zirpu

Fork

Add WP query access, and fix a few bugs.

Most importantly, make choices on RWKV work. The same technique should
work on LLaMA too.

Bump RWKV and LLaMA dependencies.

Set up ABCs and refactor RWKV/LLaMA support modules.

The LLaMA support needs to be cleaned up, but I want to first bump
dependencies in order to get at newer APIs.

Split thoughts into two agents.

Well, it works. But it sucks.

Fuck it. Nobody gives a shit anyway.

Fix a typo. A type-error thinko.

I think it all works? I'm gonna let it run overnight and see what
happens. What could go wrong~

Prototype choices, create subtagging for messages.

These go hand-in-hand; we want to be able to choose to not just reply on
IRC, but reply on a specific channel, so we need to use the channel as
subtag data.

The choice mechanism is clunky; I would very much like to just use
ReLLM, and maybe that'll be important in the future if/when we migrate
back to HF with their brand-new RWKV integration. They don't support
quantized models, though...

Do the needful.

Parameterize model path.

So that it can run on other peoples' machines!

Tweak locking for gen state.

We need to serialize tag-switched actions, since the tag-switching
mechanism acts as a prefix.

This is almost good enough to share, even though upstream dependencies
aren't yet stable.

Switch to Twisted.

Fight with asyncio.

I regret choosing this IRC bot library. It's not a bad library, but it
doesn't correctly call into asyncio, and the mismatch doesn't seem to be
something I can easily monkey-patch.

Fuck it, let's rewrite with Twisted.

Try adding a basic clock.

I think that I need to refactor the coupling between agents and minds.

Get the unified stream to work.

We are able to chat in IRC! We aren't synthesizing new thoughts yet,
though...

WIP: Just have a single unified stream of consciousness.

The idea is to incrementally spend 10-15s/line doing completions,
instead of waiting for >20min for contexts to load over and over.

Start integrating RWKV.

Use a llama.cpp with optimizations.

Whoops!

Reach the height of LLaMA on June.

June doesn't have enough RAM to go for 65B parameters, so instead we're
going to try another transformer. In particular, RWKV has several
appealing qualities.

Also, add basic joins and parts.

Try extending thoughts with a minimum of tokens.

I realized that every word I was adding was a mistake, so I took them
all away. Then I took away the punctutation, too. What remains is
hopefully the simplest thing that could possibly work.

The thought databases need to be so much richer than the old character
prompts. The payoff is palpable, though; Verdant is clearly meta in a
way which her sister isn't.

Integrate llama-cpp-python.

There's more than a few spots of jank, but it works!

Faster sentence embedding.

Simplifies SentenceIndex, no longer requires a tool to append thoughts,
and more!

Initial swing at packaging rwkv.cpp.

Actually call LLaMA correctly.

Okay, now we're good!

Become a Nix flake.

The old zirpu.py script is rotted to the point of incommensurability, so
I'm removing it. The rest of the scripts are ported, and they seem to
work; agent.py works, at least.

Try to tweak thought prompts.

I think I'll need a faster embedder, so that I can embed arbitrary lines
from chat and get reactions.

Tighten up the thought-oriented prompt approach.

There's really only two pieces to the prompts now: IRC logs and trains
of thought. Completing IRC logs leads to sending messages on IRC, and
completing trains of thought leads to more thoughts.

Give sentence indices an inigo behavior.

The exponentially-decaying behavior appears to work well with the
given parsimony metric. I spitballed the exponent, based on a couple
thought files from some bots.

Temperature really should be lower, so I tweaked that too.

Fix some newlines, and experiment with thoughts.

Trying to avoid picking exponentially-long chains of thought by
preferring shorter thoughts. First, select relevant thoughts; then,
select shortest thoughts.

Clean up indexing of thoughts.

Intercept faux IRC timestamps.

GPTs will gladly hallucinate bogus messages, complete with fake
timestamps, in the middle of replies. However, usually they come after
an agent has made a complete reply. So, we can just split on these fake
timestamps, like we'd split on missed newlines.