Hacker News Reader

tveitaon RIP, vector database

That's just a search engine, but then you're competing with traditional players like Elasticsearch and Vespa who all have built-in vector support by now, and you have to compete on attributes like price, performance, features, and who can mention 'AI' the most times on their web page.

tveitaon 37,500 border drawings: a map of the world as people remember it

I was wondering the same thing, but playing it you only fill in small border segments at a time, with the rest of the map filled in, so the projection doesn't change much.

tveitaon Making a Python interpreter in 1024 bytes

As they say in TDD, write a test, then write the simplest code that will make it pass.

Clearly supporting multiple functions starting with 'p' would be overengineering.

tveitaon What is Nueralese and Why is it Bad

The recurrent depth sounds a lot more like what is described in https://dnhkng.github.io/posts/rys/ - a way to add depth to a network without increasing the number of parameters.

e: While the actual CoT in neuralese paper is Facebook's Coconut https://arxiv.org/abs/2412.06769 - not sure if any production models use that one.

tveitaon Why are there no flow batteries with symmetric ferrocyanide electrolytes?

You might find the home made batteries of this guy interesting as well https://www.youtube.com/watch?v=eq7fR9ISuCw

tveitaon Qwen3.8-Flash-Next

It should be fine keeping the n-gram embeddings on SSD which lets you run at least the Q1 and Q2 models on 64GB

tveitaon Expert Witness to ChatGPT: "Show how 3M is 0 percent at fault"

Interesting that he pointed it to the CSB video - those are great explanations for us laymen, but I assume the LLM would prefer a written report.

Luckily it's still up https://youtube.com/watch?v=CFVUSDzHL8A

I was worried the videos would already be disappearing given Trump's attempts to shut it down https://www.chemicalprocessing.com/safety-security/risk-asse...

tveitaon AMD acquires Taalas to boost inference performance by etching models in silicon

> Not sure if modern models "think" only by outputting <thinking> blocks

That's pretty much it - a small refinement to "Chain of Thought" prompting, where you tell the model explicitly in the prompt to "Think step by step" or similar, so it writes out more steps before giving a final answer, potentially catching some errors. The "thinking" models are tuned to do that without being prompted to, and to output the "thinking" markers around it, so they can be hidden from the user.

tveitaon 2027 memory capacity is reportedly sold out

https://dam.stanford.edu/memory-prices.html

Pretty sure those devs were saying the same thing way before 2017 as well, which seems to be ~ the last time RAM was this expensive.

Now RAM in the cloud, now you're really paying the Java premium.

tveitaon NSA tries to weaken mlkem standardisation?

> Most cryptographers would still recommend the hybrid over pure ML-KEM. This RFC (for pure MLKEM) is marked "recommended to implement = N". It is purely for settings where the implementors independently want to use pure ML-KEM for some reason.

That's exactly how it was with Dual_EC_DRBG.

E.g. https://www.schneier.com/essays/archives/2007/11/did_nsa_put...

   I don’t understand why the NSA was so insistent about including Dual_EC_DRBG in the standard. It makes no sense as a trap door: It’s public, and rather obvious. It makes no sense from an engineering perspective: It’s too slow for anyone to willingly use it. And it makes no sense from a backwards-compatibility perspective: Swapping one random-number generator for another is easy.
  
  My recommendation, if you’re in need of a random-number generator, is not to use Dual_EC_DRBG under any circumstances.

So most cryptographers _recommended_ staying the hell away from Dual_EC_DRBG. But hey, harmless, no one serious about security would actually use it right?

Except as we know now, after the standardization NSA was able to persuade/bribe vendors to implement it.

RSA is still a viable cryptography vendor, after accepting money to backdoor their product for paying customers. The standardization gave them a fig leaf of plausible deniability. Honest mistake, could happen to anyone, right? If they had needed to implement a "non-standard" backdoor, or if it had been officially struck from the standard, it would have been a lot harder to row away from.

It also gives you the option of serialising the RNG states directly instead of using the counter hack.

tveitaon Google's Antigravity bait and switch

> who put the Transformer into LLMs?

Google?

> who invented neural networks

People like Geoffrey Hinton, who was notably at Google Brain from 2013 to 2023?

The people who say Google was ahead were paying attention long before you were.

tveitaon Your Most Improbable Life

I think https://jamesclear.com/great-speeches/finding-your-own-visio... is a better take.

What gives you an unique perspective and your own voice can be sticking to your thing for a long time and exploring your exact path more deeply than anyone else has. You don't need to take a million random stabs to become "improbable", and there's no reason that should lead to anything authentic.

tveitaon Google’s AI is being manipulated. The search giant is quietly fighting back

Would love to read specific examples of "the same trick being used to dismiss health concerns about medical supplements or influence financial information provided by Google's AI about retirement", but the relevant link in the article currently goes to

file:///Users/GermaTW1/BBC%20Dropbox/Thomas%20Germain/A%20Downloads%20and%20Documents/2026/And%20there's%20evidence%20that%20AI%20tools%20are%20being%20manipulated%20on%20a%20wide%20scale.

tveitaon Google changes its search box

I can get a really high hit rate by only searching for dumb trivial things I already know the answer to.

tveitaon Google changes its search box

The point of LMGTFY is to land people on either the official documentation or a curated site like Stack Overflow. Google used to be able to do that reliably.

With the power of LLMs you can Google a standard library function and get an inaccurate summarisation of a Reddit discussion where neither side knows what they're talking about

tveitaon Mojo 1.0 Beta

Is there any project that showcases Mojo for running neural network models on the GPU - like ideally something like llama.cpp that could run one or more existing models to showcase the readability and performance?

tveitaon €54k spike in 13h from unrestricted Firebase browser key accessing Gemini APIs

"Can you lay tiles until I say stop, or until it's about $250 worth, whichever comes first"

"No, as one of the top tile layers in the country I can't do that, for your own protection. What if fifty elephants came and wanted to use your bathroom all at once? You'd feel pretty dumb having to reject them instead of me simply automatically adding $1 million to your bill"

tveitaon YouTube users get option to set their Shorts time limit to zero minutes

Honestly I'd sometimes like a checkbox to ignore any results uploaded after, say 2023. Or else to only see 'verified non-AI' content. There is nothing I'd ever search Youtube for where an AI generated video would be an acceptable answer.

tveitaon Pokemon Evolution vs Darwinian Evolution

There are real examples of males and females being initially mistaken for different species, as well as for adults and juvenile forms.

e.g. https://en.wikipedia.org/wiki/Cetomimidae

  In early 2009, the Royal Society published an article detailing the discovery "that three families with greatly differing morphologies, Mirapinnidae (tapetails), Megalomycteridae (bignose fishes), and Cetomimidae (whalefishes), are larvae, males, and females, respectively, of a single-family, Cetomimidae."

tveitaon Gas Town: From Clown Show to v1.0

Building harnesses does seem like a task that's particularly conductive to psychosis. I've wondered if it's because there is no push-back; almost anything you try will be "right" in that your changes appear to make things happen. "Oh look I added a carpenter and now it's walking around and making notes about the scaffolding" So you get to stay in your flow state without reassessing if the concepts you are forming are ultimately meaningful.

Although I think the post also self-diagnoses some factors that also help:

  With the Gas Town Mayor, you feel like you’re operating at a special level, a VIP, above all the workers. You are talking to someone important: the mayor of a factory the size of a town. You have access to someone with resources, someone who gets you, someone who appreciates how busy you are.

  Working with regular coding agents just doesn’t give you that special feeling.

tveitaon Is it a pint?

https://en.wikipedia.org/wiki/Fill_line

Selling drinks in mislabeled containers should warrant a fraud report to your local consumer protection agency. A crowdsourcing app seems like the wrong tool here.

tveitaon Apply video compression on KV cache to 10,000x less error at Q4 quant

"video compression" by analogy only, what this claims to actually do is delta encode the values in each token from the previous token.

Interesting idea, but the results seem almost suspicious? even accounting for the extra bits used to store the 16-bit start value for each block - ~5% for k=64

The code does funky things, like the encoder updates the reference value for each encoded token, using the non-quantized value! [1] But the decoder just ignored all that. [2] how can this work?

[1] https://github.com/cenconq25/delta-compress-llm/commit/f185f...

[2] https://github.com/cenconq25/delta-compress-llm/commit/f185f...

tveitaon I was interviewed by an AI bot for a job

Sure, for instance, if all of them go through an 1 hour AI interview, then you might find a better candidate, at the cost of 1000 man-hours of work. You hire that person, another company opens a position, gets 999 applicants, send them all their own AI interview, and so forth.

How much would better would your hire be considering that you managed to check all 1000 of them, rather than just 50?

Assume that candidate fitness is a number normally distributed around 0 (half of them obviously being negative), that both you and the AI can perfectly pick out the best candidate, and that you picked the 50 to interview completely at random. The average actually seems to be around 40% better? Suprisingly decent. Is that improvement worth 1000 man-hours?

So attempt two here: maybe instead of each company sending candidates through an interview, there should be a common gatekeeper. All working age people take the same 1-hour AI interview, and the glorious overseer assigns them to the position they are best suited for.

(An actual answer here is you assess how important it is to get "the best candidate", and you interview enough people to get a reasonable approximation. The hour cost on your side is what keeps you honest. If wasting candidate time is free on your side, you're going to waste 500 man-hours of work for a 5% better result for you.)

tveitaon I was interviewed by an AI bot for a job

I absolutely agree in principle, but I understand that the companies are also seeing a lot more applicants trying to skate past screening and interviews with AI assistance.

Connecting verified humans for a mutually respectful chat is a trust problem that companies like LinkedIn should be creating solutions for, instead of offering both sides automated shovels to shovel slop faster.

tveitaon I was interviewed by an AI bot for a job

> They are the ones who started using AI in the hiring process

Aren't you ignoring the reports of companies receiving thousands of ChatGPT-written resumes, bots sending applications, and interviews with applicants being live coached by AI?

This is a breakdown of trust on both sides.

tveitaon Many SWE-bench-Passing PRs would not be merged

Probably more like the long tail of software - software that was created for a particular purpose in a particular domain by a single person in the company who also happened to know programming - maybe just as Excel macros.

I strongly assume the long tail is shifting and expanding now and will eventually mostly be software for one-off purposes authored by people who don't know how to code, and probably have a poor understanding of how it actually works.

tveitaon Lazy JWT Key Rotation in .NET: Redis-Powered JWKS That Just Works

There's some odd choices here.

  - 90 days is a very long time to keep keys, I'd expect rotation maybe between 10 minutes and a day? I don't see any justification for this in the article.
  - There's no need to keep any private keys except the current signing key and maybe an upcoming key. Old keys should be deleted on rotation, not just left to eventually expire.
  - https://github.com/aaroncpina/Aaron.Pina.Blog.Article.08/blob/776e3b365d177ed3b779242181f0045cd6387b3f/Aaron.Pina.Blog.Article.08.Server/Program.cs#L70-L77 - You're not allowed to get a new token if you have a a token already? That's unworkable - what if you want to log in on a new device? Or what if the client fails to receive the token request after the server sends it, the classic snag with use-only-once tokens?
  - A fun thing about setting an expiry on the keys is that it makes them eligible for eviction with Redis' standard volatile-lru policy. You can configure this, but it would make me nervous.

tveitaon Agentic Engineering Patterns

I've definitely seen Opus go to town when asked to test a fairly simple builder. Possibly it inferred something about testing the "contract", and went on to test such properties as

  - none of the "final" fields have changed after calling each method
  - these two immutable objects we just confirmed differ on a property are not the same object

In addition to multiple tests with essentially identical code, multiple test classes with largely duplicated tests etc.

tveitaon Weave – A language aware merge algorithm based on entities

> Elijah Newren, who wrote git's merge-ort (the default merge strategy), reviewed weave and said language-aware content merging is the right approach, that he's been asked about it enough times to be certain there's demand, and that our fallback-to-line-level strategy for unsupported languages is "a very reasonable way to tackle the problem." Taylor Blau from the Git team said he's "really impressed" and connected us with Elijah. The creator of libgit2 starred the repo. Martin von Zweigbergk (creator of jj) has also been excited about the direction.

Are any of these statements public, or is this all private communication?

> We are also working with GitButler team to integrate it as a research feature.

Referring to this discussion, I assume: https://github.com/gitbutlerapp/gitbutler/discussions/12274