Rohit Agarwal

Founder @usetokenless (YC S26)
Automating model selection
CS PhD Cand @PrincetonPLI | Prev. UC Berkeley EECS

[Rohit Agarwal @Rohit_Writes]

taxes are a solved problem that humans know how to do, ai is decent just not all people want to trust it with a task that’s much higher stakes for them

No one knows how to do RH, but if RH gets solved for most laypeople it’s not that high stakes

[Rohit Agarwal @Rohit_Writes]

the actual reason one may think this is because a non mathematician has a lower TRUST barrier to accepting a proof of RH than a mathematician, as it means way less to them

[Rick @rickasaurus]

AI is going to solve the Riemann Hypothesis. before it can reliably do my taxes isn’t it

[Rohit Agarwal @Rohit_Writes]

@willccbb @tenobrus not exactly algorithms but construction of better codes in coding theory have actual applications to communication

for example Arikan polar codes from 2016 are now used in 5G

[Rohit Agarwal @Rohit_Writes]

@mikeandallie cohn-elkies was shown to be better than KL-based bounds, but this is the first time we have a new “explicit best number” for asymptotic sphere packing

i agree i did dramatize a bit

[Rohit Agarwal @Rohit_Writes]

Yeah I think it’s way smarter than a grad student. For context

1. last numerical progress 1978
2. last progress 1979
3. posed 1999
4. posed 1982
5. defined 1979, first improvement since trivial bound
6. last progress 2016
7. last progress 2003, first polynomial bound
8. posed 1964
9. posed 1961, $250 prize by Erdos
10. posed 1967

[Noam Brown @polynoamial]

An internal version of Astra, @OpenAI’s next major model family, solved 10 major open problems in mathematics, quantum complexity, and theoretical computer science.

We believe it will be a major step for scientific reasoning. https://t.co/iP6cyheZ7i

[Rohit Agarwal @Rohit_Writes]

Cost per task is a good metric, but not perfect if the tasks don’t cover your space

Would highly recommend trying a mix of deepseek and other models instead eg @usetokenless

[Jen Zhu @jenzhuscott]

DeepSeek Kill Zone.

Models that are inferior and significantly more expensive should cut spending and be in survival mode ASAP.

Models that are superior but also more expensive or slightly inferior but with comparable pricing can survive a bit longer.

[Henry Yuen @henryquantum]

Some initial thoughts, and a complicated mix of feelings.

1. Wow. I mean, Erdos problems are cool (I genuinely mean that), I didn't know about the Jacobian conjecture before it got disproved. But this newest batch from OpenAI hits home in a way the previous announcements did not.

New circuit lower bounds? A simple, easy-to-describe non-sofic group? Hardness of approximation for CVP without needing a unique games-like conjecture? I didn't just hear about these problems from my friends or from seminars. I feel their importance in my bones; I deeply care about the answers to these questions.

[Rohit Agarwal @Rohit_Writes]

@adntaha_ai well it sounds like we need to make lean or whatever else a trusted kernel through rigorous fixes

if this is the case, then what one needs to do is just verify the statement of the theorem is correct, which is a whole lot easier!

[Rohit Agarwal @Rohit_Writes]

@0xbadface1 Yeah most likely

[Rohit Agarwal @Rohit_Writes]

@terry_033 actually i made a small typo it was $100 for the specific result achieved

but it was $100 in the 1960s, which was a lot!

[Rohit Agarwal @Rohit_Writes]

@ColinZwanziger sphere packing was resolved in dimensions 8 and 24. this version is asking about the asymptotics in large dimension, ie as dimension increases what does the optimal sphere packing fraction tend to?

[Rohit Agarwal @Rohit_Writes]

@RomiLifshitz hmm was the previous factor not good enough? would love an in-depth explanation

[Rohit Agarwal @Rohit_Writes]

@Ravenoil i agree these are a bit silly! but these are really important interesting conjectures that are breakthroughs in their subfields

[Rohit Agarwal @Rohit_Writes]

@Ravenoil i think which result seems most important depends on taste. i did a bunch of tcs stuff so sphere-packing and coding theory stuff feels more impactful to me

the results presented here are super super major conjectures i had heard all of

also 27 years is a long time!

[Rohit Agarwal @Rohit_Writes]

I want to emphasize these are NOT obscure problems, but deep and foundational problems that many people were either stuck or people were making small incremental progress on.

[Rohit Agarwal @Rohit_Writes]

@Do_owne how long does it take for this to run?

[Rohit Agarwal @Rohit_Writes]

@SonglinYang4 @simonguozirui take notes

[Rohit Agarwal @Rohit_Writes]

@gro_tsen Using < for subset,

NP^G < NP, P^G < P, and P <= NP, clearly you're putting two lower-bounds on two quantities and saying the lower bounds are unequal... doesnt say anything about the original quantities

[Rohit Agarwal @Rohit_Writes]

while you're here, would love to tell you guys about model routing with Tokenless https://t.co/lqBIBC625A

[Rohit Agarwal @Rohit_Writes]

This is insane. This is a VERY central problem in additive combinatorics and it’s been resolved via AI… wow

[Tencent Hy @TencentHunyuan]

For a finite set of integers (A), how much faster can (|A+A|) grow than (|A-A|)?

A 1969 theorem gave an upper bound of 2 for the exponent. For more than 50 years, the best constructions barely exceeded 1.1.

With help from our research agent Hyra and the Hy3 model, we found an explicit construction showing that the optimal exponent is exactly 2.

A 50-year-old problem, solved.

Paper: https://t.co/OU7lfH0hXR
Hyra blog: https://t.co/NeQEekUSZ3
Formal proof: https://t.co/LMbGdKqo2D

[Rohit Agarwal @Rohit_Writes]

@thomasfbloom Perhaps the proof is not very surprising--

But this was a problem I thought about when I did TCS by hand in the stone age. My puny TCS brain thought it was pretty cool

(obviously AP-free sets, sum-product and PFR are cooler)