UT Austin computer scientist Scott Aaronson: Grad students are scared after OpenAI's release of 370 math proofs they don’t know how to build a career when AI can solve the problems they’d train on. ---- Full video on "MTS" YouTube channel, (link in comment)

Full video on "MTS" YouTube channel

Now they get to formulate the theorems instead of grinding out the lemmas.

i get why they're scared, the problems you used to learn on are exactly what the models do best now. but someone still has to know which questions are worth asking and check if a proof actually holds up. feels like the training changes more than the career disappears.

Evolve or sink. It's very simple. The heavily indoctrinated will perish first.

Solution: career is for suckers.

Understanding of a domain is still going to be a superpower, the open question is a call to explore the “how?”. Teach the children! Remember a full education was always Trivium and Quadrivium for a Reason. SI/AI doesn’t “do everything” it does things for YOU!

AI solving advanced math problems isn't just a flex. For grad students, it raises an uncomfortable question: if AI can handle problems that once took years to master, what skills will actually set the next generation of researchers apart?

Those 370 math proofs still need human verification. LLMs can create mass hordes of data, but the bottleneck remains human review of error-prone predictive models.

If one is afraid that AI is going to take over one’s job, one is not really a researcher. Research is playing with that which one does not understand, which at the moment means the interaction of AI and mathematics.

guess people got scared of calculators too

20 years in IT taught me the same lesson every cycle: when tools solve the training problems, the skill shifts to choosing which problems matter. Grad students won't vanish — the apprenticeship just moves up a level. #AI #math

He's scared because his only contribution to science is to scare people.

Maybe if a subject can be solved by a $20 AI subscription, there was a real issue having someone owrk on it for 4 years

if a model drafts the proof, the student still has to pick the problem. that choice feels like the actual job.

Us humans have to pass through the gauntlet of peer review panel to even have a modicum of hope to have our analyses heard by society at large. Mindless AI bots its shareholders claim it authentically resolved questions? They default as fact until proven otherwise.

At least they could start advocating for universal income.

rough time to be a grad student

370 proofs is the demo. The hard part is the 12 that are wrong and sound exactly like the right ones.

Grad students always trained on problems with answers in the back of the book. What changes is research: if a model can draft the proof, the scarce skill is picking which conjecture is worth a month and spotting the one lemma that's quietly wrong.

the final boss of grad school is an api now

Schools should actually give some of these SI proofs out as assignments towards credit, if wanted... as it will kill two birds with one stone. (prove and teach)

Grad students training on problems AI already solved is a real career-path problem.

Grad school was always partly an apprenticeship in taste. Trouble is, taste is a lot harder to put on a CV than a solved problem.

Still train on them, graduate, then agentize the method, test and validate with at least 100x productivity for the win.

@rohanpaul_ai what's the complexity ceiling right now? are there problem classes the 370 proofs can't touch, or is it more about supervision cost at scale?

Governments have long stifled exotic tech that could disrupt whole areas of the economy. Yet the new cold war in AI seems to be an exception. I wonder what's different this time?

370 proofs at once would scare anyone in academia

Perhaps he can use AI to find a nicer shirt to wear on camera.

Why don’t they use AI to solve bigger problems?