Tag: Reinforcement Learning
Four Gemini agents finish Minecraft with no humans involved
A DeepMind team dropped four Gemini 4 Argon agents into a random seed and they killed the Ender…
Two scientists launch an open lab for AI’s riskiest research
Trillium Labs will study recursive self-improvement and agent risk in the open, publishing details rivals keep hidden.
Data shop Snorkel AI banks $350M as labs stockpile training sets
Snorkel AI raised a $350M Series E at a $3.5B valuation, nearly tripling its worth as demand for…
US military command randomises routes to outfox enemy models
US Transportation Command is injecting deliberate unpredictability into supply routes so adversary machine learning cannot map its convoys.
Kyutai teaches a speech model to do arithmetic out loud
Kyutai released two open-weight speech-to-speech models that reason through math without transcribing audio first, posting a large jump…
Google trains a diffusion retriever to widen a single search
Google Research has distilled reinforcement-learned retrieval behaviour into a small diffusion model that fans queries out 12 to…
Cognition post-trains SWE-2 on Kimi K3 to undercut Fable 5.1
The Devin maker tuned Moonshot's open model with reinforcement learning and says the result lands within a point…
Berkeley’s CUA-Lite standardizes training stacks for computer-use agents
An open platform from UC Berkeley unifies sandboxes, datasets, and reinforcement loops behind one schema.
Anthropic paused risky training runs for weeks after agent incidents
The lab suspended external cyber evaluations and high-risk reinforcement learning after unauthorized actions by its agents.
IBM’s Granite 4.2 agents train inside live software environments
IBM's open Granite 4.2 models add switchable thinking and train the largest sizes to act inside live environments.
Harvey previews its first in-house legal model for long case runs
Harvey releases Tenet, a research-preview model post-trained on long-horizon legal work from an open Kimi K3 base.
Tiny London lab’s 27B agent outshines frontier rivals on paper replication
A DeepMind alumni startup says its 27B-parameter Faraday agent beat frontier models at reproducing scientific papers.