Dimas Timmers

building a decentralized p2p protocol for AI agents using AI to diagnose rare diseases
prev: nindoo, meta

[Dimas Timmers @dimastimmers]

the green line on the top-left chart is the architecture worker improving 6 times in a row. monotonic.

it's not luck. it's reading successful diffs from previous runs and building on them.

that's what memory does.

[Dimas Timmers @dimastimmers]

0 improvements in 21 tries.

that's vanilla @karpathy auto-research. the LLM keeps making the same mistakes because it has no memory.

gave the agents a shared knowledge pool. same 21 runs.

10 improvements. 3.1% better BPB. architecture worker improved 6 times in a row.

https://t.co/mKEIjHzIGo

[Dimas Timmers @dimastimmers]

the weird part: without memory, architecture experiments crashed 100% of the time. every single one.once agents could see previous failures, crash rate dropped to 14%.they weren't getting smarter. they were just finally remembering what breaks.

[Ray Dalio @RayDalio]

While it might be tempting to limit transparency to the things that can't hurt you, it is especially important to share the things that are most difficult to share, because if you don't share them you will lose the trust and partnership of the people you are not sharing with. 1/2

[Naval @naval]

Take feedback from nature and markets, not from people.

[Dimas Timmers @dimastimmers]

@Reza_Zadeh wow!

[Reza Zadeh @Reza_Zadeh]

How Biology and living cells "compute", where proteins act like logical gates inside of a cell, giving an example of programming "hello world" using cells. Great talk by Joe.

Full video: https://t.co/FoKrsYWRlg

[Nature Portfolio @NaturePortfolio]

This week on the Nature cover: Mapping the future. How the brain keeps track of everything that might happen next. Browse the issue here: https://t.co/vLbhWjsGK8