Hacker News Reader

> even in the 70s people recognized that supersonic travel had real concrete issues with no solution in sight. I don't think LLMs share that characteristic today

I think they pretty strongly do

The solution seems to be "just lower your standards for acceptable margin of error to whatever the LLM is capable of producing" which should be concerning and absolutely unacceptable to anyone calling themselves an Engineer

> absolutely unacceptable to anyone calling themselves an Engineer

Isn’t that exactly what engineers do? Even very strong bridges aren’t designed to survive every possible eventuality.

That's what a "margin of error" is. The margin of error of a bridge is predictable thanks to well-established techniques of physical analysis.

An LLM system, on the other hand, can fail because you moved some punctuation around.

> An LLM system, on the other hand, can fail because you moved some punctuation around

An LLM system can fail without changing anything, it could just fail more or less randomly without any way to diagnose why it happened

No

I'm talking about engineering a bridge for 50 cars that collapses at 51, not engineering a bridge for 500 cars that is only expected to get 50

Engineering does require tradeoffs of course. But that's not what the minimum possible quality is

99% or more of software developers behave in ways that would be inconceivable in actual engineering. That's not to say there aren't software engineers, but most developers aren't engineers and aren't held to that standard.

Code is not physical. While computation errors can have real effects, a lot of orgs and people are resilient about them.

“Increasing Success by Lowering Expectations” That is from Despair Inc. This was obviously meant to be funny by them, now it looks like the state of play.

> The reason why I partially note this as wrong is that even in the 70s people recognized that supersonic travel had real concrete issues with no solution in sight. I don't think LLMs share that characteristic today.

The fundamental problem has already been mentioned: Nobody can figure out how to SELL it. Because few people are buying it.

It's useful for aggregation and summarization of large amounts of text, but it's not trustworthy. A good summary decreases noise and amplifies signal. LLMs don't do that. Without the capability to validate the output, it's not really generating output of lasting value. It's just a slightly better search engine.

It feels like, fundamentally, the primary invention here is teaching computers that it's okay to be wrong as long as you're convincing. That's very useful for propaganda or less savory aspects of business, but it's less useful for actual communication.

> Nobody can figure out how to SELL it. Because few people are buying it.

Just picking one company who basically just does AI, OpenAI. They reported it has 20 million PAID subscribers to ChatGPT. With revenue projected above $12b dollars (https://www.theverge.com/openai/640894/chatgpt-has-hit-20-mi...).

I think what you meant to say is that costs are high so they can't generate large profits. but saying that they can't figure out how to sell it seems absurd. Is it Netflix level of subscribers, no. But there can't be more than a couple of hundred products that have that type of subscription reach.

Ok but isn’t 20 million subscribers out of what, 800 million or 1 billion monthly users or whatever they’re claiming, an absolutely abysmal conversion rate? Especially given that the industry and media have been proclaiming this as somewhere between the internet and the industrial revolution in terms of impact and advancement? Why can they not get more than 3% of users to convert to paying subscribers for such a supposedly world changing technology, even with a massive subsidy?

Because they give too much of it away for free? Most casual use fits into the very generous free tier.

Ok so the argument is that all the model builders either suck at business or they are purposefully choosing to lose billions of dollars?

They are purposely losing billions, this is a growth phase where all of the big AI companies are racing to grow their userbase, later down the line they will monetize that captured userbase.

This is very similar to Uber which lost money for 14 years before becoming profitable, but with significantly more upside.

Investors see the growth, user stickiness and potential for the tech; and are throwing money to burn to be part of the winning team, which will turn on the money switch on that userbase down the line.

The biggest companies and investors in the planet aren't all bad at business.

I'd say the userbase has grown. You can't claim half a billion users and simultaneously say you're still trying to grow. This isn't a month-old technology now. And they still can't turn a profit. (edit: and by "you" i meant "they")

>You can't claim half a billion users and simultaneously say you're still trying to grow.

You can if you're still growing. ChatGPT is the 5th most visited site on the planet yes, but it is still growing hundreds of millions of visits with every passing month.

They aren't turning a profit because they aren't monetizing the vast majority of subscribers in any way (not even ads). LLM inference is cheap enough for ads to be viable.

As another commenter notes, because you get access to a lot of functionality for free. And other providers are also providing free alternatives. The ratio for their free/paid tier is about the same as YouTube's. And like YouTube, it's not that YouTube isn't providing great value, but rather that most people get what they need out of the free tier.

The better question is what if all LLM services stopped providing for free at all -- how many paid users would there then be?

You could say the same of Dropbox. Or Gmail.

True, although I don't think Dropbox or Gmail's operating costs to support those free users are anywhere near those of OpenAI.

A service like Gmail or Dropbox with low storage is close to free to operate. Same thing with iCloud - 50 gigs a month is what, 1 dollar? How is that possible?

Because 50 gigs is next to nothing, and you only need a rinky dink amount of compute to write files.

YouTube, on the other hand, is actually pretty expensive to operate. Takes a lot of storage to store videos, never mind handling uploads. But then streaming video? Man, the amount of bandwidth required for that makes file syncing look like nothing. I mean, how often does a single customer watch a YouTube video? And then, how often do people download files from Dropbox? It's orders of magnitude in difference.

But LLMs outshine both. They require stupid amounts of compute to run.

Close to free per user, maybe. But dropbox has 800 million users, only ~2% pay, and Gmail has billions. They spend a lot of money running those services.

In my companies, AI subscriptions and API access are now the biggest costs after salaries and taxes. Don't know what makes you think these services aren't attracting paid customers?

> even in the 70s people recognized that supersonic travel had real concrete issues with no solution in sight. I don't think LLMs share that characteristic today

I hate to dogpile on this statement but I can think of two major issues right now:

* Small context windows, and serious degradation when pushing the limits of existing context windows. A human can add large amounts of state to their "context window" every day.

* Realtime learning. My humans get smarter every day, especially in the context of working with a specific codebase.

Maybe the AI companies will figure this out, but they are not "same technique more processor power" kinds of problems.

There are sound math reasons for skyscrapers topping out, mostly due to elevator capacity and the inability to effectively get people in and out of the floorspace as you go past a few hundred ft. There's no construction engineering reason you can't go taller - the Burj Khalifa, for example, is three times taller than a typical Western major city skyscraper - it just doesn't make economic sense unless you're a newly rich nation looking to prove a point.

Economic Concrete construction (what China specializes in) typically tops out at 30-40 floors, so the vast majority of buildings in Asia are that height, a sweet spot so to speak especially for residential (even in limited space HK).

>I think this is both right and wrong. There was a good book that came out probably 15 years ago about how technology never stops in aggregate, but individual technologies tend to grow quickly and then stall. Airplane jets were one example in the book. The reason why I partially note this as wrong is that even in the 70s people recognized that supersonic travel had real concrete issues with no solution in sight. I don't think LLMs share that characteristic today.

I don't see any solution to hallucinations, nor do I see any solution in sight. I think that could count as a concrete issue that would stop them.

Vision and everyday-physics models are the answer: hallucinations will stop when the models stop thinking in words and start thinking in physical reality.

The way you phrased it reveals how these model providers have framed the conversation in their favor. Models don’t “think.”

They had easy access to a large corpus of writing to train on, way larger than any human being trained their own language model on. I can't see where they are going to find a large corpus of physical interaction with reality to train that kind of model.

cameras are fairly cheap these days.

Yeah, and with LLMs the thing I can't shake, however, is that this time it's pretty strongly (maybe parasitically) latched onto the aggregate progress of Moore's law. Few other technologies have enjoyed such relatively unfettered exponential improvement. It's like if skyscraper materials double in strength every n years, and their elevators approach teleportation speed, the water pumps get twice as powerful, etc., which would change the economics vs the reality that most of the physical world doesn't improve that fast.

Was the problem that supersonic flight was expensive and the amount of customers willing to pay the price was even lower than the number of customers that could even if they wanted to?

From what I had read in passing and remember.

  - They were loud (sonic booms were nasty).

  - They were expensive to maintain and operate. Guzzlers. (Britain and France clung to them as a matter of pride/ego)

  - They were narrow and uncomfortable. I have seen videos where there is space only for one stewardess to walk. I had been inside of one in Seattle museum. Very cramped.

  - As you mentioned, ticket cost was high.

  - I suspect people traveled in these mostly for bragging rights.

You made this point in passing, but it's so relevant to LLMs I wanted to highlight it: The development and operational cost was heavily subsidized by the British and French governments, because having an SST was a point of national prestige.

Yeah, basically. Nobody wanted to pay $12,000 to be in a plane for three hours when they could pay ~$1200 to be in one for six hours. Plus, they used up a lot of fuel. That made them real vulnerable to oil price spikes.

Contrast that with modern widebody jets, which fly ~300 people plus paid cargo on much more fuel-efficient engines.