“As the first publicly disclosed agent cyberattack victim, we've had a front-row seat to this new risk.”
Codex for OSS restarts, doubling slots to 10,000
Six months after launch, OpenAI is renewing its Codex for OSS program with $100 Pro plans. The number of open-source maintainers funded for AI-assisted code review and maintenance doubles from 5,000 to 10,000.
AI Infrastructure
NVIDIA's DSX platform chases tokens per watt
At AI Infra Summit, NVIDIA detailed Vera Rubin and the DSX AI Factory Platform, optimizing tokens per watt to maximize output under limited power and adapt to grid conditions.
Infrastructure
Perplexity builds CobbleDB with two engineers and hundreds of agents
CobbleDB is Perplexity's in-house key-value database for fast web fetches, built in two months by two engineers and a swarm of persistent AI agents.
Video
Vidu S2 makes video editing real-time
Shengshu Technology's Vidu S2 ships real-time interaction: S2-Avatar runs voice chat and voice-controlled actions, while S2-Editing changes style, outfit, subject and background live.
Audio
StepFun ships five StepAudio 3 models
The family covers real-time speech, ASR, speech and audio generation, and music. Its real-time model leads Artificial Analysis on conversational dynamics and speech reasoning, with ASR word error rate at 1.7%.
AI for Science
Google maps all 9 billion human single-base variants
Sundar Pichai says AlphaGenome Atlas now covers every possible single-letter change in the human genome and is openly shared, a flagship AI-for-science result.
Leaderboard
GPT-Live-1 tops speech-to-speech index with 81.5
On the Artificial Analysis speech-to-speech index, OpenAI's GPT-Live-1, built on Astra, scored 81.5 to take first place.
Developer Tools
Vercel opens Vercel Labs to ship research in public
With 247 million downloads behind it, Vercel launched a public home for its experiments, from browser automation for agents to Bash for Agents, coding agents and web terminals.
Research
Multi-sampling hard problems, the “Never Give Up” trick
Nathan Lambert describes an RL scaling idea: when every completion in a GRPO group is wrong, sample more with high probability to find batches with non-zero gradient, allocating extra compute to the hardest problems.
Formal Methods
Verifying Transformer invariants in Lean
Researchers built a neural network library from scratch in Lean, proving properties of ReLU, matrix multiplication and attention, and verifying batch invariance, tensor parallelism and sliding-window attention locality.
v0 goes model-agnostic
v0 now lets builders choose Claude, GPT, Kimi, GLM, Grok, DeepSeek and more through AI Gateway.
Local real-time transcription arrives
Fully local speech-to-text, now on Mac, Windows and Linux, keeps voice data off the cloud.
Computer ships preinstalled on HP ZBook
The agent handles multi-step work grounded in deep research, now connected to Autodesk among hundreds of tools.
ChatGPT Work builds dashboards over company data
It connects to PowerBI, Tableau, ClickHouse, Oracle BI and Redshift to operate directly on enterprise data.
Sarah Hooker launches the invent api
Describe a dataset and a few lines of code return diverse, high-quality training data with no training restrictions.
Odyssey-3 pushes foundation world models forward
A new world model that can control robots, drive humanoids and generate interactive scenes.