Multiverse Computing’s Sept 21 Hugging Face write-up models transformer-block deletion as an Ising optimization problem so candidates can be scored without rebuilding each pruned model. On Llama-3.3-70B-Instruct with 40 of 80 blocks removed and no retrain, it scored 76.9 MMLU vs 54.0 for a block-influence baseline (original 82.2); GitHub code is patent-pending and research/non-commercial only.
Mint, citing The Information, reports OpenAI and Anthropic are negotiating a binding pact for mutual API access to stress test each other’s commercial models for safety flaws, with a no retain data rule. The talks sit beside Altman’s support for Amodei’s push to embed independent lab evaluators and follow a looser 2025 cross test…
CNCF TOC Initiative 1746 is evaluating whether Anthropic’s Model Context Protocol can be the default Kubernetes native wire spec for distributed agents, with proposed MCP for Clusters and Agent CRD deliverables. Solo.io’s sandbox kagent already ships MCPServer/Agent CRDs as a preview; this is guidance under review, not a ratified standard…
Light Origins shipped Light O1 on Sept 21: pretrain on recovered human actions from internet video, then adapt the same base to LightBot and Unitree G1. It also released Light O1 Preview (6B, Apache 2.0) on Hugging Face for text to motion; full robot demos still need separate controllers…
Cloudflare made Python Workers generally available on Sept 21, so FastAPI, Django, and Flask can talk to R2, D1, Workers AI, Queues, and the Python MCP package without a JavaScript conversion layer. Native C/C++/Rust packages still need PyEmscripten wheels…
RuntimeWire reverse engineered Codex desktop build 9922 and found a feature gated Cloud path (vs Cloud Legacy) that lets the agent draft environments, then attach Tailscale auth keys, proxy delivered network secrets, and Azure OIDC workload identity. Public docs still describe manual legacy setup; launch timing is unconfirmed…
Spotted something we missed? Start a thread.