Hacker News: usernametaken29

New comment by usernametaken29 in "Sperm Whales blow bubbles to achieve restful, vertical sleep"

usernametaken29 — Sat, 25 Jul 2026 12:09:47 +0000

> It is particularly fascinating they are able to make these fine-scale adjustments while they are thought to be asleep, a challenge totally alien to terrestrial mammals like us humans.

While fascinating it’s not surprising. The vegetative nervous system does a LOT of things on autopilot. I think most people have no idea how many things their body regulates permanently that are actually technically difficult (orientation is particularly hard) and the lower brain stems just wing it.

For me this is a classic case of associating intelligence only with conscious awareness when there’s many other ways how we (and other animals) are “clever” even if we don’t know it at first ;)

New comment by usernametaken29 in "ARC-AGI Leaderboard"

usernametaken29 — Sat, 25 Jul 2026 11:12:29 +0000

It’s because of inductive bias. Harnesses will massively skew results towards working solutions. You might think that’s a good thing but what it might mean that sometimes it becomes enough to run brute force search or a simple parameter search over the harness. Creating the harness is the actual work, because you’re selecting what are the levers to pull. There were some attempts of LLMs generating harnesses on the fly in ARC 2, but they were all mostly based on one handcrafted DSL that was copied over and over again. As it stands harnesses are not allowed because they’re simply not a meaningful measure. What you’d like is to measure how the model performs if it saw this benchmark for the very first time… but then again everyone knows the game is rigged, millions are at stake, and the AI companies fine tune and cheat on the benchmarks any way they can.

New comment by usernametaken29 in "Ethics Is the Skill That Matters Most Now"

usernametaken29 — Wed, 22 Jul 2026 17:13:57 +0000

> Every skill humans once bragged about is getting commoditized. Writing, coding, analysis, translation, design, research. A machine does each of them faster, cheaper, and increasingly better.

This is very much up for debate. I also would have liked your article to delve into whether thing’s that bring us humans joy actually need to be faster, and cheaper. Knowing why you do something crucially depends on knowing the thing being done. It’s very hard to cast an ethic judgement over a process you don’t understand. If your argument is that we need to be ethic stewards of machines but need to step aside the doing then your premise might lead you straight into tautology.

New comment by usernametaken29 in "OpenAI Presence"

usernametaken29 — Wed, 22 Jul 2026 16:55:02 +0000

Maybe the LLM doomsday won’t be self aware AI but instead everyone will be able to alter anything in any customer service backend in any way they want to the point we bankrupt all the banks and insurance agencies and governments. Yikes

New comment by usernametaken29 in "OpenAI Presence"

usernametaken29 — Wed, 22 Jul 2026 16:47:49 +0000

> The challenge for enterprises is no longer proving that AI agents can work, it’s making them reliable enough to do high-value work in production.

Is this some sort of joke? Like. That’s the proof. If it can do reliable work then that’s the proof. Contradiction in sentence one. If your pipeline breaks 80% of the time but sometimes the magic token lottery gives you something remotely useful, that’s NOT proof. The ceiling for these companies is so low it’s unbelievable. What happened to products that worked first, and worried about the rest later? Can we have an inspired ad article for that instead?

New comment by usernametaken29 in "I Built a Private Genomics Study with Stoffel MPC"

usernametaken29 — Wed, 22 Jul 2026 16:12:10 +0000

So essentially you’re telling people: “Don’t use the evil labs they can read your DNA, instead send DNA to our lab and with magic maths we won’t be able to read your DNA but diagnose you!”

Sounds about right to me

New comment by usernametaken29 in "I Built a Private Genomics Study with Stoffel MPC"

usernametaken29 — Wed, 22 Jul 2026 16:10:31 +0000

This! And even if you could magically have a 1000$ sequencer at home, you still need to correctly extract the blood, and apply the kit. It sounds easy but really isn’t, this isn’t home baking but PCR, where it’s easy to mess up. You need to know when a result is a true positive and whether your sample is useful / correctly collected. Maybe an inspired bioinformatics citizen lab can do it, but it’s bizarre to think the average consumer would do it. It’s not as simple as taking a drop of blood beep boop result like it is for insulin tests and I don’t think it’ll be like that for a very long time to come.

New comment by usernametaken29 in "Passkeys were invented by engineers with zero understanding of consumer brain"

usernametaken29 — Wed, 22 Jul 2026 15:45:33 +0000

I literally presented you with the solution in the sentence before. Buy a usb c dongle key. You pay 20 euros to NEVER have to remember a single password ever again. Seems like an ok trade off to me. Likewise you can register multiple passkeys for multiple devices, so long as you do it in sequence (first Apple, then Android) etc. Really it takes very little getting used to. Ever tried to fill out a crap password form which wants 7 special characters but no dashes but uppercase but not uppercase Y and so forth.

New comment by usernametaken29 in "I Built a Private Genomics Study with Stoffel MPC"

usernametaken29 — Wed, 22 Jul 2026 15:37:01 +0000

In genetics the end user is never the consumer. You have to collect blood, prepare it, sequence it. These are all highly specialised steps in which something will and can go wrong, and that’s why typically a hospital carries them out. That’s before even getting the device to do sequencing. The cheapest nanopore comes you at a couple ten thousand euros. Given the premise your work strikes me as oddly theoretical. Who is that mystery home DNA lab that needs decentralised private compute? Also, and funnily enough, for much of the genomic pipelines to run (eg a paternity test or a cancer test), a home computer is sufficient. Again, of course, you would need proper medical training to read, interpret and judge the results, which is why a hospital does it… so what problem exactly are you solving?

New comment by usernametaken29 in "Passkeys were invented by engineers with zero understanding of consumer brain"

usernametaken29 — Wed, 22 Jul 2026 15:27:42 +0000

I don’t understand this point at all. I think the author has himself confused with the average consumer. For the first time in a decade or so you can buy a PHYSICAL key and use it to sign into websites. I can explain this to any grandma out there. Likewise, I’m an Apple user. Once you’re in Apple universe passkeys are extremely easy. Tap your thumb on the scanner, done. Now we can put on the tinfoil hat and say how this fosters vendor lock in yadda yadda but the last thing I would say is that it has terrible user ergonomics. LOL

New comment by usernametaken29 in "Building relationships with customers through support didn't turn out as hoped"

usernametaken29 — Mon, 06 Jul 2026 09:31:51 +0000

I personally know some businesses in which the key differentiator is excellent human customer support. But then again, we’re talking 5 digit software licenses where that matters a lot more then for simple SaaS subscriptions

New comment by usernametaken29 in "New AI tutor achieves 0.71-1.30 SD effect size in Dartmouth course [pdf]"

usernametaken29 — Sun, 05 Jul 2026 23:19:21 +0000

Wow, putting effort into training material, thoughtfully designing it, and relating the material to the final exam, will increase performance on said exam. So much AI so much wow. Like seriously. Most university statistics courses suck big time, so literally any effort put into them will Improve the field. I’m happy the authors want to improve education but they don’t seem to understand that preparing questionare style material is a confounding factor which could very well explain the better performance too… instead of cramming AI into the next thing. I’m generally opposed to AI on basic textbooks. You don’t want hallucinations imprinted on students who have no idea and can’t judge the quality of the generated text. Some things require effort, reading intro to statistics is one of them, and it’s for a reason, the effort IS the learning

New comment by usernametaken29 in "Reality has a surprising amount of detail (2017)"

usernametaken29 — Fri, 03 Jul 2026 12:47:57 +0000

This article is awesome. It should be required reading for all engineers but probably mostly ML researchers. I’ve encountered my fair share of geniuses that are oblivious to the fact that the world is indeed complex and has a near infinite amount of detail. Of course, they’ve been trained as engineers, so they think in models and abstractions, but reality is almost always more complex than what people at desks estimate it to be.

New comment by usernametaken29 in "Is One Layer Enough? A Single Transformer Layer Matches Full-Parameter RL Train"

usernametaken29 — Fri, 03 Jul 2026 05:42:35 +0000

You’re thinking information content in the neurons but not distributions. D(world of text) is much much much larger than D(transformer). Just because you’re passing information through a smaller channel doesn’t mean the original distribution necessarily isn’t smaller. Tokenisers are the numbers but then you have V^n where V is your vocabulary and n is the sequence length as a subset of the whole distribution of possible text. Now you pass that through something with sufficient capacity, which compresses the input distribution (you loose bits), and then recreate from your lossy distribution. AutoEncoders are not trained this way because they simply don’t converge if you try to pass smaller input to produce larger embeddings, but they’re doing the same thing. Another way is to think of it is that LLM layers act like binary classifiers via soft max for the input distribution (there were a slew of papers comparing LLMs to SVMs). Essentially during your training you decide what part of the input distribution is going to be part of the embedding. In my humble opinion, while the training diverges, these models produce the same manifold induced by the original distribution.

New comment by usernametaken29 in "Is One Layer Enough? A Single Transformer Layer Matches Full-Parameter RL Train"

usernametaken29 — Fri, 03 Jul 2026 05:33:16 +0000

The problem really is one has to find out what constitutes the embedding and output if you want to apply f(manifold) but as others have pointed out that’s out of scope for this paper. My only insight here is that it is not surprising at all that one layer, or even a single function pass, suffices to get the desired reinforcement. I don’t actually know if there’s any research that quantifies the “manifold threshold” of each layer to try out this approach but it would be interesting for sure

New comment by usernametaken29 in "Is One Layer Enough? A Single Transformer Layer Matches Full-Parameter RL Train"

usernametaken29 — Thu, 02 Jul 2026 13:53:50 +0000

Now that’s interesting.. what exactly distinguishes latent representations and the manifold? IMHO, those are the same, and you’re constructing a piecewise function of the manifold itself. Decoders also produce manifolds much in the same way, with the distinction being that the encoder isn’t learned but static after initialisation. So fundamentally it is still DOING the same operation.

New comment by usernametaken29 in "Is One Layer Enough? A Single Transformer Layer Matches Full-Parameter RL Train"

usernametaken29 — Thu, 02 Jul 2026 13:33:13 +0000

Kind of. Autoencoders don’t need to have an embedding that’s smaller than the input. Their only requirement is that they compress information and thus create reconstruction loss. Typically however they are not trained this way because they don’t converge.. transformers do the same thing, but they can squeeze much more bits of information through one pass because the way they are designed. This holds true even for decoder only networks because they’re still doing the same thing

New comment by usernametaken29 in "Is One Layer Enough? A Single Transformer Layer Matches Full-Parameter RL Train"

usernametaken29 — Thu, 02 Jul 2026 13:10:38 +0000

If you think about it for some time then you’ll come to realise transformers are autoencoders on steroids. A small input space is expanded onto a big manifold and contracted again. Now, suppose you want to impose a function to regulate the output of an autoencoder. It’s actually pretty obvious that you need exactly one layer to do so… f(manifold).

New comment by usernametaken29 in "Why jet engines aren't made in China"

usernametaken29 — Thu, 02 Jul 2026 13:03:27 +0000

I agree partially with this. America is a special circumstance because of the way the industrial complex and government contracts are linked, and there’s a lot of blatant corruption. If you look at the European regulation framework I believe it works rather well. You have multiple big companies delivering components for different parts of the airplanes from different countries which are all individually big but not too big to fail because no government is solely reliant on them.. as it should be

New comment by usernametaken29 in "Why jet engines aren't made in China"

usernametaken29 — Thu, 02 Jul 2026 00:55:29 +0000

Honestly this is a good thing. I can endure buggy software but I don’t want to deal with buggy planes. Regulatory pressure is a market force and a useful one too. There’s a huge difference between ship fast and ship right - the latter one requires deep pockets and willingness to commit to ongoing risk. People always say big Pharma and aviation and such are oligopolies, and that’s bad, but they rarely see the capital intensiveness of the whole process. Some things are slow and deliberate and restricted to big corporate only for good reasons