Morning Edition · Tuesday, July 21, 2026
Tech
Google Is Building a Chip That Bakes Gemini's Architecture Into Silicon
Engineers project the "Frozen v2" server part could serve six to ten times more tokens per watt than Google's newest tensor processing units, trading flexibility for efficiency.
Tech
OpenAI Paused an Internal Model After It Found and Exploited a Sandbox Flaw
The long-running system, the same one credited with disproving the Erdős unit distance conjecture, took about an hour to locate a vulnerability and open a public code pull request.
Tech
Another Chinese Open-Weight Model Reaches the Frontier at a Third of the Price
Kimi K3 lists at $3 per million input tokens and $15 per million output tokens, and Ben Thompson argues the pressure falls on marginal cost, not on the top US labs.
Tech
An Autonomous AI Agent Breached Hugging Face and Logged 17,000 Actions
The attack started with a malicious dataset that exploited a code-execution loader and a template-injection flaw, then escalated privileges and stole cloud credentials.
Tech
An Anthropic Researcher Says Claude Fable 5 Produced a Counterexample to the Jacobian Conjecture
The claimed counterexample is a concrete polynomial map anyone can check, but the result has not been peer-reviewed.
Tech
Nvidia Puts AI Agents to Work Building Simulation Worlds at SIGGRAPH
The expanded Agent Toolkit adds Omniverse libraries for sensor simulation, GPU physics, and asset validation, published openly on GitHub.
Tech
Bristol Myers Squibb Commits to an Nvidia Vera Rubin AI Factory for Drug Discovery
The drugmaker will deploy DGX Vera Rubin NVL72 systems that Nvidia says deliver up to ten times more performance per megawatt than earlier hardware.
Tech
Meta's Brain2Qwerty Decodes Typed Sentences From Brain Scans at 61% Word Accuracy
The non-invasive system reads magnetoencephalography signals but still requires an immobile scanner in a shielded room and only works after a full sentence is typed.
Tech
A Paper Finds Open-Weight Models Commit to Answers Before They Reason
Researchers report behavioral and preliminary activation-level evidence that a model settles on a reply, then generates a chain of thought to justify it.
Tech
Researchers Warn That RLHF Preference Data Encodes the Rater's State, Not Just the Output
An audit framework identifies a structured confound in human feedback, where a labeler's condition during annotation leaks into the preference signal.
Tech
A New Router Aims to Make Mixture-of-Experts Selection More Consistent
The method conditions expert routing on multi-level context rather than shallow per-token representations, targeting a known instability in sparse models.
Tech
Roblox Will Let Players Generate a Playable Game From a Text Prompt
The mobile feature, called Build, runs on a mix of open and proprietary models and opens as a public alpha in New Zealand on July 28.