AI Reading for Thursday October 29

Dario Speedwagon: Six weeks after urging an industry slowdown, Anthropic CEO Dario Amodei told investors the company would race ahead - The Washington Post

Mistakes were made: Dario Amodei defends Anthropic's compute deals and weakened pause commitments to critic Nathan Calvin, conceding some items "could have been handled better" - Don't Worry About the Vase

The agony of the early Anthropic employee - Nothing Human

Anthropic tells Newsom courts should be able to halt risky AI models; OpenAI backs shutdown controls but not court orders - TNW | Government-policy

Damodaran values Anthropic at about $1.5 trillion - Musings on Markets

Barclays' Alexander Altmann says Anthropic's IPO filing and Oracle and Microsoft earnings rattled the AI trade, raising doubts that AI spending is rate-proof - Bloomberg

Anthropic and OpenAI's cash burn may leave others, not the frontier labs, capturing most of AI's value - Fast Company

The Spy Who Loved GPU: Chinese businesswoman who had a ‘close relationship’ with Jensen Huang was assessed to be an intelligence asset passing info to Beijing. President Trump was briefed and took no action. - The New York Times

Nvidia's $12.9 billion Hugging Face deal draws DOJ scrutiny - CNBC

OpenAI and Anthropic claim AI-generated proofs disprove the 27-year-old exponential time hypothesis, upending decades of computer science research - WIRED

P ≠ NP means there are problems whose solutions can be verified in polynomial time but cannot be found in polynomial time. Seems obviously true, but has never been proved. Much of modern cryptography relies on related assumptions about computational hardness and asymmetry. It’s easy to check a solution, but enormously expensive to find one.

ETH is stronger, says the set of 3SAT problems, which is NP-complete, needs at least 2^(kn) time for some k > 0. So at least exponential, not just polynomial.

OpenAI claims to have solved a class of 3SAT problems in 2^(n/√log n) which is better. So ETH as stated is false.

Many papers say if ETH is true, then X. They become dead letters. So maybe ETH just needs to be replaced with a weaker conjecture? And 2^(n/√log n) is still superpolynomial and grows super fast.

Anyway AI is good at theoretical computer science, and a lot of theoretical computer science papers need fixes, which may or may not be simple.

Feeling unlucky: Alphabet shares fall about 10% after third-quarter sales of $111.2 billion and search ad revenue miss forecasts, reviving AI disruption fears - Bloomberg

Google Cloud revenue doubles to $30.7 billion, but slowing Search growth sends Alphabet shares down 8% - TechCrunch

Google's Philipp Schindler outlines plan to insert sponsored offers into agentic shopping as AI Mode searches monetize at lower rates - Business Insider

Googling its own grave? TFW when your raison d’être is accessing the world’s information, but if you make it too easy to access it breaks your whole business model.