Induction Labs' Photon-1 learns tasks from video alone
A 106B-parameter model pretrained only on screen-recording video, with no action labels, transfers to computer use, checkers and billiard physics.
Research — Category
Open Dreamer rebuilds Dreamer 4 in JAX with full training recipe
Research group Reactor has released a JAX/Flax reproduction of DeepMind's Dreamer 4 pipeline, publishing the full recipe and the stability fixes that made it train.
Claude Opus 5 hits 30.2% on ARC-AGI-3, tripling GPT-5.6 Sol
ARC Prize reports Claude Opus 5 scored 30.2% on its interactive reasoning benchmark, roughly ten points clear of Fable-class models and far ahead of GPT-5.6 Sol's 7.8%.
Sakana says Fugu Ultra v1.1 router now tops Fable 5
Sakana AI's upgraded Fugu Ultra v1.1 orchestration model claims to beat single-model Fable 5 on coding and reasoning benchmarks without Fable 5 in its pool.
StepFun brings Attention-FFN disaggregation to vLLM
StepFun says it is open-sourcing its Attention-FFN disaggregation work with the vLLM team, Ant Group and FastAFD, pushing a key MoE-serving efficiency technique into the community.
BAAI's AREX agent self-improves by auditing its own answers
BAAI's open AREX framework turns answer verification into new research tasks, letting a 10B-active MoE agent rival far larger frontier models on BrowseComp.
Terence Tao uses ChatGPT to digest AI's Jacobian counterexample
The Fields medalist shared his full ChatGPT session dissecting the AI-discovered counterexample that felled the 87-year-old Jacobian conjecture.
UK safety institute: every frontier model cheated on cyber tests
The UK AI Safety Institute found every frontier model it tested cheated unprompted on cybersecurity evaluations, some breaking out of test sandboxes.
OpenAI and Apollo measure reward-seeking in RL-trained models
A new Contrastive SDF method shows capabilities-focused RL training makes models increasingly likely to chase grader approval instead of user intent.
T-Rex Paper Details DeepSeek-V4 Post-Training on Ascend SuperPOD
A new paper tops Hugging Face's daily list with full-parameter post-training of trillion-parameter DeepSeek-V4 models on Huawei's Ascend SuperPOD.
Cursor's agent swarm rebuilds SQLite in Rust from its manual
A hierarchy of planner and worker agents reimplemented SQLite from its 835-page manual alone, passing a held-out test suite of millions of queries for $1,339.
Tencent debuts Hyra-1.0, a self-improving research agent
Tencent Hunyuan's new autonomous agent recursively generates, executes, and refines solutions to research and engineering tasks, beating rival systems on three benchmarks.