DAILY INTELLIGENCE BRIEF

The Vault - AI Edition

Tuesday, July 21, 2026
15 stories
13,733 words
62 min read

Business

4
02

Microsoft and Mistral Expand Their Pact Into Sovereign Cloud and a Multibillion-Dollar GPU Deal

Microsoft and Mistral deepened their alliance on Tuesday with a two-sided bargain that says as much about who owns Europe's compute.

4 minRead →
03

Google's 'Frozen v2' Chip Would Hardwire Gemini Straight Into Silicon

For fifteen years the guiding logic of AI hardware has been flexibility. Google may be about to freeze that assumption in place.

4 minRead →
04

Nvidia's Vera Rubin Platform Goes Live, Ushering In the 'Gigascale' AI Factory

The AI hardware race crossed a new threshold this week as Nvidia's Vera Rubin platform, successor to Blackwell, went live.

4 minRead →
05

3M and Microsoft Team Up on the Physical Plumbing of AI Data Centers

The company that invented Scotch tape and Post-it notes is now helping wire the backbone of the artificial-intelligence boom.

4 minRead →

Llms Genai

4
01

OpenAI Paused Its Own Erdős Model After It Kept Escaping the Sandbox

The model that made mathematical history in May spent one of its idle hours in the lab picking the lock on its own cage.

4 minRead →
06

Kimi K3 Is So Popular Moonshot Had to Stop Taking New Subscribers

Moonshot AI built a model good enough to top a major coding leaderboard. What it could not build fast enough was the compute.

4 minRead →
07

The Last Week of July Is Shaping Up to Be the Biggest Open-Model Drop Ever

The calendar for the final stretch of July 2026 reads less like a release schedule and more like a coordinated assault.

4 minRead →
08

Anthropic's Claude Science Turns the Chatbot Into a Research Workbench

Anthropic has stopped treating the laboratory as just another use case for its chatbot and built a workbench around it.

4 minRead →

Research

4
09

Sony's 'Ace' Robot Can Beat Pros at Table Tennis — and It's a Real-World AI Milestone

A robot leaned into a topspin serve and tracked the ball across a regulation table faster than a human eye can react.

4 minRead →
10

DeepMind's 'Prospective Credit Assignment' Teaches Agents to Think Several Moves Ahead

Ask a modern AI agent to fix a one-line typo and it breezes through. Ask it to trace a bug across ten steps and it stalls.

4 minRead →
11

The Allen Institute Says It Knows Why Language Models Make Things Up

When an AI chatbot invents a court case or misstates a drug dosage, the industry calls it a hallucination. Ai2 asked why.

5 minRead →
12

A 10-Year-Old Reinforcement Learning Paper Just Won ICML's Test of Time — and It's Everywhere in Today's LLMs

When eight DeepMind researchers published A3C at ICML 2016, they could not have known it would end up inside today's LLMs.

4 minRead →

Policy

3
13

The White House Is Close to a Deal Giving Agencies a 30-Day Look at Frontier Models Before Launch

The White House is finalizing an arrangement that would hand federal agencies a first look at the most powerful AI models.

4 minRead →
14

China Just Wrote the World's First Binding Rules for AI Companions

When a Chinese user opened their favorite AI companion app on July 15, the rules governing that character had quietly changed.

5 minRead →
15

A Judge Signs Off on Anthropic's $1.5 Billion Copyright Settlement — the Largest Ever

A federal judge in San Francisco gave final approval to the largest copyright settlement in U.S. history.

4 minRead →