Phillip Kerger, a UC Berkeley assistant teaching professor and former NASA Quantum AI Lab researcher, joins TwoSetAI for a one-hour conversation about research, education, and human judgment.

Important caveat: The interview discusses a purported AI-produced Navier–Stokes proof. The video’s own description calls it unconfirmed and controversial; Kerger says he is not a fluid-dynamics expert and has not read the full paper. This capture does not establish that a Millennium Prize problem has been solved.

Research changed faster than his expectations

  • Kerger recalls AI being useless for his research in December, helpful by March, and solving a problem he had pursued for two years by July.
  • His field is optimization complexity: how much computation a problem requires, and what makes a solving method efficient.
  • He credits AI with much of the work on his own reported result, rather than treating it as an ordinary writing assistant.

Intelligent search is not magic

  • He interprets the claimed fluid-dynamics result as a search for a mathematical construction, reportedly using roughly 10,000 concurrent agents.
  • The agents still need enough understanding to narrow the search and evaluate candidates; more parallel workers do not remove an exponentially growing search space.
  • Finding a construction and inventing a fundamentally new style of mathematical argument are different claims.
  • A proposed proof still needs expert checking, clear attribution, and transparent reporting of AI involvement.

Why choosing still matters

  • A mathematical question can have a definite correctness criterion; a bike design has competing goals, audiences, costs, and aesthetics.
  • AI can propose 20 designs, but building 10,000 physical prototypes is not a cheap substitute for judgment.
  • Behavioral research still requires collecting real data, choosing a useful model, and deciding what to deploy.
  • Generic AI dashboards and writing repeat familiar styles unless someone supplies distinctive taste and constraints.

Assistance is not the same as impersonation

  • Kerger is comfortable delegating coding, mathematical help, and simple scheduling, but wants approval over meaningful communication.
  • His analogy is an excellent secretary plus an excellent research assistant—not a replacement personality.
  • Human performance can remain valuable even when machines outperform humans; chess spectatorship is his example.

Quantum computing needs a concrete advantage

  • Kerger’s skeptical view concerns quantum computing for mainstream AI, not the whole field of quantum research.
  • Ask for the exact problem, algorithm, dataset, hardware, and comparison against the strongest classical solver.
  • Quantum memory, state preparation, copying restrictions, and repeated readout can defeat attractive theoretical speedups.
  • He describes a network algorithm whose useful crossover required an absurdly large network with quantum hardware at every node.
  • His counterexample to industrial hype: an Airbus loading-challenge instance solved in a reported 1.1 seconds on his Microsoft Surface.
  • Quantum chemistry and materials simulation are more natural targets because the underlying systems are themselves quantum.

Make homework serve learning

  • He grades homework for honest engagement rather than correctness, reducing the incentive to outsource perfect answers.
  • In-person paper exams supply a separate check of foundational understanding; effort itself is difficult to verify.
  • He recounts another instructor’s AI-permitted exam averaging roughly 99%, while acknowledging uncertainty about its details.
  • These are teaching experiments and anecdotes, not proof that one assessment policy works everywhere.

The AI learning paradox

  • You can look up almost any individual fact, but having no knowledge leaves you unable to judge which answer or approach fits.
  • Build a critical mass of foundations in a domain you care about; expertise makes AI explanations faster to understand and easier to question.
  • Use AI for retrieval practice—answering from memory—rather than only asking it to produce finished solutions.
  • A concrete routine: give it your lecture notes and ask it to quiz you, or practice vocabulary through voice interaction.

Motivation, physical work, and new researchers

  • Teaching should connect abstract methods to real interests, such as optimization through sports analytics.
  • He sees education as developing reasoning and judgment, even if credentials and narrow technical tasks lose some value.
  • Physical engineering and in-person relationships retain constraints that cannot be scaled like purely digital searches.
  • He thinks beginners can explore serious research problems with AI, provided an expert checks any claimed result before publication.

“You have to build like a critical mass of skills and knowledge.” — Phillip Kerger, 43:35