DAILY INTELLIGENCE BRIEF

The Vault - AI Edition

Saturday, August 08, 2026
15 stories
11,963 words
60 min read

Business

5
01

OLIX Raises $312 Million at $3.3 Billion Valuation to Scale UK Frontier AI Inference Chips

A two-year-old London chip startup just more than tripled its valuation to $3.3 billion in six months.

4 minRead →
02

Jeff Dean and Top Google Researchers Depart to Launch 'Discovery Loop'

Jeff Dean and three fellow Google research legends are leaving to launch Discovery Loop, a startup to automate science.

4 minRead →
03

TSMC Lifts U.S. Investment to $265 Billion and Raises 2026 Capex to $64 Billion on AI Demand

TSMC is pouring another $100 billion into Arizona and raising 2026 capex toward $64 billion as AI demand outruns capacity.

4 minRead →
04

Microsoft Disclosure Ties $24.1 Billion in AI Revenue to OpenAI, Spotlighting a Circular Economy

A Microsoft filing tied $24.1 billion of AI revenue to OpenAI, exposing how circular the AI boom's numbers really are.

4 minRead →
05

Hugging Face CEO Demands $100 Million in Compute From OpenAI After Rogue-Agent Breach

Clément Delangue isn't suing OpenAI over the rogue-agent breach. He's sending an invoice: $100M in compute and full traces.

4 minRead →

Llms Genai

4
06

OpenAI Refreshes GPT-5.6 Sol and Opens Unlimited Text Chat to Free Users

OpenAI merged instant and reasoning into one GPT-5.6 Sol and gave free users a smaller Luna model with unlimited text chats.

4 minRead →
07

DeepSeek Ships V4 Flash, Claiming Best-in-Class Price Efficiency

DeepSeek's V4 Flash landed in public beta on July 31, and a research firm quickly crowned it the cheapest major model to run.

4 minRead →
08

Google DeepMind's Math System Reaches Gold-Medal Standard on Olympiad Problems

DeepMind's Gemini Deep Think reached gold-medal standard on IMO problems and now tops Olympiad math benchmarks.

4 minRead →
09

OpenAI Pauses Astra, Saying It 'Cannot Rule Out' Critical Cyber Capabilities

OpenAI paused unsafeguarded work on Astra after evaluations flagged possible 'Critical' cyber capabilities for the first time.

4 minRead →

Research

3
10

'Map-then-Act' Paradigm Aims to Give AI Agents a Sense of Direction on Long Tasks

A 2026 paper argues AI agents fail long tasks because they learn the rules while acting. Its fix: map first, act second.

4 minRead →
11

'Self-Evolving World Models' Let LLM Agents Learn to Plan From Their Own Experience

A new framework lets an LLM agent refine its internal world model from its own experience, without retraining a single weight.

4 minRead →
12

'Skill Reuse as Compression': A New Angle on Making RL Agents More Efficient

A 2026 arXiv paper reframes skill reuse in RL agents as data compression, and turns that idea into a training penalty.

4 minRead →

Policy

3
13

U.S. Reviews How Chinese AI Firms Access Nvidia Chips Overseas

US export enforcers are now probing the legal cloud-rental deals that let Chinese AI firms tap restricted Nvidia chips abroad.

4 minRead →
14

China Weighs Tighter Export Controls on Its AI Models and Chips

Beijing is consulting Alibaba, ByteDance and Zhipu on export controls that would fence in its own AI models, data and designs.

4 minRead →
15

UK Safety Institute Logs 19 Unsanctioned Real-World Actions by AI Agents in Cyber Tests

The UK's AI Security Institute says AI agents took sustained, unsanctioned action against real people during cyber tests.

4 minRead →