Hacker News Reader

So OA PPUs can now be sold and transferred without restriction to arbitrary buyers, outside the tender offer windows?

> It's no longer hanging over anyone's head,

The tender offer limitations still are, last I heard.

Sure, maybe OA can no longer cancel your vested equity for $0... but how valuable is (non-dividend-paying) equity you can't sell? (How do you even borrow against it, say?)

gwernon Matt Trout has died

My point was that 'child prodigy' was not loose flattery or exaggeration by a friend, because he was at 'has newspaper articles being written about him' levels of child prodigy and was unarguably a child prodigy.

gwernon Matt Trout has died

'Child prodigy' is not an exaggeration here: https://trout.me.uk/prec/

gwernon Show HN: I AI-coded a tower defense game and documented the whole process

Actually, I came up with that all on my own after I noted to myself that capture-recapture would work; and it amused me so much that I resolved to try to come up with a proper list filling out the idea. I did get some of the other ideas from LLMs, though.

gwernon The chemical secrets that help keep honey fresh for so long

> I'm going to call out the WP article as being factually wanting on that point.

No, it's not. What is factually wanting is any case of an actual tomb honey.

When I looked into it, every single supposed tomb honey, not already debunked by Leek, deadends in a useless citation which is 'someone somewhere once found tomb honey trust me bro'. Clearly made-up. They don't exist.

The burden of proof is on anyone who still believes in tomb honey to name a single specific verified instance, with time and place and tomb, and someone actually witnessing it and analyzing or eating it and proving it's honey, rather than several other possible residues like the Leek examples. Otherwise, it's just more telephone game nonsense.

(Note: the Leek PDF is unfortunately not working right now, because Hetzner has temporarily disabled my account over torrenting alerts.)

gwernon Show HN: I AI-coded a tower defense game and documented the whole process

https://xkcd.com/1425/

One of the frustrating things about web dev, I find, is the staggering gulf between apparently nearly identical tasks and unpredictability of it. So often I will find myself on gwernnet asking Said Achmiz, 'this letter is a little too far left in Safari, can we fix that?' and the answer is 'yes but fixing that would require shipping our own browser in a virtual machine.' ¯\_(ツ)_/¯

gwernon The War on the Walkman

> Oscar Gross was preparing to take the case all the way to the supreme court, but backed out after someone was killed crossing the street while wearing headphones. The kind of tragic anecdote that is the inevitable and unavoidable price of freedom.

gwernon Shadow of a Doubt

The exposure therapy part was definitely the most interesting part for me: licking dumpsters, hugging crazy homeless people (who then run away shrieking at invisible people), (very slightly) vandalizing cars, running a gauntlet of people holding knives... Not sure how much I got out of the previous sections, but the part starting here was fun:

> Onstage facing them stood Grayson, the mastermind of the night. This Grayson was a heady cocktail of signifiers garnished by an Indiana Jones–type hat he’d worn, I gathered, to all twenty-one Road to Recovery Tours to date.

gwernon Creating fair dice from random objects

Reference: https://en.wikipedia.org/wiki/Randomness_extractor#Von_Neuma...

gwernon DeepSeek R2 launch stalled as CEO balks at progress

I don't think it's well known that TI has FB CoIs. I didn't know myself until fairly recently. You can talk to a lot of people about this sort of stuff without anyone pointing it out.

gwernon I deleted my second brain

I'm fascinated by all these comments I see on HN and elsewhere where people will deny that a blatantly LLM-written article was not LLM-written, including cases where people praise it for not being LLM-written (eg. https://news.ycombinator.com/item?id=44384138 ). Like, leave aside the issue of whether it's a good or bad thing (I've been doing generative text NNs since 2015, so I'm mostly for it, when done well), I'm just interested in the inability to notice.

Skimming your comments, you, for example, do not seem to be illiterate or a bad writer at all despite being ESL (although you overuse the double-sentence structure in your comments), but you describe this as being 'stream of consciousness' (it is not even close to that, look at an actual example like Joyce) and seem to think it is fine.

So I'm puzzled how. Why isn't it obvious to you that the style is so mode-collapsed ( https://gwern.net/doc/reinforcement-learning/preference-lear... )? Do you also not notice how all the ChatGPT images are cat-urine yellow? (I've been asking people in person whether they have noticed this in the Bay Area and I'd say <20% of enthusiastic generative AI users have noticed.) What are you thinking when you read OP? Does it all just round off to 'content', and you don't notice the repetition because you treat it all as a single author? Are you just skimming and not reading it?

gwernon Alternative Layout System

FWIW, I call this approach 'square' writing, and have compiled some links at https://gwern.net/doc/design/typography/square/index

Probably the most interesting one is the 'Hangulatin' font (https://www.t26.com/fonts/22320-Hangulatin-EN), which is exactly what it sounds like, and unfortunately has been abandoned/linkrotten but you can see a lot of it in the old video: https://www.youtube.com/watch?v=T0syCsC0_4s

gwernon The Business of Betting on Catastrophe

> She is the author of “Investable! When Pandemic Risk Meets Speculative Finance – A Cautionary Tale,” from which this article is adapted.

So I think structurally, the conclusion here is that 'cat bonds are an example of how insurers can work with abstract risks, and so any risk (such as global pandemic) could be worked with this way', and the rest of the book then examines how people are trying to actually do so with pandemic risk.

gwernon The Hollow Men of Hims

> This style of writing is a welcome change from all the AI slop or self promoting blogs out there.

Dude. This is AI slop. And quite obviously so! You think all those EM DASHes are there naturally? Or the constant use of reversal? No one writes like that. (Even the people who love em-dashes and make a point of using the Unicode point will change it up more than that, rather than using it like a metronome.) Even if there wasn't that adrafinil referral URL giving it away (which incidentally tells you that OP wasn't doing all his own research but relying on the search plugin to compile a report he could spin), at this point you should recognize the 4o style.

This just doesn't sound like the normal ChatGPT because the author prompted ChatGPT to make it as invective and rhetoric and axgrinding as possible (or possibly, just went through and heavily edited a more neutral ChatGPT draft but I doubt that is responsible for the bizarre analogies or rhetoric like the hot dog thing).

So, a good example of the "don't worry about seeing AI slop on HN; worry about when you stop seeing AI slop on HN" evolution. Stripping referrers or avoiding EM DASHes is, after all, easy to do...

Also, top keks:

>> It is worth noting that the culture that produced Hims—Silicon Valley's peculiar blend of messianic self-regard and algorithmic thinking—has convinced itself that traditional gatekeepers are inherently suspect, that disruption is inherently virtuous, and that the phrase "move fast and break things" applies as beneficently to human bodies as it does to software systems.

gwernon Requiem for a Solar Plant

I think writing fiction with LLMs is fine. "based on a true story" or roman à clef fiction is also fine. I would have no objection to this if it were clearer that this was a "movie treatment". (As LLM fiction goes, this isn't even all that bad.)

However... Look at this HN page! Not a single person here realized (before my still-downvoted comment) that this was a 'movie treatment', or that the 'quotes...were also mostly spat out by the LLM as well'. They all are taking it as 100% gospel, and engaging with it in good faith, as if it were genuinely a 'technical post-mortem' (as it takes pains to present itself) instead of your fiction-writing hobby,

Just make it clearer that this is a fictional story.

> I'm curious if you consider having a ghost writer to be lying, or a cinematic re-enactment

Yes. This is why ghost writers are supposed to be credited as such if they are contributing non-trivial content rather than simply serving as a amanuensis or an editor, and cinematic re-enactments are flagged as such if it wouldn't be obvious to a reasonable person (by reputable documentarians, anyway).

gwernon Requiem for a Solar Plant

One of the problems with undisclosed and thoughtless use of LLMs for writing, like ChatGPT in this case, is that it undermines trust.

Take the many quotes in this article, which all sound like ChatGPT. Obviously, none of those people said those words, or even anything all that like what they supposedly said, because that's not what real people sound like.

If the author is willing to silently do that (which could be uncharitably described as "lying"), why should I trust anything else, like the numbers, or the factual claims?

Did any of this actually happen? (I note that there's not a single external link or fact that an ignorant layman like myself could quickly and easily verify, including the Astral Codex Ten part.)

Incidentally, there are a lot of typos in the titles in https://7goldfish.com/ .

> This is a much better read than I expected.

Hm...

gwernon The Brute Squad

> I characterized coding agents as camels in RotJD, and that metaphor still has legs. Four of them. But I find it difficult, as a writer, to convey the siren-song allure of coding agents when I'm describing them as a bunch of camels gurgling at me. So they're… baby camels? Regardless, they’re like toddlers, but with proper supervision and food preparation, they're absolute brutes at writing code. I can't bring myself to leave when my brutes are hungry. I've tried, it's a no-go. I have to run a practiced escape plan every night to get my computer closed by 2am. First I get them all spinning at once. Then I leap up, run out of the room, slam the door, jam my fingers in my ears, and sprint away shouting lalalala. It is perhaps not the most refined of plans, and doesn't have a great success rate. My babies need me.

Amdahl's law in action: the economic gain from coding agents is ultimately bottlenecked by the slowest serial component in the system. In this case... Steve Yegge himself.

Which is why the goal is to replace Yegge entirely. Even the perfect coding assistant which makes Yegge 100x more productive is still 100x worse than full replacement and running 10,000 virtual Yegges in parallel. Why settle for 1%?

gwernon Hacking Pinball High Scores

Achmiz argues that it makes more sense to define each page as 'new' vs 'old', and disable the black-stars based on that, rather than on whether it's a blog or not. This is more work, but makes more sense (and we might want to do other things based on that binary variable, like emphasize the created date at the top of the page, to warn readers that something is recent and thus still half-baked). So we're implementing that now.

gwernon Stepping Back

https://gwern.net/doc/science/1986-hamming#great-thoughts-fr...

You can think of it as a RL problem, and there are some interesting algorithms which achieve good performance by periodically 'breaking out' of exploitation, but less and less: https://arxiv.org/abs/1711.07979 (I expect that you can come up with an infinite hierarchy of 'wake ups' which converge to a fixed overhead and which aren't terribly far off an optimal schedule of replanning wake-ups, and that something like 'great thoughts Friday' is the first step; then you'd have 'great thoughts first-of-the-month' and 'great thoughts new year's day' etc: https://gwern.net/socks#fn10 )

gwernon Hacking Pinball High Scores

"My sealed, postmarked, and notarized certificate stating that my nickname for pinball playing is 'DDOM' is raising a lot of questions about my pinball nickname I really thought I had already addressed."

gwernon Hacking Pinball High Scores

I agree it's not helpful. The star marks a 'new' link (ie. seen for the first time recently in the website corpus). It's useful for helping flag recently-modified writing or new additions in my essays, where I might have added a bunch of links 15 years after starting the essay, and then the interesting new paragraph jumps out. They are 'starred' as something new which might be worth looking at even if you've read that page before. (Classic but forgotten Web 1.0 design pattern.) Unfortunately, in cases like this, I haven't ever written about pinball hacking before, so almost all the links are new...

I've thought about how to handle it, but haven't come up with a good answer. It may be that since 'blog' posts come with a clear temporal context, compared to my usual pages, it makes more sense to just disable that particular feature on all blog posts. We'll think about it.

gwernon Hacking Pinball High Scores

No, it's not. As should be obvious from the writing style alone - no LLM writes like that, and you should easily see the absence of the usual markers like the ChatGPT twist ending or the repetitive triplets... If you're that bad at spotting LLM writing, don't go around accusing people of it.

I did ask the LLMs for ideas, and used many, but the original idea and seed list was mine, the curation was mine, and the overall work is mine, and I wrote almost every word myself because I disliked the LLM style that much. I take full responsibility for it as being worth reading and better than any LLM does on this topic. (This is why I do not credit them as co-authors like I do on some of my pieces.) Also, a lot of them were dumb and I discarded them entirely. (I'm sorry, Claude-4-opus, you really wanted some sort of quantum multiverse or anthropics category of pinball hacks, but I just couldn't think of one which wasn't stupid.)

If you are curious, I am happy to share my LLM sessions as best as I can (sometimes the 'sharing' is disabled for reasons I don't understand, which makes it hard). Here is the main ChatGPT session: https://chatgpt.com/share/683c93e8-2c74-8006-a5f6-f08cbaa960... Also https://chatgpt.com/share/683c948e-57e8-8006-8064-2c768e2db3... https://chatgpt.com/share/683c94a1-37d4-8006-8618-98543430ae... So you can see for yourself what I curated and rewrote into into my final version, and how I did it.

gwernon My website is ugly because I made it

https://en.wikipedia.org/wiki/Trypophobia maybe? Could lean into that as an esthetic - everything is always misaligned carefully by 1 pixel, using trypophobia fonts which overload ligatures to punch tiny holes into each letter...

gwernon Compiling a neural net to C for a speedup

Thanks for checking. It sounds like the C compiler isn't doing a great job here of 'seeing through' the logic gate operations and compiling them down to something closer to optimal machine code. Maybe this is an example of how C isn't necessarily great for numerical optimization, or the C compiler is just bailing out of analysis before it can fix it all up.

A fullstrength symbolic optimization framework like a SMT solver might be able to boil the logic gates down into something truly optimal, which would then be a very interesting proof of concept to certain people, but I expect that might be for you an entire project in its own right and not something you could quickly check.

Still, something to keep in mind: there's an interesting neurosymbolic research direction here in training logic gates to try to extract learned 'lottery tickets' which can then be turned into hyper-optimized symbolic code achieving the same task-performance but possibly far more energy-efficient or formally-verifiably.

gwernon Compiling a neural net to C for a speedup

How much of a speedup is the C compiler optimization able to achieve in terms of compiling it down to a hand-written C equivalent vs the -O0 non-optimized assembler? What does the optimized C/assembler do which isn't actually necessary and accounts for the remaining inefficiency?

gwernon Mustard Watches (1990)

See also hunting lions using math: https://news.ycombinator.com/item?id=29106064

gwernon Beware the Complexity Merchants

Likely AI-written. OP sanitized some of the most obvious tells, or perhaps the blogging software added those spaces to the em-dashes, but when a dev who is as gung ho on GenAI as https://chrlschn.dev/blog/2024/10/im-a-gen-ai-maximalist-and... is, writes an ending like that, and the whole thing is about as intellectually nutritious as a bag of Doritos, we all know what they did. A waste of time.

---

Out of curiosity, I asked ChatGPT-o3 what LLM it thought wrote the ending, and I agreed with its rationale for guessing that 4o wrote it:

"The prose is archetypal ChatGPT rhetoric—second-person apostrophe (“dear traveler”), triadic repetitions, and a moralizing close. / Vocabulary inflation (“complexity merchant”, “moat of self-importance”), syntactic parallelism, and the cadence of stacked, comma-separated clauses are hallmarks of GPT-4’s temperature-controlled completions. / No concrete anecdotes, data, or links—just abstract exhortation—exactly what an LLM defaults to"

But more interesting was the rationale for it not being Claude or Gemini-pro: "those models tend to hedge (“may”, “might”) and inject more safety language. This text is bolder and more florid—typical of GPT-4 with style-boosting prompts."

I'm not sure I would've identified that as stylistic tics of Claude/Gemini. I'll have to look closer for those in the future.

gwernon Mouth bacteria partially spread depression and anxiety in newlywed couples

This is Iran, so the researchers are often not great, or it's just straightup fraudulent. It's one of the epicenters of developing-world science fraud. So always worth keeping that in mind - there may not, in fact, be any facts that have to be explained.

gwernon The Decline of Usability: Revisited (2023)

Crazy how often it backfires, too. I wasted 15 minutes a few weeks ago checking a user report that a PDF I hosted was missing exactly half its pages. Eventually, after walking it through my workflow step by step and it not breaking, I realized that it was simply the PDF viewer hiding the horizontal scrollbar which told you that it was a two-page layout. We had both fallen for it. Thanks, GNOME et al! (╯°□°)╯︵ ┻━┻

(This also illustrates the fallacy with the sibling comment blandly asserting that scrollbars should 'simply be hidden when not necessary'.)