ASI pill podcast

GPT-6 Just Did Brain Surgery on Another AI

11/4/2026
0:00
22:49
Rewind 15 seconds
Fast Forward 15 seconds

This is the audio version of a very visual film. Watch it on YouTube: https://youtu.be/fWoFu38GfM0


On November 2, 2026, Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab (MIT CSAIL) published "Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights". In eight experiments, OpenAI's GPT-6 Astra, used as it is with no special training, reads raw neural-network weights pasted into its chat as text, fixes a failing network by editing its weights, writes small networks from scratch, and edits 0.016% of Qwen3.5-0.8B's weights to stop its thinking loops (MATH-500 from 7.9% to 48.1%). The authors published every turn of every experiment (CC BY 4.0).

The blog post: https://rikiyat.github.io/blog/weight-space-programming/

The complete logs: https://github.com/RikiyaT/lm_understand-logs

Rikiya Takehi's thread: https://x.com/rikiyatakehi/status/2117313742891327994


WEIGHTS ARE CODE: a network is a brain made of numbers, Karpathy's Software 2.0, the four authors, weights pasted into a chat as text. THE TRAINED BRAIN: what a policy is, 17,253 weights, PPO as a cautious golf coach, a billion steps. READING THE WEIGHTS: Astra names four games from their raw weights (the X-ray of where each network looks), draws three diffusion models' shapes, finds the copying heads in torn-open language models, misses the sycophant from weights alone. THE FIX: the Kinetix car that a billion steps of PPO never got over the ramp, hot and cold, Astra's 51-line bypass, 1,024 of 1,024. WRITTEN FROM SCRATCH: a music box of 815 pins, Breakout 517 vs 25, in-situ distillation, a 1.5 million weight Go player. THE CONTEXT WALL: four tokens per weight, an 872,000-token window, Kimi K2's five million windows, the MRI scan, a swarm of surgeons. BRAIN SURGERY: Qwen3.5-0.8B's thinking loops, 158 turns, 131,072 weights, the epilepsy analogy, Gemma, Lucas Beyer's catch. OLD HAT: Jürgen Schmidhuber's 1991 fast weight programmers, whiteboard vs stone, the DeltaNet twist, the verdict. THE LOOP: self-surgery, open weights vs the vault, the sleep scaling law, recursive self-improvement, Maitreya.


CHAPTERS

0:00 Can one AI do brain surgery on another AI?

1:33 01 Weights Are Code

3:52 02 The Trained Brain

5:14 03 Reading the Weights

8:13 04 The Fix

10:06 05 Written from Scratch

12:32 06 The Context Wall

14:12 07 Brain Surgery

17:12 08 Old Hat

19:17 09 The Loop

21:51 Outro and the Maitreya prayer


MORE ON THE ASI PILL

Recursive Language Models, Alex Zhang: https://youtu.be/HZyklT50Rdw

How an AI Thinks, Visualized: https://youtu.be/-Pi7EfVeLyg


All 42 sources are linked in the pinned comment.


CREDITS

Script written with Claude from the blog post itself, the authors' published logs (CC BY 4.0), and the posts and pages linked above. Narration in the ASI Pill's own words. Every number and every motion on screen comes from the authors' logs and figure data: the four game networks' real first-layer weights, the car's recorded runs before and after Astra's edit, the recorded Breakout episode of the network Astra wrote and its 815 nonzero weights, Qwen's 64 edited columns per layer, the 158-turn trajectory, the Go and math results per forward-pass budget. The pictures of Rikiya Takehi, Jayden Teoh, Alex Zhang, Omar Khattab and Jürgen Schmidhuber are stylized cartoon illustrations made with GPT pictures from their public photos; their real photos are shown beside their names. Blog pages, posts and web pages are shown briefly for commentary and criticism. Logos name the models we talk about; no endorsement. 3D scenes rendered with three.js; Parzival's voice is a clone made with ElevenLabs.

Narration: Parzival, the host's own cloned voice. AI-generated pictures: the illustrations and the closing still of Parzival. Photos of the people quoted: see the pinned comment.


#AI #MachineLearning #GPT6

More episodes from "ASI pill"