AI in Technology

OpenAI’s page title points to engineering for the physical world, but the material reviewed provides no details about the project or its results.

A November 2 research post reports GPT-6 Astra reduced Qwen3.5-0.8B’s MATH-500 score-limiting loops by editing two weight tensors.

The New Stack reports message threads are speeding up Claude Code, while token costs remain unchanged. Performance and availability details are not provided.

A headline describes Anthropic research on provable inference for language models, but the method, results and practical scope are not available.

A headline-only report says AI contributed to 26% of next-generation Claude development, but provides no methodology or further details.

A report says Astra changed 0.016% of Qwen3.5-0.8B’s weights, lifting its MATH-500 score from 7.9% to 48.1% in the reported test.

Anthropic is reported to say it may be nearing recursive self-improvement, but the available report provides no details to verify the claim.

OpenAI has published a page titled “Scaling Git at OpenAI,” but the material available here does not include its text or technical details.

A ByteDance Seed-attributed headline says Chinese AI labs are pursuing self-improvement without changing model weights, but provides no evidence or methods.

A headline reports that Z.ai’s founder sees self-improving AI as a step toward AGI, while stressing that safety remains key.