Sign in

Shubhendu Trivedi

@shubhendu.bsky.social
1.1K followers 280 following 5.6K posts

shubhendu-trivedi.org

PostsRepliesMedia
Reposted by Shubhendu Trivedi
Sam Power @spmontecarlo.bsky.social · 8h
Some elementary (but a bit neat) stuff: github.com/sampower88/P... "Slow Motion of Gradient Flows from Bad Initialisation"
0121
Shubhendu Trivedi @shubhendu.bsky.social · 8h
You can sometimes tell by looking at the price action. Sometimes it looks so bad that it can't just be business as usual.
020
Shubhendu Trivedi @shubhendu.bsky.social · 8h
Turns out this was indeed not "just a feeling," it was an AI focused hedge fund called VARA.
100
Shubhendu Trivedi @shubhendu.bsky.social · 20h
I came across some slides by Blum from 1997 or 1995 when I was an undergrad, and they used Comic Sans (a variant I think, not the original). This also uses the same. Pretty impressive stuff.
050
Shubhendu Trivedi @shubhendu.bsky.social · 22h
Check out our paper on whether first-price auctions can efficiently allocate queries among strategic AI agents (in a decentralized serving-time market): arxiv.org/abs/2610.20587
2336
Shubhendu Trivedi @shubhendu.bsky.social · 23h
People always make up reasons like CPI, Anthropic numbers underwhelming etc., but someone big got liquidated today. Hopefully not Situational or similar again.
100
Shubhendu Trivedi @shubhendu.bsky.social · 15/10/2026
Not too many of these, though.
000
Shubhendu Trivedi @shubhendu.bsky.social · 15/10/2026
Louis Maddox you will always be famous
010
Shubhendu Trivedi @shubhendu.bsky.social · 13/10/2026
The OP thing basically always translates to: Big Professor, Genius Award, Big University, Now Accelerating at Big Lab, says no one is ready (the fact that these things have a lot of influence on people says people still care about experts in a twisted way; not just that fear sells).
021
Shubhendu Trivedi @shubhendu.bsky.social · 13/10/2026
Being grounded and forward looking is important to construct a better future. Otherwise, you are setting the stage for horrid outcomes (which will have nothing to do with the initial reason for the fear). Expansive varieties of social imagination don't come from fear or denial (often also fear).
111
Shubhendu Trivedi @shubhendu.bsky.social · 13/10/2026
But I guess a plea to be forward looking without being delusional, and not be either fearful (w/o ever saying anything useful), or be in complete denial (say via ornithomorphizing LLMs), is too much to ask for. The ship for this probably sailed when narrative capture around AI happened 10-13 y ago.
110
Shubhendu Trivedi @shubhendu.bsky.social · 12/10/2026
Perhaps it is just because I think it'll be good, but all attention should focus on human flourishing. Don't scare people by repeating slogans day and night and how have been visited by God and handed over the truth no one has seen. PS: I have a few threads here about reactions post AlexNet.
120
Shubhendu Trivedi @shubhendu.bsky.social · 12/10/2026
And to be clear, the zealotry almost always translates to a fearful register ("no one has seen the truth yet, update yourself, wake up!"), which is what I am whinging about. There also exist people who have been seeing it coming since AlexNet, and they never went there. I just don't get it.
100
Shubhendu Trivedi @shubhendu.bsky.social · 12/10/2026
If anyone is thinking: No I am not subtweeting J. Nelson, I mostly agree on the substantive point. But it reminded me of the contours of the constant "I was a skeptic k months ago, now I have seen the truth." But going from there to zealotry? The speed of this is the fastest in the TCS/COLT crowd.
100
Shubhendu Trivedi @shubhendu.bsky.social · 12/10/2026
You can actually believe that there will be large dislocations if people do not start to adapt, if policy responses are lacking. Things will indeed change very fast, but you have to say something constructive and useful, and for once be a little normal. I have started losing respect for that area.
130
Shubhendu Trivedi @shubhendu.bsky.social · 12/10/2026
I will not lie that I find the genre of TCS person "AGI confessions" very funny: I did not believe two weeks ago (was asleep for ten years, though), but now I do, you should too. This has happened again and again and again, and they all become zealots.
140
Reposted by Shubhendu Trivedi
Alondra Nelson @alondra.bsky.social · 12/10/2026
A great essay indeed. Come for the thesis, stay for the endnotes. It is in dialogue also with my IAS colleague Akshay Venkatesh's prescient 2022 essay which was a call to awareness to the mathematics community. www.math.ias.edu/~akshay/rese...
math.ias.edu
0145
Shubhendu Trivedi @shubhendu.bsky.social · 12/10/2026
Hahaha, yes, had to let it out of the system! I am very proud of my grandparents and their larger family for how they were.
100
Shubhendu Trivedi @shubhendu.bsky.social · 12/10/2026
and yet after a while you see that it is of no use -- just led to a señorito satisfecho -- just a tech bro type very pleased with himself. This is also bad (for example a bad role model for students, feeds into the "smug professor" sort of thing). Anyway, sorry I sound too crazy around here.
010
Shubhendu Trivedi @shubhendu.bsky.social · 12/10/2026
Yes, it is really very sad. Although I must say "anything else" "knowing a lot" does not really capture it. In fact one does not have to. It is more about the orientation and formation. For example, there is this person who wrote an unreadable book on topology, seems v cultivated and knowledgeable
110
Shubhendu Trivedi @shubhendu.bsky.social · 12/10/2026
*4 STOC papers, and 8 NeurIPS papers, a gnawing sense of insecurity / inadequacy, and no clue of anything else.
100
Shubhendu Trivedi @shubhendu.bsky.social · 12/10/2026
Next time I am in NYC, I know who i am meeting.
100
Shubhendu Trivedi @shubhendu.bsky.social · 12/10/2026
It wouldn't be in the situation that it finds itself in. I have a long theory about this, but it would turn into an abrasive screed (and I want to avoid that), so I will not indulge.
100
Shubhendu Trivedi @shubhendu.bsky.social · 12/10/2026
Gamburd recently appeared on twitter likely to promote his recent article. He doesn't post, only makes laconic comments, but they are so good. Such a learned man. Imagine if academia was full of such people and not hiring based on 4 STOC papers, and 8 NeurIPS papers, and no clue of anything else.
150
Shubhendu Trivedi @shubhendu.bsky.social · 12/10/2026
There is a plausible and quick Dr. Strangelove solution to this, but we don't have a S. R. Hadden type of billionaire to put in the resources to force it to happen (there could even be a windfall profit in the process).
010
Shubhendu Trivedi @shubhendu.bsky.social · 11/10/2026
*non-ML person shared (from my time consulting for pharma companies) I think it might have to do with the fact that people keep bringing up progress in medicine and compare it to mathematics re: understanding. Probably he got influenced by it.
110
Shubhendu Trivedi @shubhendu.bsky.social · 11/10/2026
Yes, it's very LW-brained. Gowers and to some extent Tao are really bad at this. But regardless, I am still surprised at the uncharitable readings (someone even shared this with on whatsapp; saw it 200 times on Twitter).
120
Shubhendu Trivedi @shubhendu.bsky.social · 11/10/2026
What I am worried about is losing an entire generation of future mathematicians because of the atmosphere that people (not just science; it's general, even humanities folk I talk to think math is solved) are creating, besides the sudden change in incentives (now it's all a competition) and craft.
150
Shubhendu Trivedi @shubhendu.bsky.social · 11/10/2026
Knightian uncertainty over entire trajectories (or universes) of mathematical (even civilizational) development.
130
Shubhendu Trivedi @shubhendu.bsky.social · 11/10/2026
moving FV path), but it can't remain detached for too long. Because after all, one way of thinking about pure math is to think of it not just as the production of mathematical knowledge, but also a civilization-scale mechanism to account for
120
Shubhendu Trivedi @shubhendu.bsky.social · 11/10/2026
But anyhow the reason I have been thinking of lately (which I need to expand carefully) is because despite possible near-time disruptions, it's like in the principle of reflexivity. The fundamental value of something might itself temporarily change (think Keynesian notions of equilibration with a
120
Shubhendu Trivedi @shubhendu.bsky.social · 11/10/2026
The long term importance of pure math isn't going anywhere, nor is it going to be done automatically in LHC-like undertakings (this has become a popular opinion on twitter, the other place where you can get groundbreaking opinions on AI). That level of investment is math would be great.
130
Shubhendu Trivedi @shubhendu.bsky.social · 11/10/2026
Tbc, I'm not worried about mathematics despite short term turbulence due to a surprisingly large % of science people thinking it is "solved." What is getting "solved" is that it can be used in regular engg applications more easily – where you don't need arcane math, and verification is trustable.
160
Shubhendu Trivedi @shubhendu.bsky.social · 11/10/2026
But he could have chosen a less confusing and ignorant-seeming analogy. Regardless, the role of understanding is very different in the two domains. In pure math it is about transformation since it has no real end goal like a drug that can be tested via a RCT. The goal there is to produce new math.
010
Shubhendu Trivedi @shubhendu.bsky.social · 11/10/2026
He is trying to draw an analogy to "prove statement X, do a breakthrough, pass all the tests." Are people F dense?
210
Shubhendu Trivedi @shubhendu.bsky.social · 11/10/2026
It looks like a weirdly misinformed analogy, but he is trying to make a different point than about pharmaceuticals. Note he is talking about a magic scenario where you prompted "pass phase three trials."
140
Shubhendu Trivedi @shubhendu.bsky.social · 11/10/2026
AGI AGI AGI AGI AGI -- that's what happens!
100
Shubhendu Trivedi @shubhendu.bsky.social · 11/10/2026
Open weight models and data centers don't necessarily pull in opposite directions. A lot of companies that now use open models use it on the cloud, they just have to. Compliance is much easier in that case. On prem is still something that people only in banking and law do (health usually on cloud).
100
Reposted by Shubhendu Trivedi
Joshua Grochow @joshuagrochow.bsky.social · 10/10/2026
…so, I conclude (this is me now, not the quote) we should just not accept these as part of the record in the first place.
5485
Reposted by Shubhendu Trivedi
Joshua Grochow @joshuagrochow.bsky.social · 10/10/2026
“We’re now expected to either (1) check all this garbage that OAI is producing or (2) concede that all of these conjectures have been proven,” one algebraic topologist wrote to me. “I believe in politics this is known as ‘flooding the zone.’ We simply don’t have the resources to correct the record.”
57628
Shubhendu Trivedi @shubhendu.bsky.social · 10/10/2026
Good Taste and Critical Thinking™
030
Shubhendu Trivedi @shubhendu.bsky.social · 09/10/2026
In the 3-4 books I have read (sadly including Eros the..), that is. It starts off so good and you are like woah, and then it just keeps repeating in a way that it gets stale. Doesn't manage to keep up the tempo. Someday this will change.
010
Shubhendu Trivedi @shubhendu.bsky.social · 09/10/2026
Eros the bittersweet is interspersed with sappho (I have a love hate relationship with Carson, she always built up a lot of hope and possibility and then always disappointed).
100
Shubhendu Trivedi @shubhendu.bsky.social · 09/10/2026
But hopefully the GPU node thing works out!
130
Shubhendu Trivedi @shubhendu.bsky.social · 09/10/2026
Note that none of this says anything about AI being a bubble. But there is a lending bubble, which is definitely fragile.
120
Shubhendu Trivedi @shubhendu.bsky.social · 09/10/2026
which can trigger a bunch of failures (including at least one of the two "Labs").
120
Shubhendu Trivedi @shubhendu.bsky.social · 09/10/2026
Not at all, if "we all believe it" the opposite will happen and "the we" will get short squeezed out of existence. One scenario that could happen is that the bond market contagion ends up driving corporate bonds over a cliff almost overnight, then there are two very weak links, which can fail
120
Shubhendu Trivedi @shubhendu.bsky.social · 09/10/2026
There is a meme about this with a bald guy -- there are dozens... dozens. I mean, I share the opinion too, fwiw.
010
Shubhendu Trivedi @shubhendu.bsky.social · 09/10/2026
I was thinking of that too!
000
Shubhendu Trivedi @shubhendu.bsky.social · 09/10/2026
The point about Hayek as a mere theoretician becomes the most clear when you talk about the conditions for success of the gold standard (another long soliloque). It's so amazing to me! Yet, Keynes the market genius is the degenerate socialist and Hayek is the inveterate capitalist.
040