Runway's Project Continuum Turns Your Desktop Into Live AI-Generated Video
Runway Labs is prototyping a desktop where parts of your screen are generated video, not code, powered by its Solaris interface world model.
20 stories
Runway Labs is prototyping a desktop where parts of your screen are generated video, not code, powered by its Solaris interface world model.
Ai2 open-sources Olmo-core 3, a redesigned MoE training stack that hits 2.7x throughput and scales past a trillion parameters.
Kyutai trained its 100M-parameter on-device text-to-speech model with a new one-step generative objective, hitting sub-1% WER without Jacobian-vector products.
Boston Dynamics unveiled a new 13-DOF hand for Atlas, directly actuated and purpose-built for high-fidelity simulation and sim-to-real reinforcement learning.
A new discrete flow matching framework shrinks proteins like GFP and Cas9 while juggling multiple drug-like properties under hard biochemical constraints.
Fermion Research shipped a 164 MB English speech recognition model that matches a 2.5 GB teacher's accuracy and transcribes an hour of audio in 20 seconds on a MacBook Air.
A new Agent Skill called yomiyasu rewrites AI-generated Japanese by rebuilding sentence structure instead of blocking buzzwords, and ships with a lint tool.
A Princeton-led team released UniMate, a single diffusion transformer that animates any rigged skeleton from text, with code, dataset, and interactive demo.
Adaption's new Invent a Dataset system generates post-training data from a plain text description, beating five frontier model APIs on quality and diversity.
A new study finds reinforcement learning post-training trades solution coverage for consistency on agentic tasks, and proposes an adaptive sampler that avoids the hit.
DeepMind's SynthID Bio embeds invisible signatures into AI-designed proteins and 3D structures, letting DNA synthesis labs verify provenance without blocking research.
A RoBERTa classifier from Unitary scores comments across seven toxicity and identity-attack labels while actively minimizing demographic bias during training.
LOOM diagnoses why looped Mixture-of-Experts models stall after two iterations and fixes it, scaling stable recurrence to 9-12 loops with real gains.
A new intrusion detection prototype swaps text-generating LLMs for a typed-output model, cutting cost and latency while catching more zero-day attacks.
A five-level shortcut audit cracks 77 of 115 video benchmarks, then distills 505,518 questions down to 840 that actually require watching.
Upstage's new 35B MoE reasoning model hits a Pareto-optimal sweet spot for intelligence per active parameter, but verbose outputs make tasks slow and costly.
A new preprint shows that forcing looped Transformers to share one KV cache across recursions cuts memory 76-79% and improves perplexity.
A single reinforcement learning policy retargets hundreds of human hand demonstrations to dexterous robot hands, hitting 90% success with 100x less compute than prior methods.
A new study finds that when an LLM states a confidence in words, that number is tightly linked to the probability distribution over its next token.
A new open-source Claude Code plugin turns the agent into a full modding pipeline: engine recon, reverse engineering, fal-generated assets, in-game testing, and showcase video cutting.