Lead story: 🧠 AI’s inner monologue goes public

Researchers from OpenAI, DeepMind and Anthropic, plus a slew of other AI bigwigs including Geoffrey Hinton and Safe Superintelligence CEO Ilya Sutskever, have published a paper-cum-open letter arguing that every step of an AI’s internal reasoning—its chain of thought (CoT)—be captured and made auditable. The signatories argue “glass box” CoT logs should be not a nice-to-have but a requirement to operate safely. Research shows that a smaller AI model can effectively monitor a more powerful model’s chain-of-thought. However, powerful models may still learn to obscure their reasoning, so this approach doesn’t settle the AI safety debate.

Nevertheless, the safety argument makes sense: researcher’s get valuable data, enterprises gain error forensics and regulators get a breadcrumb trail for safety checks. [Tomek Korbak, arXiv]

Key signals, quick scan

A 30-second scan of four secondary signals that hint at where the curve is bending.

• 📝 China filed more AI-patent applications in 2024 than the US, EU and UK combined. Quantity is being matched with quality: Chinese research accounts for more than 40% of global citations in 2024 which is 4x greater than the US and EU individually, and 20x more than the UK. [TechWireAsia]

• 🚗 Waymo’s self-driving vehicles have travelled more than 100 million miles, a figure no US rival can match, giving Waymo an orders-of-magnitude advantage in real-world edge-cases (e.g. swerving past a dog) and regulatory credibility. But competition is around the corner: Uber has announced a partnership with Baidu that covers the Middle East and Asia—and perhaps soon, the US. [Inc., The Information]

• ✈️ Delta plans to price a fifth of its airline seats dynamically using AI by the end of the year. Using models developed by Fetcherr, an Israeli dynamic price startup, Delta is aiming to squeeze revenue from every flight in real time, and plans to remove static pricing completely. It’s only a matter of time before other industries follow. [Fortune]

• ⚠️ A Google autonomous bug-hunting model discovered and responsibly disclosed a critical SQLite flaw ahead of any exploitation, proving that AI can now outpace human and malware scanners in discovering vulnerabilities. This shifts cyber security from reactive patching to proactive discovery. [The Record]

Highlights from my conversations: Frontier or perish

Kevin Weil, CPO of OpenAI, shared with me Sam Altman’s test for startups building today:

If you are building at the frontier of model capabilities and you cannot wait for our next model because you know it will make your product sing, then you are probably building in the right place. If instead you are afraid of our next model because it might not have those weaknesses, that is a bad place to be building. The thing to do is reimagine use cases from first principles, building them from the ground up with AI.

Watch the full episode on YouTube. I go live on Substack every Friday at 5 pm UK time | 12 pm Eastern | 9 am Pacific.