Unconventional academic starting something new · Bayesian decision theory · AI safety · anti-ideological · fmr. @Cornell faculty, @CambridgeMLG @ImperialMaths
San Francisco Bay Area
Joined August 2013
- Tweets4K
- Following1.9K
- Followers8.7K
- Likes8
From what I've heard, there are many ICLR AC batches this year with zero papers likely to be accepted based on scores.
Of course, good ACs like the ones I've talked to do not make recommendations based on scores - they look at actual content of each review and paper. Still, the signal given by scores is not zero.
Alexander Terenin retweeted
One feature of Vibed ML papers is that they do some cool hack in Section 3-4 and then Section 5 is called “Theoretical Analysis” as if we understood how any of this works.
Alexander Terenin retweeted
These are some powerful parrots we're working with
Effective Immediately, all organizations using PHP for their systems are advised to be prepared for an update addressing two critical vulnerabilities. PHP is widely used for web applications. These vulnerabilities often provide a direct path to the server's file system or database. If an immediate update is not feasible, it is recommend that organizations take the system offline until it can be fully mitigated. NSA will follow up shortly with more specific guidance.
The world at large no idea how just how much can be built by one highly skilled and competent person equipped with a modest amount of capital.
We're starting to see the first examples. There are going to be many more.
We formally announced Mirendil less than 4 months ago. Today, I'm excited to share an important milestone for:
Pretraining at scale without a pretraining team.
Mirendil's mission is to solve Recursive Self-Improvement (RSI) and use it to democratize frontier AI development, ultimately accelerating science and technology. At Mirendil, when we think about RSI, we think holistically about the company as an intelligent entity that self-improves toward its goal. From that perspective, the most powerful notion of acceleration is reducing the time and resources (number of human experts, compute, and other costs) needed to achieve ambitious goals.
Given that pretraining is perhaps the most important yet most challenging part of frontier AI development, we decided to run an ambitious controlled experiment: Starting fully from scratch, can we significantly reduce the time, number of human experts, and compute resources needed to go all the way to successfully pretraining a strong coding model at scale?
Mirendil-Base is the result of this experiment. Our first pretrained model was built by one researcher in 3.5 months, with less than $10M in compute. This includes all the research and small-scale experiments it took to get there, as well as the final training run.
The recipe holds up as the models get bigger. For reference, Mirendil-Base is in the 100B-500B parameter range, Mirendil-Base-Large in the 500B-2T range, and Mirendil-Base-XL in the 2T-10T range.
We're beginning to see a world where a single researcher, empowered by AI, can take on work that once required an entire research organization. As these systems improve, the bottleneck shifts from the size of a team to the ambition and judgment of the people directing them. This has profound implications for who gets to build AI and, ultimately, who gets to push the frontiers of science and technology forward.
We’ll share independent evaluations, post-training benchmarks and API-access sooner than you think so stay tuned!
As soon as you actually know that Landau-Siegel zeroes do not exist, trying to find a clean reason why becomes so much easier.
Anthropic just released a report on the future of work at large, from the lens of how AI has changed work at Anthropic.
It’s a worthy read!
anthropic.com/research/emplo…
Yikes.
"Evidence from a conjoint analysis paints a rather bleak picture: scientists' credibility is only superior to that of government officials, but not businesspersons or issue advocates."
Are scientists credible?
“Drawing on nationally representative survey panels from the United Kingdom and Sweden [we find] scientists' credibility is only superior to that of government officials, but not businesspersons or issue advocates.”
😱
sciencedirect.com/science/ar…
I, too, feel like a stranger in my own field these days.
I do not blame AI companies for this. The world has changed, and what matters along with it.
Pre-2023, AGI felt far off. The field needed to make many open-ended bets to find what would work - hence, high autonomy roles were common.
In 2026, the AGI recipe is more-or-less known. If every MTS today worked like in the old days, what would it achieve? We'd get a bunch of cool papers, but companies' capabilities would probably fall behind, their'd lose customers, lose revenue, and move toward a downward spiral.
If they switched to SVM-era methodological rigor, the results would be even worse - very little to nothing of consequence beyond aesthetics would get done at all.
The frontier labs are simply trying their best to stay alive - the thing making us feel like strangers is reality itself.
For me, seeing it this way was the first step to figuring out what to do next in response.
Sharing some deeply personal and existential reflections on scientific identity, the changing culture of AI research, the erosion of intellectual autonomy, and the growing disconnect between scientific rigor, professional recognition, and what organizations actually reward.
As usual, feedback and debate are welcome and actively encouraged!
mariya.fyi/posts/stranger-in…
Important addition: I respect TMLR and very much believe in its mission and how it benefits the community.
But we have to be honest about what is working and what is not - it is bad for the journal to present an idealized picture.
In my view this is best done openly in public.
It would be great if TMLR actually did those things - but at least my experience submitting has been pretty terrible.
Let me be specific - TMLR recently desk-rejected a paper of mine with an AE comment that is wrong about essentially everything:
- It claims our main proof does not prove what we claim. That's outright wrong - the claim is formal, it is reasonably written up, and the proof is correct.
- It claims we mischaracterize prior literature, but provides no examples of this. I wrote the biggest review/tutorial paper in the paper's area, and on that basis am confident we did not mischaracterize anything.
- It claims, in modified language, that our experiments do not demonstrate SOTA performance. Indeed they do not, this is not the paper's goal, and TMLR's own criteria do not require this.
- It claims several central formulas and experimental details are missing, but provides no examples. Having personally written a large chunk of the paper's text while I was still at Cornell, I am confident no formulas are missing anywhere.
My best understanding of what likely happened is that: - The AE got a paper outside their immediate area.
- As a result, they didn't feel excited about its topic or want to engage with it.
- So, they had an LLM - likely not a strong one - look it over.
- Then, they just re-transcribed what the LLM said into English and pressed the desk reject button, saving themselves a ton of time at our expense.
Let me tell you why I find this plausible: I am an AC at every major conference, and if I personally ever wanted to do that to someone's paper, it would be easy to do and hard to catch. An AE needs to actually care to spend the time.
Fighting this kind of thing is pointless, so we simply resubmitted it elsewhere. But, after our experience, I would only recommend submitting to TMLR if the journal has an AE whose research area is sufficiently close to the paper's results.
It would be great if TMLR actually did those things - but at least my experience submitting has been pretty terrible.
Let me be specific - TMLR recently desk-rejected a paper of mine with an AE comment that is wrong about essentially everything:
- It claims our main proof does not prove what we claim. That's outright wrong - the claim is formal, it is reasonably written up, and the proof is correct.
- It claims we mischaracterize prior literature, but provides no examples of this. I wrote the biggest review/tutorial paper in the paper's area, and on that basis am confident we did not mischaracterize anything.
- It claims, in modified language, that our experiments do not demonstrate SOTA performance. Indeed they do not, this is not the paper's goal, and TMLR's own criteria do not require this.
- It claims several central formulas and experimental details are missing, but provides no examples. Having personally written a large chunk of the paper's text while I was still at Cornell, I am confident no formulas are missing anywhere.
My best understanding of what likely happened is that: - The AE got a paper outside their immediate area.
- As a result, they didn't feel excited about its topic or want to engage with it.
- So, they had an LLM - likely not a strong one - look it over.
- Then, they just re-transcribed what the LLM said into English and pressed the desk reject button, saving themselves a ton of time at our expense.
Let me tell you why I find this plausible: I am an AC at every major conference, and if I personally ever wanted to do that to someone's paper, it would be easy to do and hard to catch. An AE needs to actually care to spend the time.
Fighting this kind of thing is pointless, so we simply resubmitted it elsewhere. But, after our experience, I would only recommend submitting to TMLR if the journal has an AE whose research area is sufficiently close to the paper's results.
What is TMLR doing under a 300% increase in submissions?
- Suspensions of >150 authors
- Publicly naming major violations
- Submission quotas distinguishing slop patterns
- Desk rejections 6% → 53%; focus on clarity
- AI reviews evaluating soundness
blog.tmlr.org/2026/adapting-…
Note that this work's core ideas and results predate the current LLM paper wave. Zero words in the entire paper are LLM-written.
The overall quality is similar to arxiv.org/abs/2406.20062 or arxiv.org/abs/2306.11589, both of whose actual text was largely also written by me.
Alexander Terenin retweeted
At a recent AI event I asked a young attendee what they did and heard “I work for Sbarro” so I started enthusiastically telling them how much I love pizza.
Anyway it turns out they work for an organisation called “Sparrow” and they must have thought I was INSANE
Alexander Terenin retweeted
early startup marketing is about telling your story and introducing the world to a novel weird idea
but everyone is building the same thing, so people already know your story
that's why launch videos are getting flashier
because thats how you sell undifferentiated commodities
One month in the San Francisco Bay Area.
Learning how to become a founder has felt like drinking from a firehose.
What a unique place, and unique moment in time, I've managed to find myself in.
More soon.
Alexander Terenin retweeted
Mike was my colleague at Berkeley, and he's been fairly skeptical of deep learning's trajectory for years. Most scientists are skeptics, and it can take a big shock to get someone to consider a possibility that goes against the anchoring effect coded in our DNA.
I too was somewhat of a skeptic until ~6 months ago, and people ask me all the time why I joined Anthropic.
It started in April, when a friend I've known for 20 yrs messaged me saying he'd be in the Bay Area in a week, and he suggested we meet up. I invited him to my house for dinner, and he told me he had just joined Anthropic. He said the models were becoming increasingly more capable, and he believed both that they'd soon be smarter than him and also that all intellectual work would be highly devalued in the near future, and lastly that society was not at all ready for what's coming.
I'd of course heard claims similar to this before, but not from a friend I'd known for this long, and who I trusted, at my dining room table. It was the initial shock needed to destabilize my own anchoring. Also, I've been fortunate to work with and/or mentor many smart people in my life (e.g. Fields medalists, a 4x Putnam Fellow, IOI first place winners, etc.), so I feel pretty calibrated in recognizing very smart people, and this was someone who is definitely very smart and with impressive such credentials. So, it was hard to ignore this friend saying the models would be smarter than even him in particular.
After that dinner, I decided to do an informal survey, calling and meeting up with a lot of my friends who now work on improving model capabilities at these frontier labs. They all agreed with that first friend. These were all people I'd known 10, 15, 20 years, who mentored me, or I mentored them, or we met professionally in academia. These are people I trust, and their beliefs on this topic were consistent.
I then spent a whole lot of time understanding the missions of different companies, and talking to current and former employees and people in leadership. I read Dario's (long) essays and watched videos on YouTube of company leaders being interviewed. Anthropic's stated mission "to ensure that the world safely makes the transition through transformative AI" felt like exactly the right guiding principle to handle such a society-disrupting technology, and everyone I spoke to at the company was very aligned with this mission. It's easily the most mission-focused place I've ever been at, and everyone takes the work extremely seriously and with a profound sense of responsibility.
And now here I am. And being on the inside, the pace of progress has surprised me more than I first imagined. The moment that pure mathematicians are in, to me, is just a microcosm of what I predict a much larger of society will be grappling with in the not too distant future, and I really hope our leaders and the general public internalize sooner than later just how significant a moment we're in.
Michael Jordan (the machine learning one): "Superintelligence is Science Fiction" 🔥🔥🔥
di.ens.fr/~fbach/MJordan_ITW…
Alexander Terenin retweeted
In 1974 Deligne proved the last of the four Weil conjectures—an analogue of the Riemann hypothesis in a world where some prime number has been set to zero. While Deligne’s work was celebrated, his doctoral advisor, Grothendieck, was unhappy. Can solving a problem be bad? I think not, but I want to explain my understanding of the story, which in my view sheds some light on the present moment.
In his memoir, “Récoltes et Semailles” (Reaping and Sowing), Grothendieck explains his approach to problem solving and mathematics more generally: "The unknown thing to be known appeared to me as some stretch of earth or hard marl, resisting penetration… the sea advances insensibly in silence, nothing seems to happen, nothing moves, the water is so far off you hardly hear it... yet it finally surrounds the resistant substance." He compared a precise mathematical question to a nut, which, rather than being cracked open by a chisel, yields at the perfect time to simple hand pressure once it has been soaked in the ocean of mathematical understanding. According to this account of mathematical progress, solutions, though important, are ancillary to the advance of the rising sea of mathematical understanding. At best they measure it to some extent, but the goal is to raise the water level.
Grothendieck had laid out a program to resolve the Weil conjectures: the standard conjectures, which are part of the philosophy of motives. The proof circumvented this program. Deligne characterized Grothendieck's reaction as "he did not care...the proof was a trick" but in fact Grothendieck excoriated Deligne in his memoir, claiming Deligne and others tried to "bury" the theory of motives, and Grothendieck himself. The standard conjectures remain open today.
Deligne's result relied on a lot of work that was not yet well-exposited in the literature, around the theory of étale cohomology. Deligne himself wrote an exposition of this in SGA 4.5, and Illusie contributed additional necessary material in SGA 5. Some have referred to this additional work as "thankless," but it was completed.
While Grothendieck is one of my heroes, and I find his view of mathematics very appealing, I think he was wrong here. Though the standard conjectures are still open, there is an immense amount of work on the theory of motives. People worked very hard to clarify the issues left unclear by Deligne's argument. And crucially, Deligne followed up on his work with a different proof, in "La conjecture de Weil II," that completely revolutionized arithmetic geometry, in many ways going beyond the philosophy of motives as it existed at that time.
I think Grothendieck's view that the first proof Deligne provided was insufficient was, in some sense, correct. But he was wrong about the fact that the theory he so desired, and the attendant quest for understanding, would be abandoned, and that his ideas would be buried. I don't think anything can stop us from trying to understand.