ABOUT THIS EPISODE
Recorded live on June 22, 2026.
Jared Duker Lichtman, a Stanford number theorist, proved a conjecture Paul Erdős posed in 1986: among all sets of whole numbers in which no number divides another, the primes score highest on the measure Erdős invented for such sets. The radio hour stays with the craft. Jared explains primitive sets with quarterback ratings and jersey numbers, describes the year he spent alone in his room during lockdown until the key definition arrived, and tells what happened when a 23-year-old amateur working with ChatGPT opened a 60-year-old Erdős problem Jared had chased for seven years. His reaction was joy, and he brought the amateur, Terence Tao and others into a joint paper. The podcast hour moves from Aristotle and Leibniz to what Jared calls verified superintelligence: machine output checked against a specification, the way a program either compiles or does not. Erik pushes hard on whether any specification can bind a mind smarter than the person who wrote it, and the two work it through to a shared answer about locality: a guarantee holds only as far as you can seal the problem off from the rest of the world. Jared closes on the idea that "genius" is a pointer society attaches to work after the fact, and that most of the work never gets one.
Key Topics
- [00:00 - 02:29] Cold open: Erik and Guy introduce their first mathematician: an Oxford PhD under Fields medalist James Maynard who proved a conjecture Erdős left open for decades
- [02:29 - 12:17] Erdős and the primitive set conjecture: Erdős as the itinerant mathematician who paid for his lodging in co-written papers, with about 1,500 papers and more than a thousand open problems; primes as atomic building blocks; the twin prime conjecture as a simple unsolved question; what a primitive set is (no member divides another); Erdős's 1935 bound and his conjecture that the primes are optimal, "too beautiful not to be true"
- [12:17 - 23:43] Four years on one problem: first hearing it as a Dartmouth senior; an undergraduate paper with partial progress; COVID sends him home from Oxford for two years; going back to Erdős's original 1935 paper; a new definition after a year in lockdown; the mountain-top view; Guy on watching the problem seize him ("it sounds like love"); checking the proof with his undergraduate mentor, then with Maynard; a 22-page paper with a proof of about eight pages
- [23:43 - 37:31] Maynard's record, the amateur, and the craft: last digits of primes and arithmetic progressions; improving on Maynard's methods; the amateur whose ChatGPT session opened a 60-year-old problem Jared had worked on for seven years; the joint paper with Terence Tao; the father-at-the-birth analogy; craft must adapt rather than disappear; taste and goal choice as the human role; number theory's accidental gift to cryptography; math as a backstop against hallucination
- [37:31 - 39:25] Radio close: where to find Jared, the Stanford Future of Mathematics Symposium, and the "Erdős #1196 and beyond" preprint
- --- [Act 2: Extended Conversation] ---
- [39:25 - 44:03] From Aristotle to Leibniz to legal contracts: logic distilled from ordinary argument, echoed in chain-of-thought training; Leibniz's attempt to encode language in mathematics; why contracts define their terms and courts exist for the rest; Erik on being a lawyer who looked at smart contracts early
- [44:03 - 52:37] Sending ships into the ocean of math: Newton's pebbles on the shore; AI agents exploring the "dark matter of mathematics" and reporting back; "post scarcity of mathematics"; interpretability in AI and in neuroscience; taste as an interpretive framework, like building telescopes; Erik on letting his own coding agents make conjectures from a roadmap; proofs as bridges between islands; tokens, efficiency and compression
- [52:37 - 1:04:57] Verified superintelligence: a program either compiles or it does not; specifications as the box an AI must stay within; humans at the center on goal specification; laws as specs for a society of strangers; Erik's argument that we cannot chain a future god to our values; how well Claude understands human values today
- [1:04:57 - 1:14:19] The thermostat argument: Erik's robot arm and 80-degree thermostat against Jared's distinction between checking what happened and proving what will always happen; math versus physics; locality and "vectors of infiltration"; the shared landing on expanding the ground you can really hold as truth, in a lineage from Aristotle to Leibniz to today
- [1:14:19 - 1:27:11] Datasets, relevance and the question of waking up: how one website of all the Erdős problems steered the AI labs toward them; why the conjecture called to him; invariants and latent spaces; Guy on relevance realization and whether AI could come to care; "stochastic parrots" giving way to new solutions; truth as an attractor of intelligence; goalposts of inevitability and of hope; weak coupling between physics, chemistry and biology
- [1:27:11 - 1:34:03] Education as hope, and genius as a pointer: the more people understand the ingredients (data centers, energy), the more voice they have; why AI was missing from the last presidential debates; Van Gogh and Nietzsche recognized late; prizes and gold medals as pointers to the work that earned them; the discoveries history never recorded, including by women and others shut out of the institutions
- [1:34:03 - 1:38:39] Intelligence, power, and the amateur's reaction: Guy on why intelligence and power are not the same thing; a cure for cancer falling in the forest; how the 23-year-old amateur took it; a breakthrough "alters your reward pathways"; close
Guest Bio
Jared Duker Lichtman is a Szegő Assistant Professor in the Department of Mathematics at Stanford University, working in number theory, especially the prime numbers. He earned his AB and AM in mathematics at Dartmouth College in 2018, a MASt in pure mathematics at Cambridge as a Churchill Scholar in 2019, and his DPhil at Oxford in 2023 as a Clarendon Scholar, supervised by James Maynard, who received the Fields Medal in 2022. In 2022 he proved the Erdős primitive set conjecture: call a set of whole numbers greater than 1 primitive if no member divides another; Erdős proved in 1935 that the sum of 1/(a log a) over any primitive set stays below a fixed bound, and asked in 1986 whether the primes attain the largest such sum. Lichtman showed that they do (Forum of Mathematics, Pi, 2023). He also set a new record for the level of distribution of primes in arithmetic progressions (66/107, beyond Maynard's earlier 3/5). He joined Stanford in 2023 as an NSF Postdoctoral Fellow, co-organized the Stanford Future of Mathematics Symposium in May 2026, and co-authored, with Terence Tao, Liam Price and others, the May 2026 preprint "Primitive sets and von Mangoldt chains: Erdős Problem #1196 and beyond." On X he is @jdlichtman.
Notable Moments
- [~11:25] Why the conjecture pulled people in. "...it seems almost like too beautiful not to be true." Jared on Erdős's guess that the primes are the best possible primitive set.
- [~17:57] The breakthrough, after a year in lockdown. "Once you get to the top of the mountain, you kind of see the whole vista all at once."
- [~30:13] On being scooped by an amateur with ChatGPT. "If a solution comes by any means, and if it's a problem you care about, then it's just like a moment of joy." He compares himself to the father who arrives for the birth after the labor.
- [~46:09] Scouts. "And we can just, you know...