HNHacker News
TopNewBestAskShowJobs

currymj

3,822 karma · joined March 27, 2016

submissionscomments
currymj··on GenAI Art Is the Least Imaginative Use of AI Imaginable
i'm sort of pleased to see that with DeepSeek and the o-series, the hype/energy is shifting back towards coding and mathematical tasks, and away from generating images and videos. Programmers seem to be mostly happy with LLMs, unlike artists and musicians who almost all hate them. It seems possible that the incentives will continue to be aligned well, because it's staff at AI labs automating the skills that make them valuable (rather than putting others out of a job).

It's also not so unpleasant. I don't want to see AI images any more. I certainly don't want to listen to AI music. People are right to call it "slop". I understand how the field of machine learning got to image generation as a task, and that the products can be sometimes useful and fun as a novelty. But there is just a giant externality caused by the existence of these products.

currymj··on No Bitcoin ETFs at Vanguard (2024)
to the extent that executing transactions on a fully decentralized, trustless database or distributed computer is something people demand, and transaction fees must be paid in Bitcoin or Ethereum or Solana or whatever, cryptocurrencies do have some amount of actual use value, so are sort of like "commodities".

I believe this is a correct argument in theory. In practice, at the moment, everything is exclusively driven by speculation, and if you took the degenerate gambling out of the system, I don't know what would be left.

It's just gambling at all levels. There are legitimately "useful" services on blockchains, but the useful services they provide are mainly facilitating more gambling on stupid garbage coins. And even such useful "picks and shovels" services often still choose to make their own token for people to speculate on, because crypto people aren't interested in anything else. And then the L1 cryptocurrencies also have their prices driven by speculation.

there is a core of actual technical innovation, which is almost totally irrelevant to the phenomenon of "crypto" today. But there may be a future world where cryptocurrencies provide real value. In that world where all the true believers have their wildest dreams come true, the price of a cryptocurrency would be much higher than zero, but even then, it might be much lower than they are right now.

currymj··on Stargate Project: SoftBank, OpenAI, Oracle, MGX to build data centers
there are many many problems with the EA movement, but they do generally support unconditional cash transfers.

cash transfers are seen as the "default" baseline. the bar for charity is that it must be better than cash transfers. they do find some such charities that they claim are even better than cash transfers, but they are totally comfortable with giving people unconditional cash.

currymj··on Open Socrates by Agnes Callard review – a design for life
i like to think about how the influences flow in all directions.

for example at one time there were a bunch of weirdos arguing over some abstruse philosophical question about the fundamental nature of true logical statements, and one of them came up with the “Turing machine”.

then once it was actually built (by engineers, by way of physical and mathematical advances), it started to cause all kinds of confusion in philosophy of mind and other areas.

currymj··on More than 40% of postdocs leave academia, study reveals
generally you are somewhat right, but it's complicated by country. for example it would be less surprising that a UK-style 3 year PhD might want a couple more years training after.
currymj··on ROCm Device Support Wishlist
it's an interesting question. the unhelpful answer is Vulkan didn't exist when Tensorflow, PyTorch (and Torch, its Lua-based predecessor) were taking off and building GPU support. Apparently PyTorch did at one point prototype a Vulkan backend but abandoned it.

My own experience is that half-assed knowledge of C/C++, and a basic idea of how GPUs are architected, is enough to write a decent custom CUDA kernel. It's not that hard to do. No idea how I would get started with Vulkan, but I assume it would require a lot more ceremony, and that writing compute shaders is less intuitive.

there is also definitely a "worse is better" effect in this area. there are some big projects that tried to be super general and cover all use cases and hardware. but a time-crunched PhD student or IC just needs something they can use now. (even Tensorflow, which was relatively popular compared to some other projects, fell victim to this.)

George Hotz seems like a weird guy in some respects, but he's 100% right that in ML it is hard enough to get anything working at all under perfect conditions, you don't need fighting with libraries and build tools on top of that, or the mental overhead of learning how to use this beautiful general API that supports 47 platforms you don't care about.

except also "worse is better is better" -- e.g. because they were willing to make breaking changes and sacrifice some generality, Jax was able to build something really cool and innovative.

currymj··on ROCm Device Support Wishlist
users mainly use PyTorch and Jax and these days rarely write CUDA code.

however separately, installing drivers and the correct CUDA/CuDNN libraries is the responsibility of the user. this is sometimes slightly finicky.

with ROCm, the problem is that 1) PyTorch/Jax don't support it very well, for whatever reason which may be partly to do with the quality of ROCm frustrating PyTorch/Jax devs, 2) installing drivers and libraries is a nightmare. it's all poorly documented and constantly broken. 3) hardware support is very spotty and confusing.

currymj··on Is the world becoming uninsurable?
adaptation to hurricane winds has largely been done in many parts of Florida; adaptation to storm surge is possible and some cities are beginning to.

the issue for Florida is that the state is made of permeable limestone, so it’s not possible to engineer around sea level rise. not so much an insurance issue exactly though, because it’s not a one-off disaster.

currymj··on Is the world becoming uninsurable?
you cannot win this argument with the average person who lives in a chilly European country. it just does not compute.

there are whole important cultural lifeways related to opening and closing windows at proper times for efficient cooling and ventilation. these work really well — in Europe — and are treasured traditions.

getting people to accept AC is sort of like trying to convince the average American to go grocery shopping on a bicycle. some may accept the idea but only the most European influenced already.

currymj··on Is the world becoming uninsurable?
A lot has to do with infrastructure.

In most of South Florida basically anything left standing is pretty well built to withstand hurricanes.

A category 1 storm hitting NYC or North Carolina is an unbelievable disaster. A category 1 storm hitting Broward County is usually disruptive to everyday life but that’s it.

currymj··on TikTok preparing for U.S. shut-off on Sunday
obviously bad policy for many reasons, but as a geriatric millennial I'm selfishly happy. As long as the ban continues, I will never have to sit on the bus and listen to those horrible robot voices blasting nonsense out of someone's phone speakers.
currymj··on JPMorgan Workers Ponder Union in Wake of Return-to-Office Mandate
I generally think for many jobs, in-person is actually better.

SWE in a place with good process, remote is probably fine. Generally similar for jobs with clear deliverables on that kind of time scale. if it's more research oriented, there is a huge benefit to being in-person in front of a whiteboard. For a job with shifting requirements day to day (like some legal or banking work), it is easier to coordinate everyone in the same office.

unclear if this overcomes cost to employees of the commute, housing, etc., but the value is there.

however there is also "worst of all possible worlds" RTO, where you have to commute to the office because the office is the place you are required to sit for your 4+ hours of Zoom calls per day with other colleagues in other offices. I expect a lot of companies are going to do this, which is totally stupid.

currymj··on Perplexity got ads
because they have to pay for queries to other APIs for each search, COGS is relatively high I think, which is counterintuitive for a web product like this.

my sense is their fixed costs are relatively low.

currymj··on Perplexity got ads
for each Kagi search they hit APIs of other search providers etc., so they need to cover that cost.
currymj··on A path to O1 open source
many people are dismissing this paper because it has errors in spelling and grammar.

this is a terrible heuristic for evaluating AI papers. If you use it, you will miss a lot of good work by very strong researchers with below-average English writing skills.

I have not read this paper carefully so claim nothing one way or the other about its quality. It superficially seems like a pleasant and timely survey although a little flag-planty.

currymj··on I sensed anxiety and frustration at NeurIPS 24
this is true overall, but for the last decade there have been many research-oriented positions at large tech companies for which the PhD degree is a formal job requirement. others don’t make it a formal requirement, but demand research experience which can be gotten most easily as a PhD student.
currymj··on i sensed anxiety and frustration at NeurIPS 24
i think PhD programs and advisors did make some implicit or explicit promises to recruit students to work for them, very hard, for years, at very low salaries.

the PhD is not like a master's program, it's an apprenticeship relationship where you essentially work for a senior professor, and your work directly financially benefits the department as well.

so there is a deal here, but if one end of it is collapsing, the students in some departments might very well feel betrayed -- not by society but by their advisors and programs.

currymj··on I sensed anxiety and frustration at NeurIPS 24
the physical conference itself is now an excuse for companies to throw lavish parties, but NeurIPS as a publication venue is still mainly for academics.

for research that translates directly to an industry setting, look someplace like KDD. venues like that are where data science/ML teams who do some research but are mainly product focused tend to publish.

currymj··on FTC bans hidden junk fees in hotel, event ticket prices
i guess it's technically true that you can go to the movie theater, buy a ticket, and then get back in line to refund the ticket and leave, but this seems like an odd thing to do.

whereas buying a ticket in advance online, and then later refunding it (but losing the convenience fee), is common. "peace of mind knowing your spot is reserved" is what i mean by an option on the seat.

now maybe i'm in fact wrong and all convenience fees are exactly passing through the credit card/fandango fees etc., but in practice it sure looks like an option.

currymj··on FTC bans hidden junk fees in hotel, event ticket prices
when you buy a ticket online, you are actually buying an option on the ticket, because you can go in and refund it at any time before the movie starts. the “convenience fee” is the option premium.

it would be fairer if this were opt-in. Some e commerce sites now allow you to pay a few extra dollars to have free returns, something similar would work for movies.

currymj··on When a Crystal Ball Isn't Enough to Make You Rich
a real world version of this experiment: from time to time there is a hacking case where a criminal group steals earnings reports from companies before they are made public. some of these conspiracies have been quite profitable, but i believe all of them had at least a few trades go against them.

a more extreme "crystal ball" were certain life insurance policies written by the French insurance company Aviva. They allowed customers to purchase shares today at last week's prices. This sounds incredibly stupid, but that's what the contract said, although the insurance company would like to get out of it. The legal battles have dragged on for a couple decades at this point.

currymj··on What Is Entropy? A Measure of Just How Little We Know
i think those convenient properties are the sign of some deeper law. you do want to work with the exponential and logarithm, same as in so many other places. you don’t necessarily want some other generalized “entropy”.
currymj··on Willow, Our Quantum Chip
specifically avoid resources written by and for physicists.

the model of quantum mechanics, if you can afford to ignore any real-world physical system and just deal with abstract |0>, |1> qubits, is relatively easy. (this is really funny given how incredibly difficult actual quantum physics can be.)

you have to learn basic linear algebra with complex numbers (can safely ignore anything really gnarly).

then you learn how to express Boolean circuits in terms of different matrix multiplications, to capture classical computation in this model. This should be pretty easy if you have a software engineer's grasp of Boolean logic.

Then you can learn basic ideas about entanglement, and a few of the weird quantum tricks that make algorithms like Shor and Grover search work. Shor's algorithm may be a little mathematically tough.

realistically you probably will never need to know how to program a quantum computer even if they become practical and successful. applications are powerful but very limited.

"What You Shouldn't Know About Quantum Computers" is a good non-mathematical read.

https://arxiv.org/abs/2405.15838

currymj··on Self-Funding Harberger Taxes
realistically there should just be some kind of homestead exemption for one primary residence. several states already have such a system, where you get reduced or frozen property taxes, and protection during bankruptcy.
currymj··on What happens if we remove 50 percent of Llama?
i think the younger generation who came up post deep learning, has a very very low bar for “understanding” because they never knew a world where SotA models worked in a way that made sense.
currymj··on Engineering Sleep
website marketing materials for Quviviq make a big deal that it should be used daily for best results.
currymj··on Taxing unrealized gains has caused an entrepreneurial exodus in Norway
they separate out oil activities in their national accounts. “Mainland Norway” GDP is also not bad compared to neighboring countries.

although it doesn’t measure indirect effects of oil wealth on other sectors. but still, “petrostate” isn’t really accurate.

currymj··on Robot Jailbreak: Researchers Trick Bots into Dangerous Tasks
your sentence is correct but we have no idea what a comprehensive model of the world looks like, whether or not these systems have one or not, what harm even means, and even if we resolved these theoretical issues, it’s not clear how to reliably train away harmful behavior. all of this is a subject of active research though.
currymj··on Undergraduates with family income below $200k will be tuition-free at MIT
actually ETHZ and EPFL are very good and highly ranked, and have cheap tuition and open enrollment. i don’t know how they do it. I guess things just work better in Switzerland.
currymj··on Undergraduates with family income below $200k will be tuition-free at MIT
the ideal is that college should be very expensive for rich people and cheap, free, or at least more affordable, for less wealthy people.

american universities get closer to this ideal than you might expect. the days of outrageous student debt are thankfully fading away, at least for undergraduate degrees.

it would make more sense to do this redistribution through taxes if possible, but many US institutions are private so that doesn’t really work. so the colleges basically have their own privately-run means testing programs, and like all such programs there are flaws and loopholes.

← PreviousPage 5 of 34Next →