Re: [spam][ot][spam][ignore] Enlrge Ur Pen1s
what seems nice for continuing might be considering the idea of mapping logits to how logoish they might be. one approach to that would be expanding a logo parse tree through all the tokenizer's vocab. then with some parallel data, simply backsolving from that parse tree, i'm thinking it could work. it's no longer a quick approach but it nicely demonstrates how the comic links the patterns of machine learning and software engineering together.
atm I am no longer at my raspberry pi and my phone is at 6% battery, so the project is on pause. basically result from walking away from the system when I saw it had somehow become late afternoon.
oh and for completing the vocab: I suppose you'd pass the whole recording through the original model and then add the output to the tokenizer. it'll get enough words I think. then the random noise stuff would preserve the non-code text but more interesting to map state machines to logit matrices
I'm thinking on how the machine-generated output likely has highly dense regions of precisely matching audio. I'm thinking on identifying them. For example, if we were to do an autocorrellation, they would stand out in some way. Once we have similar audio, we could then compare with the output of any old speech2text model. This could be used to form an association between audio and letters. --- We could possibly take a break from pursuing an xkcd-related goal there, and try to build a graph that describes some of the content of the language model and its tokenizers .... --- Would it be hard to build a compressor out of a speech2text model? The speech2text model operates on very low-density data. --- What could be built would be an open-source app that transcribes things live. There's likely one of those already, but I haven't stumbled on it myself, and it's probably not very good. ---- The tether here is the xkcd task, and we do want some measure of success. Using the model and having it produce a correct example transcription from the new recording is of course success. What is the next success? --- This pursuit was a big struggle. Not sure how it might continue. [what was the reason for the goal to change? just the length of the struggle, and lots of engagement ... what was the reason to _do_ the goal? ... opportunity, maybe ...] [so not a strong validity like other goals, unfortunately :S] [a part was representing this goal. it's probably for AI-like shapes. we're trying to make public hyperintelligence, roughly. this goal likely helps with that.] [okay, so it's good to pursue this in a way that shares with others the ability to do fancy stuff. that supports the addition of graphs and whatnot. makes the approach more reusable.]
so i dunno why it's relevant, but there's an idea that if something used autocorrelation, which could be made into a neural network layer and done with some feedback efficiency [amnesia] [...] that it could um produce like multiple avenues, multiple guesses of what patterns underlie the data. and as it is exposed to more information, some of those guesses could become more relevant, and some could become less. that sounds helpful to display to others. xkcd is a web thing. is javascript helpful? [a javascript logo was revealed as the one used in development] [we're not familiar with a javascript way of doing this kind of thing. likely python is what's up, although python may be web-compilable.] [it's hard to find a good resonance after all that engagement. maybe the random noise would work?] [there were a couple notes too...] [the virtual machine transcription looks kind of slow and sidetracky, not sure of course [the whole project does][yeah]] [it's supposed to be possible to do this without additional training data. is this true or false?] [it gets meta-near if you take that all the way. it has to derive speech, writing, and cartesian graphics [and i think it has no reason to pick our representations of those]] [it's funny to imagine it forming interpretations of the recording that produce completely different artwork. i imagine they eventually go away since they weren't as intended as te real ones. it doesn't sound like something that would complete fast at all.][is it even possible for it to complete before we die?][yeah everything is if you get meta enough] [there are some parts that are much simpler than that. the 'speech' is machine-generated so it is more like printed tokens than speech. the logo code, similarly, is in a very small phrase-space.] [so for example, without additional data, an algorithm could: - identify all phonemes in the generated speech and their sounds and locations - identify all tokens in the logo code - differentiate the logo code from the dialog, and separate them] [but it would have no idea how to write the tokens in the logo code.] [it might be able to figure the artwork out, though, from the numeric nearness of the coordinates, with a little seeding around cartesianness]
[it's unfortunate to lose track of the value around pursuing the task, so it to pick in what way to pursue it.] [what if we were to just produce the output, and not try to find any return? we could add just one or two extra things to try to increase the return, maybe. how would you produce the output in the simplest way possible?] [i guess i'd use the human transcriptions as a train/test set, and then add heuristics to improve the output if needed, in feedback, as considered before] [well, that's really similar to the original plan, so maybe we should go ahead with the heuristics.] [ok i think. i think i was getting confused around processing the way the current model transcribes letters and symbols as similar-sounding words.]
[if we weren't trying to add any additional return, how would we get the letters/symbols right?] [with the heuristics, obviously.] [oh ... it's more distant. different kind of thinking.]
[so when we pick heuristics, we're thinking of ... things that are simple to implement, for now. maybe also simple to come up with or reuse, but mostly simple to implement and effective.] [what heuristics have we come up with? - autocorrelation to associate letters with sounds. this could be a direct feedback path. - random noise to update tokenizers - checking parsing using a parsing tool (even if the text fails this, on average it helps) - some kind of likelihoodness, like how correctness might be associated with distance - things we come up with after trying something, an incredibly normal approach - hand-added data ... ] [a basic idea here was using multiple channels of confidence with feedback so as to rapidly increase quality with only a little bit of added structure. if this idea is kind of held nearby, it could possibly help. with the thinking issues, maybe we could do a little, then look at the idea, then do a little more, bopping back and forth, unsure.] [too much for some of us]
[ok maybe we can do the idea. right now we want the code to make symbols instead of words. thinking on process of recursive confidence] [1 to start you'll need a way to discern some basic improvement, such as something being more right if it has a given change, or something being wrong in a certain way 2a later i suppose we'll need a way of providing for changes to resonate off that information: for example transcribing it as an automated metric. 2b the system needs some way of discerning the attribute. whether it is relevent, or what it might be, or testing if it is there. some way to engage it repeatedly.] [so how does the system tell anything about symbols? is there any way to know the symbols are there?] [there is almost certainly a logits pattern that indicates the symbols. they're near different other words. right in the output.] [we could look in model internals too, but it's expected to be about the same since the input data is expected to be so consistent.] [we could check the input data too.] [so we could train the model on that.] [theoretically but training takes data. we've been conditioned to think about training, trying to connect with mainstream work. instead, we want an algorithmic pattern that helps, or just a little data.] [okay um - the letters probably have patterns that indicate them - we could resonate those patterns by identifying some data is this reasonable] [or is the resonance too vague?] [it could have gone other ways, it's a hybrid] [ok ummm it sounds like it would be quite helpful for it to identify the onset. this is resonance, but for it to be _confidence_ resonance we'd need to be more precise around how it would work. what the confidence would be. and we'd want multiple channels.] [ok so that actually looks like a good approach because autocorrelation and similar things are very general and can apply to a lot of data. we could try to find patterns that let it learn to predict which areas are code and which are text. the patterns would be found from feature-resonance between areas. but we want to formalise out the idea of feature resonance, to produce a confidence metric.] [ok so we would give it some kind of heuristic axiom around patterns in general, that would have some hardcoded confidence.] [now we have hard or inhibited part around the patterns inside [looks inhibited]] [let's do one part at a time. any pseudo-draft of 'feature resonance'?] [for example, you could train a model to simply predict what data comes next in the stream. [oops two branches here]] [we were going to consider feature resonance but then realised that such a model would already be trained to recognise the change, because it would have to learn it to do the prediction appropriately. there are different token frequencies in the two sets] [ok such a model would indeed produce confidence. what was the earlier idea though?]
[there was some earlier idea around something ... maybe the idea of sequential prediction can be a little cognitively surprising. we could instead tr ...] [classifier. classify the data.] [oh this old idea. randomly tag the data, randomly pick tags to be train or test, and then update the tests with the predictions produced from the trains, until the data is broken into categories based on its patterns. this is probably what you were considering something near.]
[that sounds like a weaker approach] [i like how it [has whatever properties you are valuing] but it seems we do need ...]
[i guess it's nice to just think of this. i'm imagining features emerging from the data, based on their utility for predicting it. one approach is to actually train a prediction model.] [another approach is what? remember this data has precisely-same audio components, scattered everywhere in patterns to be mapped.] [features emerge from data, based on utility for predicting it ... without sequentialness?] [there are a number ways of approaching this. what's important is that it isn't being publicly discussed, and would be incredibly useful.] [features in the data ...]
[okay i'm imagining like a small transformer that could learn a feature if given data to do so. how do we bind it to the utility for predicting?] [so maybe 2 small transformers? one to predict other data, and one to produce a feature useful for predicting it] [maybe! you could start with really small transformers so more use structure emerges.] [you'd want something about the location?] [this could start having a lot of models. i do like the idea of taking apart a deep transformer model and turning it from a single stack into a useful graph.]
[well i don't know. i think i want to generalise and daydream around all the parts, but they're too big for the working/active memory in this state of mind.]
[the goal was to help people build ais on their own, by exposing them to a way to do something new] [okay, goals are helpful]
[and we were thinking of confidence-resonance] [so maybe the prediction model is fine? it produces strong confidence. it doesn't need new data] [ok i kind of see! ....]
[to simplify all the thoughts, we could pull that out and consider the whole recording, avoiding the specifics of symbology. we could fix errors in the recording, by training a model to predict it in resonance with the one that attempts to transcribe it. there is a _lot_ of predictable behavior in there.] [nice i like that. we could fix up the symbology some easy way, like using the output of the other work online. how would we fix up their straight line errors?] [i guess we'd need to parse the output numbers into integers and floats that would be fed back into the model]
[the idea is of making a small class that repairs errors made in data by providing an interface for the user to inform it of general information about the data. it might be reasonable to make some simple form of 'confidence-resonance' prominently visible in the implementation, showing how data from multiple sources could significantly augment a model.] [we don't actually know probability, so the function would be a heuristic for time efficiency. it could have a comment asking for people who do know probability to improve its accuracy. we do suspect it only needs to trend the right way to succeed, due to the effect of feedback loops.]
[distracted thinking of the value around inferences similar to autocorrelation. been really inhibited around this for years. barely remember the reason for it. maybe it's a way to [do a valued logging behavior].] [we switched the order of things, kind of. we're focusing more strongly on logging now. but we do generally spend time [this way]] [reasons really help. the above goal found a reason to help others. do we have a reason to help ourself? we probably need code generation or robotic assistance, more than data analysis. unless you can write something that fixes corrupt disks.] [oh yeah this approach could fix corrupt disks] [it could? we could use that.] [yeah basically the disk is the audio data.]
[so the parsing attribute is similar to e.g. a filesystem on a disk. the data has a certain rigorous structure.] [mm] [let's start with the confidence resonance structure maybe? and keep the daydream around disk repair open?] [i'm worried we might not have analogous behavior this close later] [we're really under duress, or at least have been, and a warning rose. maybe think of it later? unsure] [hrm really limited working memory around topics.]
[how about the idea of encoded layers of data. we have _some_ rote encodings that we can display to the system, but not all.] [okay, kind of like 'views' of it, maybe? i dunno.] [that's a little helpful for how the ideas could be brought closer. let's think about the confidence approach. what were the two examples of extra confidence? a prediction model and ....] [for the new idea, let's use the parser. we seem to like the parse. let's assume it's a process we can call to convert correct data into some other representation.] [ok i think i see the edge spot here. we basically need some correct data to work off of. we can theoretically generate that, in which case we need some way to move toward it or discover it. but it makes sense here to require the transcription to have a little compilable logo. that basically means providing some correct logo data with timestamps.] [there is a way to discover it if the model understands letters. for example, the _existing model_ may have been trained off of recordings of people saying the names of letters, e.g. reading a form or something. if so, then the letters or symbols would have logits with weights to them, in the output.] [it sounds a little fun]
[this repeats possibly-helpfully some value around clear access to the noncorrect outputs of a model, with their strengths. near the idea of 'confidence-resonance']
[we'll want an approach that could conclude faster so as to produce a result that's useful for anything.] [can we repeat a way to reach the goal, without regard to the parts?] [it's expected that if we train a model on the existing data, it could likely produce the rest of the data. if some is wrong, hand-updating some of it is likely to propagate fixes quickly.]
[given that has propagation of improvement already, and multiple sources of input (the human could be abstracted into a datasource), maybe it's valuable just to add the parser, with its similarity to filesystems]
[this big-inhibition currently-experiencing is the kind of thing we deal with. it's why we need fast-good-result, basically.] [so now we'll want a nearby task that can succeed. it is really helpful it moves toward a larger goal with expected success.]
[[[non-code text transcribed at https://www.explainxkcd.com/wiki/index.php/2601:_Instructions/Audio_Transcri... . the existing code transcriptions were made with amazon and youtube services. ]]]
[ok i'm imagining the code as completely opaque. no data on it! but in reality we have data on it. people have hand-transcribed it, and made quick heuristic transcriptions.]
[let's just sit here for a bit.] [we were earlier considering using the existing data, which is incomplete. this provides information on the code. likely enough for e.g. the existing model, to output it as desired.]
[i like fixing the heuristic description. that's really similar to the idea of fixing errors and stuff. some it is right, and the right stuff is scattered all around.]
[maybe time for some other task] [can u try to add small thing before go. so much time spent on. even the randomness-training.] [ok maybe i'll plan to add a training loop to the code that was so hard to expand the generate function in, again. training loop could be useful for making models based on other data.]
[i broke it. i started wondering what i could use it for, and i couldn't come up with anytihng, and then realised it was likely cause of the cognitive issue.]
[usefulness: sometimes we take voice recordings during times we have amnesia. often the audio is muffled. one of them was of somebody we really valued strongly what they said. this kind of approach could recover some words.] [i am so beat i want to just daydream that. could we reconstruct higher quality streams of audio of the speaker? i know we could, i dunno how. maybe 1-5 minutes of daydreaming this?] [... maybe? how would you do it ...?] [say we have a model that can produce tokens. we would then run it backward to make audio again?] [hmm yeah i don't know if models are reversible (they look like they would be? maybe i am totally wrong) but you could always train a new model as the inversion of the old.] [either of those ways, how would you isolate a single speaker?] [well that's different, separating channels] [ok let's expand to multiple people talking. can you transcribe different people?] [how would you identify them?] [you'd need isolated snippets of them talking] [hrm ... umm .. this is clearly doable but looks more compelx than interesting! i'd start by reversing one person talking.] [what about in an audiorecording, where who is talking switches] [i think i saw some models trained around this, identifying the speaker in each person] [say the model could identify the speaker. could you extract a profile of their voice? or a sample of them talking? or tag all the times they talk?] [yeah take those in reverse order and each one would produce the next] [huh] [so identifying the speakers would be important] [let's go back to reproducing the audio][just for fun i guess] [i dunno how to do that] [you'd train a model with two microphones] [but training a whole new model] [maybe you could od it with just a little of two microphoners, and mostly just invert the existing model?] [you'd need some extra channels for voice profile information] [where would those go if you were reversing the model?] [i guess you'd train both together, and they'd be additional outputs and inputs?] [maybe something we don't know here] [back to task 5 minutes up] [ok ummmm make a training loop]
ok, Confused Person needs help using GPG. Rock That Stands Tall Amidst Waves That Cannot Reach It: Use GPG! Confused Person: Uhhh.... Tall Rock: USE GPG! Confused Person: Ok um ... i need to run it somewhere. maybe right here? Tall Rock: USE GPGP PROPERLY! Confused Person: Ok um I have a number of airgapped machines but i dunno if I'll remember that I'm using it. Tall Rock: WE CAN DO THIS! Confused Person: DOES IT MATTER IF IT'S AIRGAPPED OR NOT. ok um ... I dunno. I experience more frequent misbehaviors on systems that are networked and heavily used. I also experience misbehaviors on offline systems. I do not experience misbheaviors on systems that have been completely isolated from online devices. Tall Rock: Let's use the airgapped systems. Tall Rock: Set up a GPG key on an airgapped system! It's okay if it misbehaves a little bit. Confused Person: Hey, I have some holiday-gifts-phones I haven't used for anything. Maybe I can use one of those. I also have some raspberry pis I could set up. Last time we went maybe 8 hours or so, and only did the first step. The following steps didn't happen. So maybe watch the time, and use network if it takes a few hours.
goal: gpg on phone. curious: can phone read qr codes? can it send them? --- 0812 concern: set up phone, may not have working camera. has 2 cameras. apps seem to crash when using them, could be wrong
--- phone looks awkward. maybe not best solution. could migrate microsd cards instead of camera? .....
participants (1)
-
Undiscussed Horrific Abuse, One Victim of Many