Will A.I. Be a Bust? A Wall Street Skeptic Rings the Alarm

What jobs are being lost? Show us the data.

https://en.wikipedia.org/wiki/Brandolini%27s_law

> People forget that there was a shit ton of investment in the early internet that also didn't pay off

Yes, those were called busts too. It’s not a value judgement, it’s just the term for that moment when lofty expectations don’t pan out.

That's definitely responsible for billions of dollars in salaries. But also it's a bit of an arms race. Like the standards will just go up for everything evenly right?

Investors are throwing unfathomable amounts of money at AI right now, and frontier models are very expensive to build. There’s also the issue of these AI-as-a-Service companies requiring an ever increasing amount of GPUs and electricity, and because frontier models keep getting bigger, the problem keeps getting worse.

There is a serious risk to these companies. A frontier model doesn’t sit at the frontier for very long. Given a year or two, open source models can catch up for less than 1% of the money spent to build a frontier model. The moat doesn’t last long, so you need to keep burning tens of billions to stay ahead.

But as open source models continue to improve, we run the risk of reaching a moment that looks like the current smartphone market. People aren’t upgrading their phones as often anymore because a 4 year old smart phone is still pretty great. There’s no need to buy every year. Well, at some point an open source model will be close enough to a frontier model that many users will just stick with it. They’ll prefer it because there’s no monthly fee to run it on their own hardware, because there are no rate limits, and because they prefer the privacy and flexibility.

Right now you need a big cluster of GPUs to run top tier models. But with every passing year we get closer to good-enough open source models running on gaming GPUs.

And when that happens, would you want to be one of the investors shoveling tens of billions of dollars into building frontier models with declining subscription numbers where you’ve got no moat?

I don’t think the economics of this are going to pan out in the long run. Everyone is still hyped about ChatGPT, so investors are happy to throw money at it. But I think we’re in a very big AI bubble right now. The economics only make sense if someone achieves AGI within the next couple of years. If they don’t, this whole thing is going to decline sharply as open source models provide a better free alternative.

o1 cannot do my job, and hallucinated immediately.

chatgpt has been here for what ? 24 months ? Unemployment rate is the same as ever, maybe a bit on the lower side if anything. If it was the miracle they promised we would see it everywhere: gdp, unemployment rate, productivity, &c.

Have you heard the good news? The price per million tokens has fallen from $36 to $0.25 in the past 18 months.

> actually having to abide by copyright

Copyright is being eroded, first by the internet and its copying capabilities, then by the social networks and sharing, and now by generative AI. You can't make a profit owning expression in this distributed and interactive system, copyright was invented in the age of passive consumption of content. When you do a simple search you can find many alternatives to choose from. Any new work has to compete against decades of accumulation.

On top of that, AI is not really infringing, or it's a bad infringement system - it's much slower and more imprecise than just copying, if infringement is what someone seeks. It's a guided remixing tool, there's a human shaping its output. How can a model 1000x smaller than its training set be really infringing, it has no space for all the exact details.

OpenAI accumulates trillions of tokens in chat logs, eliciting experience from its users, it is crawling the tacit knowledge reservoir of its user base. That means it collects better information than web scrapers, information that would be normally lost. They can just sit there and people will bring data, guidance and feedback to the model. Do you see the business value in collecting practical experience and repackaging it?

As an ML engineer and AI developer, I don’t see the real value at all, not to mention the added cost of using LLM

I think it's the latter. They were absolutely right about the internet changing the world.

Uses != pays. I personally know only one person who pays for it, and dozen who use it for free

Spotify by comparision has 250 million paying subscribers and is currently valued at 75 billion. OpenAI is at 150 billion for 11 million subscribers ?

We're "parroting" things we don't really understand. When we take a pill, we don't know what's inside. When we go to the doctor, we don't study medicine first. When we write an app, we don't know the minute details of the operating system or frameworks we use. As software developers we know the limits of abstraction and that no abstraction works everywhere.

Our society is functional, not based on genuine understanding. Can't be any other way. There is no central understanding, not in the brain and not in society. Remember the parable of the Elephant and the blind men? That's us.

I pay OpenAI and Anthropic directly for API access that is hooked into a plugin on my IDE. I also use Chatbox, an app that lets me chat with them, using the API directly.

No idea if they are making money on my use of their services, they haven't made that information public.

If the price was higher I would still pay it. Given that they keep offering lower prices on newer models, I am inclined to say that they are making money on the incremental use of the models (which is basically the cost of renting a GPU). No idea if they are making enough to recoup the cost of research.

What novels are you reading?

This is fascinating to me, because the world is quickly becoming a place where we have to choose which information from the unlimited information stream to consume. It feels like unlimited opportunity cost. I, for one, don't think I'll ever have enough time to watch every Academy Award nominated film (let alone all of the winners). And that's just one type of information.

You're going after some obscure (?) stuff. What brought about the interest?

Do you personally know Mandarin well enough to judge the quality of a translation? Because you can't do that from just looking at the output.

It is freaking amazing that this works almost perfectly. Seriously, it’s mind blowing. The problem is, when you keep in mind how the technology works, you realize that the “almost” can never be removed. That’s fine for some use cases but not for others. I understand that human translators make mistakes but they have a conception of truth and correctness, that matters.

If you go to plain ChatGPT, a system not specifically designed to translate languages, and tell it to translate "以后你再使用我照片请使用这两张任何一张都可以这是我们结婚的照片" to English, you get a better result than any machine translation from just a few years ago. For example, it gets from context that "这是" has to be translated to a plural phrase in English. Even right now, Google Translate still gets this wrong.

I'm worried that a lot of the impact these technologies have will eventually turn out to be overwhelmingly bad. Google Photos is already partially broken by the amount of shitty AI images it returns. But the fact that they do have a huge impact can't be denied.

I don't know what exactly qualifies something as "changing the world", but if LLMs don't qualify, then not a lot of things do.

1) Most people have no need for translation software

2) Before LLMs we already had decent free translation in the form of the free Google Translate, using pre-transformer NN models

Personally I still use Google Translate as my go-to for translation, rather than using Sonnet 3.5. Maybe Google now use an LLM under the hood, but I haven't noticed any increase in quality in last few years.

Stock traders are so abstracted from real human value, that it is not a stretch to say that their jobs do not add value as a whole to the world. If we had half as many people working in financializing the world, capital allocation would not suffer. At least that's my belief.