HNHacker News
TopNewBestAskShowJobs

EvgeniyZh

9,025 karma · joined April 27, 2016

https://evgeniizh.com/ https://t.me/j_links https://twitter.com/evgeniyzhe
submissionscomments
EvgeniyZh··on U.S. bombs Iranian nuclear sites
3. You can use highly enriched uranium (HEU) for things other than weapons, but the weapons are the only things that require HEU. Medical isotopes and propulsion can be done with LEU. For instance, Argentina produces medical Mo99 from LEU [1]. US Navy wants to switch to LEU for submarines [2]. One of the reasons for these developments is exactly proliferation risk management.

[1] https://inis.iaea.org/records/fe51q-17w28/files/35015774.pdf

[2] https://fissilematerials.org/library/doe16.pdf

EvgeniyZh··on U.S. bombs Iranian nuclear sites
> the Muslim world, one beset with colonial extraction for centuries

Surely you mean on the side of extractors? The Ottoman Empire practiced mass movement of people (sürgün), basically settler colonialism; earlier Umayyad and Abbasid caliphates are among the largest empires in history, and their population was mass converted to Islam.

EvgeniyZh··on New US visa rules will force foreign students to unlock social media profiles
So out of 9 members of "the Squad", 7 were reelected and only 2 outed? Quite inefficient suicide in my book.
EvgeniyZh··on New US visa rules will force foreign students to unlock social media profiles
> He didn't say that.

Sure, you replace "Jews" by "Zionists" and then every trope is ok. It's Zionists who have the power to influence or direct US government behavior (i.e., control it).

> Two members of congress? Out of hundreds?

Well yeah, to show something possible it's enough to show one example.

> Yes!

How is being elected to Congress political suicide?

> It's not even debatable that the US is absurdly pro-Israel.

I'm not debating that. I'm debating the idea that this is somehow doing of evil Zionist lobby or that repeating antisemitic tropes is not antisemitic because you euphemize Jews by Zionists

EvgeniyZh··on Quantum Computation Lecture Notes (2022)
There has been steady progress in metrics, that, if keep improving, will eventually lead to fault tolerant quantum computing. We also do not know any fundamental reason why we won't be able to keep improving them. It's true that the progress is not fast, because the problem is hard, but I don't see why would you call it scam (there are definitely scammers in the field, but there are in any field).
EvgeniyZh··on Vision Language Models Are Biased
Well you send a vaguely worded email like "I think you may find our work relevant" and everyone knows what that means and adds the citation
EvgeniyZh··on Deepseek R1-0528
> I wasn't trying to invent anything. Just describing what you would obviously have to do if you were to take a "scientific" or "objective" approach: Sound experiments, reproducible, free of financial incentives.

But how is it different from what arena or matharena does?

> That sort of thing is not showing broad intelligence anymore than a person both knowing a chess player and a poet is having broad intelligence.

The claim is that these problems require somewhat broad intelligence by themselves, as opposed to specialization into specific task while unable to do anything else.

EvgeniyZh··on Deepseek R1-0528
> But those benchmarks are in general fairly narrow. They don't really measure the "broader" intelligence we are after.

I think a general model that can

- finish nethack, doom, zelda and civilization,

- solve the hardest codeforces/atcoder problems,

- formally prove putnam solution with high probability, not given the answer

- write a PR to close a random issue on github

is likely to have some broader intelligence. I may be mistaken, since there were tasks in the past that appeared to be unsolvable without human-level intelligence, but in fact weren't.

I agree that such benchmarks are limited to either environment with well-defined feedback and rules (games) or easily verifiable ones (code/math), but I wouldn't say it's super narrow, and there are no non-LLM models to perform significantly better on these (except some games); though specialized LLMs work better. Finding other examples, I think, is one of the important problems in AI metrology.

> So, yes, we are kind of left with vibe checks, but in theory, we could do more; take a bunch of models, double-blind, and have a big enough, representative group of human evaluators score them against each other on meaningful subjects.

You've invented an arena (who just raised quite a lot of money). Can argue about "representative," of course. However, I think the SNR in the arena is not too high now; it turns out that the average arena user is quite biased, the most of their queries are trivial for LLMs, and for non-trivial ones, they cannot necessarily figure out which answer is better. MathArena goes in opposite directions: narrow domain, but expert evaluation. You could imagine a bunch of small arenas, each with its own domain experts. I think it may happen eventually if money flow into AI continues.

EvgeniyZh··on Deepseek R1-0528
There are quite a few benchmarks for which that's not the case:

- live benchmarks (livebench, livecodebench, matharena, SWE-rebench, etc)

- benchmarks that do not have a fixed structure, like games or human feedback benches (balrog, videogamebench, arena)

- (to some extent) benchmark without existing/published answers (putnambench, frontiermath). You could argue that someone could hire people to solve those or pay off benchmark dev, but it's much more complicated.

Most of the benchmarks that don't try to tackle future contamination are much less useful, that's true. Unfortunately, HLE kind of ignored it (they plan to add a hidden set to test for contamination, but once the answers are there, it's a lost game IMHO); I really liked the concept.

Edit: it is true that these benchmarks are focusing only on a fairly specific subset of the model capabilities. For everything else vibe check is your best bet.

EvgeniyZh··on 'Strange metals' point to a whole new way to understand electricity
Theoretical condensed matter here (though not an expert on strange metals specifically). The idea is that in a quantum many-body system, a (low-energy) state can often be represented as a set of independent excitations of the field, which are called quasiparticles. The "particle" picture for the Fermi liquid already operates in a framework of quantum fields. "No quasiparticles" means that there is no such representation, i.e., you can't say something like "these two states differ by an addition of a particle with such-and-such properties."
EvgeniyZh··on Data manipulations alleged in study that paved way for Microsoft's quantum chip
The point still stands, Microsoft didn't build their research on top of that, but started their research because of these problems.
EvgeniyZh··on Data manipulations alleged in study that paved way for Microsoft's quantum chip
> I agree with you, its a terrible idea to base your work on someone else's when it hasn't been well confirmed in independent research.

It's not about whether it's good or bad idea. To make follow-up experiments you need to first reproduce the original experiment. That's why faking "big" experiments like Schön could never work.

> Microsoft built so much work and spent so many resources building on top of it. It's not foundational to the entire field but it is foundational to a lot of follow-up research.

Will all due respect, a single group (even large one) doing a single type of experiments (even important and complicated one) is not a lot of research. Also, Microsoft knew about data manipulation, that why they moved the experiments in house. They didn't do experiments under assumption that the early Majorana papers are correct, then they wouldn't need to develop their own complicated (and somewhat controversial) protocol to detect Majoranas. It was quite clear for everyone that regardless of data manipulation people were too optimistic interpreting Majorana signatures in these early papers.

EvgeniyZh··on Data manipulations alleged in study that paved way for Microsoft's quantum chip
> won't be the last time we hear about quantum research that has been foundational to a lot of work

This research wasn't foundational to a lot of work. Most of important/foundational works in quantum (doesn't matter if computing or general, I'm not sure which one you meant) are verified. How can you possibly base your experimental work on someone else's work if you can't replicate it?

EvgeniyZh··on Data manipulations alleged in study that paved way for Microsoft's quantum chip
There was for this year's sigbovik [1]

[1] https://fixupx.com/CraigGidney/status/1907199729362186309

EvgeniyZh··on I'd rather read the prompt
> In American universities where your GPA from your in-class assessments forms part of your final grade? Yes, absolutely.

Whether not doing everything to maximize your GPA is "stupid" (literally or figuratively) is a good question too.

But even if your assignments influence your GPA it's rarely the only thing that does, and not doing assignments will harm your ability to perform in midterm/exam/whatever.

EvgeniyZh··on I'd rather read the prompt
YMMV, in my uni there is more or less zero impact of cheating in your homeworks in most classes: the homework has low weight in final grade and are graded very generously.

I agree that making assignments not designed with external sources in mind significantly impact the final grade is not ideal. I think this is minor and easily fixable point rather than some failure of the whole education system.

EvgeniyZh··on I'd rather read the prompt
> they would be stupid not to if their classmates are using it.

Why would they be stupid? Were people before LLMs stupid for not asking smarter classmate/parent/paid contractor to solve the homework for them?

Large part of education is learning about things that can be easily automated, because you can't learn hard things without learning easy things. Nothing conceptually changed in this regard, like Wolfram Alpha didn't change the way differentiation is taught.

EvgeniyZh··on DeepSeek-Prover-V2
These are still need to be formalized in Lean which can be harder than solving the problem sometimes
EvgeniyZh··on DeepSeek-Prover-V2
The solutions weren't published anywhere. There is also no good automatic way to generate solutions as far as I know, even expensive ones (previous sota was 10 solutions and one before was 8 using pass@3200 for 7b model). Potentially the developers could've paid some people who are good in putnam-level math problems and lean to write solutions for LLMs. It is hard to estimate likelihood of that but it sounds like waste of money given relatively marginal problem/benchmark.
EvgeniyZh··on Sycophancy in GPT-4o
The solution is make a public leaderboard with scores; all the LLM developers will work hard to maximize the score on the leaderboard.
EvgeniyZh··on PhD Timeline
> Do you actually prefer if people with a different political view than yourself are just… removed?

Do you think that the whole "diversity statement" and DEI shticks were exactly tools of removal of people with different political views from academia?

US academia political activity is what made it attractive target for Trump administration. I'd prefer that it didn't happen or that US didn't have a president willing to destroy the academia, but I'd also prefer that universities would act as non-partisan entities. I think Boaz Barak summarized it well [1]

On a separate note, US searching/detaining/refusing entrance to other countries' nationals happened all the time [2,3,4,5,6,7], I don't think it's newsworthy or should change your position on visiting US.

[1] https://www.thecrimson.com/article/2025/3/28/barak-harvard-m...

[2] https://www.straitstimes.com/asia/east-asia/us-entry-horror-...

[3] https://firrp.org/wp-content/uploads/2024/09/September-2024_...

[4] https://immigrantjustice.org/press-releases/three-travelers-...

[5] https://www.vice.com/en/article/the-us-keeps-mistakenly-depo...

[6] https://www.propublica.org/article/can-customs-border-protec...

[7] https://thepienews.com/indian-government-gives-student-trave...

EvgeniyZh··on Jagged AGI: o3, Gemini 2.5, and everything after
Is your research AI-, or more generally, CS-related? Because I feel that it is still quite bad (by researcher standards) in physics for example.
EvgeniyZh··on Palestinian activist arrested by ICE while expecting U.S. citizenship interview
It sounds plausible that Mahdawi is the second person on Khalil's case memo [1]. If that's the case, the claim is that "alien whose presence or activities in the United States the Secretary of State has reasonable ground to believe would have potentially serious adverse foreign policy consequences for the United States" is deportable.

[1] https://www.documentcloud.org/documents/25894225-dhs-documen...

EvgeniyZh··on Neutron Stars Hint at Another Dimension
No they don't
EvgeniyZh··on Microsoft unveils Majorana 1 quantum processor
> I’m an experimentalist at Microsoft Quantum who was involved in the work presented in our recent Nature publication.

It is very cool to hear from you!

> In my opinion, the citations above do not represent a balanced view of the Majorana field status but are rather negative.

That's true, but the goal of the citations was to demonstrate there are some negative opinions too. Maybe together with positive OP these form a balanced view.

I understand that it can be very unpleasant to have people like Frolov or Legg trying to prove you're wrong, but I think it shoudn't be personal (from either side). Trying to find alternative explainations is part of science. And Frolov did turn out correct in past, and we did think we found Majoranas when in fact we didn't, and this part of the story can't be just ignored. Citing Feynman "The first principle is that you must not fool yourself and you are the easiest person to fool". While it's tempting to dismiss the critics as broken record, I think it would both increase the credibility of the studies and improve the science if their criticism was taken at the face value. Answering specific points publicly would also create more balanced picture. I'm not aware of the responses to the cited opinions that I could cite to "balance out."

> We published two experimental papers recently that went through a rigorous peer review process.

Peer review is important, but is not the answer to specific claims, eg that TGP accuracy is overestimated, or (if we take Henry's word for it) the promised errata that never came out.

> Finally, we have exciting new results that we just shared with many experts in the field at the Station Q conference in Santa Barbara.

I've read about it from Das Sarma's twitter [1]. It does indeed sound exciting. If you're able to manipulate, store, and read out quantum data from qubit, then I think people will have easier time to agree you have one. There is of course question of non-Clifford gates, but that's a separate problem.

> We will share more broadly at the upcoming APS March meeting.

I look forward to hearing about it. If you (or someone from your team) are interested, I'd love to meet and chat at MM. My contacts are in bio.

Edit: I've also now seen Chetan Nayak's comment in Scott Aaronson blog with some details [2].

[1] https://x.com/condensed_the/status/1892595693002293279

[2] https://scottaaronson.blog/?p=8669#comment-2003328

EvgeniyZh··on Microsoft unveils Majorana 1 quantum processor
That's fair, and I don't like their marketing too (and others too, look how the reviewers pushed back on misleading claims in the paper). I was talking specifically on Majoranas being behind sc/trapped ions/cold atoms. If they manage to make 1 and 2 qubit gates and it will have good error rates in couple of years (and that's a big if), they'll be approximately where for example Google is expected to be. And the question what will be easier to scale is very unclear and will decide who wins eventually
EvgeniyZh··on Microsoft unveils Majorana 1 quantum processor
I'm doing theorical research in the topological quantum computing.

The idea behind topological quantum computing is to utilize quantum materials whose low-energy physics looks like an error correcting code. Since these systems are very large (macroscopic number of atoms), the error rates are (theoretically) very low, ie the qubit is fault tolerant by construction, without any additional error correction. In reality, we do not know how good these qubits will be at finite temperature, with real life noise, etc.

Moreover, these states do not just occur in nature by themselves, so their construction requires engineering, and this is what Microsoft tries to do.

Unfortunately, Majoranas in nanowires have some history of exaggerated claims and data manipulation. Sergey Frolov's [1] twitter, one of the people behind original Majorana zero bias peaks paper, was my go-to source for that, but it looks like he deleted it.

There were also some concerns about previous Microsoft paper [2,3] as well as the unusual decision to publish it without the details to reproduce it [4].

In my opinion, Microsoft does solid science, it's just the problem they're trying to solve is very hard and there are many ways in which the results can be misleading. I also think it is likely that they are making progress on Majoranas, but I would be surprised if they will be able to show quantum memory/single qubit gates soon.

[1] https://spinespresso.substack.com/p/has-there-been-enough-re...

[2] https://x.com/PhysicsHenry/status/1670184166674112514

[3] https://x.com/PhysicsHenry/status/1892268229139042336

[4] https://journals.aps.org/prb/abstract/10.1103/PhysRevB.107.2...

EvgeniyZh··on Microsoft unveils Majorana 1 quantum processor
The claim/hope is that topological qubits are fault tolerant or at least suffer from much lower errors (very roughly you can think of topological qubits as an error correction code built of the atoms, ie on scale of Avogadro's number). If, for example they could build a single qubit even with 10^-6 error rates that would in fact put them __ahead__ of all other attempts at the path to fault tolerance (but no NISQ).

It is unfortunately unclear how good the topological qubits practically are.

EvgeniyZh··on Jane Street's Figgie card game
A better strategy would be to estimate card value based on your starting hand (i.e., probability that each suit is goal suit). It requires some non-trivial combinatorics but I guess if you take the game seriously you just memorize those once. Then you buy cards if the price are lower than your estimate and sell if it is higher. You also should pay a somewhat more for the card that gives you majority (4th/5th).

The updates are in this case, at least assuming others are doing something similar, is to increase expected price if you see sell for higher price and reduce if you see the sell for lower price. To figure out how much to update would be the hard part.

EvgeniyZh··on NSF starts vetting all grants to comply with executive orders
SAT/ACT scores are the best predictors of undergrad GPA as well as probability to have high income in the future, largely independent on the socio-economic background [1].

[1] https://opportunityinsights.org/wp-content/uploads/2024/01/S...

← PreviousPage 3 of 14Next →