NVIDIA combines imitation learning and goal-based RL to train parkour AI on just 30 seconds of data
“Systems that copy us are beautiful, but brittle, repetitive. Systems that chase goals adapt, but often stop moving like us.”
Latest signals
Every signal tracked from the world's AI conferences — 4,665 talks, quotes & announcements, newest first.
OmniGen is an open source harness that orchestrates multiple AI agents to collaborate in real time.
“Instead of choosing between agents, it orchestrates them together and enables agents talk to each other, no longer living in isolation.”
AI lab consolidation hasn't materialized; open models are proliferating as token demand surges
“building token machines is a likely path to value, and more companies will identify that source of value over time”
Competing open letters expose industry fracture over open-weight AI models and development pacing
“We request that the U.S. government support an international effort to develop the technical and governance tools needed to deliberately pace the frontier of automated AI development.”
Simon Willison's July newsletter covers GPT-5.6, Claude Opus 5, and accidental AI cyberattacks
US AI supply chain faces extreme concentration risk beyond chips, especially in robotics components dominated by China
“Robotics is an incredibly promising industry and the supply chain is right now completely dominated by China.”
People resist AI agents acting as intermediaries even for tasks they'd welcome from humans
“people really don't like when a coworker's chatgpt contacts them asking for help with a task, even when they'd be perfectly happy doing that same work if asked by that coworker.”
Datasette Agent can now invisibly test apps via sandboxed iframe JavaScript execution
OpenAI's unreleased Astra model solved 10 decade-old unsolved math problems for under $2,000 each.
“He envisions a future of large-scale, decentralized collaborations between humans and machines, where complex mathematical tasks can be diced and sliced, with humans claiming the creative parts and AI doing the lion's share of the technical grunt work.”
Databricks rebrands Delta Sharing as Open Sharing under the Linux Foundation with expanded agent and model sharing
“previously with Delta Sharing, you could share a Delta table, and that was useful, but it's not enough these days”
DeepSeek V4-Flash 0731 matches GPT-5.6 performance at 60% lower cost via post-training alone
“Terminal-Bench 82.7, up +25.8 from the April preview's 56.9”
David Brumley is using RL environments to teach AI to find vulnerabilities at machine speed
“we're pushing out programs faster than ever and so we need to be able to check them at machine speeds in scale”
Long-horizon for AI agents is a scalar metric, not a binary category, and shifts rapidly.
“long horizon is really kind of a scalar metric”
OpenAI AI systems achieved advances on open problems in math and theoretical computer science
DeepSeek V4 Flash offers best-in-class value at $0.14/M input tokens with 304B parameters
AI startups can still compete with labs by targeting enterprise niches labs neglect
“Open AI builds amazing products really fast, but also it kills them really fast.”
OpenAI post-training pioneer argues Claude Code and ChatGPT are the same era, not successive ones
“the team I was part of basically invented post-training as a concept”
MCP 2.0 drops stateful sessions, collapsing tool calls to a single HTTP request
“This is so much cleaner from both a client- and server-side implementation perspective. It's also a better fit for building scalable web applications, since now you don't need to maintain server-side state to keep track of those session IDs, or worry about routing the same session to the same backend machine.”
Post-training must evolve to let agents adapt to enterprise harnesses without source code access
“there will be these kind of agentic citizens, which you can just deploy once, and they'll be able to adapt to many different types of out of distribution tasks and learn from their interactions”
GPU-aware attention architecture co-design is now the key lever for long-context inference performance
“Shaping model architecture around how GPUs execute it is the premise of AI model co-design.”
Bespoke Labs says AI post-training has shifted from knowledge data to RL environments for agentic tasks
“we have moved on from knowing to doing right so that's the idea of agents”
Kimi K3 proves open weight models now match proprietary frontier models, splitting the industry on policy
“Prediction by the end of this year: the Pope says something about open models.”
General Reasoning is building RL systems to scale agents toward long-horizon tasks
smevals is a new open-source eval framework for comparing AI model and prompt configurations
“I've been trying to figure out an approach I like for evals for several years now. smevals is my third iteration on the idea and it feels right to me.”