Frontier AI strikes the best team deal just 14% of the time
The cooperation test Stanford, Oxford and MIT just open-sourced — plus a lab-video-to-sim method and DeepSeek's next model moving to Huawei chips.
Copy markdown
A teamwork eval you can run this weekend
Stanford, Oxford and MIT's Collaborative Intelligence Benchmark runs 18 models through 240 cooperation tasks: frontier models reach the best joint deal in just 14% of conflicts and accept a weaker offer 81% of the time — even after asking what the teammate needs in 99% of runs — while more reasoning effort often made it worse (Opus 5.5's concession rate rose 90%→96%). They open-sourced 160 tasks plus a 40-task lite subset (Kendall's τ=0.95 vs the full set), so you can stress-test your own agents' handoffs cheaply instead of trusting capability leaderboards as a proxy for teammate quality.
Turn a lab video into a runnable simulation
Lab2Sim (South China Univ. of Technology + Hong Kong PolyU) tests whether a model can watch an experiment video and emit a browser-runnable web simulation; its EduVideo2Code pipeline — extract procedure, compile to a state machine, run, then self-repair — lifted apparatus fidelity by up to ~24 points and hit a 100% runnable rate. The paper is on Hugging Face now, with code and the 218-video dataset promised at the project page — a near-term unlock for auto-generating interactive explainers from any how-to clip.
DeepSeek's next model is going all-Huawei
DeepSeek is reportedly standing up at least 160,000 Huawei Ascend 950DT chips at a new Ulanqab data center to begin pre-training its V5 model as early as late 2026. For the many builders leaning on DeepSeek's cheap open weights, it signals the supply behind those drops is decoupling from Nvidia — which affects how reliably future DeepSeek releases ship under US export controls.