> “I give my wife a lot of credit. She was, like, you should never apologize,” Langley said. “Sure, people make mistakes. I've made a lot of mistakes in the last nine years. I made a lot of mistakes this summer. But we do good work.”

So this guy admits he did things that he shouldn't have then contends that he shouldn't have to say sorry for them? Am I reading this right? Isn't the point of an apology to acknowledge the bad outcome and signal to aggrieved parties that the offender will try avoid repeating the negative behavior? If yes and this guy is refusing to do it then what does that say about his world view?

I feel there's a misconception in a lot of the comment threads about what an LLM-from-scratch tutorial is for. Of course you're not going to come out of this kind of tutorial being able to train a useful LLM from scratch. (It requires too much compute, and datasets you don't have access to, and there are a lot of tricks and operational problems to solve for real training runs.) If that's your bar for "impracticality", then sure, this is impractical.

But that's not the point; the point is understanding these things better by seeing how they are made (or grown, if you prefer). Building something at a level further down the stack than you will ever go forces you to confront details that shape the layers above that you do deal with. This is one reason why Crafting Interpreters is such a beloved book here on HN.

LLMs are now at the core of tools we all use everyday... don't you want to know how they are made? Maybe not everyone has this itch to understand the platforms and tools they use, but I sure do, and it's helped me over and over again. Sure, LLMs aren't like operating systems or compilers. They're a quasi-organic mass of billions of parameters learned during training and honestly nobody (not even the mech interp folks) really deeply understands why they work so well. But that's no reason to completely give up and there is still a lot you can learn.

Take RLVR ("Reinforcement Learning from Verifiable Rewards"). This is the training technique that made LLMs get so good at programming and math in 2025 and 2026. A lot of people who confidently talk about LLM capabilities either haven't heard of it (and thus end up parroting out-of-date factoids about LLM training derived from the pre-2025 era) or have only a very dim idea of what it is. In this course you actually do RL training of an LLM to get it solving math problems better. You're right down in the details. It's interesting for its own sake but it will also give you a grounding for discussing questions such as: "why did LLM writing get so weird and bad in mid-2026?" and "why is it so hard to train LLMs to be good at things that aren't like math and programming?"

Anyway, Sebastian's two books are great and highly recommended; I worked through the first one a year ago and it was one of the best investments of my time in all of 2025. If you prefer video for learning this is great as well (although I'd still say - try the books first).

All the spam this generates probably sucks for 404 Media, but the underlying problem is that ChatGPT is right. If you have any kind of problem with these overgrown, understaffed companies, your only chance is to get media attention.

Apple famously put "if you run to the press and trash us, it never helps" into their App Store Guidelines for developers, but the only reason they did have to do that is because it's a lie. Running to the press is the only thing that helps, which is why people do it.

"anti-ai"

> https://github.com/DilawarShafiq/nietzschian-debugger#setup

Step one to setup: "Set your Anthropic API key"...

I cringe with AI "colonialism".

The article states three reasons for doing so, but all of them are very cringe baity words.

Low paid workers are due to local economy government doing jack to improve their wages in the first place. In China, competition is high, thus low pay (with other reasons such as real estate crashes, etc)

Visualization makes sense but does not mean it's colonialism as colonialism requires forced intervention, and there is no such play here.

Cool project bad political motivated wordings in the post.

This is exciting news, and I would like to say "hello" to Richard Kenneth Eng, the sole user of Pharo in the world.

maybe peta can offer themselves as human test subjects to study disease and medicine etc., and their kids. It might sound harsh but i really do not know how they envision people developing this type of thing without such tests.

ofcourse its not great, but afaik we still can't do much better. Perhaps if universities and their medical centers got proper budget they could grow humans and human parts to test on. theres many things like artificial wombs and 3d printing of tissue etc. but afaik these will be hugely expensive and way outside their budget.

Or does peta prefer only big pharma to be able to do research and sell simple medicines for ludicrious prices (as we've all seen them do already).

Wonder what other people's take on it is.

i also wonder why they dont fight for human rights instead of animal rights. i mean, sure animals are essential, and sentient, and know what is suffering and incarceration, but so do many humans around the globe, even in 'developed' countries.

maybe they should put them in a childrens' hospital full of kids with incurable diseases and ailments -_-. i've been there and seen _that_ and its a ton more depressing and horrible and a few monkeys suffering (which is still horrible but magnitudes less i can tell u).

try looking one of those kids in the eyes and tell them you are advocating for something that will prevent their chance of a cure how slim it might be. im quite sure u need to be pretty unempathic to be able to do that to say it lightly....

this kind of advocating causes litterally violence to some of these researchers... completely bonkers.

Ctrl+F "ProMotion" 0 results

non-starter

Ctrl+F "ProMotion" 0 results

non-starter

What pediatricians and many others actually say:

0-3 yrs old: no screens at all. No TV playing when they are in the room, even if they can only hear the sound. Refrain from using your phone too long when they are in the room with you, they need to see your face react to them to learn human interaction.

3-6: no screens used alone. No internet. Interactive content with a parent is ok, not regularly, not in the bedroom.

6-9: may discover safe internet spaces, with help from parents.

9-12: can learn to manage their screen time, protect their private life, develop critical thinking.

12+: no screens at night, talk to your kid about what they do online, more autonomy.

This is what we follow (refraining from using the phone ourselves is prolly the hardest to do), they have never watched cartoons alone, they love old point'n'click games on the PC with me, we share the mouse, and I thought they would ask for more and more screen time, they don't. They actually want to read books with parents, paint, draw, build, go outside, ride a bicycle, invite friends. Maybe we're lucky. They don't throw tantrums in their seat, they sometimes do while grocery shopping, it's difficult, but I would never use a screen to stop it.

So, I don't see how this kind of device can be useful and safe at any age. It seems designed to cause the infamous "iPad kid" syndrome or get them hooked on Amazon products even younger than before.

But there's even worse in stores today: AI assistant integrated plushies.

I believe that we'll see a lot of "x is good for your brain" news (with X just being any activity that doesn't involve a screen), now that attention spans are so short and everyone feels like their brain isn't the same anymore.

The same way we got gyms when people realized physical inactivity is super unhealthy.

Interestingly enough the “before it was cheap” version was distributed for free when this one is for sale for half the price of the game this is supposed to run on.

Why do people try to vibecode PDF engines that already exist in Rust? https://github.com/yfedoseev/pdf_oxide which this project builds on completely misses https://github.com/fschutt/printpdf which was already developed with tons of fixes over years (both with and without AI). MIT license and also builds on WASM with a demo running on https://fschutt.github.io/printpdf/

One Google search (yeah I'm old, sorry, nobody "Googles" anymore) for "rust pdf library" shows this. Yet people now redo and redo and redo everything again just to have "their" semi-broken copy of a HTML -> PDF renderer (and then run into the same edge cases that were already fixed before). There are lots of edge cases in typography especially, I recently had a lot of fixes as I was trying to ocr-and-cleanup a 1732 scan of some books with Claude auto-extracting the font. There were a ton of edge cases in the layout, esp. related to column, footnote and sidebar handling, or things like "Hebrew RTL text accidentally disappears if text breaks weirdly", "1-bit image compression", etc.

I have nothing against AI, I use it heavily myself, but what is annoying is people completely ignoring the years of work of existing projects that do the exact same thing - instead of just forking, but instead re-tokenizing the exact same work again. Most likely the author didn't research "does this already exist" and that's annoying, then you get 100 of similar projects reinventing the world with tokens again and again. Sometimes its beneficial and can lead to faster exploration of new approaches, but in most cases its not and misses edge cases that were solved before.

I'm stunned that "mapping out" blocks of bad RAM is a thing you can do. It reminds me of the old days when hard disks had to be "low-level formatted" first (ask your parents, kids) and it would ask you for a list of bad sectors to map out. Sometimes hard disks would even come with a list of known bad ones printed on the top.

I don't think "judge writing by its quality instead of its source" is coherent. Certainly, the value of reading is in the thoughts it produces in the reader's mind. So if you want me to ignore the source, you're asking me to value the thoughts I get from the writing without regard for the thoughts that produced it. This seems dubious, if not actually oxymoronic. As a reader, it feels much like you telling me to eat something you made, which you enthusiastically assure me is high quality, but which you can't be bothered to eat yourself, and in fact you've never eaten anything like it before. Plus it smells like spray-cheese.

Gaming used to be done in arcades for 25c a play*.

Over time NeoGeo realized they could sell you the hardware to play at home (“on-prem”) at arcade quality albeit at a price out of reach to most people. Alongside them, Nintendo and Atari gave you 80% of the fun at 20% of the cost. History will view games arcades as a flash in the pan compared to gaming at home.

Namco, Capcom, and Konami went into selling software instead of hardware but it feels hard to imagine buying a home copy of Fable** for $X,000 any time soon.

⸻

* That’s for video games. In fairground arcades you’d literally pay for tokens then use tokens for play. The analogy almost writes itself.

** The model, not the video game.

> "son of a failed coup organizer" is an absolute false statement

But, his dad:

> Former Brazilian President Jair Bolsonaro is serving a 27-year prison sentence after being convicted by Brazil's Supreme Court for plotting a coup following his 2022 election defeat

I don’t see how to read it any other way. Explain?

One major element in this is the rise of earnings from pass-through income. Self-employed or nearly self-employed individuals who would have been counted as employees of their small businesses 30+ years ago increasingly take "profits" from their business rather than "wages," even though they are the same thing. The "profits" get tax advantages. This makes labor share decline despite dentists, consultants, etc. still doing the same amount of work and taking home the same pretax income. See: https://www.ericzwick.com/labshare/labshare.pdf

I feel like the average tech enthusiast just assumes that most people have computers, but they do not. Most people have a phone as a primary computing device. Tablets have gone up in popularity, but for a significant portion of people, their only device is a phone. Those people are buying phones on carrier financing or other deals, so they don't even see the full cost of their device at all. So for people doing this type of upgrade cycle, a folding phone is quite compelling: an expanded screen size for watching video, consuming content, or heck, even maybe productivity, while being effectively a marginal upgrade in cost on a monthly payment.

Also, just for the record, some carriers in Japan are offering financing that gives you the option of paying over 48 months and Apple offers 36 months. Is that quite absurd to most people? Absolutely, but for the average person who wants a bigger screen sometimes, the cost is not even going up much on a monthly basis.

For the record, especially in Japan, folio cases where you open your phone like a book are already the default. Effectively most people on the train have their phone open in some "Duo-ish" form-factor already. So the jump from a folio case to a folding phone is already substantially lower.

Obviously the gimmick nature of folding is not the compelling reason why people want folding phones. The question is more about tradeoffs, and I think that especially in asia, the demand for folding phones is quite obvious: upgrade your primary PC to have a huge display when you want it that way, otherwise it stays compact enough to be with you all the time.

I'm 27. Posts like this fill me with a weird kind of... melancholy? Sadness? A feeling of having missed out?

At some level I know I'm seeing these pictures through rose-colored glasses, and these early-ish tech folks had plenty of gripes about office life. But man, I came into the workforce just in time for the idea of "working with people in person" to die. And I like working from home, but... man, working on truly new stuff, in person, with people who also care, looks like it was so much fun.

It's nice that this article goes beyond the "context too long" scapegoat diagnostic.

There is more to it than just reducing content in context. LLMs attention mechanism works by finding graphs of related words and concepts and it works best when the relevant relationships are close together in the text. Length matters but how well the information is organized matters as well.

It's useful to be familiar with the concept of dependency length. The two sentences in this image ( https://benoitessiambre.com/img/dependencylength.png ) have the same words and same meaning but sentence C is organised to minimize the distance between related words.

This can also apply to paragraphs and sections. Ordering them so that related information is recursively closer together, helps AIs think better (also helps humans, Dependency Length Minimization, DLM is an idea from psycho-linguistics).

This is one reason progressive disclosure and sub agents are important. Including unnecessary information spreads apart surrounding information and makes it more difficult for AIs to figure things out.

For eg. optimizing skills, it can help to put more important information at the bottom of a skill. Usually after a skill is loaded, the AI will start “thinking”, talking to itself to try to reason through a problem. The thinking block will have an easier time considering the text closest to it, the last part of the skill, than information farther up.

Thinking blocks can produce a lot of text (a lot of it often hidden from users) that also spreads information apart. Sometimes, when an early instruction is still relevant but there’s been a lot of thinking and discussion added to the context since, it’s worth repeating a summary the early information ( keeping in mind that too many reminders can also spread other information apart ). I think however, LLMs are tuned to overweight very early information so that they don't disregard system prompts. You might be able to exploit this by placing some important facts near the very beginning of a conversation.

If you've ever written compression algorithms where repeated blocks are referenced instead of encoded, you'll be familiar with how related information being clustered allows for better compression. It's another example of the parallels between intelligence and compression and it hints at the information theoretic aspect of DLM.

This is why Anthropic and OpenAI are pushing for more “regulation”. You can’t show ROI to investors by cutting inference costs.

Regulation might inconvenience them a little, but the upside is huge if it means open models out of China get banned.

Same thing happened in Hungary earlier this year, after 16 years of Orbán (and FWIW, the two are / were "good friends").

Interesting times.

This isn’t as useless as it seems. It is a visual representation of how old the tab is. And I control what “old” is.

To put it another way, it helps me recognize which tabs I am using. And which I am saving for later without having to evaluate the content of the tab.

One tweak… the coverage percentage can it be adjustable? Max = 100% looses all information about the pages contents, except the title.

Also also… if the images can be made translucent… you can make old timey effects like yellowing and decay.

Man, I was about to go on a proper rant and claim that the site didn’t have every battle because it’s missing the battle of Maiden Castle…

I claim no great historical authority here but, since I grew up mostly within sight of the hillfort, we were taught about its history at school to an extent that, at the time, felt unrelenting, cruelly repetitive, and utterly tedious. In reality though, it probably lasted for a single term at the fag end of 1985 when we learned about the Celts[0].

But what we learned was that General Vespasian captured Maiden Castle after a brutal one day battle in AD 43 in which Maiden Castle’s Celtic defenders were slaughtered in what amounted to a one-sided rout due to the superior weapons, tactics, and discipline of the Romans. This is what I remembered, but I wanted to look it up to ensure my facts were straight.

Welp… TIL because subsequent investigation and analysis has shown that it didn’t happen at all so, no, dear reader, my “facts” were not straight.

There never was any big battle.

There might have been the odd smaller skirmish during the Roman period, but the bodies we found that were all supposed to have died in the big battle actually span several generations, dying at different times over a period of a couple of hundred years, and the evidence of violence in their deaths was in the vast majority of cases inter-tribal and occurred before the Romans even showed up.

Maiden Castle had already fallen into decline by the time the Romans arrived and was mostly abandoned for a long period until a temple (that you can still visit the ruins of) was built early in the 4th century.

So the reason the battle of Maiden Castle doesn’t appear on the site isn’t because the site is carelessly built AI-slop but because no such battle ever happened.

I have been… disillusioned. Hats off to the human behind the AI that built the site.

[0] I remember this with such precision because it was the first period of history we learned about at middle school, and matters didn’t improve significantly, because next it was the domestic matters of the bleeding Victorians, and lots of episodes of “How we used to live”.

The reality is that unless your in the US or "most" of Europe, you're going to WANT Unicode domain names - people in china are mostly fine with http://www.china.cn but they're going to really want to use their own language https://xn--fiq8ituh5mn9d1qbc28lu5dusc.xn--fiqs8s or https://xn--mgbeiue0iwb.xn--ngbc5azd - https://xn--wgv71a119e.jp

Demanding that everyone use our alphabet because we're too dumb to recognize the difference between google.com and xn--ggle-55da.com is the worst kind of "security through superior firepower"

(the fact that HN shows these correctly in the EDIT window but have to dump to punycode in display is sad, tbh)

Note that the trial was for “stage IIB–IV melanoma that had been surgically removed prior to entering the trial.”

> lost my dad to cancer because of cancer medication

Please don’t phrase it like this. Loss is hard. But desperate people have skipped treatment because they misread something online.

> Why does AI-generated imply "not bothered" to you?

1. AI generated text is just hard to read and verbose, if you worked so hard on your docs/specs, why would you want your README to be unreadable?

2. There's so much garbage written with AI that I'm just going to assume yours also is. Would you sign up to a gym with broken windows? It could be the best gym ever, but I wouldn't count on it.

My 5 year old just got his last imagination library book. It showed up right after Dolly passed away.

There was a letter from Dolly to him that ended “I will always love you”. Made my wife tear up.

Gibson published Agency nearly three years before ChatGPT's release, and six years before the emergence of today's increasingly autononomous AI agents.

...and fifteen years after Stross' Accelerando opened with a man who spent every waking hour doing what his autonomous AI agents told him to do via his smartglasses...

...and twenty-seven years after AT&T's "You Will" ad campaign sold a future including agents, teleconferences, gps navigation, video on demand, and many more things, pretty much none of which were actually brought into daily life by AT&T...