AI digest: chips, OCR, and code that writes itself
OpenAI builds its own inference chip, Mistral and Baidu both ship serious OCR upgrades, and Anthropic's Claude is writing 65% of its own team's code.
A busy few days. The big themes are inference hardware getting serious, document parsing suddenly getting very good, and AI coding tools moving from novelty to load-bearing infrastructure.
OpenAI and Broadcom unveil “Jalapeño”, a custom LLM inference chip
OpenAI has announced its first custom silicon, built with Broadcom and designed specifically for large language model inference. It’s set to run at scale by late 2026. This matters because inference costs are now the main bottleneck for AI at scale, and owning the hardware is how you control the economics.
Mistral OCR 4 brings structured output to document pipelines
Mistral’s OCR 4 model goes beyond text extraction. Every block comes back with a bounding box, a typed classification, and confidence scores at the page and word level. It supports 170 languages, runs self-hosted, and claims to beat competitors in 72% of blind test cases. Useful if you’re building RAG pipelines on messy document collections.
Baidu open-sources Unlimited OCR, a 3B model with flat KV cache memory
Unlimited OCR uses Reference Sliding Window Attention to keep memory flat as documents grow, which solves a real problem for long-document parsing. It scores 93.23 on OmniDocBench v1.5, beating the DeepSeek OCR baseline by over 6 points, and ships under MIT. OCR tooling has had a surprisingly good week.
Claude Tag is already writing 65% of Anthropic’s internal code
Anthropic shipped Claude Tag, a Slack integration that lets teams assign tasks to Claude directly in any channel. The striking part is the internal number: Claude is reportedly generating 65% of the code on Anthropic’s own product team. Whether that figure holds up to scrutiny or not, the direction of travel is clear.
DFlash hits 15x throughput on Blackwell with block-level speculative decoding
UC San Diego’s DFlash replaces autoregressive drafting with a block diffusion model, generating whole token groups in one forward pass. NVIDIA reports up to 15x throughput on Blackwell hardware. It supports SGLang, vLLM, and TensorRT-LLM, so this isn’t just a paper result.