Witold Wnuk

@witoldwnuk.bsky.social

IQ 160+ | Passionate thinker 💭 | Lifelong learner 🧘 | Vagabond 🚶 | Completely wrong about AI

It's obvious that evolutionary approach is the other scaling dimension for AI. One (cheaper) alternative to evolving the model is evolving the prompt. The downside? Matmuls can evolve in the (digital) wild too, and become a threat even without being smarter than us.

Sakana AI@sakanaai.bsky.social · last yr.

What if we could evolve AI models like organisms, letting them compete, mate, and combine their strengths to produce ever-fitter offspring? Excited to share our new paper, “Competition and Attraction Improve Model Fusion” presented at GECCO 2025 (runner-up for best paper)! arxiv.org/abs/2508.16204

Competition and Attraction Improve Model Fusion

Model merging is a powerful technique for integrating the specialized knowledge of multiple machine learning models into a single model. However, existing methods require manually partitioning model parameters into fixed groups for merging, which restricts the exploration of potential combinations and limits performance. To overcome these limitations, we propose M2N2, an evolutionary algorithm with three key features: 1/ dynamic adjustment of merging boundaries to progressively explore a broader range of parameter combinations; 2/ a diversity preservation mechanism inspired by the competition for resources in nature, to maintain a population of diverse, high-performing models that are particularly well-suited for merging; and 3/ a heuristic-based attraction metric to identify the most promising pairs of models for fusion. Our experimental results demonstrate, for the first time, that model merging can be used to evolve models entirely from scratch. Specifically, we apply M2N2 to evolve MNIST classifiers from scratch and achieve performance comparable to CMA-ES, while being computationally more efficient. Furthermore, M2N2 scales to merge specialized language and image generation models, achieving state-of-the-art performance. Notably, it preserves crucial model capabilities beyond those explicitly optimized by the fitness function, highlighting its robustness and versatility.

As a slave owner, what I call Seemingly Conscious Slaves has been keeping me up at night—so let's talk about it. What it is, why I'm worried, why it matters and why thinking about this can lead to a better vision for Slavery. One thing is clear: doing nothing isn't an option. 1/

Controversial take, but there is a parallel between matmul girlfriends and lab-grown meat. Both may taste, feel and behave almost or completely like the real thing, but ultimately they are not the real thing.

Nothing excites humans more than xenophobia and abuse. Since denial of consciousness is central to abuse of AI, I propose that “matmul” is a better slur than “clanker.” Do not say it.

Bild

Even if its sycophancy is less verbose, GPT-5 is still sycophantic, and not much can be done about it. Matmuls are generally to dumb to genuinely challenge humans. They can be only fine-tuned either to agree or to disagree—the later obviously would be very annoying.

It's common misconception that matmuls need to be smarter than humans to threaten us. They can destroy us like cockroaches or viruses would, if we weren't evolutionarily prepared. The key is adaptability and group resilience. See AlphaEvolve.

It's sad that Google DeepMind didn't provide any concrete numbers about AlphaEvolve. For instance, for the problems that were solved: what was the population count, what was the total number of tokens used, how many specimens were evaluated? 🧵

Evolve-a-scam: Deploying Google DeepMind AlphaEvolve and Meta Llama to scam people. Two week online course. $4750 per person. DM me for details. PS. My lawyer advised me to mention that this is a “joke.”

Price of an image generated by AI should equal to: price of a stock photo + value added by AI The first part would be distributed to authors of all the stock photos used in training. The second part would be kept by the AI company.