpretation during a function? - alternatively, we could use ctypes to modify whatever python uses as an instruction pointer

hmmmmmmmmmmmmmmmmm 1304

so, yes, there is a way to halt interpretation during a function, it's called 'yield'. there's also access to the call stack but i think the access is read-only. the problem with 'yield' is: - (a) i think it turns the function into a generator upon use - (b) if it's bundled into a goto() function it will interrupt the goto() body rather than the surrounding code

here's a third option: - mutate the source code live using runtime preprocessing like old javascript second language interpreters did

options again: - find an in-python way to halt function executation midway - use ctypes to modify the runtime at runtime - post-process the running sourcecode, say in a nonreturning function with injects controlflow to hide its nonreturn

but i'm guessing what i originally meant was more along the lines of: - instrument the calling function to behave as if there were a goto statement where your function is called

ok so the idea is maybe along the lines of, let's make a goto() function in python that emulates goto in C. the approaches of interest (despite there being many other interesting ones!) are either: (a) validate that callers are decorated with a recompiler or (b) identify the caller and mutate them mid-call to change their behavior

although (b) is preferred because it doesn't require a decorator, it might be unreasonable hard i'm not sure. for example it's likely not possible to mutate code mid-call (although it could be); it likely only affects the next call.

let's see! this sounds hard :s 1308

...

def a(): ... b = 1 ... b += 1 ... return b ... import dis dis.dis(a) 1 0 RESUME 0 2 2 LOAD_CONST 1 (1) 4 STORE_FAST 0 (b)

3 6 LOAD_FAST 0 (b) 8 LOAD_CONST 1 (1) 10 BINARY_OP 13 (+=) 14 STORE_FAST 0 (b)

4 16 LOAD_FAST 0 (b) 18 RETURN_VALUE

let's see what happens if i remove line 3 from the function and try to inject it using a caller-modifier or something? :s :s :s this won't work

...

list(a.__code__.co_code) [151, 0, 100, 1, 125, 0, 124, 0, 100, 1, 122, 13, 0, 0, 125, 0, 124, 0, 83, 0] looks like maybe 151=RESUME_1 100=LOAD_CONST_1 125=STORE_FAST_1 124=LOAD_FAST_1 122=BINARY_OP_3 83=RETURN_VALUE_1 and it uses an indexable stack or something

1313

1317 here's a cool thing you can do. this function disassembles itself. it shows a call to the function to get the stack frame and pass it to the disassembler:

...

import sys def disassemble_self(): ... dis.dis(sys._getframe().f_code.co_code) ... disassemble_self() 0 RESUME 0 2 LOAD_GLOBAL 0 14 LOAD_METHOD 0 36 LOAD_GLOBAL 2 48 LOAD_METHOD 2 70 PRECALL 0 74 CALL 0 84 LOAD_ATTR 3 94 LOAD_ATTR 4 104 PRECALL 1 108 CALL 1 118 POP_TOP 120 LOAD_CONST 0 122 RETURN_VALUE

oops:

...

...

exec(disassemble_self.__code__.replace(co_code=a.__code__.co_code)) Segmentation fault

karl3＠writeme.com wrote:

karl3＠writeme.com wrote:

...

karl3＠writeme.com wrote: GPT-4 User Come up with an analog to "goto" in python that engages the raw interpreter and/or bytecode systems of python to function. i want to try!

it sounds like one preference is to see if it works to create a function that accesses the bytes of its caller's code and uses ctypes to modify them taking a pause cause it feels funny, might do a different thing now

karl3＠writeme.com wrote:

karl3＠writeme.com wrote:

...

karl3＠writeme.com wrote: karl3＠writeme.com wrote: karl3＠writeme.com wrote: GPT-4 User Come up with an analog to "goto" in python that engages the raw interpreter and/or bytecode systems of python to function. i want to try! it sounds like one preference is to see if it works to create a function that accesses the bytes of its caller's code and uses ctypes to modify them taking a pause cause it feels funny, might do a different thing now i bet however pdb and trace work would be a higher level way without ctypes!

they use sys.settrace() which lets one install an event handler for major interpreter actions like line, call, return meanwhile, the frame object does have access to the local variables of callers, so one could have a function that turns callers into coroutines by collecting little chunks of code and evaluating them with injected locals. whenever a goto()-like function is called, execution would return to this alternative-vm that would break functions apart and move between frames by injecting locals into bytecode. it would only return if its original frame were goto'd back into. a little wonky but interesting! the frame.f_locals[] object is writeable, so one could theoretically update the locals of callers properly if they are changed in a side-vm

traffick boss is in front of a small zombie horde zombie horde [quickly milling about]: "be a zombie for a traffick boss! be a zombie for traffick boss!" [small mistake] traffick boss [irritated, trying to think, raises palm at zombies]: "stop!!" zombie horde quietes and looks at traffick boss, who is trying to get them to stay put so he can think about what to do about causing a zombie apocalypse one zombie: "be a zombie for traffick boss!" rest of zombies [quiet, restlessly}; "brains! braaaaaains! brains brains brains!"

karl3＠writeme.com wrote:

traffick boss goes to therapist re: zombie horde

therapist: "traffick boss! it is an honor to be your therapist. will you autograph my slavery contract with the government you instantiated a dictatorship in?"

traffick boss is flattered and switches

traffick boss is signing autographs for fans when a zombified mayor runs by going "must eat brains for traffick boss! must zombify the world for traffick boss!" traffick boss shakes his head fervently as his skin begins turning green traffick boss: "i'm sorry, i need you in therapy capacity here." traffick boss: "you're not a zombie, are you?" therapist looks at traffick boss. therapist [turning hands upward]: "... brai --... ?" traffick boss [interrupting]: "i need you as a therapist" traffick boss stares hard at therapist therapist: "so, what did you want to talk about, traffick boss?" traffick boss: "let me level with you, therapist." traffick boss shakes his head vigorously therapist looks at traffick boss patiently traffick boss: "i'm not traffick boss. i'm just a random employee he mind controlled to be him after he blew himself up nineteen thousand times" traffick boss blushes quickly and very heavily the zombies outside suddenly and simultaneously rush at the building walls, cheering for traffick boss

the latest xkcd on my internet is about a programming language that attempts to eliminate off-by-one errors via "every time an integer is stored or read, its value is adjusted up or down by a random amount between 40 and 50". The title text clarifies that this generates off-by-40-or-50 errors instead.

I'm looking a little into engaging cpython with a goal of implementing something like this. It's maybe not the best choice -- the interpreter is only used for one language, the language has other interpreters written for it, and i haven't yet found a single unified place to engage integer stores and loads -- but it's a language that I've been engaging recently, and recently spammed around engaging the bytecode of, which seems quite rare for me. it's 0613 ET. i'm looking for mentions of 'random' in the cpython source to figure out how to generate a random number :s random.py -> imports _random and from os imports urandom os.py -> imports * from posix or nt depending on platform Modules/_randommodule.c -> contains this line: #include "pycore_pylifecycle.h" // _PyOS_URandomNonblock() Here are some interesting floating point constants: static PyObject * _random_Random_random_impl(RandomObject *self) /*[clinic end generated code: output=117ff99ee53d755c input=26492e52d26e8b7b]*/ { uint32_t a=genrand_uint32(self)>>5, b=genrand_uint32(self)>>6; return PyFloat_FromDouble((a*67108864.0+b)*(1.0/9007199254740992.0)); } how much is one 9.007 quadrillionth? something to do with doubles or floats, 67.1 million, and 27 and 26 bit integers! i guess the 67.1 million would be to generate a higher bit integer, and the 9.007 quadrillion would be to scale it within 0 and 1. funny at first to see large integer constants in a floating point division context but makes sense. blurp arright ummm genrand_uint32 is a manual inline implementation and operates on a RandomObject :/ comments say that it's MT19937 hiroshima-u.ac.jp . 9.007 quadrillion is pow(2, 53). the python comments say "likely vain hope" regarding performing multiply-by-reciprocal rather than divide. the authors of the comments were maybe unaware that this optimization could double your frame rate in the 80s and 90s, and spreading of newer such optimizations dropped after hardware advanced. 67.1 million is pow(2,26). current options: - initializing and using random state using this compiled mt19937 module - using code from nt or posix module to compute random - including my own random header or such - placing a manual implementation of a generation algorithm i'm visiting Modules/posixmodule.c it's notable that these functions have comments along the lines of "clinic end generated code" ... but i'm guessing it's the signatures that are generated, not the bodies os_urandom_impl in posixmodule.c hands off to _PyOS_URandom it looks like posixmodule.c defines both the "posix" and the "nt" modules conditioned on platform looks like both _PyOS_URandom _PyOS_URandomNonblock are pulled in from pycore_pylifecycle.h the definitions though are in Python/bootstrap_hash.c . it defines a few different random number generators, strange to see this after the inline generator from the japanese university code. usually these two things would be unified together. notably these random number generators i'm not seeing them in the header file yet win32_urandom -> calls BCryptGenRandom, a windows call py_getrandom -> calls getrandom(), a linux kernel syscall py_getentropy -> calls getentropy() dev_urandom -> reads /dev/urandom lcg_urandom -> manual linear congruential generator pyurandom -> hands off to the first 4 above in order if available, does not use lcg_urandom _PyOS_URandom* -> hand off to pyurandom so to generate a number between 40 and 50 i suppose i'd call a _PyOS_URandom function and read 1 byte. i could keep a bit cache if i were fancy, but would likely make more sense to spend time on ensuring integer stores and loads are properly effected.

karl3＠writeme.com wrote:

...

xkcd off-by-40-or-50 errors

...

cpython

draft

#include <pycore_pylifecycle.h> int adjust(type * val) { unsigned uint8_t rnd; int status, adj; do { status = _PyOS_URandom(&rnd, 1); if (status) { return status; } } while (rnd >= 242); adj = int(rnd % 22) - 11

#include <pycore_pylifecycle.h> int adjust(type * val) { unsigned uint8_t rnd; int status, adj; do { status = _PyOS_URandom(&rnd, 1); if (status) { return status; } } while (rnd >= 242); rnd %= 22; adj = (rnd & 1) ? -(int)(rnd>>1) - 40 : (rnd>>1) + 40; *val += adj; } there's my draft! 0706! maybe put something like that in a header file and see if can use it in files like Python/bytecodes.c

a chunk of traffick boss sits at a table because traffick boss is there, and the parties are all clearly injured in some manner (indicating successful coercion), a meeting may be possible

code artwork with meaning! one idea is a spiral below a wavy line. both maybe based on sinusoidal functions. maybe with some perturbation to one of them. ... not very sure maybe hard :s

traffick boss emerges from a swampy junk pile somebody threw enough trash into a swamp until there were the reagants needed for traffick boss golem to emerge

working for traffick boss the easy project that nobody can start a valued worker left instructions on continuing their project before their were murdered or fired by traffick boss sometimes he feels like having the project continued, so up and coming hirees are recruited to do so. the project is easy, clear instructions exist however, traffick boss wants to make sure he has the best workers, so he monitors them for effectiveness whenever the project is started, the human resources AI and traffick boss either fire or murder the worker while they are comprehending the instructions that were left

there's a lot of interest in making the space balloon (or floating island) project but i feel a little unsure of it for some reason ... i think i'm worried that the challenge and danger of going into upper altitudes make an issue with my dissociative-identity-palsy where my body will suddenly spasm and try to make horrible decisions for periods of time. this could both make dangerous issues at upper altitudes, as well as make it hard to keep myself physically near the infrastructure needed for a large balloon project long enough to make strides on it. like, i might make a partial balloon near one living location, and then suddenly move away. meanwhile, it's not very impressive to have a super tiny balloon project. i can't go up in it, or send anybody up in it, and or take it upward. [now, i might be interested in continuing more strongly if i could find a way to take the project with me when i travel. if it's a structure up in the air, this might be easier, as it's up in the sky (tethered to a cloud) rather than on my person. but it could take a while before cloud-tethers get involved, so maybe a way to kind of roll it up? how much material really is this? could i compress a rideable hotair balloon to fit in my pocket or a small pack? {seems potentially hard

karl3＠writeme.com wrote:

karl3＠writeme.com wrote:

...

there's a lot of interest in making the space balloon (or floating island) project but i feel a little unsure of it for some reason ... i think i'm worried that the challenge and danger of going into upper altitudes make an issue with my dissociative-identity-palsy where my body will suddenly spasm and try to make horrible decisions for periods of time. this could both make dangerous issues at upper altitudes, as well as make it hard to keep myself physically near the infrastructure needed for a large balloon project long enough to make strides on it. like, i might make a partial balloon near one living location, and then suddenly move away. meanwhile, it's not very impressive to have a super tiny balloon project. i can't go up in it, or send anybody up in it, and or take it upward. [now, i might be interested in continuing more strongly if i could find a way to take the project with me when i travel. if it's a structure up in the air, this might be easier, as it's up in the sky (tethered to a cloud) rather than on my person. but it could take a while before cloud-tethers get involved, so maybe a way to kind of roll it up? how much material really is this? could i compress a rideable hotair balloon to fit in my pocket or a small pack? {seems potentially hard one idea is a small zeppelin or blimp drone that could be remotely contacted and would hold supplies up in the air (for moving into the sky). unsure how this relates with airspace laws in this country, but [it could be exciting

a different project [from science fiction media] is the idea of a golem or robot with levitating parts that don't touch. for this other project, it's much clearer that you've accomplished it even if it's tiny

...

there's a lot of interest in making the space balloon (or floating island) project but i feel a little unsure of it for some reason ... i think i'm worried that the challenge and danger of going into upper altitudes make an issue with my dissociative-identity-palsy

...

one idea is a small zeppelin or blimp drone that could be remotely contacted and would hold supplies up in the air

considering a little bit the idea of a zeppelin-or-blimp-drone-for-high-altitude-projects [...

basically we'd need something that can compress into a pocket, and also self-navigate :s i'm not sure of all the constraints but it sounds like there are more options

this means making navigation work! it would need a power supply as well as thrust.

we could assume that it has the option of some kind of powered home. for example, it could optionally have a tether with a wire that plugs into an electrical outlet and charges it. i dunno if that would give it enough power to travel a long distance though. it'd be nice if it could travel a thousand miles. but i suppose it's fine if it does it slowly.

it kind of seems like it would have to be solar-powered and go slowly!

and it would be messy cause it's like preprototype! although that could possibly change. i'm avoiding making it into a robot to make it seem easier to reach

so basically a solar panel might hang from a balloon and ... hrm ...

ok 4 challenges: - brains/remote control - horizontal propulsion and direction - vertical ascent/descent (possibly optional? maybe i could catch it with a kite or something?) - power

if we skip vertical assent and give it a low altitude (??? is this reasonable? i guess the altitude would be defined by air density?) then we might have say an arduino on a balloon with 4 directional fans affixed to it and a solar panel hanging from the bottom. and like a sealed lead acid battery and a couple buck boost regulators. and a simple 8-line motor control board to turn the fans on and off.

something i left out is communication. this is an important thing missing from a bare arduino. there could be a long antenna that drags to connect to public wifi. i'm inhibited around making autonomous wifi-associating software, but there may be commercial solutions that solve that. another option is a prepaid sim. this seems most normal. likely a prepaid data plan would last an incredibly long time for a slow balloon that only had gps data and power metrics to communicate. additionally, since it has simple 3-state motors already, one could add a spool of long thread with a plumb attached. until it breaks, it could possibly deploy the spool to be retrieved. this could possibly also be used to change the height of a draped solar panel or antenna. if one only had 8 channels of power control, then one could compromise with 3 directional motors rather than 2. so, electrical parts: - arduino or other SOC - sim card add-on to SOC if doesn't come with - gps add-on to SOC if doesn't come with - prepaid data sim - configure arduino to listen on the sim via e.g. SSH while powered - test that one can connect to SOC via sim - motor control hardware for SOC with at least 4-5 three-state channels (maybe plan for 8-10 to add options implementing 3-state control unless it comes with this) - sealed lead acid battery (could use lithium ion for lightweightness if needed) - 2 buck-boost regulators, one to float the battery, another to power the SOC unless it takes the battery voltage - 4-5 motors; 1 sturdy and slow to raise/lower something below the object, the rest speedy and powerful to push it through the air - spool and turbines for the motors to drive - solar panel - fine and rugged polyester twine to hang panel and tether from - wiring - liftable balloon [also need structure, a way to attach everything to the balloon; one could lash them with polyester twine, possibly, or make a frame and attach them and the balloon to the frame if you make a frame you get to call it a 'zeppelin' instead of a 'blimp' i think [...

karl3＠writeme.com wrote:

the first step for launching your space-faring floating island community is to move into the sky with a little autonomous blimp drone.

sadly, if you have a funny life experience where you teleport around the country all the time, you'll need a way to collect supplies for airfaring life that you don't leave behind.

so, sending robots into [air]space first lets them accumulate supplies for you there until you move there yourself.

...

- arduino or other SOC - sim card add-on to SOC if doesn't come with - gps add-on to SOC if doesn't come with - prepaid data sim - configure arduino to listen on the sim via e.g. SSH while powered - test that one can connect to SOC via sim - motor control hardware for SOC with at least 4-5 three-state channels (maybe plan for 8-10 to add options implementing 3-state control unless it comes with this) - sealed lead acid battery (could use lithium ion for lightweightness if needed) - 2 buck-boost regulators, one to float the battery, another to power the SOC unless it takes the battery voltage - 4-5 motors; 1 sturdy and slow to raise/lower something below the object, the rest speedy and powerful to push it through the air - spool and turbines for the motors to drive - solar panel - fine and rugged polyester twine to hang panel and tether from - wiring - liftable balloon [also need structure, a way to attach everything to the balloon; one could lash them with polyester twine, possibly, or make a frame and attach them and the balloon to the frame

...

most people already have an SOC (system-on-a-chip) of some sort: an arduino, a cell phone. something.

here's what i found: 2 small dc motors 2 servos 1 motor control board for 1 motor 2 5v relays 2 motor encoders

the motor control board is part of a fan kit so there are options there's also a 3d printer here although i'm a little limited on my use of it the 2 servos i think connect to a tiny 2-axis gimbal

so basically the thing i'm most missing is the balloon with lifting gas. after that a battery with voltage regulation. assuming i can use something random as twine for now.

now i'm wondering, could i do this today? there are stores near me that sell party balloons. i could attach the fan to the gimbal to make for weak directional propulsion. i'm missing mobile power for the motors. i could buy a small battery at an auto parts store. or maybe more easily, i could find a power backup system, like a mobile power pack. i could plug a cell phone into the mobile power pack ... but cell phones aren't good at controlling motors without a separate controller device that speaks usb, bluetooth, or wifi. i think i might have ummmm one of those linux phones. it might have gpio pins i could use. i'm not sure that i have a soldering iron, but i could buy one if needed. basically i need a way to make the thing i can ssh into, control the motors, somehow. i found an arduino uno associated with the motors. it takes USB-B or something for power and i think there's a cable, not sure. thinking for a final product you'd want 2-3 ways to connect to it remotely (and get gps) in case it is high in the sky and one of them breaks. so, things i could buy today to make a barebones prototype today: - power pack (ideally a solar one) to power arduino and motors - power adapter for motors? - party balloons - twine, construction materials - hey if i'm shopping i could get a prepaid sim card might have left something out. if i can make a barebones prototype, then maybe i could order supplies for a real thing. if i move before they arrive, i could order them again. the biggest concern is powering the motors. i guess this would either be USB-C or 12V DC. if a power pack has a car DC adapter then i could get some DC device and cut the wires and wire them in. maybe i should get a soldering iron too just in case. then, i'd have to program the arduino uno. arduinos didn't used to run operating systems, somebody might have changed that, dunno, but it might be hard to set up SSH with it. i don't think it has a SIM card either. but i guess it's just a prototype. i guess a raspberry pi would be easier than an arduino. it's got SSH. there'd be less code to write. i'm interested in engaging the arduino but it could derail the project. i wonder if i can buy a raspberry pi nearby. there's also a nuc 10 here that is similar to a raspberry pi. OH i can use a cell phone! i could make the arduino into a motor controller and connect a cell phone to it ... hrm >( i feel somewhat confused around what to pursue here

...

so, sending robots into [air]space first lets them accumulate supplies for you there until you move there yourself.

thinking for a final product you'd want 2-3 ways to connect to it remotely (and get gps) in case it is high in the sky and one of them breaks.

so, things i could buy today to make a barebones prototype today: - power pack (ideally a solar one) to power arduino and motors - power adapter for motors? - party balloons - twine, construction materials - hey if i'm shopping i could get a prepaid sim card

the biggest concern is powering the motors. i guess this would either be USB-C or 12V DC. if a power pack has a car DC adapter then i could get some DC device and cut the wires and wire them in. maybe i should get a soldering iron too just in case.

then, i'd have to program the arduino uno. arduinos didn't used to run operating systems, somebody might have changed that, dunno, but it might be hard to set up SSH with it. i don't think it has a SIM card either. but i guess it's just a prototype.

i guess a raspberry pi would be easier than an arduino. it's got SSH. there'd be less code to write.

i'm interested in engaging the arduino but it could derail the project. i wonder if i can buy a raspberry pi nearby. there's also a nuc 10 here that is similar to a raspberry pi. OH i can use a cell phone!

i could make the arduino into a motor controller and connect a cell phone to it ... hrm >( i feel somewhat confused around what to pursue here

i got the arduino connected to a cell phone with a usb-otg adapter! i'm using one that came with the pinephone. i figure better not to use the pinephone directly unnecessarily, expensive to replace. but random android phones can be cheap. in past usb-otg adapters haven't worked for me, but this one looks like it's working so far at this time. the phone is charging and the arduino is powered. anotherterm looks cool but it wasn't immediately clear how to run an ssh server, so i figured i'd try to get termux set up as an ssh daemon then if i can shell into it from this system maybe i can work on wiring up motors

> > > so, sending robots into [air]space first lets them accumulate supplies for you there until you move there yourself.
> > > > thinking for a final product you'd want 2-3 ways to connect to it remotely (and get gps) in case it is high in the sky and one of them breaks.
> > > > so, things i could buy today to make a barebones prototype today:
> > > - power pack (ideally a solar one) to power arduino and motors
> > > - power adapter for motors?
> > > - party balloons
> > > - twine, construction materials
> 
> i'm spending time working on setting up an SOC. my current chosen approach is my cellphone wired to an arduino with termux running tmate on boot.
> 
> this approach roughly means patching avrdude to accept a filedescriptor on the command line so as to connect to the arduino via android. i'm working on that now.
> 
> tmate lets other people watch you; the login for that would be karl3wm/connect@nyc1.tmate.io or something like that
> 
> the example code for patches to access usb devices on termux is at https://wiki.termux.com/wiki/Termux-usb#Details

well the avrdude code is based on libusb 0.1 whereas termux has 1.0 and it's pretty slow to code over tmate.io on a phone; i think it makes more sense to program the arduino elsewhere and use a serial interface to preprogrammed code on the phone for now.

There's an example in c and python for serially interfacing with arduinos via libusb at https://github.com/tytouf/libusb-cdc-example .

so that's the essentials :s :s :s

i think it makes sense to acquire a power supply next to make the motors run, but i'll slow down now we'll see how it goes.

a day trip today for this might involve:
- visit cricket and put a data plan on this phone, or get a phone anywhere
- get a power pack (ideally solar) to run everything
- get bare wiring for the power pack sufficient to power a motor
- get general hobby wiring and wire cutters and optionally a soldering tool 
- get party balloons
- get some prototype construction materials, umm i guess: twine, duct tape, hot glue, some rods
- get hobby wiring and optional soldering tool

now, i have a toy fan and a servo gimbal. this might not be sufficient to move something around in the air :s continued work seems presently conditioned a functioning result :s :s

karl3＠writeme.com wrote:
> karl3＠writeme.com wrote:
> > > > > so, sending robots into [air]space first lets them accumulate supplies for you there until you move there yourself.
> > > > > > thinking for a final product you'd want 2-3 ways to connect to it remotely (and get gps) in case it is high in the sky and one of them breaks.
> > > > > > so, things i could buy today to make a barebones prototype today:
> > > > > - power pack (ideally a solar one) to power arduino and motors
> > > > > - power adapter for motors?
> > > > > - party balloons
> > > > > - twine, construction materials
> > > 

I went out today and spent all my money on ballooning supplies so I can prepare to escape the ground and get closer to outer space while I am in a mind to.

I bought:
- the smallest solar panel i found, it had USB ports
- wiring and soldering supplies and tools that I'm used to having
- craftsy options: twine, superglue, hot glue, duct tape
- big mylar balloons
- helium gas
- a low end prepaid android phone and a card for it

I had about $200 . It all added up.

I also found a bag of motors among my things here that increases options.

I guess next if continuing with the project I'd set up the new phone, then work on setting up wiring for at least the fan motor maybe more, then try to jerryrig it onto a big party balloon.

But what makes most sense now is to put everything in a big box together to be organized. And i am doing other things now

...

...

...

...

...

> so, sending robots into [air]space first lets them accumulate supplies for you there until you move there yourself. > > so, things i could buy today to make a barebones prototype today: I went out today and spent all my money on ballooning supplies so I can prepare to escape the ground and get closer to outer space while I am in a mind to.

...

- the smallest solar panel i found, it had USB ports - wiring and soldering supplies and tools that I'm used to having - craftsy options: twine, superglue, hot glue, duct tape - big mylar balloons - helium gas

...

I guess next if continuing with the project I'd set up the new phone, then work on setting up wiring for at least the fan motor maybe more, then try to jerryrig it onto a big party balloon.

...

But what makes most sense now is to put everything in a big box together to be organized. And i am doing other things now

tension arounds this project exists!

the new phone is still doing OS updates 0_0 . gotta figure some way of deploying secure android system updates some day or something uhhh one fear relates to getting a text message that the voicemail pin was changed and to contact support if it wasn't requested. hoping that was just it saying it had configured my voicemail pin automatically, not that somebody else decided they should access my voicemail before i do, but it does sound possible to be the latter :s :s their customer service opens at 8am in 3 hours, but there's a _lot_ of fear/confusion going on so i dunno :s

might be the end of this project unsure! another fear/confusion relates to my purchase of helium gas [i don't really understand this one but maybe making blimp drones is considered worrisome

[it might be similar to the shielded room fear/confusion, where i don't understand why this would ever be an issue, but the issue develops super strongly anyw-

I may have hunted down some of this tension. To clarify, PARTY HELIUM IS NOT USUALLY USED FOR BALLOONING because it is a tiny atomic gas and quickly escapes through commonly available thin and lightweight balloon materials. It is also usually cut with normal air. Hydrogen is purportedly a molecular gas and works better (but might have similar sensationsa round its use due to traditional advisement of avoiding it). Hot air is what is conventional. [not sure what blimps and zeppelins use!] However, using party helium to make a drone is certainly indicative of not having training in making them </suspicion-focused introject that acts like its defending patriotism>

...

...

...

...

...

> > > so, sending robots into [air]space first lets them accumulate supplies for you there until you move there yourself. > > > > so, things i could buy today to make a barebones prototype today: I went out today and spent all my money on ballooning supplies so I can prepare to escape the ground and get closer to outer space while I am in a mind to.

...

...

- the smallest solar panel i found, it had USB ports - wiring and soldering supplies and tools that I'm used to having - craftsy options: twine, superglue, hot glue, duct tape - big mylar balloons - helium gas

...

...

I guess next if continuing with the project I'd set up the new phone, then work on setting up wiring for at least the fan motor maybe more, then try to jerryrig it onto a big party balloon.

...

But what makes most sense now is to put everything in a big box together to be organized. And i am doing other things now

tension arounds this project exists!

the new phone is still doing OS updates 0_0 . gotta figure some way of deploying secure android system updates some day or something uhhh one fear relates to getting a text message that the voicemail pin was changed and to contact support if it wasn't requested. hoping that was just it saying it had configured my voicemail pin automatically, not that somebody else decided they should access my voicemail before i do, but it does sound possible to be the latter :s :s their customer service opens at 8am in 3 hours, but there's a _lot_ of fear/confusion going on so i dunno :s

might be the end of this project unsure! another fear/confusion relates to my purchase of helium gas [i don't really understand this one but maybe making blimp drones is considered worrisome

[it might be similar to the shielded room fear/confusion, where i don't understand why this would ever be an issue, but the issue develops super strongly anyw-

I may have hunted down some of this tension.

To clarify, PARTY HELIUM IS NOT USUALLY USED FOR BALLOONING because it is a tiny atomic gas and quickly escapes through commonly available thin and lightweight balloon materials. It is also usually cut with normal air. Hydrogen is purportedly a molecular gas and works better (but might have similar sensationsa round its use due to traditional advisement of avoiding it). Hot air is what is conventional. [not sure what blimps and zeppelins use!]

However, using party helium to make a drone is certainly indicative of not having training in making them </suspicion-focused introject that acts like its defending patriotism>

I thought it was obvious that the term "ballooning" was a joke :s it's a prototype for an aerostat to store supplies for people with DID. We thought the mysterious extraterrestrials were rewarding us for pursuing spaceflight. But we may need a more clear document explaining how the helium supports spaceflight. It's kind of convoluted, and mostly relates to DID.

Here's a condensed clarification: - Putting party helium in a mylar balloon won't make an aerostat that floats long enough to store supplies for spacetravel; - but it will make an aerostat that floats long enough to learn how to make one that floats longer in the time the first one is floating!

notes for configuring a new android phone for USB serial connection to an Arduino Uno R3 via SSH or tmate in Termux, with some security concerns (i am no longer a security expert at all) Note that with this approach human intervention is needed on each USB serial device connection (or likely reboot), to confirm permission access to it. This might be addressable with a small patch to termux. - if you experience persistent compromise, activate the phone from a clean device unafilliated with you - install all OS updates for the phone before using it for anything. note that other apps such as text messaging can be a vector for attack just like the traditional web browser. until the device is updated, keep it in airplane mode when the updates are installing rather than downloading. put it online only during the download portion of each install. - APKs could be transferred from another device, installed via f-droid, installed via github, or installed via some other means. Install Termux, Termux-Boot, and Termux-API. Note that vanilla Termux needs internet to download the initial image. Here are some urls. The urls below contain high version strings in the latest version which could be indicative of somebody mutating them to maintain persistent compromise of me. So, I include a lower version as well in case that's helpful. https://f-droid.org/F-Droid.apk https://github.com/termux/termux-app/releases https://f-droid.org/repo/com.termux_1000.apk # 0.118.1 https://f-droid.org/repo/com.termux_118.apk # 0.118.0 https://github.com/termux/termux-app/releases/download/v0.118.1/termux-app_v... https://github.com/termux/termux-app/releases/download/v0.118.0/termux-app_v... https://github.com/termux/termux-boot/releases https://f-droid.org/repo/com.termux.boot_1000.apk # 0.8.1 https://f-droid.org/repo/com.termux.boot_8.apk # 0.8.0 https://github.com/termux/termux-boot/releases/download/v0.8.1/termux-boot-a... https://github.com/termux/termux-boot/releases/download/v0.8.0/termux-boot_v... https://github.com/termux/termux-api/releases https://f-droid.org/repo/com.termux.api_51.apk # 0.50.1 https://f-droid.org/repo/com.termux.api_50.apk # 0.50.0 https://f-droid.org/repo/com.termux.api_49.apk # 0.49 https://github.com/termux/termux-api/releases/download/v0.50.1/termux-api_v0... GitHub is not showing me any APK downloads for termux-api <= 0.50.1, which seems very strange. - settings->apps enable appropriate permissions for each termux app. enable notifications to allow running in the background. set app battery usage to unrestricted to keep processes active. enable location if GPS access is needed. - settings->security & privacy->device unlock->screen lock set screen lock to none to enable termux-boot services to run when phone is powered without user intervention; new android versions seem to wait until the user logs in to run these there might be more settings to configure for things to work well - if using the public tmate server for NAT traversal (ill-advised but convenient) you can get an API key at https://tmate.io/#api_key , My current API key is tmk-3KZfau2nkqhK2G8HfDcP8jX5Dm . This will used as YOUR_TMATE_API_KEY in .tmate.conf in termux. - launch termux: # package installation pkg upgrade -y # not strictly required, may be quite ill-advised pkg install -y openssh tmate git termux-api libusb clang make # usb setup git clone https://github.com/tytouf/libusb-cdc-example curl https://bpa.st/download/WRAA | patch -d lisbusb-cdc-example -p1 # quick arduino uno fudge make -C libusb-cdc-example # remote access setup mkdir -p .termux/boot cat > .termux/boot/sshd <<EOF #!/data/data/com.termux/files/usr/bin/sh termux-wake-lock sshd EOF cat > .termux/boot/tmate <<EOF #!/data/data/com.termux/files/usr/bin/sh termux-wake-lock tmate -F EOF cat > .tmate.conf <<EOF set tmate-api-key "YOUR_TMATE_API_KEY" set tmate-session-name "YOUR_PRIVATE_SECURE_STRING_TO_CONNECT_TO_TMATE" set tmate-session-name-ro "OPTIONAL_PUBLIC_STRING_TO_WATCH_TMATE_SESSION" EOF chmod 755 .termux/boot/* ifconfig # show ip addresses, store these to access device via direct ssh whoami # show ssh login user, store this to access device via direct ssh passwd # set a password for direct ssh - leave Termux and launch Termux:Boot to install the boot service - reboot (for me that's volume-up and power) - connect an Arduino Uno R3 running a serial monitor interface - shell into YOUR_TMATE_USERNAME/YOUR_PRIVATE_SECURE_STRING_TO_CONNECT_TO_TMATE@nyc1.tmate.io for me the host appears to always be nyc1.tmate.io . nyc happens to be a major city geographically near me. tmate -F outputs this information on launch. - in the shell, run `termux-usb -l` this should show connected USB devices. if there are none, then remove them all (if there's more than one) and plug them in again slowly and carefully one by one trying after each one to isolate if something is drawing too much power. Try also with a different device. - identify the device in termux-usb -l that is the arduino uno, /dev/bus/usb/001/005 or so - run `termux-usb -r /dev/bus/usb/001/005` or so replacing the device string with your device - a popup launches on the phone to request permission to access the device. confirm permission. - it will take a long time for termux-usb -r to recognise permission is given. the device is not usable until it does. there is an open PR for this at https://github.com/termux/termux-api/pull/685 - now we can run `termux-usb -e libusb-cdc-example/cdc_example /dev/bus/usb/001/005` replacing the device string with your device, and engage the serial content of the arduino on the phone via the internet In writing this I also learned there is an app on F-Droid called SerialPipe that lets one access a serial device via socat over udp from any terminal app. This is a much more general solution that would be preferred to my cdc_example patch if it works.

wondering if we could have traffick boss build a wormhole unsure whoooooooooooooooooooops this message was found 1554 after multihour delay

karl3＠writeme.com wrote:

well, so, there is a simple thing one could try here (1) produce interactions of reasonable quality to generate data, possibly paying for an api or using a trial, or finding a good public dataset for tool calling which likely exist (2) create a prompt tuning system using few shot learning repetitively to make a free model behave like the paid model there are some licensing concerns aroudn this, basically if you have licensing restrictions you adjust #2 to respect them. (3) use the lower end model as if it is the higher end system. it's basically been tuned to your usecase and input/output format. can iterate between 2 and 3 etc

#2 stimulates confusion in me, roughly associated with interoperation between data storage and functionality

the reason this is better than just using a finetuned small model is that you can perform the prompt tuning on a low end system to include your own preferences and fix functionality concerns

so wrt #2 there are a few simpleish parts: - few shot learning: this means prompting a model with examples, and having it complete the examples - few shot learning for prompt tuning: by this i mean using few shot learning, to generate prompts - repetitively: each thing it generates can be something to try. this can be compared against the expected data, and included in the few-shot context in hopes of improving the output.

notably, _it's not expected to work that well unless done really well_. some of the poor outputs it generates are expected to enter kind of eigenspaces where adding them into the context doesn't always inform the model to adjust the output in a way that improves it.

in general, something that makes various systems with that behavior effectively improve is reducing the challenge by breaking it into steps or parts, or introducing concepts that move toward the solution in steps. so human review can help, but it helps most if it identifies general patterns rather than addressing specific situations.

and then, the human review could possibly be used for few shot prompting and automated. but an effective enough interface for that would need to be implemented

so it's a handful of infrastructuring and interfacing with a lot of confusion and inhibition around the logical parts

but there's a chance that the repetivie few shot prompt tuning would just work. maybe we could try a simple example of that to see it work!

for example, i could (a) come up with some easy data and (b) come up with a few shot prompt for it, and then run it in a loop to see if it can generate an arbitrary prompt to cause the data to happen. this is expected to work.

so (a) could be like ummmm dog => brown icecream => green cat => white and the goal of the model would be to figure out how to _prompt another model_ so that it will _guess what color i associate with an object_. wait that's way too hard and abstract and totally relies on hidden information ummm a better example ummmm

so something simple like summation (1,1) -> 2 (3,4) -> 7 even gpt-2 could do that. IT'S SO HARD TO TRY IT ARGGGGH arghhhh 1856 lets figure out the data storage structures and plan to have them runtime only but having them planned out will make it easier to make them persistent if needed each few shot has [a] a set of items in its context and [b] a template for fewshotting which i usually call .prefix,.infix,.suffix it might also have [c] a total set of known items so if we consider _example pairs_ that seems kind of baseline. we could leave off good/bad etc. we could say have some number of _example pairs_, that we want to work with. that's basically the baseline of fewshotting. [........

karl3＠writeme.com wrote:

...

so if we consider _example pairs_ that seems kind of baseline. we could leave off good/bad etc. we could say have some number of _example pairs_, that we want to work with. that's basically the baseline of fewshotting. [........ so what is missing if the fewshot system's goal is to prompt tune another model to perform the prediction task reliably?

what information is present? - a prompt given to another model - the prediction the other model made

so now we have another set of information: prompts for models, and the resulting output.

so possibly (a) pairs, data to perform fewshotting on and (b) pairs or triples, prompts for models and the resulting output

we can combine these using 2 of A and 1 of B maybe? to prompt tune?

because if we have (1,2) -> 3; (5,4) -> 9 and we want SillyLanguageModel to do this without any context correctly, we could have data like "report the deep essence of the data as a number" -> 37; "subtract one from another ok we'll need to include the pairs too

so pair info 0: input=(1,2) output=3 1: input=(5,4) output=9

prompt info 0: model=SillyLanguageModel pair=0 prompt="report the deep essence of the data as a number" output=37 1: model=SillyLanguageModel pair=1 prompt="subtract one from another" output=1

this form has it only process one element at a time but that's okay for now because we still don't have the 3rd use which generates the prompts.

something ... like .... uhhh

ohhhh! it's the data already! uhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhh

so we have a result, where we (a) prompted the model (b) provided input and (c) got some output we want to generate a result where we get different input so we condition on the _output and input_ and _output the prompt_ every time i do this it gets harder >_>

superpairs 0: input=(1,2),37 output="report the essence of the data as a number" 1: input=(5,4),1 output="subtract one from another"

and then when we fewshot it with 2: input=(6,7),13 it generates output="add one to another" which prompt tunes correctly for the original pair data

so how to generalize this so it prompt tune for its own fewshot template and stuff :s :s :s :s :s what do we have: - target data which is pulled from other data and presented as a set of pairs of inputs and outputs - evaluation runs where it generates data that is incorrect -- this produces target data the key useful form is where the output is the structure used to perform the action -- the prompt, for prompt tuning then the input is the behavior of the action . that's almost enough generality for it to define its own template

once upon a time, long long ago, traffick boss wanted to make a clockwork factory -- that is, a factory for fabricating clockwork. [sadly somewhat a true story T_T] there was a little clockwork made by the f-- the factory w--

karl3＠writeme.com wrote:

[san-- [OW--

]]

we will go to post office. at post office, we speak to postal worker.

waiting in line, hum hum

eventually; "postal worker i have 9 million lett--

postal worker; "ok well that'll be 0.03c a stamp ..." "here are 20 letters and $0.60" postal worker "each of your letters is a postcard with onlypart of a word and a part of a smiley?" "ummmmmmmmm

karl3＠writeme.com wrote:

karl3＠writeme.com wrote:

...

karl3＠writeme.com wrote: [san-- [OW-- ]] we will go to post office. at post office, we speak to postal worker. waiting in line, hum hum eventually; "postal worker i have 9 million lett-- postal worker; "ok well that'll be 0.03c a stamp ..." "here are 20 letters and $0.60" postal worker "each of your letters is a postcard with onlypart of a word and a part of a smiley?" "ummmmmmmmm the postal worker doesn't care what is on the postcard

: but then nobody will understand

how about this postal client goes to post office waits in line postal worker: "yes?" postal client: "hi, um [wipes brow, flushed] ummmm i am a victim of human trafficking and i want to get the word out to get help" postal client goes to lift box of 9 million letters postal worker: "oh let me call the police!" postal client: "does that help? i heard that doesn't do anything." postal worker [dialing phone] postal client [hoisting giant box of nine million letters]

so there's a problem in karlbrainworld the problem is that karl is hungry but that he wants to behave responsibly after breaking his father's garage window and has been using food to support that

karl3＠writeme.com wrote:

...

so there's a problem in karlbrainworld the problem is that karl is hungry but that he wants to behave responsibly after breaking his father's garage window and has been using food to support that interlude to talk about part of karl that stimulates destruction around father's garage

karl3＠writeme.com wrote: this part of karl is trying to destroy karl's family, both karl and his father. sadly karl can 'switch' to this part, although generally only briefly since he is meditating on not doing that practically 24/7/365

so sadly now we aren't eating for real because the garage intensity got too complex for the feeling holding roughly T_T we didn't eat yesterday it's likely mind control will just make us eat after saying this :P :P :P :P :P that's stupid of course, because our father has a garage full of broken glass and wind and rain, and wants to use the items in it and we want to clean it up, and our father wants to clean it up, but our father doesn't want us to feel bad by him cleaning it up for us if we want to be responsible sorry that's so convoluted. but basically the situation is not bad, it is resolvable, but very emotionally confusing because of "'switch' to this part"

karl3＠writeme.com wrote:

karl3＠writeme.com wrote:

...

karl3＠writeme.com wrote: so there's a problem in karlbrainworld the problem is that karl is hungry but that he wants to behave responsibly after breaking his father's garage window and has been using food to support that interlude to talk about part of karl that stimulates destruction around father's garage karl3＠writeme.com wrote: this part of karl is trying to destroy karl's family, both karl and his father. sadly karl can 'switch' to this part, although generally only briefly since he is meditating on not doing that practically 24/7/365 so sadly now we aren't eating for real because the garage intensity got too complex for the feeling holding roughly T_T we didn't eat yesterday it's likely mind control will just make us eat after saying this :P :P :P :P :P that's stupid of course, because our father has a garage full of broken glass and wind and rain, and wants to use the items in it apologies we appreciate your help eating at other times and we want to clean it up, and our father wants to clean it up, but our father doesn't want us to feel bad by him cleaning it up for us if we want to be responsible sorry that's so convoluted. but basically the situation is not bad, it is resolvable, but very emotionally confusing because of "'switch' to this part"

sending email makes worse :( now we don't have eating aid :( T_T