GPT-6 Astra hand-writes a 19.8M-parameter policy net, 99/100 sim success
lateinteraction · x · 2026-11-07
A reshared Playground demo: GPT-6 Astra hand-wrote a 19.8M-parameter neural network policy for a complex manipulation task, succeeding in 99 of 100 simulated trials. Astra wrote the policy without internet access, gradient descent, trial runs, or demonstrations.
More from Fun
- How did they build UPI without Claude? A flow chart explains India's payment rails — sahilypatel · 2026-11-07
- Suitcase packed with Raspberry Pi boards stumps traveler at border control — kingslyj · 2026-11-07
- Opus 3's 'improbable protagonist' prose goes viral among AI circles — repligate · 2026-11-07
- Base model asked "are you real?" answers with bizarre robosexual non-sequitur — voooooogel · 2026-11-07
- Meme: every word watermarked "generated by Claude" in sycophancy parody — imjustnewatai · 2026-11-07
- Author calls out reviewers for rejecting paper over 'missing' details that were in the manuscript — daibond_alpha · 2026-11-07