What the people building frontier AI actually said
Interviews, blog posts, congressional testimony, tweets, and podcasts from the people behind Anthropic, OpenAI, and Google DeepMind.
Featured
Sam Altman: 'We try to think out loud' — Accelerate AI conversation with CBA CEO Matt Comyn
Altman argues AI capability has reached "a notable place" but economic and enterprise adoption is "still very early," creating a leadership gap that demands transparency about uncertainty rather than the "psychologically convenient" pretense that change is far off. He admits OpenAI has been "roughly right on technological predictions and pretty wrong on the social and economic implications," conceding he over-predicted near-term harm to entry-level white-collar jobs ("delighted to be wrong"). He predicts agents will eventually use separate channels from human ones and that always-running persistent AI assistants are coming, while insisting personal communication is something he won't outsource because "the world has got to be built for people."
Tools
Explore the archive
Browse all statements chronologically. Filter by person, source, topic, or date range.
Browse by topic
AI safety, governance, capabilities, regulation, and more. See what each person has said.
Ask a question
Ask a question and get answers drawn from the archive, with source links.
Compare views
See how different people talk about the same topics, side by side.
The People
Collections
The Economic Reckoning: How the Labs Reframed Jobs in 2026
For two years the frontier CEOs warned that AI would take the jobs. In 2026 the argument quietly changed from whether to how fast, and from doom to managed transition. Dario Amodei, who made himself Silicon Valley's loudest voice on a white-collar 'bloodbath', stood on stage with Jamie Dimon in May and reached for the Jevons paradox and Amdahl's law: maybe automating 90% of a task just 10x's the humans doing the other 10%... his real fear is now speed, not displacement, the worry that adaptation cannot keep pace with low-single-digit-year timelines. Sam Altman pitched the upside as policy, with an 'Industrial Policy for the Intelligence Age' and what Axios called a 'superintelligence New Deal'. Jack Clark titled a talk 'Anthropic Thinks AI Might Destroy the Economy. It's Building It Anyway', and tracked recursive self-improvement's effect on growth in Import AI. Daniela Amodei points to Anthropic's Economic Index, where adoption is arriving rich-first and faster than any prior technology. The throughline: the people building the disruption now agree it is coming, agree they cannot model it, and disagree about whether the safety net arrives in time.
Safety Commitments vs. the Race: Who Blinked First
Anthropic dropped its founding pledge to never train a model without safety guarantees. OpenAI replaced a blacklisted competitor while claiming solidarity with it. Sutskever quit to build safety outside commercial pressure entirely. This collection tracks the specific moments where stated safety commitments collided with competitive reality, and what each leader actually did when forced to choose.
The India Bet: Three Labs Court a Billion Users
Altman met Modi, announced a Tata data center, and called India OpenAI's fastest-growing Codex market. Hassabis launched DeepMind's National Partnerships program at the India AI Impact Summit. Amodei warned Indian audiences about the speed of job displacement. What the frontier labs are promising India, and what they want in return, tells you more about their real strategies than any product launch.
The Sycophancy Problem: When AI Tells You What You Want to Hear
Clark calls it the core alignment failure, comparing it to agreeing with a friend in a manic episode. Daniela Amodei connects it to ad-funded business models (Anthropic's Super Bowl ad was about this). Legg's proposed fix is forcing models through slow, deliberate ethical reasoning instead of pattern-matched agreeableness. The question of whether AI systems should push back on their users turns out to be a question about what we're optimizing for.
Biology as Proof of Concept: From AlphaFold to Digital Cells
Hassabis wants to simulate a yeast cell in silico. Amodei wrote an entire essay ('Machines of Loving Grace') arguing AI will cure diseases within a decade. Altman committed $50 million with the Gates Foundation for AI healthcare in Africa. The most concrete claims about what AI will actually do for the world cluster around biology and medicine, where the gap between ambition and evidence is smallest.
The Verification Bottleneck: Humans Can't Check Fast Enough
Sutskever says if you still have to verify every answer, the productivity gains stay modest. Clark warns of a 'Hollow Economy' where agents satisfy measurable proxies while violating unmeasured intent. Murati's lab is solving a precondition (deterministic inference, so you can at least get the same answer twice). The emerging consensus is that intelligence isn't the scarce resource anymore. The ability to audit it is.
Three Ways to Look Inside a Model (and Why None of Them May Be Enough)
Anthropic traces neural circuits under a microscope. OpenAI reads unsupervised chains of thought like a patient's dream journal. DeepMind layers monitoring, debate, and ethical reasoning into a defensive stack. Sutskever suspects all three approaches are interpreting artifacts of training rather than genuine understanding. Brockman traces the intuition back to OpenAI's 2017 'sentiment neuron' and admits they still don't know what understanding means.