DAILY INTELLIGENCE BRIEF

The Vault — AI Edition

Wednesday, August 05, 2026
15 stories
12,409 words
60 min read

Business

5
01

Anthropic Signs $10 Billion Compute Deal With Startup Volta for a Norway Supercomputing Site

A company that did not exist eight months ago just became one of the most important names in Anthropic's supply chain.

4 minRead →
02

Anaconda Buys Enkrypt AI to Secure the 'Trillion-Token Enterprise'

Anaconda has spent a decade as the plumbing beneath enterprise Python. Now it is betting on the next problem: not whether AI agents move fast, but whether anyone can prove what they are doing.

4 minRead →
03

Google Consolidates AI Power in California as Hassabis Steps Up to Chairman

Google reshaped the command structure of its most valuable asset, elevating a little-known engineer to run its AI empire and moving Demis Hassabis into a chairman's seat.

4 minRead →
04

BYD Unveils Its First Humanoid Robot as Automakers Pile Into Embodied AI

The world's largest maker of electric vehicles wants to sell you a car. Increasingly, it also wants a robot to do the selling.

4 minRead →
05

Google.org Commits $15 Million to a New Digital Futures Fund Cohort

Google.org has poured another $15 million into the debate over how to govern artificial intelligence, naming a fresh cohort of think tanks and academic institutions.

4 minRead →

Llms Genai

4
06

'Tokenmaxxing Is Not What We're Optimizing For': Microsoft Puts Divisions on AI Token Budgets

For two years, Microsoft's pitch to the corporate world was to put a copilot in front of every employee and let it run. This month, it started telling its own engineers to slow down.

4 minRead →
07

Simon Willison Ships LLM 0.32, Deepening the Open-Source CLI's Tools and Fragments

When Simon Willison pushed version 0.32 of his open-source llm tool on August 4, he called it the most significant new version since the project's launch.

4 minRead →
08

'Exploratory Sampling' Aims to Make LLM Outputs Genuinely Novel, Not Just Reworded

Ask a large language model for ten story ideas and you usually get one idea wearing ten outfits. Closing that gap is the target of a 2026 decoding method called Exploratory Sampling.

4 minRead →
09

'Tool-DC' Uses Divide-and-Conquer to Fix How LLMs Pick the Right Tool

When an AI agent has three tools to choose from, picking the right one is trivial. Give it thirty, and something quietly breaks.

4 minRead →

Research

3
10

Latent World Models Let Language Models Plan Against Compressed State, Not Raw Tokens

A language model asked to plan ten steps ahead has had only one place to think: the token stream. A wave of 2026 research argues that is the wrong substrate for planning.

4 minRead →
11

MAS-Orchestra Trains Multi-Agent Systems With Function-Calling RL and Debuts MASBENCH

The pitch for multi-agent systems has been intuitive: if one language model is smart, a team should be smarter. The reality has been messier.

4 minRead →
12

'Act As a Real Researcher': A New Benchmark Suite Tests LLMs Across the Full Research Lifecycle

Give a modern coding agent a well-specified experiment and it will often run the thing end to end. A new benchmark asks a harder question: would you trust it as a research intern?

4 minRead →

Policy

3
13

Ninth Circuit Frees Perplexity's Comet Agent to Shop Amazon, Vacating an Injunction

A federal appeals court handed the AI industry its first major appellate win on whether software agents can roam the web on a user's behalf.

4 minRead →
14

White House to Host OpenAI, Anthropic and Google on a Voluntary AI Safety-Testing Framework

The Trump administration gathered the biggest names in AI at the White House to walk them through a finished national framework for voluntary safety testing of the most powerful models.

4 minRead →
15

Rust Publishes an LLM Policy: Bots Can Review and Refine, but Not 'Create'

When the Rust project finally wrote down its rules for AI, the entire philosophy fit in a single sentence: help, refine, review, but do not create.

4 minRead →