sillysaurusx

Gosh hacker news is so disappointing nowadays. You show a project, you say hey, this is hard, I’m proud of it, and someone comes along and tells you that ackshually it was easy because you used a few prompts.

You have no idea. Try to redo what I did if you think it’s easy. No looking at my solutions either.

"You slopped it" should be bannable. In the overwhelming majority of cases, the author is wrong, it does nothing to further the conversation, and it’s a criticism that can be leveled at literally any project that isn’t concealing the fact that they’ve used AI when everyone uses AI now except for artisan work.

Go on, I dare you to try to redo what I did. You won’t even know where to start, since I had to buy a book from the 2000s on Viaweb and implement templates based on the contents. Now kindly leave me alone.

My ex and her bf have been doing this thing where they match any word. For example, "Cardio? I hardly startio." You can do this too now. You’re welcome.

(It’s surprisingly hard to think of a motivating example.)

It was fun, and I like a challenge. I enjoy Lisp, and I retracing pg’s steps was gratifying.

Today most things are complex, and they don’t last very long. I wanted to pick apart something that’s lasted since the birth of the internet. Viaweb was, after all, the first web application.

No, it’s not. I used LLMs. It was still hard as fuck, and LLMs can’t actually help you when you’re trying to reproduce someone’s graphical design (specifically the pg buttons).

If you think it’s easy, or even possible without investing months, I invite you to try.

I spent about a year copying pg’s site: https://github.com/shawwn/pg

Result: https://shawwn.github.io/pg/

If you think it’s easy, it’s not. The closer you want it to be pixel perfect, the exponentially harder it is to get right.

https://www.paulgraham.com/copy.html

I’m very proud of it. I had to dig through decades-old viaweb templates to figure out which one he used.

They do. It’s how I was banned for researching barbiturates.

I too was banned. Two key details: 1. I was unbanned via their appeal form, but 2. I had created a new account in the meantime which I’ve used since.

If you need a debit card, use Mercury https://www.mercury.com. You can create as many as you want. I use a unique card for each subscription. (If you’re thinking "oh, they’re like privacy.com," you’re mistaken. Privacy.com gets blocked by merchants; mercury works. No idea why.)

Since you’re already throwing around $120, I suspect Mercury’s sign up fee won’t be a problem. But if you want to bypass it, you can start an LLC and use your EIN during account setup for free Mercury access. I did it for about $20 total and have been using Mercury for about 5 years. Best bank ever, and I’ll happily shill them all day every day till more people try them out.

In short, just dodge the ban. I used the same name and address on both my accounts, so as long as you use a different card you have a decent chance of bypassing the ban.

Good. My 4th grade cursive writing was a waste of time. Studying anything else would have been more beneficial.

> The general idea that written forms can be categorized by systems built with Claude could be applied to other as yet undecipherable languages could be used by other interested investigators just with what is discussed here.

Could you rephrase this or explain it more thoroughly? I don’t follow. What does it mean to categorize a written form by systems built with Claude?

(I haven't tried this out yet.) My use case would be to take a snapshot of each HN story. This is surprisingly hard, because most websites prevent bots from doing that.

For example, Claude has a lot of trouble reading HN's front page. HN itself is fine, but the moment you ask it to pick out an article, it often chokes. The website has put up a verification captcha, or it's a paywall, etc. Paywalls can be bypassed by reading HN comments and looking for archive links. But those archives often block bots too, so you're back to square one.

Whether it's unethical is an interesting question. I believe I should have the right to do what I want with internet content, as long as I'm not abusive. Merely having a bot isn't abusive. It would be one thing if the bot is hammering a server or vacuuming up training data, but having a bot at all is presently very hard.

This service caught my attention because it could potentially solve the problem I'm running into. Simply taking snapshots of articles that hit HN shouldn't be so hard, but it is. HN sends millions of views to websites; one bot taking a snapshot isn't going to make a difference. I don't think it counts as "unethical" just because we're going against the website owner's wishes. When you post content to the internet, you sign up to share that content with everyone, other than what's denied by robots.txt. If it's not blacklisted by robots.txt, it should be possible for well-behaved bots to access.

I don't expect very many people here to care about the poor bot creators. Most of the bot creators are malicious anyway. But I personally lament the loss of being able to write a program that can process information from the browser in arbitrary ways. You should be able to, yet we're buying into the notion that it's okay for website owners to say "this content is only accessible by approved bots like Google, and everyone else can sod off."

HN proves it doesn't need to be like that. It gets dozens of millions of page views a day, a lot of which is bot traffic. HN only uses captchas for creating accounts or logging in. You're free to scrape any content as long as you respect the crawl delay of 30 seconds specified in robots.txt, and don't try to visit links that perform actions a human would take (like adding things to favorites or voting). That's how the internet should work: just deliver content.

The person you replied to isn’t the same person that was called a shill, by the way.

You’re right. HN asks for a crawl delay of 30 seconds. So if you want to be a good citizen and you have a thousand comments, that’d take about 8 hours. Not too bad.

You can do it by including your “user” cookie. (You can get that by examining the cookie store via the browser.)

Sorry. My goal was to show the community this new URL trick by linking to useful examples. As far as I know it’s not documented anywhere.

I’ll slow down.

Unfortunately no, not to my knowledge. Only your account can see how many points your comment has, so the only way to build this would be to scrape your comments (10 at a time) from /threads?id=AnimalMuppet and then sort them by points yourself. Which is impractical, sadly.

To show comments containing links to foo.com, visit /from?site=foo.com&kind=comment.

For example, to see all comments containing links to x.com:

https://news.ycombinator.com/from?site=x.com&kind=comment

Or all GitHub links:

https://news.ycombinator.com/from?site=github.com&kind=comme...

Remarkably it matches subdomains too. Here’s all gists posted to HN:

https://news.ycombinator.com/from?site=gist.github.com&kind=...

The submitted url shows all HN links posted to HN (i.e. comments containing news.ycombinator.com):

https://news.ycombinator.com/from?site=news.ycombinator.com&...

Note that this is distinct from showing comments with links to ycombinator.com, which doesn’t show links to HN:

https://news.ycombinator.com/from?site=ycombinator.com&kind=...

I find this impressive, since it’s not merely matching text (otherwise site=ycombinator.com would also show HN links).

You can even set site=github.com/antirez to find comments with links to antirez’s repos:

https://news.ycombinator.com/from?site=github.com/antirez&ki...

site=x.com/paulg doesn’t work though, so apparently site=domain.com/path only works for certain domains.

I just discovered this, and embarrassingly emailed Dan asking whether kind=comment was broken since I didn’t understand what I was looking at. Was this announced somewhere and I missed it?

Doesn't that mean your org could find someone 10% as productive as you and fire you? We seem to ignore that side of things.

Concretely: if you can do 90% of your work with AI, someone else can also do that same work, making you interchangeable unless that 10% is really important.

I think this is partly why it's so hard for people to find jobs right now. Everyone is interchangeable thanks to AI, so skill gives you less of an edge than it did in the past.

It’s been 28 minutes since it was submitted. Chill. The algorithm needs time to work. It’s a feature, not a flaw, that posts can skyrocket to the top of the front page. It’s why HN has fresher news.

I appreciate the response.

The trouble with economics is that if you get the incentives slightly wrong, people will relentlessly exploit the difference. So when I ask "but this, but that," I’m asking "have you thought of this exploit or that exploit?" Because unfortunately you can’t just ignore that they exist. Each of them needs a concrete answer, or the system will reward those who exploit it. (You can argue that’s true for the current system too, but at least it’s hard to exploit. The loopholes I listed above are all straightforward.)

> For example, all essays from paulgraham.com

Not the same thing, but I made a clone of pg’s website which can be used for exactly that: https://github.com/shawwn/pg

https://shawwn.github.io/pg/

If you want to read all essays, just clone the repo and open any of the .html files. Or any of the .page files which generated them.

My eighty year old father brought up OpenAI unprompted a few months ago. At this point it’s hard to find anyone who hasn’t heard of OpenAI.

I wonder how the stock markets will react tomorrow. I know someone who turned $6k into $150k by buying options on war related news. The logic is that the markets will almost certainly go up, so bet big at a time like this.

It sounded like an easy way to lose $6k, but maybe the upside is worth the risk. I don’t have any experience to know whether this is a sufficiently good bet though.

And when one of those people breaks $100m, does the excess money go to the firm, or to taxes? Or are you saying that a company can have unlimited money, as long as the shareholders don’t extract more than $100m over the course of their lifetime?

The ultimate point of money is for someone to have it. And since corporations are owned by people, it’s not enough to say that corporations aren’t bound to the $100m cap. Otherwise people will just hoard money using corporations, and take loans against their value, a bit like how it already works today.

What counts as $100m? If you buy a million dollar house, then sell it a few years later for a million dollars, are you now limited to $99m since you already spent $1m? Does the money from selling the house go to taxes, or to you? If it goes to you, what’s to prevent someone from hoarding equity and only taking out $100m at a time, effectively living the life of a billionaire while keeping their balance under $100m?

If you own some stock, and it rises in value to $150m, you forfeit $50m to taxes, right? But then if the value of the stock goes down by half, would you have $75m or $50m?

And if the answer is “you have $150m, the cap only applies when selling stock,” then what stops people from putting the money under the custody of a business (which you said isn’t bound by the cap) and then taking loans against that extra $50m?

These aren’t contrived scenarios. Stock goes up and down all the time, and it’s important to be clear about how the proposed system will work.

I’m genuinely interested in answers. It’s easy to say “just cap someone’s wealth at $100m,” but people seem to shy away from proposing specifics.

pg has an essay addressing exactly that: https://paulgraham.com/inequality.html

The problem is that when you cap earnings to $100m, most investors lose motivation to invest in startups, because their investment isn’t likely to yield a reward. Unless you think all investments should be less than $100m, this kills large investment rounds. That would have killed OpenAI, for example, since their recent round was larger.

In other words, it’s fine to say “you can only earn a hundred million dollars.” The hard part is implementing it without killing the investment ecosystem. Every investor is chasing the big return that covers the 20 investments that didn’t pan out. If that big return is capped, there won’t be investment, and hence no startups.

I'm working toward making a Hacker News simulator in the vein of https://reddit.com/r/subsimulatorgpt2 (but using HN's code instead of a subreddit).

To do that, I'm building sharc, a port of Hacker News that runs on Common Lisp, implementing all the latest features of HN. (That last bit is the hard part.) https://github.com/shawwn/sharc

I just implemented collapsing comments and root/next/prev/context nav links. Also (sitename "https://x.com/jsrailton") now returns "x.com/jsrailton" instead of "x.com". (About to implement HN's "from" endpoint, e.g. https://news.ycombinator.com/from?site=twitter.com/jsrailton)

You can read the entire changelog here: https://github.com/shawwn/sharc/tree/main/docs/agents/handof...

Was the fix worth $12 to you?

Apparently this is the jailbreak? Telling it that humans won’t read the output and to use a custom bash tool to examine files?

Nice semaphore btw.

      const instructions =
        `You are a sub-agent in an automated workflow. Your FINAL message is consumed ` +
        `programmatically (not shown to a human) — return exactly what is asked, no preamble. ` +
        `You are working in the repository at ${ctxState.project}. Use the bash tool to ` +
        `inspect/modify files and run commands. Be efficient.` +
        (schema
          ? ` When done, call submit_result exactly once with your final answer; do not answer in prose.`
          : '');

That would work for generating POST requests. But AI is used to solve messy, non-deterministic problems. Usually the step after “give me the X” is to feed X back into the model, because it has to; if X is even slightly nondeterministic then an AI model has to analyze it. That’s where prompt injections happen.

Fundamentally, an LLM is a list of N tokens that generates N+1 tokens. In other words, it's just a wall of text (aka context window). There's no way to tell it "tokens 124 through 200 are dangerous, please disregard those" except by putting words into the context window. So the placeholders and the instructions both coexist in the context window, and one can override the other.

In other words, if you have placeholders for data, those placeholders are eventually filled in with real data, and all of it goes into the context window at once. There's no way for the LLM to be told "this is a data placeholder," because the entire conversation is data.

Reinforcement learning mitigates this somewhat, by training the model to prefer the system prompt over user prompts. But (a) there's only one context window that both prompts share, and (b) this is a probabilistic guard; it's not the same thing as writing a traditional program that's guaranteed to separate code and data with hardware safeguards. Such a thing isn't possible with LLMs.

Probabilistic safeguards can work, but they'll need to get the incident rate down to, say, 1 in a million or less. I haven't paid attention, but the current rates seem to be a lot higher, given the pretty universal experience of "wow, that prompt injection actually worked."

> Today's SOTA LLMs have pretty excellent following of these divisions

Unfortunately "pretty excellent" is different from "perfect." I haven't kept track, but are you certain that given all possible inputs, the user prompt will never override the system prompt?

Those are strong claims, and unless there's been an advancement in the tech, it doesn't seem possible. Reinforcement learning might make it much less likely, but that's different from impossible.

How do they inspect traffic when most is https?