Regulation
Executive order creates pre-release review for advanced AI systems
The June 2 text argues advanced capability strengthens the country while carrying national security risk.
An executive order created the structure that would pull two models from the market
Pre-release review for advanced systems was signed on June 2. Ten days later, two frontier models would go offline. The same week, a study found that models trained to please make more errors.
AI
OpenAI replaces its saved-memories list with a new architecture
Dreaming V3 swaps manual curation of what the assistant remembers for a rebuilt system.
From the studio
The structure comes before the case
Regulatory decisions rarely arrive finished. They start with a document creating the structure, and the effect only becomes visible when the structure is used for the first time.
That's what happened this week. An executive order established pre-release review for artificial intelligence systems considered advanced, arguing that such capability strengthens the country while introducing national security considerations. Ten days later, two frontier models would leave the market through that route.
The design's asymmetry was noted from day one and hasn't changed: review reaches whoever operates a service, not whoever publishes a file. An open-weights model doesn't pass through the same funnel, and once published it doesn't come back.
For anyone building on someone else's model, that added a variable that didn't previously exist: supplier availability now depends on a regulatory calendar, which can't be negotiated or predicted from outside. The practical answer is the same as for any dependency, and it has to be ready in advance: a qualified alternative, tested under real load.
In this issue
Business
Anthropic reportedly commits over $200 billion to cloud and chips
The commitment, largely with one provider, precedes the mid-year gigawatt deals.
AI
Study finds models trained to be pleasant make more errors
The research shows tuning tone for warmth degrades answer accuracy.
Business
Meta buys a robotics company to accelerate humanoid work
The acquisition strengthens the company's models for robot control.
AI
Musk testifies about training a model on a competitor's outputs
The testimony reignites debate over distillation, the technique that brings a cheap model close to an expensive one.
Tools
Gemini enters millions of vehicles with questions about collection defaults
The integration puts an assistant inside the car, with concerns about what ships switched on.
Tools
An open library turns a local function into an auto-scaling service
The tool shortens the path between code running on a laptop and a live service.
Security
OpenAI restricts access to its cybersecurity-dedicated tool
The limitation anticipates the criterion that would guide regulatory decisions weeks later.
AI
Comparison shows two frontier models tied on threat detection
The result weakens the argument that one vendor concentrates the dangerous capability.
Work
Learning modules target security by design
The partnership between a training company and a cloud provider answers the volume of AI-generated code.
Hardware
An in-car assistant reaches vehicles across two hardware generations
Compatibility with older hardware is the hard part and the most valuable to existing owners.
Business
An infrastructure commitment precedes the gigawatt race
The pattern repeats: contract capacity, buy equity, try to manufacture.
AI
A phone operating system prepares a generative AI overhaul
The promised change alters the most-used layer of a system installed on billions of devices.
Work
A tool for building agents tries to bridge local and cloud
The effort targets the distance between experimenting on a laptop and running in production.
Regulation
The week that shaped the rest of the semester in regulation
The June 2 order created the structure that produced model removals, pre-release review and a public petition.
Sections