Eyvind Niklasson

pleased to share 'DiffLogic CA', made of the finest blend of NCA and binary-gate circuits! adopting @FHKPetersen's differentiable circuit learning, we learn *stateful* circuits, recurrent in space & time. by @PietroMiotti, me, @zzznah, @RandazzoEttore 1/5 google-research.github.io/self-organisin…

"The Cursive Transformer" is a project I finished about a year ago but forgot to post about. We trained a Transformer to write cursive! greydanus.github.io/2025/03/30/cur…

if you have something cool at the intersection of self-organisation and evolution, submit it to the Evo-Self workshop at @GeccoConf, extended deadline April 3rd! organised w/ @nisioti_eleni @miltonllera @risi1979 @MarcelloBarylli @RandazzoEttore @zzznah @mayalen_etc

adaptive efficiency is always beautiful

Mammals have hundreds of joints and muscles. Controlling them individually would be nearly impossible. How does the nervous system organize such complexity into coherent actions? Our new study explores this question through a natural behavior: jumping. 1/15 🧵

The next major breakthrough will branch out at a much lower level than deep learning model architecture. It will be a new approach. A better model architecture can lead to incremental data efficiency & generalization gains, but it won't fix the fundamental issues of the parametric learning paradigm.

Neural cellular automata are the retinal waves for LLMs? 🤔

Very neat work on the compute-only path to intelligence:

Can language models learn useful priors without ever seeing language? We pre-pre-train transformers on neural cellular automata — fully synthetic, zero language. This improves language modeling by up to 6%, speeds up convergence by 40%, and strengthens downstream reasoning. Surprisingly, it even beats pre-pre-training on natural text! Blog: hanseungwook.github.io/blog/nca-pre-p… (1/n)

@rabrg << what our Universe looks like to God

Not every research project has to end up in a conference publication, but some are still worth sharing. This is one of them! Working on this made us deeply question the fundamentals of deep learning (specifically backpropagation through time) so we wrote a blog post about it 🖥️

Working on the new simulator. I just wanted to see what Atari2600 fetching data from ROM looks like at CMOS FET level (@tinytapeout TT09 Atari circuit by @__ReJ__)

You can write a self-replicating physical program in just 45 tokens (RNA bases). That's small enough to emerge spontaneously via brute force recombination at scale.

🤔

🔜

Growing Graphs demo is finally out! 🕸️✨ 🔗 znah.net/graphs/ Videos from a few months ago finally meet a finished implementation, thanks Gemini for doing the boring parts. Inspired by Paul Cousin's Graph-Rewriting Automata: like a Game of Life, but cells can split if they want to #GenerativeArt #WASM #SwissGL

In this new paper, Gary Lupyan and I address ongoing controversies regarding how to best think of what LLMs are doing: are they a language mimic, a database, a blurry version of the Web? arxiv.org/abs/2601.11432

#genuary9 Crazy automaton. Cellular automata with a small language model as rule. The model is trying to recover the corrupted text. #genuary2026 #genuary

Survival of the fittest code. Core War (1984) is a game where programs must crash their opponents to survive. Warriors written in an assembly language called Redcode fight for control of a virtual machine. Our new paper: Digital Red Queen: Adversarial Program Evolution in Core War with LLMs, explores what happens when LLMs drive an adversarial evolutionary arms race in this domain. We task LLMs to write Warrior programs in Redcode that must out-compete a virtual world full of such programs. Core War is a Turing-complete environment where code and data share the same address space, which leads to some very chaotic self-modifying code dynamics. This approach is inspired by the Red Queen hypothesis in evolutionary biology: the principle that species must continually adapt and evolve simply to survive against ever changing competitors. In our work, programs continuously adapt to defeat a growing history of opponents rather than a static benchmark. We find that this adversarial process leads to the emergence of increasingly general strategies, including targeted self-replication, data bombing, and massive multithreading. Most intriguingly, it reveals a form of convergent evolution. Different code implementations settle into similar high performing behaviors, mirroring how biological agents independently evolve similar traits to solve the same problems. I think this work positions Core War as a sandbox for studying Red Queen dynamics in artificial systems. It offers a safe controlled environment for analyzing how AI agents might evolve in real world adversarial settings such as cybersecurity. By simulating these adversarial dynamics in an isolated sandbox, we offer a glimpse into the future where deployed LLM systems may start competing against one another for limited resources in the real world.

Announcing New Pareto Frontier on ARC-AGI 27.5% for just $2 333x cheaper than TRM! Beats every non-thinking LLM in existence Cost so low, its literally off the chart Vanilla transformer. No special architectures. Tiny. Trained in 2 hrs. Open source. Thread: