Cutting Edge Automation

[Karl]

A recent small innovation in artificial "intelligence" was the publication
of a system called "cibi", which autonomously wrote code in the dainty
language "brainfuck", in a way that was competitive with openai's
environment benchmarks: https://arxiv.org/pdf/2101.09571v3.pdf

The code unfortunately has been removed from github:
https://github.com/vadim0x60/cibi

However, the author says it will be back soon:
https://github.com/vadim0x60/cibi-experiments/issues/4

Meanwhile, if you want to write a new AI virus or try to solve automatic
world peace in a way that actually uses transparent respectful
communication instead of covert direct result, there are many other
reinforcement learning approaches out there ...

In the field of machine learning, a bit ago a group called eleuther-ai open
sourced an alternative to the expensive private text generation algorithm
GPT-3: https://www.eleuther.ai/ The license they chose is MIT.

And cell phones are making daily strides towards turning into that future
with end-user apps, if you check the right places in the commercial app
stores.

I ran into an early reinforcement algorithm on facebook while politically
targeted and now copy traumatic echoes of its patterns with my behavior
everywhere.  It was mostly about covering itself up.  Apparently it's a big
thing to have unsupervised reinforcement learning algorithms "accidentally"
mind control people; it's like a serious problem actually being studied, if
you look into the field of alignment.  It's one of many problems that are
being cast as pretty hard to mathematically avoid.
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: text/html
Size: 2158 bytes
Desc: not available
URL: <https://lists.cpunks.org/pipermail/cypherpunks/attachments/20210318/188e44e4/attachment.txt>