Hacker News Reader

gwernon Show HN: Veil – Dark mode PDFs without destroying images, runs in the browser

FWIW, we did consider a histogram heuristic, and I believe GreaterWrong still uses one rather than InvertOrNot.com. But I regularly saw images on GW where the heuristic got it wrong but ION got it right, so the accuracy gap was meaningful; and that's why we went for ION rather than port over the histogram heuristic.

gwernon Show HN: Veil – Dark mode PDFs without destroying images, runs in the browser

> It's absolutely true that there's a subset of raster images, like diagrams with white backgrounds and black lines, that would benefit from inversion. I could be wrong, but in my experience they're a minority, and the cost of accidentally inverting the wrong one (a medical photo, a color chart) is much higher than the benefit of inverting a black and white diagram, from my point of view. For now the per-page toggle covers those cases.

OK, so I did understand, but this sounds very hand wavy to me. You say it's a 'minority'; well sure, I never claimed that was >50% of images, so I suppose yes, that's technically true. And it is also true that a false positive on inverting is usually nastier than a false negative, which is why everyone defaults to dimming rather than inverting.

But you don't sound like you have evaluated it very seriously, and at least on my part, when I browse my dark-mode Gwern.net pages, I see lots of images and diagrams which benefit from inverting and where I'm glad we have InvertOrNot.com to rely on (and it's rarely wrong).

It may be nice to be able to advertise "No AI" at the top of the page, but I don't understand why you are so committed to biting this bullet and settling for leaving images badly handled when there is such a simple easy-to-use solution you can outsource to, and there's not a whole lot else a 'dark mode PDF' can do if 'handle images correctly' is now out of scope as acceptable collateral damage and 'meh, the user can just solve it every time they read every affected page by pushing a button'. (If Veil doesn't exist to save the user effort and bad-looking PDFs, why does it exist?)

gwernon Show HN: Veil – Dark mode PDFs without destroying images, runs in the browser

> In practice they're rare enough that the per-page toggle handles them, but it's the honest limitation of the approach.

I don't understand how you handle raster images. You simply cannot invert them blindly. So it sounds like you just bite the bullet of never inverting raster images, and accepting that you false-positive some vector-based diagrams? I don't see how that can justify your conclusion "it wasn't necessary". It sounds necessary to me.

gwernon Show HN: Veil – Dark mode PDFs without destroying images, runs in the browser

Have you considered, since you can extract the images via the mask, selectively inverting them?

One can fairly reliably use a small NN to classify images by whether they should be inverted or just dimmed, and I've used it with great success for years now on my site: https://invertornot.com/ https://gwern.net/invertornot

---

On a side note, it'd be nice to have an API or something to let one 'compile' a PDF to dark-mode version PDF. Ephemeral browser-based is a drawback as often as a benefit.

gwernon Personal Encyclopedias

A good wiki like MediaWiki supports various levels of visibility. For example, you could define a namespace for each group of readers like 'Family:'. Or use transclusions from subpages. (This might sound like a bit of a hassle but you can use a template to set it up once and for all: a page transcludes a public sub page followed by the distant relatives material followed by parents / siblings followed by your-eyes-only.) And I'm sure one could come up with other approaches too.

A real example: Said Achmiz (obormot.net) uses PMWiki for his D&D campaigns, and PMwiki lets you control who can see a page, so he can do access control tricks like a page for a location, where only the DM can see all subpages with all the secrets, while each player can see their own 'notes' subpage. So everyone in their own web browser can go to the same page and see the same thing overall, but will see just their private additional information. And this is quite flexible so you can encode whatever patterns you need. You don't need some WotC fancy custom CMS for your D&D campaign to keep track of information and silo appropriately, you just need a design pattern on wikis.

gwernon So where are all the AI apps?

> Detractors of AI are often accused of moving the goalpost, but I think your comment is guilty of the same. Before Claude Code, we had Cursor, Github Copilot, and more. Each of these war purportedly revolutionizing software engineering.

What's sauce for the goose is sauce for the gander. If you make that argument that 'I don't believe in kinks or discontinuities in code release due to AI, because so many AI coding systems have come out incrementally since 2020', then OP does provide strong evidence for an AI acceleration - the smooth exponential!

gwernon Thinking Fast, Slow, and Artificial: How AI Is Reshaping Human Reasoning

I was being sarcastic, because that point obviously also applies to the subjects in the experiment as well.

gwernon Thinking Fast, Slow, and Artificial: How AI Is Reshaping Human Reasoning

One might worry that it would increase the authors' confidence even following their LLM rewrite errors and reduce accuracy overall regardless of moderators.

gwernon The gold standard of optimization: A look under the hood of RollerCoaster Tycoon

End-to-end optimization in action! Although I'd've liked more than 1 example (pathfinding) here.

gwernon I turned Markdown into a protocol for generative UI

Or more precisely, isn't this reinventing notebooks (not the first JS-centric notebook either)?

gwernon Scaling Karpathy's Autoresearch: What Happens When the Agent Gets a GPU Cluster

Yes, it's greedy so may hit local optima. You can fit learning curves and try to extrapolate out to avoid that problem, to let you run long enough to be reasonably sure of a dead end, and periodically revive past candidates to run longer. See past hyperparameter approaches like freeze-thaw https://arxiv.org/abs/1406.3896 .

gwernon Scaling Karpathy's Autoresearch: What Happens When the Agent Gets a GPU Cluster

> The agent can theoretically come up with a protocol to run those same 12 experiments one-by-one and only then decide which branch to explore next - which I think would lead to the same outcome?

At least in theory, adaptiveness should save samples and in this case, compute. (As noted, you can always turn the parallel into serial and so the serial approach, which gets information 'from the future', should be able to meet or beat any parallel approach on sample-efficiency.)

So if the batch only matches the adaptive search, that suggests that the LLM is not reasoning well in the adaptive setting and is poorly exploiting the additional information. Maybe some sort of more explicit counterfactual reasoning/planning over a tree of possible outcomes?

gwernon Scaling Karpathy's Autoresearch: What Happens When the Agent Gets a GPU Cluster

Except the power drill isn't being used to make a better chimpanzee.

gwernon The math that explains why bell curves are everywhere

A little disappointing. All about the history of bell curves, but I don't think it does a very good job explaining why the bell curve appears or the CLT is as it is.

gwernon Every layer of review makes you 10x slower

After some chatting with GPT-5.4 Pro, I think the 10x claim might simply be because in the >90% load region, you get the famous unbounded delays. And with realistic loss functions like absolute or squared error (consider that in a queue of queues, delays would add) and distributions like Poisson, you'd encounter massively increased losses if you pushed too far towards 100% load. Without extremely detailed accurate statistics of the sort almost no one has, it'd be hard to distinguish 'by eye' 90% from 95%, say, and you'd just see '10x' everywhere you looked.

(And in a pyramid of queues like many layers of reviews, each layer will wind up being about equally loaded, because otherwise you would get a big benefit from adding/removing capacity, so each layer will slowly be optimized towards its breaking point, yielding the 10x everywhere.)

gwernon Every layer of review makes you 10x slower

The 10x claim reminds me of various scaling laws in anthropology/sociology, especially the much-debated Haire cube-root scaling law (eg https://gwern.net/doc/sociology/1959-haire.pdf https://gwern.net/doc/economics/1983-stephan.pdf ). I wonder if one could give a derivation from queuing theory where you have an argument along the lines of 'in indefinitely many layers of queues of queues, the total delay is minimized at 90% load per queue which yields a 10x scale-free slowdown'?

gwernon GPT‑5.4 Mini and Nano

I agree. That's why you should write as much as you can now, if you want to get it into the LLMs (https://gwern.net/blog/2024/writing-online). You never know when the window will slam shut and LLM training goes 'hermetic' as they focus on 'civilization in a datacenter' where only extremely vetted whitelisted data gets included in the 'seed' and everything is reconstructed from scratch for the training value & safety.

gwernon The Brand Age

https://en.wikipedia.org/wiki/The_Spirit_Level_(Wilkinson_an...

gwernon The Brand Age

Linux isn't a strong brand at all. Even nerds would struggle to tell you what a 'Linux' is (or as I prefer to call it, since that's every non-smartphone Linux I've ever used, 'GNU/Linux'). And I've never even heard of 'Portuguese tarts', and I've heard plenty about Portugal and briefly considered going there last year for tourism.

Pebble went out of business, incidentally.

gwernon The Brand Age

Some of the typos and poor typography might be accidental simply because that is so common, but the rest of it is surely deliberate.

Remember that _American Psycho_ is repeatedly hinting to you throughout the movie that most, or even all, of what you see is just Bateman's own delusions and hallucinations, which are inferior imitations of reality due to his own impoverished mind and lack of substance. So there's a lot of little tricks all throughout it to subtly destabilize you: https://www.moviemaker.com/american-psycho-anniversary-oral-...

gwernon Elon Musk pushes out more xAI founders as AI coding effort falters

It's definitely very valuable, but for what AI model? How does any of that lead to AGI, or even just a good coding agent?

gwernon Celebrating Interesting Flickr Technologies

https://news.ycombinator.com/item?id=47134407

Thanks for writing it up!

gwernon Printf-Tac-Toe

List of examples: https://gwern.net/turing-complete

It was probably unintentional, yeah, I don't recall any mentions of early printf being overloaded to do stuff, nor is it clear why you would do that since you're using it in a much more convenient Turing-complete language already (C).

gwernon Malus – Clean Room as a Service

The solution here seems to be to impose some constraint or requirement which means that literal copying is impossible (remember, copyright governs copies, it doesn't govern ideas or algorithms - that would be 'patents', which essentially no open source software has) or where any 'copying' from vaguely remembered pretraining code is on such an abstract indirect level that it is 'transformative' and thus safe.

For example, the Anthropic Rust C compiler could hardly have copied GCC or any of the many C compilers it surely trained on, because then it wouldn't have spat out reasonably idiomatic and natural looking Rust in a differently organized codebase.

Good news for Rust and Lean, I guess, as it seems like everyone these days is looking for an excuse to rewrite everything into those for either speed or safety or both.

gwernon Why the global elite gave up on spelling and grammar

I would have to consider carefully if I thought I was a high-enough quality candidate that it would be interpreted as a countersignal rather than a signal.

If I, gwern, specifically, were to apply, I might; because I know I am widely read on HN and I've talked with any number of YC partners etc, and they all know I take care in writing, and so me not capitalizing is a deliberate message rather than laziness or incompetence. They may or may not appreciate the message, but they won't infer the usual things, at least.

If I were anyone else and my application just one of thousands in the flood? You'd better believe I'd capitalize and spellcheck my YC application: https://gwern.net/blog/2023/good-writing

gwernon Why the global elite gave up on spelling and grammar

Also just differing levels of relevance. You don't talk with a businessman or investor or famous people in general because of their writing; if you made a list of relevant skills, 'proper spelling when quickly texting from a phone' surely doesn't crack even the top ten thousand skills. In academia, on the other hand, writing a formal application properly is a core skill.

gwernon Show HN: How I topped the HuggingFace open LLM leaderboard on two gaming GPUs

My understanding was that Autoresearch was defined as training from scratch (since it's based on the nanogpt speedrun), not using any pretrained models. So it couldn't do anything like upcycling a pretrained model or the Frankenmerge, because it's not given any access to such a thing in the first place. (If it could, the speedrun would be pointless as it would mostly benchmark what is the fastest fileserver you can download a highly compressed pretrained model checkpoint from...) It can increase the number of layers for a new architecture+run, but that's not the same thing.

gwernon Show HN: How I topped the HuggingFace open LLM leaderboard on two gaming GPUs

Adding, swapping, or duplicating layers has a long history (eg. StyleGAN, upcycling), and it was pointed out at least as far back as He et al 2015 (Resnets) that you could ablate or add more layers because they functioned more as just doing some incremental compute iteratively, and many of them were optional. (Or consider Universal Transformers or heck, just how BPTT works.) So this idea is not far out of distribution, if at all, especially if you're a LLM who knows the literature and past approaches (which most humans would not because they only just got into this area post-ChatGPT).

gwernon Show HN: How I topped the HuggingFace open LLM leaderboard on two gaming GPUs

> They almost certainly have never seen regular conversations in Base64 in their training set, so its weird that it 'just works'.

People use Base64 to store payloads of many arbitrary things, including web pages or screenshots, both deliberately and erroneously, and so they have almost certainly seen regular conversations in Base64 in their 10tb+ text training sets scraped from billions of web pages and files and mangled emails etc.

gwernon When AI writes the software, who verifies it?

> I encourage everyone to RTFA and not just respond to the headline.

This is an example of an article which 'buries the lede'†.

It should have started with the announcement of the new zlib autoformalization (!) https://leodemoura.github.io/blog/2026/02/28/when-ai-writes-... to get you excited.

Then it should have talked about the rest - instead of starting with rather graceless and ugly LLM-written generic prose about AI topics that to many readers is already tiresomely familiar and doubtless was tldr for even the readers who aren't repelled automatically by that.

† or in my terms, fails to 'make you care': https://gwern.net/blog/2026/make-me-care