DAILY INTELLIGENCE BRIEF

The Vault — AI Edition

Thursday, August 20, 2026
18 stories
17,999 words
86 min read

Business

5
01

Marvell Hands Google a $12.2 Billion Stake Option to Win Its Custom AI Chip Business

For years Marvell Technology was the chip designer that Google did not call. On Wednesday, August 19, Marvell fixed that with a warrant worth up to $12.2 billion.

5 minRead →
02

Together AI's $800 Million Bet Lands at IBM as Capital Keeps Shifting From Training to Inference

When Together AI closed its Series C, the number that made headlines was 800 million dollars. Six weeks later came the number that actually explains it: 240 million.

4 minRead →
03

The Cohere Round That Isn't: Chasing a $400 Million Rumor Into a $20 Billion Reality

The number circulating this month is 400 million dollars: a growth round that supposedly pushed Toronto-based Cohere past 2.5 billion in lifetime funding. The Vault could not verify it.

4 minRead →
04

London's OLIX Raises $312 Million to Put Light Inside the Inference Chip

The largest semiconductor venture round in European history has gone to a company that has not yet shipped a chip.

5 minRead →
05

Unitree's CEO Says Robots Are Nearing Their ChatGPT Moment. The Shipment Numbers Are Already Moving.

Chinese robot makers filled the World Robot Conference in Beijing with humanoids sorting parcels and packing phones. The shipment data partly backs the hype - and partly undercuts it.

5 minRead →

Llms Genai

4
06

Meta AI Lands on the Mac With Screen Sharing, and a Permissions Prompt Worth Reading

Meta shipped its first Mac app on Wednesday, and the most consequential thing in it is not the chat window. It is the macOS dialog box that appears the first time you ask the assistant to look at your work.

5 minRead →
07

Thinking Machines Shrinks Inkling by Nearly Three Quarters and It Scores Higher

Fifteen days after putting a 975-billion-parameter model on the internet, Thinking Machines published a second one roughly 72 percent smaller that beats the first on four benchmarks. Model families are not supposed to work that way.

4 minRead →
16

OpenAI Cut Its Cheapest Model by 80 Percent and Reset the Floor Under Inference Pricing

The cheapest way to run a frontier-lab model through OpenAI's API is now 20 cents per million input tokens. It was one dollar until July 30.

4 minRead →
17

Ornith-1.5 Writes Its Own Homework, and Lands on Claude Opus's Doorstep

An outfit almost nobody had heard of two months ago published three open-weight models on August 19 and put a number on the board that frontier labs have spent the year defending.

5 minRead →

Research

4
11

Coding Agents Keep Rediscovering the Same Repo. SkillForge Makes Them Take Notes.

A paper posted August 19 proposes a fix that is less about smarter models than about homework: make the agent practise on a repository before any real bug report arrives.

5 minRead →
12

An AI Picked Two Superconductors. Then a Lab in Texas Actually Made Them.

A machine-learning screen pointed at two compounds nobody had bothered to cool down and check. Both are now on the list of confirmed superconductors - because a lab actually made them.

5 minRead →
13

Generated Video Falls Apart After a Few Seconds. A Training-Free Method Stretches It to Two Minutes.

Ask a modern video model for five seconds and you often get something close to photoreal. Ask for a full minute and you can watch it die.

5 minRead →
18

A 65,000-Parameter Model Beats Frontier LLMs at Finding Which Agent Broke

When a team of AI agents botches a task, someone has to work out which agent broke it. The field's reflex answer has been to throw another language model at the wreckage.

5 minRead →

Policy

5
08

Nobody Passed: The First Independent Scorecard on Frontier AI Control Gives OpenAI and Anthropic a C+

Five of the world's most powerful AI companies have been graded on a deceptively simple question - can they keep their own AI systems under control? The best mark anyone earned was a C+.

5 minRead →
09

Beijing Tells Washington Not to Make Countries Pick an AI Side

Asked about an American plan to tell dozens of governments they cannot belong to both AI camps, Chinese spokesperson Lin Jian never once said the words United States. He did not need to.

5 minRead →
10

A Senate Committee Advanced a Right to Your Own Likeness. The Floor Is Where It Gets Hard.

The vote that would hand every American a property right in their own face did not happen in August. It happened on June 18, in a markup that ended in a voice vote with no recorded dissent.

5 minRead →
14

Colorado Wrote America's First AI Law, Then Repealed It Before It Ever Took Effect

Colorado's landmark AI statute never regulated anyone. Signed in 2024 as the nation's first comprehensive state AI law, it was postponed twice, sued, and finally repealed by the legislature that wrote it.

5 minRead →
15

Germany's Machine Builders Want Europe to Treat Humanoid Robots as Industrial Policy

Read closely, the August 19 statement is a position, not a policy paper. It names no subsidy instrument, no procurement quota, no local-content threshold.

5 minRead →