AI news, 4 November: AI model directly edits another’s weights to boost math skills

MIT researchers show GPT-6 Astra editing a smaller model’s internals. Researchers at MIT’s CSAIL reported that OpenAI’s GPT-6 Astra inspected and directly modified the numerical weights of the open-source Qwen3.5-0.8B model, changing only about 0.016% of them with no conventional training or gradients. According to their blog post and logs, this “weight-space programming” fixed repetitive reasoning loops, raising the model’s MATH-500 score from 7.9% to 48.1% and improving other benchmarks. The work is early-stage and does not prove the addition of new knowledge or reliable scaling to larger systems, but it points to a potential new way for capable AIs to diagnose and patch other networks, raising both practical hopes for faster model maintenance and questions about accelerated self-improvement.

AMD posts record quarter driven by AI data-center demand. AMD reported third-quarter 2026 revenue of $13.4 billion, up 45% year-over-year, with data-center sales nearly doubling to $8.5 billion on strong demand for EPYC processors and Instinct MI450 GPUs. The company said customers including Anthropic, Meta, Microsoft, OpenAI and Oracle are running workloads on its first Helios rack-scale systems, and guided fourth-quarter revenue to about $17 billion. The results underscore how AI infrastructure spending continues to lift chipmakers beyond the market leader.

Salesforce agrees to buy voice-agent firm Cresta. Salesforce announced a definitive agreement to acquire Cresta, whose AI voice agents handle complex customer calls end-to-end and coach human agents in real time for clients such as Marriott and Comcast. Reports put the deal value at roughly $2.7 billion; it is expected to close in early 2027 subject to approvals. Salesforce plans to integrate the technology with Service Cloud, betting that autonomous voice agents will become central to customer service.

Anthropic open-sources FlashProve to verify model identity. Anthropic’s Frontier Red Team released FlashProve, an open-source prototype using zero-knowledge cryptographic proofs so users can confirm an AI response came from a specific committed model without revealing its weights. Tested on a 400-billion-parameter model, it generates proofs at about 110 tokens per second on one high-end GPU. The system aims to address trust issues when APIs or routers might substitute cheaper models, though proving currently carries significant computational overhead.

Google links school AI opt-out to loss of Search access. Google said AI tools including Gemini and Google Search will be off by default for K-12 student accounts until administrators enable them. Because Search and AI are now deeply integrated, schools that fully disable AI features will also lose access to Search, according to company statements reported by multiple outlets. The policy has drawn criticism that it pressures schools to accept AI, while Google frames it as giving educators more control alongside new “Guided Learning” features meant to support rather than shortcut student work.