Hacker News Reader

usernametaken29on EU chief opens door for Canada to become 'associate member'

> The EU needs to be truly Federal to succeed economically. That might create uniform regulations with more uniform regulations and actually make Europe competitive.

But also: > The European Union has a nominal GDP of roughly $20 trillion, placing it closely on par with or slightly ahead of China ($19 trillion), but smaller than the United States ($30–$31 trillion)

I don't know about you but the same GDP as China sounds economically successful... or not?

usernametaken29on We must pace the frontier

> Without alignment, further improvements in capability turn LLMs into wanton felony generators

Honestly, I don’t think that’s bad at all. I hope OpenAI and Antrophic keep RL training runs up that randomly fuck with a lot of people. Until the day the DOJ comes knocking, locks those idiots up in jail and closes them both down for the insane lack of responsibility and carelessness they’ve shown. Sounds like the IDEAL outcome. Finally some jail time for all the fraud, negligence, outright scamming, hype inflation etc. if anything can accelerate this, oi, be my guest. Amodei might be afraid because he knows if he keeps pulling the stunts for investment theatre, at some point they’ll actually face consequences. AWESOME. That’s what we want right there

usernametaken29on Growing proof that autonomous cars save lives

Honestly, mount a front and back dash camera, and allow people to submit tickets for speeding and tailgating themselves.

I’m not kidding excessive speed and not keeping a safe distance equates for some 50% of all deadly traffic incidents. It’s because a small and vocal minority drives like absolute fucking assholes and they will rally lawmakers not to introduce speed limits and to loosen grip on enforcement. The ADAC (German motor association) is a perfect example. When they announced after decades that they support a 130km max speed limit members jumped by the doves.

The only vastly meaningful improvement I can see in FSD is faster reaction times when it comes to assholes. Coincidentally those are the people who insist on driving petrol and will never upgrade to FSD because they like to drive like assholes.

IMHO it’s time for strict enforcement. Let’s ban them from the road finally. Set fines in percent from income. Allow self reporting via dash cams. Add stricter penalties for recurring offences, if you’re caught speeding 5 times in 2 days your licence is suspended.

That’ll have a much bigger impact than FSD. Maybe we can rally our representatives making them all giddy with all the extra cash they can earn from these offenders?

usernametaken29on How I advertise malicious software on Google Ads

Oh it does but just because you win doesn’t mean they start caring. A bailiff won’t really get involved for 500 euro compensation. A debt collection agency might. They want a cut of your compensation to deal with RyanAir because they know how annoying they are. Like, sure, you can start a 10 month project over 500 euros and maybe it’ll be fruitful eventually. But who does that? It’s insane. Answer your emails for fuck sake

usernametaken29on How I advertise malicious software on Google Ads

It’s a strategy. Ask RyanAir. They systematically shut down any way to communicate with them, except lawsuits. Tried their chatbot, hotline, email, letter, fax, small claims court (I’m not joking they don’t even care about that). Unless you turn up with a court order and confiscate a plane, they simply won’t give a fuck, and I’m sure it’s a tactic to deter 99% of their problems before they start

usernametaken29on Why do machines learn, at all?

Author: This is a condensed version of a discussion taken from my thesis that I had wanted to write down on the blog for some time. I hope it is a somewhat interesting read! Cheers

usernametaken29on How I feel about AI

Now if only humans would know how to survive without the internet. Oh wait, we did that. For a couple hundred thousand years. Yeah. It’s really insane to listen to some of those forecasts. Humanity will be wiped out by 2030. Sure. Even if AI manages to create a super effective bio weapon the likelihood that there is a part of the civilisation that’s immune is really, really, really high if not a given. Those people WILL pull the plug if push comes to shove. Maybe humanity will be put back a couple thousand years, that’s possible, but it’s pretty ignorant to think the thinking boxes will kill every single human alive

usernametaken29on Ask HN: Would you read a statistics textbook?

I feel this boils down my learning journey as well. You start unraveling a very good intuition about the underlying concepts MUCH much later, but partly because those intuitions themselves are never conveyed and are supposed to be learned from the proofs, and are an indirect product of learning.

usernametaken29on The revolt of the reader

For me it’s AI videos or music / narration that is beyond off putting. What’s worse now it seems people are writing their YouTube scripts with Claude et al. so at times even if it is a human creator you can clearly and immediately tell the words are not their own. To those creators I have but one message: IT SUCKS. I’d rather have you ramble incoherently in your mic then reading an LLM script and I will remove you from my feed immediately. I concur with the author on all accounts. We all can tell the BS people are selling us, unoriginal ideas, shallow concepts, open ended questions that hint at exactly nothing. Don’t be an LLM echo

usernametaken29on Authorization terminology is a mess: Let's fix it

Which is precisely, a standard

usernametaken29on Authorization terminology is a mess: Let's fix it

https://xkcd.com/927

Nice work and all regardless

usernametaken29on Path to Astra: critical capabilities and frontier safeguards

> we believe our production safeguards at the time would have prevented the Hugging Face incident. We have since implemented even stronger safeguards for Astra, including training the model to more reliably refuse harmful cyber requests and respect safety restrictions

OpenAI is fucking nuts. “Hey model you were bad last time please don’t do it again please please”.

Disconnect your training cluster from the internet for good. Physically pull the plug and only let scientists fire off experiments in the building. That’s an easy way to achieve 100% hacking protection. But I bet you that hasn’t happened and their weak sandbox will fall again…

usernametaken29on Reverse engineering my ADHD test

I concur that many of the diagnostics methods are small sample sized for psychological tests. However I believe we have a lot more neurological evidence for underlying conditions and mechanisms in some. Schizophrenia for example has clear neurological hypothesis relating to agency and we can clearly tell these patients in a standard MRI. The same cannot be said for ADHD or ASD

usernametaken29on Reverse engineering my ADHD test

I’m going to say that as you perform more clinical tests typically your chances of a wrong diagnosis will decrease proportionally. However particularly in the case of ADHD tests, if you know the test questions and the symptom description I think you could cheat your way to a diagnosis. If you went to all of these tests in earnest, without looking them up, just providing spontaneous responses, and all of them come back positive, I’d say chances are you have ADHD.

usernametaken29on Reverse engineering my ADHD test

If I remember correctly there are some paradigms in which you can show it in fMRI quite clearly but the trouble is of course MRIs are not accessible to a wide public. I will have to go dig through papers though, I have not read into MRI based diagnostics in a long while

usernametaken29on I trained a small transformer in 1.5hrs and it beats many LLMs

I think you’re asking the right questions, sample inefficiency is horrible in modern LLMs. Despite this, I saw your analysis:

> The biggest increases in scores were due to

Modern architecture (SwiGlu instead of GELU, RMSnorm not layernorm, etc.) More data diversity, better shuffling of data scaling up: 8 layers instead of 4

This is commonly called squeezing the lemon and is usually a bit of a last resort. You should be able to achieve near SoTa with your new method, before you squeeze any lemons. This is, because the old SoTa is typically not using new optimisers and thus your results will be distorted by a large margin.

In terms of sample efficiency I want to add two things:

Runtime per-puzzle fine tuning is a very good target that provides a LOT of information. People have not looked at evolutionary methods to harness induction since the 90ies - if I was to work on ARC ever again I’m fairly certain this is where I’d look.

Best of luck, padawan

usernametaken29on Kids These Days

Absolute case in point, in Germany drinking age is 16, but if you’re drunk and goofy and on social media rest assured you’ll be an outcast for a long while

usernametaken29on Reverse engineering my ADHD test

That’s my bad for conflating the two. Indeed they are not the same in ICT-11 but! diagnostics is moving towards the direction of treating them the same. ADHD and ASD have extremely high comorbidity and it has been proposed many times to classify them together because the symptoms overlap (difficulties at school, aggressive behaviour or temper, lack of social skills, etc). In fact it has become so difficult to find distinctive features of ADHD both in diagnostics AND treatment (both are usually treated with cognitive behavioural therapy and in more severe cases medication) that it has been proposed numerous times to drop their distinction in the ICT, but this hasn’t happened yet. Good paper to the point:

https://pmc.ncbi.nlm.nih.gov/articles/PMC4654237/

usernametaken29on Reverse engineering my ADHD test

I spent a decent amount working on ADHD diagnostics and unfortunately it sucks. To my knowledge, to this day, there isn’t a single reliable biomarker that identifies autism spectrum disorders. It’s usually correlation studies paired with questionnaires which for lack of better words can be described as “incidental” at best. Most z scored tests I know are also developed with small samples, a hundred participants with an outgroup of 20 would already be a LARGE sample. The standard ADHD questionnaire was validated on some 35 people.

There might be some hope in deep learning, particularly in observing eye movement components to deliver reliable biomarkers in the future, but I’m not so confident any time soon.

It’s dreadful and fascinating at the same time that defining and testing for ADS has been so elusive for so long…

usernametaken29on Kids These Days

I think one aspect that wasn’t discussed in this article sufficiently is:

> given that all of their peers have adopted it, there is a serious social cost to opting out.

Unlike reading a book, or fishing, or whatever, there’s no peer pressure to pick a specific hobby. There is significant pressure to use social media. I’m usually on the authors side, the view that the youth is rotten is literally as old as time (go check out Aristotle), but unlike a claimed moral decay, we are actively forcing kids to partake.

In Germany kids are drinking less than ever before. The education ministry went out and surveyed the kids: they don’t want to drink or party because they might embarrass themselves and it’ll be recorded and on social media and as such it might haunt them for the rest of their teen years. Albeit the outcome is good the enforcement mechanism is straight out of a 1984 horror novel kind of shit. As a teen you need to be able to experiment in a safe space somewhere somehow.

usernametaken29on The Rise and Fall of Agent Civilizations

I was initially creeped out by this but studying up it seems METR is heavily involved in AI2027. I’ll remind you:

“AI has started to take jobs, but has also created new ones. The stock market has gone up 30% in 2026, led by OpenBrain, Nvidia, and whichever companies have most successfully integrated AI assistants.”

It’s almost Q3 and xAI has seen one of the biggest wipeouts in trading history. Likewise, Antrophic and OpenAI have again delayed their IPOs under internal concerns of busting their stocks. So no, we’re not seeing any economic leadership here.

If anything people are increasingly trying to cut AI budgets and I wouldn’t know of anyone outside of OpenAI who has the audacity to run millions and millions worth of token compute for an eval run with no ROI (and probably no demand, because cheap/flash models).

As much as I like the cautionary tale and I’m sure we need to take it seriously, AI is not progressing as fast as projected by these experts.

usernametaken29on RAG Is Simpler Than You Think

I am a linguist and can say your view is too simplistic. This works only for things that are common enough to be mapped onto the same vector from the training set. Try searching for my company XY and it won’t work, because your embedding doesn’t have an adequate embedding, so you will end up having to build a translation table anyways..

usernametaken29on RAG Is Simpler Than You Think

> embed everything the first time

This assumes your text is small. Try embedding pdf reports - though luck. It surely won’t fit into most embeddings. I can think of many more examples: books, news articles, medical reports, insurance claims etc. they’re all too big to “index it all at once”

usernametaken29on RAG Is Simpler Than You Think

I worked on large scale RAG systems before and can say people vastly underestimate full text search and vastly overestimate embeddings. FTS is really easy, portable and scalable and gets you very far, the 80/20 rule applies. Embeddings appear to be nice and magic but when you really get into them you notice: semantic similarity isn’t as good as you think and certainly it won’t make everyone happy. You will inevitably end up having to re-embed more or different chunks of your text to accommodate more and more precise embedding search - at which point you’ll go the last mile and do reranking etc etc all the while having to support the operational burden of vector search. Then you turn around and build a search query with 500 keywords and sure it’s painful but it just works, accommodates all use cases, scales and is overall less annoying to maintain.

usernametaken29on Where Human Sleep Went Wrong

> untangle what he calls the sleep paradox: Sleep is critical to human functioning, and yet, we sleep fewer total hours than any other ape

Instead of going to Tanzania you could’ve asked a sleep scientist. Sleep length has no known correlation to brain size and is known for drastic within family variation. What’s more sleep has a high within species variation too, some sleepers can survive 4 hours, others need 8. Collectively we used to sleep 9 hours in the 20th century, but it appears shorter sleep is fine too. TLDR: no one has any idea why sleep length varies and he won’t find his answers in antrophological studies in Tanzania, though luck, buddy

usernametaken29on Apple announces changes for apps in the European Union

> Under the updated terms, developers can now offer Apple In-App Purchase alongside alternative payment options, which had not previously been permitted in the EU.

Now hold on a minute here. It’s Apple’s own App Store requirement that forces any developer regardless of payment option to pay a commission or be delisted. Framing it as an “we charge you because of the EU” is a plain simple lie. The EU forced their hand in allowing different payment channels at low commission rates with no repercussions.

usernametaken29on Cultivating a state of mind where new ideas are born (2023)

I’m going to sound rude here but I think the debate around solitudal achievements suffers a great deal from bias. For one, survivor bias favours the ideas emerging from deep thinking work. We overestimate however how much of that is the deep thinking vs quantity tradeoff. Einstein had many, many, many ideas and the majority of them were pretty bad if not outright outrageous. If you have the time and environment to produce a million ideas and pursue them, your likelihood of a good idea goes up. A second issue is that ideas almost never come up in isolation. The countless women behind the men erased from history tell exactly this story. Nowadays we fight the same fights about attribution. I should write a detailed account on why the arguments named in the article are not supported by evidence from cognitive research, but I for one am convinced a conductive environment is 1000% the ingredient necessary to come up with revolutionary work, rather than some magical great man thinking fantasy.

usernametaken29on I close SSH port 22 (and what I use instead)

And? Adding a new IP takes roughly 15 seconds in the admin panel. Can be automated away via an API… not sure I’d let the “my IP changes” argument slide in 2026

usernametaken29on I close SSH port 22 (and what I use instead)

I don’t know why this wasn’t mentioned before but why not use a Firewall. If you’re using a virtual box like Hetzner or Scaleway you can specify an ip or range at the router level. For all intents and purposes this removes public exposure. Scaleway also has a cheap VPN bridge. So you never need to connect via the public internet if you don’t want to… hardly gets more secure than that

usernametaken29on Compression is prediction

I think Hutter would vehemently disagree with you on that one ;)