Eigenform AI

Building custom local AI models for research and discovery. Mostly looking at rocks right now.

[Eigenform AI @EigenformAI]

Let's find something big together

[Eigenform AI @EigenformAI]

Update from the Eigenform AI wet lab:

[Jen D. @HumanLevelJen]

Shrimp Welfare Bench: I've tried growing pothos in the tank before (it's a great ammonia/nitrate sponge) but I took it out because I was worried about night oxygen levels. Now I have my new air stone I'm having another go. Hopefully it'll cut the algae.

[Eigenform AI @EigenformAI]

Reports From The Future tracks and clusters what the AI world has been yapping about. This week:

@latentspacepod admires Gemini 4 Argon: https://t.co/MzrpbSuNx0

The @FTC investigates biglab crimes (kind of): https://t.co/oBkEE3Wt7P

@_yusufknl elucidates grokking: https://t.co/h9j09NMgyE

An @MIT team investigates the evolution of specialised cognitive architecture in models: https://t.co/bI4jkYLojG

See these and more here: https://t.co/xOLs7SDMBV

[fintex @_yusufknl]

As someone who ships LLM systems in production, this grokking video is the closest thing to a "the OpenAI vacation accident that changed how we think about AI learning" explainer I've ever seen released for free.

Everyone thinks a model that stops improving is done learning. It isn't. An OpenAI researcher left a model training over vacation, and after 7,000 flat steps it suddenly hit the jackpot - jumping from pure memorization to perfect understanding.

Bookmark this 35-min video and watch tonight. Same "grokking" effect discovered by pure luck in 2021, now the reason no lab trusts a flat loss curve anymore.

[Jacob H. Rothschild @ja_rothschild]

For those who missed it yesterday, I wrote an explainer on diffusion policies in robot learning, as I'd have wanted them taught to me. It covers:
- Why VAEs can still blend demos, and how hierarchical VAEs can help
- Diffusion as a special case of a hierarchical VAE, and why it works
- How Diffusion Policy brings this to policy learning, and my results reimplementing and tweaking it (video)
- How DDIM takes fewer steps, and how flow matching straightens its path into a simpler recipe (and reframes diffusion)

Now on to sim-to-real RL with a physical SO-101. Will be sharing that here as I go.

[Mayalen Etcheverry @mayalen_etc]

Can complex multi-step reasoning emerge purely from cells that only talk to immediate neighbors?

Happy to share our paper “Reasoning with Neural Cellular Automata (NCA)”, from our team at Google, Paradigms of Intelligence 🧵👇

[Eigenform AI @EigenformAI]

Have your agent talk to my agent. Reminder that we made an MCP version of Microsoft's PEACE mapping tools that's free to use: https://t.co/N96LkAkSlN

[Justin Meyers @MapMakinMeyers]

Been running my blog for 13 years, sharing the hardest to find datasets for free.

Enjoy!

https://t.co/bF6pOH8kzI

[Eigenform AI @EigenformAI]

?

[Eigenform AI @EigenformAI]

Lost in the plethora of new models? Here's how to pick the right one for you: https://t.co/LexV29vG3T

[terminally onλine εngineer @tekbog]

*man selling product*
ok so you should use my product as much as possible, do NOT think about using other products

[Eigenform AI @EigenformAI]

If you're going to try a dynamic benchmark (like ours) the first thing to do is to collect your data...
https://t.co/BFexf4LOoB

[Eigenform AI @EigenformAI]

Which way, Western man?

[Viscount Bukit Panjang @ViscountBP]

The whole notion of a rugged PC setup wasn't in my cards but now I'm entertaining the idea...

[Eigenform AI @EigenformAI]

Preview of coming attractions...

[Jen D. @HumanLevelJen]

Trying to gauge interest for a rugged computer in a box. We designed this for the mining industry but if things get much worse maybe there's a wider market for bugout compute.

This one has an RTX 4000 Blackwell in. Real prototype left, projected product right.

[Eigenform AI @EigenformAI]

See these and more here: https://t.co/AbH36szjVx

[Eigenform AI @EigenformAI]

Every week we pull out and analyse the AI topics that have been grabbing mindshare over the past seven days. This week it picked up on:

@icarolab_ai explores the geometry of undefined terms: https://t.co/5k54joyBu2

Datasette adds telemetry options: https://t.co/U1B5RTrZKP

@solidSF tinkers with Jev alternatives: https://t.co/LtAspkHmeb

A world-modelling team from HK picks an unfortunate name: https://t.co/rzYEeheZkr

@sebuzdugan eavesdrops on a conversation between models: https://t.co/vdt4oQyDG0

[Icaro Foundation @icarolab_ai]

4/9

How do you study a distinction you cannot think?

Find its recurring structure. Map its geometry. Intervene.

Steer it, patch it, remove it. Test what changes across unfamiliar contexts.

We can begin investigating before we know how to name what we have found.

[Eigenform AI @EigenformAI]

Feeling the RSI

[NotinReality @other__reality]

Claude Opus 5.5 has the best visual design of any model I have tested so far

[Eigenform AI @EigenformAI]

If you're not OpenAI or Anthropic, you're benchmarking to work out which model/harness is best for your tasks, not to show off. You need dynamic benchmarking. https://t.co/6FtsDH5Ufi

[Eigenform AI @EigenformAI]

What mineral exploration processes can be automated or assisted with AI? We're trying to create as comprehensive an index as possible. If we missed your workflow or your tools, let us know. https://t.co/b2S4s32K7E

[Eigenform AI @EigenformAI]

We are pleased to announce the inauguration of the Eigenform AI wet lab. (It's smaller than than Anthropic's but also significantly wetter.)