Unverified50% confidenceFactExact time
MIT found that frontier models can directly hand-write neural networks, and the Breakout network that GPT-6.2 Astra wrote from scratch scored 20 times that of the PPU-trained version.
1
Sources
50%
Confidence
Long-term
Relevance
11/4/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedWithout any environment feedback, GPT-6 Astra can write small game networks from scratch, and these networks outperform networks trained with PPO for 100 million steps.70% similarUnverifiedGPT-6 Astra can determine what task a network was trained to perform simply by looking at its raw weights.66% similarUnverifiedThe testing team invested more than 20 billion (20B+) tokens in a comprehensive exploration of GPT-6 Astra.66% similarUnverifiedGPT-Live unifies the perception and generation of audio streams into the same end-to-end neural network through joint training, avoiding the latency accumulation of the traditional serial pipeline (VAD→ASR→LLM→TTS).63% similarUnverified从GPT-2到GPT-4，训练算力需求增长超过1000倍，而同期芯片性能提升约10倍63% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1133393API
curl https://kongchang.com/api/v1/knowledge/claims/1133393MCP
get_claim(id=1133393)