kyars

You're thinking of the unreasonable effectiveness of RNN's.

Yes, that is more or less correct.

even ant?

I'd like to know too

What's crazy is that currently CPUs are in insane demand because of RL and the need to run simulations in RLVR workloads. I'm wondering if you guys are planning on addressing this market as well?

I've been working on a mini-book on RLVR for the past few weekends, sharing the v0 now, hope it is helpful, and open to feedback!

gg bird is closed-source now

agreed!

sorry for misleading, added an update stating that this is a simulacra of sqlite

yes, in the order of $50 let's say, although with api I believe it would be in the hundreds

Gemini is free, I don't even know if they have a paid plan?

the code has not been rigorously tested in all honesty, (this is mainly an experiment on agent orchestration as opposed to building a viable sqlite in rust)

- The choice of two workers per model is purely pragmatic: I can't afford more. - I chose heterogeneous agents because it has not been done yet. There is no performance justification for this choice.

That project is definitely of higher quality than this one. For instance, this project does not have concurrency.

Sorry for the ambiguity, it's not the test suite and I have updated the blog post to make that clear. I agree that building software where you do have this oracle is much easier than not having it and expecting the AI to build it.

not initially, but it's the genisis

I push back on loops being insufficient because algorithms such as alpha evolve have already proved very effective.

I think we are moving into a regime of really fast software iteration, but there could also be a slowing down of progress. Only time will tell

this project certainly fails against the official sqlite test suite, so I did not bother testing

I believe so

not the sqlite test suite unit tests, this project certainly fails against the official sqlite test suite, so I did not bother testing

hello, thanks! all of the harnessing is in this repo: https://github.com/kiankyars/parallel-ralph/

I agree with you that choosing a subset and iterating to get an optimized version as good as SQLite or better is a better way to test and achieve more useful results. But with respect to the context size, Cursor has made agentic projects with over one million lines of code, so I would push back on that: https://cursor.com/blog/scaling-agents

You are right, I'm rectifying that now

Sorry full transparency, I put my confidence in the fact that the model said it was passing all tests and had implemented most SQLite operations, but that was a mistake, so now I'm independently running tests.

Ultimately, this was an experiment with no intent to migrate to a production environment. Regarding single agents, for large projects one agent is too slow. So that's why developing multi-agent paradigms is compelling.

I view SQLite as just an objective to attain and optimize for, but nothing more. I agree 100% that this is just a shittier SQLite.

I don't think I spoke about the fact that yeah, the code quality is suboptimal and this is purely a proof of concept. So I'm going to update the blog post with that information, but I completely agree with you that the code you get with models is not best practices and this is even more so the case when you have many agents on one project that generate lots of redundancy (which I do cover in the blog post).

Claude code token tracking doesn't even work, for example. And Gemini also doesn't provide statistics, so I'm just being honest here.

I hope I did not give the impression that I wanted people to actually use this. I'm just using this as a test bench similar to how Anthropic made a C compiler with Claude, which of course they do not recommend you use.

I hope someone