DAILY INTELLIGENCE BRIEF

The Vault — AI Edition

Wednesday, July 22, 2026
15 stories
14,359 words
64 min read

Business

4
01

Meshy Raises Nearly $400M at a $1.5B Valuation, the Largest AI-3D Round Yet

For most of the generative-AI boom, 3D was the corner of the field that investors admired but rarely bankrolled at scale.

4 minRead →
02

Humanoid Becomes Europe's Newest Robotics Unicorn With a €133M Round

London's Humanoid raised €133M at a €1.1B valuation to become Europe's first pure-play humanoid robotics unicorn.

4 minRead →
03

Moonshot AI Seeks a $50B Valuation Ahead of a Hong Kong IPO

Beijing's Moonshot AI is in talks to raise at up to $50B ahead of a Hong Kong IPO, powered by its top-ranked Kimi K3 model.

4 minRead →
04

Microsoft Commits $10B to Build Out AI in Japan

Microsoft will pour $10B into Japan through 2029, betting sovereign AI infrastructure is the next competitive moat.

4 minRead →

Llms Genai

5
05

Google Ships Three Gemini Flash Models — but Still No 3.5 Pro

Google shipped three Gemini Flash models on July 21 — cheaper, faster, more efficient — but the delayed 3.5 Pro was a no-show.

4 minRead →
06

Google Says It's Begun Its Most Ambitious Pretraining Run Yet, for Gemini 4

Google concedes Gemini 3.5 Pro is still late while touting its most ambitious pretraining run yet for Gemini 4.

4 minRead →
07

Gemini 3.5 Flash Cyber Turns AI Security Into a Three-Way Race

Google gated Gemini 3.5 Flash Cyber to governments and partners, joining Anthropic and Microsoft in a rare convergence on restraint.

5 minRead →
08

Claude Cowork Turns Screen Recordings Into Reusable Skills

Anthropic's new Cowork feature lets users record and narrate a task once, then have Claude turn it into a reusable skill.

4 minRead →
09

Block Launches Buzz, an Open Workspace Where AI Agents Are First-Class

Block launched Buzz, an open-source Nostr workspace where AI agents get their own cryptographic identities and sign their own work.

5 minRead →

Research

3
10

MIT Gives Chip-Based Lidar a Wider View Without Any Moving Parts

MIT engineers redesigned a lidar chip's antennas into different shapes to widen its field of view with no moving parts.

5 minRead →
11

Substack Adds AI Detection Just as Research Documents Its Limits

Substack rolled out Pangram AI detection the same week Epoch AI showed detectors miss up to 18% of style-imitated text.

4 minRead →
12

A New Study Says Self-Correction, Not Size, Is What Makes Reasoning Models Work

A July preprint says self-correction, not model size or chain length, is what makes reasoning models succeed.

5 minRead →

Policy

3
13

AI Lobbying Spending Hits Record Levels in the Second Quarter

AI labs poured record sums into Washington in Q2 2026, with Anthropic's $1.97M outspending Nvidia as rules loom.

4 minRead →
14

Meta Says Its AI Moderates Better Than Humans. Users Who Lost Their Accounts Disagree.

Meta says its AI moderates better than humans. At the scale of billions, 'better on average' still means countless wrongful bans.

4 minRead →
15

A Federal Judge Dismisses xAI's Trade-Secret Suit Against OpenAI, With Prejudice

A federal judge dismissed xAI's trade-secret suit against OpenAI with prejudice, finding no evidence of knowing misappropriation.

4 minRead →