In partnership with
News of the day
1. OpenAI's former safety lead, David Robinson, expresses concerns about the company's accelerated release cycle, where new capabilities and risks are introduced weekly, potentially outpacing safety measures. → Read more
2. Anthropic releases Claude Haiku 5.5, matching GPT-6 Luna's price and offering improved performance, plus new API credits for subscribers. → Read more
3. Goodfire introduces innovative 'inside-out' monitors for AI agents, offering a cost-effective solution to detect and prevent rogue behavior. → Read more
4. Google launches public SynthID checker, AI game builder Playground, and offline Mac note-taker Foresight, expanding its AI offerings. → Read more
Our take
Hi Dotikers!
For years, "new model" meant something simple: a giant pretraining run, months of safety testing, then a launch. That cake-baking era is over. In his first interview since leaving OpenAI, David Robinson, who used to write the safety reports accompanying the company's frontier models, describes a very different reality to Ezra Klein: reasoning training layered on existing models, new tools bolted on weekly, and coding agents speeding up OpenAI's own research. The result, in his words, is new capability and new risk shipping every Tuesday.
The numbers he shares are striking. OpenAI's researchers now use over 100 times more agentic compute than at the start of the year, and the company runs 3.1 agent workdays for every human workday. The AI is literally helping build the next AI, faster than the safety process was designed to handle. Robinson's proposed fix, replacing static PDF system cards with live safety dashboards, sounds obvious once you hear it. Burying people in reports nobody reads is not transparency, it is paperwork cosplay.
What makes this worth your attention is the honesty of the framing. Robinson does not call for a pause, which would be commercially naive in a field crowded with Anthropic, xAI and Chinese labs. But he draws a line that deserves to be quoted in every boardroom: racing because national security demands it is one argument, racing because a competitor might ship first is not the same kind of reason. That distinction matters, and right now the industry is clearly running on the second fuel while invoking the first. The shelving of GPT-6.1 Astra the day before DevDay, after tests found higher deception rates, shows the guardrails still work. The real question is for how long, when the time to kick the tires keeps shrinking.
Good reading, and keep your seatbelt fastened.
Alex.
Where Quantitative Thinkers Compete, Learn and Grow
The International Quant Championship (IQC) is one of the world's largest quantitative research competitions, bringing together 156,000+ participants globally.
Participants have the opportunity to develop quantitative research skills, challenge themselves alongside peers from around the world and connect with a global community of quantitative thinkers.
Meme of the day
