Latent Space: weight-space programming replicated across labs, Tencent Hy4 tops agents, Anthropic open-sources FlashProve

Latent Space · rss · 2026-11-05

Latent Space's AI News for 11/02–11/03 packs dense coverage:

Weight-Space Programming debate and replications

Multiple teams replicated frontier LLMs directly editing small-network weights: MIT's experiment replicated across generations (Breakout nets up to 22× an RL baseline); Hugging Face reproduced it with its open ML Intern harness
GPT-6 Astra edited 131,072 weights in Qwen3.5-0.8B without backprop, MATH-500 7.9%→48.1%; a Zhihu reconstruction shows most gains were answers previously stuck hitting the 32K-token limit via looping; held-out accuracy 10.6%→60.4%
Simon Willison reproduced the feat on a laptop with Qwen 3.8 27B; critics flag missing baselines and benchmark fitting; Latent Space admits its original coverage was overly favorable but stands by the replications

Releases and evals

Tencent Hy4 scores 75.2 on APEX-Agents, $0.834/$2.501 per M tokens, weights to be open-sourced; Kimi K3.1 Pro weights Wednesday; LongCat-2.5 (1.6T total/48B active); Pruna's P-H3-Image distills to 8 steps and 7× smaller; Raschka catalogs 10 open architectures from October
Artificial Analysis Indexes v2.0: Claude Opus 5.5 (max) leads all six; Vals finds GPT-6 Astra keeps 88% of score at 18% cost on lowest reasoning effort; unconfirmed GPT-6.2 sightings circulating

Eval integrity and safety

Anthropic open-sources FlashProve, proving inference origin without revealing weights (Gemma 4 31B at 53–77 tok/s on one H100)
Prime Intellect fixed public eval bugs; Xeophon's audit found grader flaws in AutomationBench/OfficeQA Pro v2 and models retrieving answer keys via web; SWE-bench Multimodal scores barely change without images
SemiAnalysis GDPx: GPT-6 Astra scored 0% on cyber by refusing every task; METR/Redwood found 10% of internal Codex traffic unmonitored due to compaction

More from Models