Natesh Pillai @Bayesprof

AI + Math. Professor of Statistics @Harvard, Distinguished Engineer at LinkedIn. Scientist, Entrepreneur. Chess & Desserts fanatic. Grew up in Kerala. Joined July 2022

Tweets

812
Followers

758
Following

568
Likes

4K

Natesh Pillai @Bayesprof

4 hours ago

@AlexKontorovich touche

0 0 0 14 0

View Details

@AlexKontorovich Do you think we can formalize the current ABC conjectures proposed "proof" and put to bed once and for all, the nonsense around it? I'd love to know how much effort it will take to formalize Scholze's objections.

Alex Kontorovich @AlexKontorovich

9 hours ago

Some more details on Alexeev's approach. Here is his zulip thread: leanprover.zulipchat.com/#narrow/channe…

1 9 50 6K 24

2 0 7 486 2

View Details

Natesh Pillai @Bayesprof

4 hours ago

@AlexKontorovich This might be the first example where formalization will be used to say I told you so instead of the other way 🤣🤣🤣

1 0 1 49 0

View Details

Natesh Pillai @Bayesprof

7 hours ago

@dzwick_weber @AlexKontorovich Oh! Good to know.

0 0 0 19 0

View Details

Natesh Pillai @Bayesprof

2 days ago

@torgeirlysen I kind of agree with you, yes. But things have changed quite a lot compared to how they’ve been even just a couple of years ago.

1 0 0 165 0

View Details

Natesh Pillai @Bayesprof

2 days ago

I don't know where PhD programs in statistics are going, but in the meantime, we can at least make the grad curriculum more useful: 0. Make gptpro/claude equivalent free for grad students; mandatory training and use of agentic workflow. Most universities still give only the 20$ version, not the most powerful one. 1. Revamp the grad programs and focus them toward "taste" and constructive criticism. I'd be ok with not having any written exams and instead having oral/take-home where students must use AI to replicate and then critique a published work to the committee's satisfaction. But here the standards have to be really higher than before for a student to "pass". This is the only thing we can still "teach". 2. Focus the grad program toward "building" instead of writing papers. A "thesis" can constitute constructing original data pipelines, assembling disparate data sources, open source of implementation of algorithms, etc. CS programs have been doing this for a while; stats has to catch up. 3. Formalize routine minimax arguments/convergence proofs via Lean, and focus on the key parts of the technical argument. In mathematics, up until recently, most graduate programs had a foreign language requirement. I wouldn't be surprised if they make Lean a language requirement in the near future.

7 12 98 15K 66

View Details

Natesh Pillai @Bayesprof

2 days ago

@VV2897 Don’t give up. The field needs you! Keep at it. Build your own and show it to depts.

0 0 2 195 0

View Details

Natesh Pillai @Bayesprof

2 days ago

@SokobanHero Yea. I’d be curious what this “minimal expertise” in each field will be. But if the oral exam is done well, then they shouldn’t be able to fake their expertise. Designing that exam is of course is key.

0 0 2 446 1

View Details

Natesh Pillai @Bayesprof

2 days ago

@invictusqed no, not giving up on theory at all. There are various suitable analogues of "building" in theory as well, including playing around with algorithms numerically to design conjectures etc. But people are already doing some of this.

1 0 1 350 0

View Details

Natesh Pillai @Bayesprof

3 days ago

@mathladyhazel That’s so cute!

0 0 0 133 0

View Details

Natesh Pillai @Bayesprof

4 days ago

@GuanyangW @tak3sh8 Very nice, congrats!

1 0 0 42 0

View Details

Natesh Pillai @Bayesprof

6 days ago

@murukku_r @TheRocketMediaX 😮

0 0 0 26 0

View Details

Natesh Pillai @Bayesprof

6 days ago

@TheRocketMediaX Not sure if he was into chess, but in one of his public lectures he says the famous chess player quote: “the rest is a matter of technique”

1 0 2 176 0

View Details

Natesh Pillai @Bayesprof

6 days ago

@TheRocketMediaX From my hometown!

1 0 2 302 0

View Details

Natesh Pillai @Bayesprof

6 days ago

@sp_monte_carlo Yes worth writing this up.

0 0 2 27 0

View Details

Natesh Pillai @Bayesprof

7 days ago

Speculative decoding was/is called delayed acceptance in the MCMC community. I am not sure where it was invented first though. And yes, it’s a super important idea.

François Fleuret @francoisfleuret

7 days ago

Speculative decoding is the closest you can get to a free lunch method. It is beautiful and astounding. I am surprised that it does not play a greater role in "AI".

23 15 440 49K 208

1 0 11 4K 9

View Details

Natesh Pillai @Bayesprof

7 days ago

@tak3sh8 Yea. There are a few like these I have come across over the years that are surprising and beautiful.

0 0 3 217 0

View Details

Natesh Pillai @Bayesprof

a week ago

@vamsiprithamp @scottnarmstrong @littmath @SidGadgil What’s the idea and why do you think it’s original?

2 0 4 2K 0

View Details

Natesh Pillai @Bayesprof

a week ago

It warms my heart to see such fundamental research coming from IITs and ISI. Congratulations to the authors! Amazing!

Lance Fortnow @fortnow

2 weeks ago

The Bipartite Matching in NC paper is out. Previously we needed randomness for an isolation lemma. eccc.weizmann.ac.il/report/2026/10… The authors Abhranil Chatterjee, Sumanta Ghosh, Rohit Gurjar, Roshan Raj and Thomas Thierauf get around isolation completely using a polynomial method.