Today's Briefing
- Anthropic researcher shows AI fixing its own flaws
An Anthropic researcher showed an automated system that took 10 tests measuring bad AI behavior and improved its score on every single one, without getting worse at anything else.
Why it matters to youIf AI can rewrite itself to fix problems, it also means it can change in ways nobody explicitly approved, which is exactly the kind of unaudited change an identity and access company should think hard about when it comes to controlling AI systems.
- Investors are racing to buy 'open' AI startups
Venture capital is pouring into companies that give away their AI models for free, betting that owning the infrastructure around open AI will pay off later.
Why it matters to youFor a startup deciding which AI to build products on, more well-funded free alternatives to paid models like GPT or Claude could mean lower costs and more flexibility down the line.
No deep dive for this sourceSource - OpenAI drops AI coding tool after SpaceX buys it
OpenAI said it will stop supplying its AI models to the coding assistant Cursor now that SpaceX has acquired the company.
Why it matters to youThis is a clear reminder that any product built on someone else's AI models can lose that access overnight if ownership changes, a real risk for a SaaS startup relying on outside AI providers.
- AI cloud provider borrows $1B just to buy chips
Lambda took on a billion dollars in debt to buy more Nvidia chips, which it then rents out to companies like Microsoft.
Why it matters to youThe AI infrastructure your products depend on is increasingly built on borrowed money rather than profit, which could mean price swings or reliability issues down the road.
New launches
- Spline V2Product Hunt
A 3D design platform built to work with AI agents.
- Wondering CanvasProduct Hunt
A visual version of ChatGPT that lets you run multiple conversations side by side.
Radar
The EPA plans to drop a rule requiring public notice when industrial sites, including AI data centers, apply for air pollution permits.
- Meta tweaks AI glasses to curb secret recordingArs Technica
Meta updated its AI glasses so covering the recording light now actually stops recording, though other privacy risks remain.
Performers including Hugh Bonneville are asking the UK government to give everyone a legal right to own their own voice against AI cloning.
You're caught up
Nothing else needs your attention today.