HNHacker News
TopNewBestAskShowJobs

querez

976 karma · joined April 11, 2020

submissionscomments
querez··on How and why we acquired our .com domain
No, I mean the owners sold the company to an (apparently not super well reputed) other company: https://news.ycombinator.com/item?id=35080777
querez··on How and why we acquired our .com domain
Well for one I appreciated their free mailboxes, which are now gone (https://news.ycombinator.com/item?id=35080777), for another, I've seen reports of their customer service not being very good anymore (https://news.ycombinator.com/item?id=22001959). Just in general, I picked Gandi initially because of their reputation, which seems to have become worse and worse over the years.
querez··on How and why we acquired our .com domain
Slightly off-topic, but: Gandi was my registrar of choice, but I read some months ago that they've sold out. Who do people go to these days for registering domain names?
querez··on Sundar Pichai received $226M compensation, boosted by a triennial stock grant
>> fired thousands unprofessionally > How do you fire thousands professionally? I thought ensuring market and shareholder satisfaction was the top priority for CEOs and they get paid to make those unpopular decisions as/when deemed fit?

While true, it's fairly obvious that the Google firings have been badly mishandled. There are stories of SREs getting fired/locked out while still on duty. And in general, sending firing emails at 4am in the morning is very bad class in a company that claims where "don't be evil" and "trust your employees" are pillars of the company is, ... well, not good for PR.

querez··on Google DeepMind
Both DeepMind and Brain use Jax, so they will definitely use Jax. However, they use different high level frameworks: All of DeepMind uses Haiku, while on the Brain side there are competing frameworks, with flax currently being the most often used one AFAIK. I'm not aware of anyone using trax there, and I would not expect it to get more adoption, on the contrary.
querez··on Google DeepMind
DeepMind clearly is a household name. Think of AlphaGo or AlphaFold, those were legendary. Google Brain as well is a household name. Think of the Transformer, or BERT. Those are legendary, as well.
querez··on Show HN: A fully open-source (Apache 2.0)implementation of llama
IANAL, but this seems very fishy to me: 1) I don't understand how this isn't a derivative work of the original code, as I very highly doubt you've done a clean room implementation. I doubt this would hold up in court.

2) Doesn't the original FB license also apply to the weights? Just re-implementing the code would not change the license on the weights. So while THE CODE may now be re-licensed, the weights would still fall under the original license.

I'd love if someone with more legal understanding could shed some light on this.

querez··on Tell HN: I stopped caring about personal development in tech after seeing GPT-4
My comparison was merely in terms of how the technology evolves, not it's applications. Sure, Ether and zero-trust coins seem to add a lot, but from a 10.000 feet it's pretty much more of the same. Also, agreed that GPT1->4 didn't involve any breakthroughs. In my eyes, there's not much diff between them.

> Unless you have a strong source for this, I find it hard to believe

Like I said, this is just my opinion. I do have a world-expert level understanding of the technology, but at the same time I'm a strong believer that even experts are bad at predicting the future, so make of this what you will. Also, my impression of what constitutes a breakthrough and what doesn't might vary a lot from yours.

querez··on Tell HN: I stopped caring about personal development in tech after seeing GPT-4
As an AI researcher: don't forget that we're at the peak of the hype-trend: a lot of what LLMs can do looks extremely impressive, but most of it falls apart on more detailed inspection. Unless you ask about mundane things, they will tend to get a lot of tiny (and not so tiny) details wrong. At the very best, these systems will allow us to automate some boiler plate things, but they're not good enough to do complicated stuff without lots of supervision. And what's even more important: I think a lot of people look at the current impressive steps and think "oh wow, if it continues at this pace we'll have the Singularity by autumn". But the thing is: we're not able to keep this pace of progress. What you're seeing now is as good as it gets (in terms of big breakthroughs, there will still be lots of small and medium ones). The next few months (and years?) will see a ton of incremental improvements and many, many, many people trying to apply these new technologies. But I personally (as a decently successful AI researcher who's been part of these developments for over 10 years) don't see a way forward to keep making many of the big strides we've been making. As an analogy: we're having a Bitcoin moment: imagine the first blockchain was just released for the first time: now, there'll be lots of people trying to understand the tech, come up with their own variants, make some (fundamental?) improvements. But the actual fundamental tech/idea is out now, and it's not really going to change much.

TL;DR: I think your job is safe.

querez··on Strife at eLife: inside a journal’s quest to upend science publishing
as a long-time NeurIPS reviewer: the problem is that most submissions are kind of "meh", and everyone knows it. They're usually well-done papers (in the sense that the math checks out and the empirical evaluation is is done well enough), but it's pretty obvious that the paper isn't going to be influential. Mostly, these are papers of the "this is a cute idea that'll give you 0.1-2% improvements on some benchmarks people usually look at" variety. So depending on how strongly a reviewer believes that this might actually become influential if someone REALLY, REALLY digs down on this idea or not, you'll give it a "barely passing" or a "barely rejecting" grade. That's what most of these 60% usually are.

I don't think the problem here is actually the "broken peer review system". I think this is a fairly natural development: if your PhD advisor insists that you have a few "good papers" before they'll agree to let you graduate, and "publication venue" is used as a proxy for "is a good paper" (which is very reasonable, because it might take years before better proxies such as "citation count" give a good signal), then you'll submit your paper to top tier venues and hope for the best. I don't really know how to fix this: Establishing things like TMLR (Transactions of Machine Learning Research -- a rolling-release journal meant to have roughly the quality of NeurIPS) might be a good way forward. Students get their stamp of approval by publishing there and NeurIPS can significantly increase their acceptance threshold. But once you do that, you'll risk that TMLR doesn't count as "good enough venue" anymore....

querez··on Generative AI is overrated, long live old-school AI
It's not wrong. It's how modern systems operate. E.g. look at Google's SayCan (https://say-can.github.io/) which operates exactly like this (an LLM ordering a Robot around).
querez··on Use GPT-3 incorrectly: reduce costs 40x and increase speed by 5x
> from itself

That is a horrendously bad source. ChatGPT doesn't know how ChatGPT works, because nothing about how ChatGPT works was published on the internet at the time when ChatGPT was trained. Frankly, there's still nothing on the internet about this.

> Of course im looking in from the outside with no validity to my claim but if i were to write a bot from scratch thats also how i would do it.

ChatGPT wasn't "written" as much as trained. Writing a bot and training one require completely different mentalities.

querez··on International Security and Estonia 2023
No they shouldn't. Russia has played the "we must protect the Russian minority in that region" card before to justify military action. Its' the whole reason we have a war in Ukraine. Estonia must make ensure that no such pretense can exist in their country. Denying citizenship to people who don't denounce their Russian nationality and don't fully accept "becoming Estonian" is the right thing to do, IMO.
querez··on Path to a free, self-taught education in Computer Science
The English name "Software Engineer" isn't protected in German speaking countries, but their German equivalents are:

- The German equivalent would be "Diplom-Informatiker" (which roughly translates to "certified computer scientist"), which you're awarded after completing a graduate degree/MSc in Computer Science.

- In Austria you'd be awared a "Diplom-Ingeneur" instead, so a "Diplom-Ingeneuer in Informatik/Software-Entwicklung" would be a "Certified Computer/SW Engineer".

querez··on Ask HN: What were the papers on the list Ilya Sutskever gave John Carmack?
A lot of other posts here are biased to recent papers, and papers that had "a big impact", but miss a lot of foundations. I think this reddit post on the most foundational ML papers gives a lot more balanced overview: https://www.reddit.com/r/MachineLearning/comments/zetvmd/d_i...
querez··on Codon: A high-performance Python-like compiler using LLVM
In my code this comes up often, e.g. when I use tuples instead of namedtuple or a dataclass.
querez··on Comparing Google and ChatGPT
yeah good luck with that, it's going to be a very tall order to integrate PageRank with neural networks. It's not just something you can do in a year or two.
querez··on ‘Why am I talking to 10 guys?’ The rise and fall of dating apps
In general I absolutely agree with you, I also never understood this either, and have chalked it up to being an "American-ism" (meaning: it's not something that people do in my part of the world (=Europe)). But I still prefer to switch to a "normal messenger" when on tinder (e.g. WhatsApp is what most people in my area use) quickly. Mostly it's just part of playing the dating game: you don't use Tinder to talk to your friends or other people, so moving off that platform is a natural "next level". It also helps when meeting up (I can share my current location so it's easier to find me in a public place once we actually meet up). But in general I don't see this as an important step of dating.
querez··on ‘Why am I talking to 10 guys?’ The rise and fall of dating apps
> When I was single I absolutely hated the mandatory 1-3 days chatting that was necessary until it became acceptable to ask someone on a date.

This was not my experience at all. 1-3 days of chatting would destroy any chance of ever meeting someone for me. Having a nice intro chat immediately followed up by a question to meet or exchange numbers within the first 10 messages worked significantly better than trying to have meaningful chats over tinder. I find that the first excitement of a match very quickly tapers of, so either you build on that initial momentum to transition into a date, or interest will fade quickly. Because let's face it, you'll both match other people in the mean time, and then those new matches become more exciting than coming up with new small-talk when the first few messages have become stale.

Getting to know a person works better face to face, so I'd rather be quick to move to that stage, and be more selective about who I think I might want to spend a first-date hour with. (My go-to was to grab a beer at a bar and then see if there's enough interest to spend the rest of the evening together walking through the city).

querez··on Economic growth no longer requires rising emissions
It seems like you didn't read the article thoroughly. Why the cheapshot remark?
querez··on Undetectable very-low frequency sound increases dancing at a live concert
Isn't this addressed in the 2nd paragraph?

> To confirm that the VLFs were not consciously detectable, 17 new participants (one of whom participated in the concert experiment) completed a two-alternative forced choice task using the same VLF speakers in the LIVELab. On each trial, participants heard two pairs of 3.5 s excerpts from the concert audio and indicated which pair’s excerpts were different (all excerpts in the trial were identical except for the presence or absence of VLFs in one excerpt). Participants performed at chance (mean 49.8% correct, SD = 4.56%)

querez··on Low Energy Chest Fridge
Oh, that's a good point! Indeed, from what I can see American fridges are indeed larger. Mine has a volume of ~250 liters (of which 180L are fridge and 70L are freezer).
querez··on Low Energy Chest Fridge
> My chest fridge [...] consumes about 0.1 kWh a day. This fridge is 10 to 20 times more energy efficient than typical household fridges on the market today.

According to a test report by Stiftung Warentest (the German equivalent to Consumer Reports), my vertical fridge consumes ~ 0.4-0.5 kWh / day. While that is more, it's nowhere close to the 10x-20x given by this guy.

querez··on Ask HN: Where do you escape for non-clickbait thoughtful/informational content?
any suggestions on who to follow on Mastodon?
querez··on Ask HN: Working in tech for climate?
As an AI researcher, I feel like this is misguided advice. The field is overrun with people, and it's almost impossible to make a meaningful dent anymore. Even if you have a group of experienced researchers and large compute reserves, most people are content with chasing SOTAs, because the green-field low-hanging fruit of yesteryear is over (and has been over for a while). Sure, you can learn some AI and then look for climate-applications[0], but then again, why not enter that field as a software engineer in the first place. Learning AI and is just an unnecessary detour if you want to end up in climate change.

[0] https://arxiv.org/abs/1906.05433

querez··on We're going to need a lot of solar panels
true, except most people don't live in new builds. As explained in other comments here, adding insulation to old buildings is highly non-trivial.
querez··on Orkut
> Google settled the case out of court, and we all know what that means.

What does that mean?

querez··on Stanford’s war on social life
To be honest, I often felt that a lot of Diversity and Equity campaigns lead to more sterile environments: when I (Not American) joined my current (American) corporation, DEI training hammered down the point to cut out any type of joke and remark unless it was super sterile. This was definitely not something I was used to before. I can see that such things are a necessity unless you want to end up in litigation all the time, because a joke or prank that might seem harmless to 99.9% of people will still offend that one outlier, and at the scale of huge corporations (or huge college campuses) you will find that one person who perceives this as yet another micro-agression at their expense. So everything that isn't bland and non-offensive gets increasingly outlawed, or at least discouraged. It definitely leads to a lot of self-censoring. Am I even allowed to poke fun at my own geekness, or make fun of stereotypes of my own nationality (obviously I wouldn't make fun of other people's traits these days), or will that offend any bystander that also shares these features with me?

Of course, the end result is what you pointed out: in a litigation-happy place (like the US) everything that offends people is going to be dragged into the courts by someone, and is going to cost actual money, and everyone wants to avoid that.

querez··on Nvidia adds liquid-cooled GPUs
If you can run your inference on CPU and it only takes 10x longer than on GPU, then it sounds to me like a 3090 will be vast overkill. Have you considered an inference-only GPU like the T4 or the A2? From what you're saying it seems to me like they're made for use cases exactly like yours.
querez··on One App – Two Worlds: This Is TikTok in Russia and Ukraine
I don't understand how they're able to just "change their IP" to something that's located in those 2 cities. How does that work, do they have someone running a proxy for them in those 2 locations?
← PreviousPage 3 of 7Next →