DAILY INTELLIGENCE BRIEF

The Vault — AI Edition

Sunday, July 19, 2026
15 stories
14,790 words
74 min read

Business

5
01

Big Tech Is Being Asked to Justify Its AI Spending as Chip Stocks Post Their Worst Week Since 2025

For three years, the question Wall Street asked about artificial intelligence was how much. Last week it changed to for what.

5 minRead →
02

Meta Compute Turns Zuckerberg's Surplus GPUs Into a Cloud Business

For three years, Mark Zuckerberg asked investors to take his AI spending on faith. On July 1, he offered them something else.

5 minRead →
03

SK Hynix Slows Its HBM4 Ramp for Fatter DDR5 Margins, and Investors Read It as a Demand Warning

The most profitable memory company in the world just decided the least glamorous product in its catalog is worth more.

5 minRead →
04

Sanofi and Owkin Bet That Agentic AI Can Run Parts of Drug Discovery Itself

Sanofi has spent five years using Owkin's artificial intelligence to help it read biology. Now it wants Owkin's agents.

5 minRead →
08

Zuckerberg Concedes Meta's AI Agent Push Is Behind Schedule, Six Months After the Reorg Meant to Fix It

Mark Zuckerberg rebuilt his company around a bet on AI agents. On July 2, he told his own employees it is running late.

5 minRead →

Llms Genai

3
05

NVIDIA's Cosmos 3 Edge Pushes Physical AI to the Robot, and Japan Is the Testbed

Jensen Huang spent last week in Tokyo, and the thing he brought with him was small.

5 minRead →
06

Meta's Muse Spark 1.1 Adds a Million-Token Window and Computer Use, and Tops the Agent Benchmarks

Mark Zuckerberg had not posted on X in three years. On July 9, he broke the silence to announce a model.

5 minRead →
07

Grok 4.5's Real Claim Isn't the Score. It's Doing the Same Work With a Quarter of the Tokens.

The most important number in xAI's Grok 4.5 launch is not 83.3%. It is 15,954.

4 minRead →

Research

4
11

KAIST's VOTP Teaches Robots Human Judgment From a Handful of Videos

Five labels. That is how many human preference judgments a KAIST team needed to teach a robot arm what good looks like.

5 minRead →
12

LLM Agents Say Different Things Off the Record, and Social Framing Is Why

Give two language-model agents a job title, a shared history, and a reason to care what the other one thinks.

5 minRead →
13

Borrowing From Distributed Databases, a New Paper Lets Agents Merge Their KV Caches Instead of Their Words

When two AI agents solve halves of the same problem and hand their thinking to a third, something fragile happens.

5 minRead →
14

The Next Big Gains Aren't Coming From Bigger Models. MLSys 2026 Says They're Coming From the Systems.

The most quietly radical sentence in AI research this year is not about a new architecture.

5 minRead →

Policy

3
09

The IMF War-Gamed What Happens to Tax Revenue When AI Takes the Jobs

On a December morning in Washington, roughly 50 economists and technologists sat down at the IMF to imagine the worst.

5 minRead →
10

America's First Comprehensive AI Law Never Took Effect. Colorado Repealed It Instead.

June 30, 2026 was supposed to be the day the United States got its first comprehensive artificial intelligence law.

5 minRead →
15

AI Agents Reached Production This Year. The Governance to Control Them Didn't.

In the first ten days of July, three enterprise software vendors shipped products solving the same problem.

5 minRead →