Pratyusha Duvvuri

Stealth repeat founder. Exploring what’s next in AI and what’s worth building with it.
SPC, Galileo (acq. Cisco), Meta, Apple, Two Sigma. CMU.

[Pratyusha Duvvuri @pratduvvuri]

@thsottiaux

[Pratyusha Duvvuri @pratduvvuri]

@Yuchenj_UW @Yuchenj_UW if they can actually resell that capacity at 2–3x, each training run has to justify using GPU hours they could otherwise sell. that changes how you choose experiments while you're still training (which defeats the spirit of getting GPU capacity early).

[Pratyusha Duvvuri @pratduvvuri]

@bookwormengr that extrapolation still depends on what customers will pay when all that capacity arrives.

even a busy cluster can earn less per GPU if token prices fall faster than throughput improves. cheaper inference can be great for the apps and still disappoint the people financing the GPUs.

[Pratyusha Duvvuri @pratduvvuri]

@theo

[Pratyusha Duvvuri @pratduvvuri]

@rauchg @rauchg give candidates a task with an agent and watch what makes them intervene.

when it takes an unfamiliar approach, do they test it or immediately rewrite the instructions? when it misses a requirement, do they catch it?

[Pratyusha Duvvuri @pratduvvuri]

@IntCyberDigest

[Pratyusha Duvvuri @pratduvvuri]

@adonis_singh we’ve achieved artificial adolescence

[Pratyusha Duvvuri @pratduvvuri]

@adamshuaib I've worked in quant, big tech and startups. switching between them forced me to notice when my old instincts become expensive.

in quant, moving too fast can lose money. in a startup, waiting for permission kills opportunity.

a nonlinear CV matters when you can change how you make decisions, not just where you work.

[Pratyusha Duvvuri @pratduvvuri]

@Jhaddix @Jhaddix expertise matters, but where it lives matters too. if an expert can encode a repeatable decision in the harness, the next user doesn't need to rediscover it.

the hard part is knowing which decisions transfer and which still need an expert on every run.

[Pratyusha Duvvuri @pratduvvuri]

@IntCyberDigest

[Pratyusha Duvvuri @pratduvvuri]

@adxtyahq love the dojo man, nothing like learning through doing

Here's a challenge idea. two RAG systems giving the same wrong answer. one never retrieves the evidence, the other retrieves it and ignores it. let people inspect the traces and choose the fix.

for KV cache - have people predict what happens to cache memory when they double the context length, then run it and explain any mismatch.

both make people show their reasoning before seeing the answer.

[Pratyusha Duvvuri @pratduvvuri]

@alexmashrabov @JensenHuang @alexmashrabov owning demand when the same workload can move between models - is a great place to be :))

the labs then compete for the workload you already have, rather than each winning the customer from scratch. potential bargaining power as well as distribution.

[Pratyusha Duvvuri @pratduvvuri]

@higgsfield

[Pratyusha Duvvuri @pratduvvuri]

@Polymarket

[Pratyusha Duvvuri @pratduvvuri]

@Polymarket “pay this supplier $500” shouldn't give an agent permission to pay a different supplier, change the amount or send it twice (speaking from personal experience)

the payment system needs to enforce what I approved, even when the agent decides something else would be helpful.

[Pratyusha Duvvuri @pratduvvuri]

@jarredsumner

[Pratyusha Duvvuri @pratduvvuri]

@RoundtableSpace

[Pratyusha Duvvuri @pratduvvuri]

@ArtificialAnlys @Kimi_Moonshot

[Pratyusha Duvvuri @pratduvvuri]

@Kimi_Moonshot

[Pratyusha Duvvuri @pratduvvuri]

@AndrewCurran_