been reading sakana’s mass. after each cycle it fine-tunes its own open model (qwen3.6-27b) on its best-ranked runs, with the workflow text stripped out of the orchestrator’s prompt. in effect the workflow it found gets distilled into the weights. the model carries it without being told it.per cycle: one shared lora (rank 64), sft on ~120 trajectories (8 tasks x 15 kept of 18 runs, ranked pairwise by an llm judge), ~6 h100s for ~19 hours.after two cycles: scienceagentbench 28.8% → 31.7%, paper delivery 57% → 76%. slight drops on swe-bench and terminal-bench, and no packaged ablation that separates the tuning from the rest of the procedure.feels like the step after “a search procedure finds the flow”: search finds it, then training absorbs it. what i want to know is whether it holds past the tasks it was tuned on arxiv.org/abs/2610.12176

this is the right direction — @huggingface huggingface.co/spaces/witnes… a central list of alignment incidents that have happened — hope frontier labs get involved here @eliebakouch @Thom_Wolf

been on a deep build of rsi repo for openkernelchallenge — almost done. looking to climb soon.

they get it

good morning — back on the hunt for the kwisatz haderach neuron.

those are neurons when triggered causes the LLM to start acting in your way my task tomorrow morning is to send 500-1000 ais to reverse engineer and find the kwisatz haderach neuron

wadafaq

some notes on harness design and specifically the different task trajectories in play. the simple duality is reading task vs long horizon. many systems today choose and optimise for one. there’s opportunity here to develop an rsi layer that evaluates and reconfigure itself based on task trajectories. meta harness with rsi seems like the way to go — hope to benchmark some of this work over the next few weeks and publish this harness as an open source project

actually how the world works youtube.com/shorts/BJnP7K0Os…

?????

been thinking today about how to approach a bounded study on swarm coordination initially i thought to approach it from a mechanistic interpretability pov. but thats a mega problem. we are still building the tools for mechinterp at scale. so i’ve taken a step back to measure it directly through harness design: a general-purpose harness built around two fundamental primitives - send(b, m): message actor b, only if you hold b’s address - create(code, budget): start a new actor with part of your budget and get its address the idea is to give a model just these two primitives, run autoresearch on a bounded goal, and study its native behaviour: what structure it designs (tree vs dag vs mesh vs ...), how it coordinates inside that structure, how it uses environments such as repls, and how all of that differs between models this takes a different view from prime agent or dynamic workflows. both leave the structure to the model, but each hands it a fixed vocabulary: dynamic workflows a script of agent(), parallel() and pipeline(); prime agent a repl with rlm() sub-calls and messages only between parent, sibling and child. when it comes to multi-agent harnesses, my gut feel says we should build harnesses that tap into the emergent properties of the model. every model is different as well. why not have self-designed harness on the fly.