Can machines help us find what nobody thought to look for?
Machine intuition, hypothesis generation, experiment, and the stubborn need to test what sounds clever.
Co-Scientist uses specialized agents to generate, debate, rank, and refine research hypotheses.
AI is moving upstream from summarizing science toward proposing what scientists should test. That can widen the search for ideas, but validation remains human work—and may become the new bottleneck.
LLM search can enlarge the space of questions while shrinking the visible space of answers. What happens when discovery itself is delegated?
Will AI agents help us do more science or better science?
Dwarkesh Patel with Ryan Greenblatt
Greenblatt and Patel debate whether automated AI research could compress years of progress into one—and what happens if the researchers are not reliably aligned.
Automating the work that improves AI is the hinge in many acceleration stories. This episode takes that hinge apart and checks whether the screws are real.
Open the note →Demis Hassabis
Demis Hassabis discusses what today's models still lack: robust reasoning, planning, world models, and scientific discovery.
A frontier claim becomes more useful when its missing capabilities are named: reasoning, planning, world models, and discovery can then be examined rather than merely awaited.
Open the note →Dwarkesh Patel with Dylan Patel
Patel follows AI down into logic, memory, power, fabrication, capital, and geopolitics—the physical systems hidden beneath the cloud.
AI is not disembodied intelligence. It is also fabs, memory bandwidth, electricity, financing, and several countries watching one another very carefully.
Open the note →Dwarkesh Patel with Ilya Sutskever
Sutskever argues that current models generalize far worse than people and that the next gains may require ideas beyond brute-force pretraining.
It puts the central technical dispute plainly: are today's methods an unfinished recipe, or are they missing an ingredient?
Open the note →David Runciman with David Spiegelhalter
David Spiegelhalter explains how a once-suspect theory of probability became a practical way to revise belief.
AI systems turn uncertainty into outputs. Bayesian thinking offers a longer history of how evidence should change confidence—and why confidence is not the same thing as truth.
Open the note →Ed Zitron
DeepSeek becomes a challenge to assumptions about capital intensity, innovation, and the competitive moat of frontier AI laboratories.
A more efficient model tests whether giant capital commitments are an enduring moat or an expensive habit with unusually good public relations.
Open the note →Dwarkesh Patel with Demis Hassabis
Hassabis describes adding search and planning to language models, connecting the AlphaGo lineage to scientific discovery and more capable agents.
It connects the AlphaGo-to-LLM-to-agent story: prediction gains plans, plans gain tools, and a clever system begins to look suspiciously like an institution.
Open the note →No Such Thing As A Fish with Olga Koch
Joni Mitchell, competitive eating, the Dawn supercomputer, and conjugation meet without adult supervision.
Our surrealism generator can combine nouns. It still cannot reproduce the confidence with which this show moves from Joni Mitchell to supercomputing.
Open the note →No Such Thing As A Fish
An absurdly capable jumping robot arrives alongside clown heads, endangered languages, and the rock-and-roll side of pension planning.
A responsible robotics discussion eventually reaches clown heads and pension planning. Other podcasts merely stop recording too early.
Open the note →Google DeepMind
A machine plays a move no human expected. The board remains still; our idea of intelligence does not.
AlphaGo made machine discovery legible: a system trained on human play found a move that expert intuition had not prepared to see.
Open the note →