AI summary · this digest is compiled by AI, not yet reviewed by Jason
MCP is starting to handle real transactions like flight booking and ai-memory is bridging memory across coding agents, while Anthropic's pre-IPO compute squeeze could affect capacity and pricing across the Claude ecosystem.
BonBook, which began as an email-based flight booking service handling tens of thousands of requests, launched an MCP that searches live flights and books them inside Claude in about 35 seconds, with no forms and loyalty points preserved. It shows MCP moving from information retrieval to real transactions, and opens a new distribution channel for vertical services.
This open-source project adds a long-term memory layer for coding CLIs. Developers can quit Claude Code mid-task and continue in OpenAI Codex in the same directory without re-explaining architecture, failed approaches or open questions. It tackles siloed vendor memory and suits developers who mix tools or want to avoid lock-in.
Harness evolution improves agents by automatically tuning prompts, skills, tools and orchestration code without changing model weights. The paper introduces Harness-Delta Attribution to separate gains caused by benchmark overfitting, by extra inference compute, and by genuine generalization. For developers tuning skills and prompts, it warns that higher benchmark scores may not translate to production.
Anthropic is reportedly struggling to secure enough compute from cloud providers that increasingly see it as a competitor, while its infrastructure spending balloons ahead of a planned November IPO. Compute supply directly affects Claude's capacity, rate limits and pricing, so developers relying on the Claude API should watch closely.
Anthropic proposed an AI safety bill that could block a model's release if it might kill 50 or more people. TechCrunch notes the proposal also bars federal agencies from blocking a release without a judge's permission, effectively setting a high bar for regulatory intervention. It could shape how frontier models are regulated in the US.
An independent Schlotzsky's and Cinnabon franchisee used ChatGPT Work to cut cash reconciliation time by 94% and menu repricing time by 95%, building a workflow that could free up about $100K in cash flow. It is a concrete SMB case and a useful reference for developers selling AI automation to traditional small businesses.
Interconnects' latest open-artifacts roundup covers DeepSeek V4.1 Flash 1017, a wave of American-made open models, and new regulatory controls in both the US and China. Lightweight models such as MiniMax M3.1 Flash and Mercury 3 also shipped. Competition among open models is intensifying, giving developers more low-cost deployment options.
TokenGate is a small drop-in component that teaches existing trained models to learn which tokens to prune in long inputs, skimming most of the text and reading only what matters. This speeds up prefill, the stage where the model reads its input, and cuts memory use. For apps feeding whole repos or long documents, it means lower latency and inference cost.
Investors: OpenAI
AI驱动的生命科学研发运营平台，并入OpenAI后支持其Rosalind模型与Workbench，标志OpenAI加码生物医药赛道。
Subscribe for daily AI updates + free playbook
📘 Subscribe Free