textninja

He’s proposing using LLMs (which model human behaviour) to study humans so the distinction is pedantic. You don’t call it speadsheetology just because someone opened Excel.

The server that delivers the page never receives the content, never knows which site you are viewing, and has no way to find out.

Let me tell you about a thing called JavaScript.

A site that was never put on a server can never be taken off one.

If you post a link on HN and the content is embedded in the link itself then HN is the de facto server.

Doooooooooo dooodeeedoooodeeee doooooooooo doooooooooooo bshshhhhhzhhhhhhzhhhh

Anyone?

The purpose seems clear to me from the explanation provided. Here's what I read between the lines.

1. Send out thousands of letters expecting some to be returned. They may be returned due to deliverability issues, or they may be returned with a reply attached or (probably less commonly) scrawled on the pages of the letter itself. Replies to letters are of course common whether they're expressly requested or not.

2. Give each letter a unique number in your database so you can cross reference the letter to the recipient information (including but not limited to the address) you have stored in your system. The letter may be returned with something else (e.g. another letter) attached so it's important to keep that information correlated.

3. Scanning the original letter is a low cost way to maintain this correlation. When the letters are returned you scan them then send them through a program you have set up to update the system accordingly. The program uses some primitive OCR and probably a checksum to automatically recognize the codes in the original letters. I can imagine this being used to automatically mark bad addresses if a letter is returned without additional context, but its main purpose is probably to route the letter - and any attachments, like other letters - to the appropriate agent.

To support a workflow not unlike the one described above, it is requested that the unique number that identifies the letter be left unobscured. This way OCR can do its job, deliverability issues can be flagged with minimal human involvement, and replies to letters can be put in front of the right person without creating too much organizational overhead.

I took that to mean savings in developer hours (reduced, not eliminated) and assume the subscription price will be nominal at its lowest tier / usage.

No, you’re right, it was chosen because “trust me bro”.

Look, it may well be something he believes, and he’s free to prognosticate (or market) however he likes, but I see absolutely nothing to support the number outside of his own opinion.

Besides, there’s no time limit on p(doom), so it’s completely unfalsifiable (“on a long enough timescale…”), and it’s about the destruction of humanity which means it’s unprovable as well. That, in my view, makes his 70% guess a sensational statement lacking scientific merit.

when compared against other 4b and 8b parameter models I would genuinely champion the quality of their responses

You clearly have some very specific models in mind. Even if the latest 4B and 8B models don’t move the needle on the “results you would champion” metric, this does not advance your argument that the state of the art hasn’t significantly progressed from 5 years ago.

I would legitimately argue

I’ll bet you would!

He added a lesser option, catastrophically harming humanity, so whatever he meant by the first is immaterial (“there’s a 70% chance of a hurricane or strong winds”). Furthermore, if it wasn’t a high number chosen for dramatic effect the estimated percentage would be completely arbitrary.

No, the number is made up and the facts don’t matter so the statement can easily be reimagined as an ad lib.

There’s a [arbitrary number] percent chance that [technology] will destroy or catastrophically harm humanity

Try these: social media, the Internet, the large hadron collider, Starlink, Neuralink, iPhones, iDrones, quantum computers, regular computers, the 2038 bug, the Y2K bug, electric cars, gasoline cars, the great firewall of China, the not so great firewalls of asbestos, mRNA technology, gain of function research, nuclear bombs, nuclear energy, paper clip manufacturers, scissors.

I’m not saying it’s true that these have a 70 percent chance of destroying or catastrophically harming humanity, but couldn’t you make the argument?

Sam Altman is revered as a tech god in this forum

I don’t think that’s true.

I will likely get downvoted, but there's something deep within his character that doesn't seem genuine or sit well with me

That’s actually an extremely popular opinion; see for example just about every recent article that’s been posted about him.

Can a collection of around 1.5 billion interconnected cells that predictably respond to signals in their environment using simple rules? How about 86 billion? 36 trillion?

These are ballpark counts of cells in crow’s brain, a human’s brain, and a human body. The question is, is it the cells themselves doing the reasoning and planning, or are they just the machinery this disembodied process happens to be running on? I’d argue intelligence is a distributed phenomenon that our DNA is as much a party to as our brains.

Certainly the question of whether humans use DNA to reproduce or DNA uses humans is a matter of perspective.

Well, yes, reasoning and planning abilities exist on a spectrum, so it isn’t so much a matter of where to draw the line as a question of degree. As for LLMs, I think their reasoning and planning is some of the most powerful and human-like we’ve seen so far, even if the hidden mechanisms and constraints are different (in some cases, more limited, but in others, vastly superior).

Our brains however are highly modular (a “committee of idiots”) so who’s to say a portion, and even a significant one, doesn’t operate on similar principles?

It’s been downgraded to “a preponderance of evidence” from “beyond a reasonable doubt”.

They absolutely can reason and plan; how do you suppose they predict the next token?

That they’re not autonomously solving complex tasks is a bit of a straw man though, and with a bit of creativity we can easily imagine them being combined with models and modalities that do provide executive function and autonomy.

My tin foil hat is telling me that no publicity is bad publicity and LLMs are great at spinning dramatic yarns.

A “SSHal credit score” tied to a pooled resource, yes, that will work out well! Kind of like how a used car purchase should come with all its tickets!

EDIT: To this feature’s credit, it’s not federated centrally, so a DDOS to nuke IP reputation would have its blast radius limited to the server(s) under attack.

We used to use them for this all the time!

Yea, taking 5 seconds to rinse the bag out in the sink is _such_ an inconvenience.

The old "you're just not doing it right" chestnut, where have I heard that one before! Hey, by any chance have you noticed how many of those silly face masks made it (or didn't make it) to landfills?

I stopped buying meat . . . Guess I'm just a stupid hippie.

You said it not me!

There’s huge business interest pushing the usage of environmentally damaging products forward because it generates money.

The converse (pushing green products) is also true so we have to consider whether the environmental impact is as big as is claimed, whether it's worth it even if it is (single issue reductivism is a dangerous way to craft policy), or whether banal financial and power incentives are the true moving force behind the lobbied solution.

I can't trust a plastics poll that doesn't also ask who wants the old straws back.

I’m taking this with a grain of salt because 85% of people adopting such a hardline environmentalist stance defies reason. I think the poll’s design and authors probably had something to do with it. Context matters!

Haha, sure, you can call it that if you want, but foolish is cousin to fun, so one application of this tech would be as a comically overwrought way of communicating subtext to an adversary who may not be able to read between the lines otherwise. Imagine using all this highly sophisticated and expensive technology just to write "you're an asshole" to some armchair intelligence analyst who spent their afternoon and monthly token quota decoding your secret message.

Seed for the message above is 42 by the way.

(Just kidding!)

If you make it do double duty as a poor-man's encryption, you are going to have a bad time.

For the serious use cases you evidently have in mind, yes, it's folly to have it do double duty, but at the end of the day steganography is an obfuscation technique orthogonal to encryption, so the question of whether to use encryption or not is a nuanced one. Anyhow, I don't think it's fair to characterize this elaborate steganography tech as a poor-man's encryption — LLM tokens are expensive!

I was imagining the message encoded in clear text, not encrypted form, because given the lengths required to coordinate protocol, keys, weights, and so on, I assumed there would be more efficient ways to disguise a message than a novel form of steganography. As such, I approached it as a toy problem, and considered detection by savvy parties to be a feature, not a bug; I imagined something more like a pirate broadcast than a secure line, and intentionally ignored the presumption about the message being encrypted first.

That being said, yes, some of my assumptions were incorrect, mainly regarding temperature. For practical reasons I was envisioning this being implemented with a third party LLM (i.e. OpenAI's,) but I didn't realize those could have their RNG seeded as well. There is the security/convenience tradeoff to consider, however, and simply setting the temperature to 0 is a lot easier to coordinate between sender and receiver than adding two arbitrary numbers for temperature and seed.

I misspoke, or at least left myself open to misinterpretation when I referred to the LLM's weights as a "secret key"; I didn't mean the weights themselves had to be kept under wraps, but rather I meant that either the weights had to be possessed by both parties (with the knowledge of which weights to use being the "secret") or they'd have to use a frozen version of a third party LLM, in which case the knowledge about which version to use would become the secret.

As for how I might take a first stab at this if I were to try implementing it myself, I might encode the message using a low base (let's say binary or ternary) and make the first most likely token a 0, the second a 1, and so on, and to offset the risk of producing pure nonsense I would perhaps skip tokens with too large a gulf between the probabilities for the 1st and 2nd most common tokens.

The weights of the LLM become the private key (so it better be a pinned version of a model with open weights), and for most practical applications (i.e. unless you're willing to complicate your setup with fancy applied statistics and error correction) you'd have to use a temperature of 0 as baseline.

Then, having done all that, such steganography may be detectable using this very tool by encoding the difference between the LLM's prediction and ground truth, but searching for substrings with low entropy instead!

What you described sounds like a very cool idea - LLM-driven text steganography, basically - but intentional obfuscation is not the problem this tool is trying to solve. To your point about secrets with entropy similar to the surrounding text, however, I wonder if this can pick up BIP39 Seed Phrases or if whole word entropy fades into the background.

I wonder if human learning will be similarly impaired.

My guess is the models will be ideologically driven and will robotically enact the governing agenda without humanity, compassion, or understanding. Hmmm.

unskilled, underpaid employees who don’t care

One of these things is not like the others. The models may not ask for a raise but they’ll no doubt go to work finding all sorts of other excuses to raise taxes.

I find the self-serve UI and overall experience with GCP to be superior, so to me it’s a bit like the premium you pay for a Mac over a PC. Granted I never needed or wanted to speak to a human, but their infamy in that respect deserves a special nod of acknowledgment.

It still takes a surprising amount of labour and artistry to get an AI to give you exactly what you want (see also: Pareto principle). Consumer expectations scale proportionally to technological progress, so demand for premium assets and brand differentiators won’t be going away anytime soon; the state of the art has advanced but stock photo business will advance right along with it.