Who Tests the Most Powerful AI Before It Ships? Britain Pays Far More Than the US, but Not “10 Times” More

A sealed glass test chamber on a white laboratory bench with a softly glowing turquoise cube inside and two pressure gauges
Britain just gave its AI Security Institute an extra £120 million, spread over four years. The US testing center at NIST runs on about $15 million a year. Per year, Britain spends roughly 8 to 9 times more, not “10 times” from the top-up alone. What the testers found in 2026, why access matters more than money, how the EU AI Office differs, and what it means for companies and users.
View Post

OpenAI Says Its AI Proved the Exponential Time Hypothesis False. Here Is What Is Actually Proven

Aerial view of a large white maze with one straight turquoise path cutting diagonally through its walls
For 25 years, computer scientists bet that a famous puzzle can never be solved much faster than by trying everything. OpenAI says its AI proved that bet wrong and that a machine checked the proof. Explained simply, with everyday examples: what is proven, what is only claimed, why it is not P vs NP, and what it means for companies and users. Full details for the curious below.
View Post

GPT-6.1 Terra Is Cheap per Task. In Our Newsroom Test, Its Max Setting Was the Worst

Six brushed aluminium rotary dials in a row on a bright white desk, some with turquoise tops, a laptop out of focus behind them
OpenAI’s GPT-6.1 Terra costs $1 / $5 per million tokens and sits between Luna and Sol. Artificial Analysis scores it at 48 for $0.40 per task at max effort, against $0.72 for Sol and $5.98 for Claude Opus 5.5, and its six effort settings span a 7-fold cost range. In my own newsroom test (12 articles per language), Terra without reasoning matched GPT-6.1 Sol, and max was its worst setting. Why price per token is the wrong number, and what companies and users should do.
View Post