Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

How the AI Bubble Will Pop

derekthompson.org

Post reply on HN
Print view View original

331–340 of 531 posts

Re: How the AI Bubble Will Pop

#331

Post by ACCount37 »

emp17344 wrote: post #276
ACCount37 wrote: post #260

Earlier quoted context omitted.

LLMs can have surprisingly strong "theory of mind", even at base model level. They have to learn that to get good at predicting all the various people that show up in conversation logs. You'd be surprised at just how much data you can pry out of an LLM that was merely exposed to a single long conversation with a given user. Chatbot LLMs aren't trained to expose all of those latent insights, but they can still do some…

Do you have evidence to support any of this? This is the first time I’ve heard that LLMs exhibit understanding of theory of mind. I think it’s more likely that the user I replied to is projecting their own biases and beliefs onto the LLM.

Basically, just about any ToM test has larger and more advanced LLMs attaining humanlike performance on it. Which was a surprising finding at the time. It gets less surprising the more you think about it.

This extends even to novel and unseen tests - so it's not like they could have memorized all of them.

Base models perform worse, and with a more jagged capability profile. Some tests are easier to get a base model to perform well on - it's likely that they map better onto what a base model already does internally for the purposes of text prediction. Some are a poor fit, and base models fail much more often.

Of course, there are researchers arguing that it's not "real theory of mind", and the surprisingly good performance must have come from some kind of statistical pattern matching capabilities that totally aren't the same type of thing as what the "real theory of mind" does, and that designing one more test where LLMs underperform humans by 12% instead of the 3% on a more common test will totally prove that.

But that, to me, reads like cope.

Re: How the AI Bubble Will Pop

#332

Post by throaway1988 »

throwaway-0001 wrote: post #201
Diti wrote: post #190

Earlier quoted context omitted.

LLMs cannot think on their own, they’re glorified autocomplete automatons writing things based on past training. If the “AI figured out something on your mind”, it is extremely likely the “thing on your mind” was present in the training corpus, and survivorship bias made you notice.

Tbh if Claude is smarter than average person, and it is, then 50% of the population is not even a glorified auto complete. Imagine that, all not very bright.

idk, how many people in the world have been programmed with a massive data set?

Re: How the AI Bubble Will Pop

#333

Post by km144 »

jstummbillig wrote: post #88
therealmarv wrote: post #46

It reminds me what I said to somebody recently: All my friends and family are using the free version of ChatGPT or something similar. They will never pay (although they have enough money to do so). Even in my very narrow subjective circles it does not add up. Who pays for AI and how? And when in the future?

People are always so fidegty about this stuff — for super understandable reason, to be clear. People not much smarter than anyone else try to reason about numbers that are hard to reason about. But unless you have the actual numbers, I always find it a bit strange to assume that all people involved, who deal with large amounts of money all the time, lost all ability to reason about this thing. Because right now that…

I think it's a bit fallacious to imply that the only way we could be in an AI investment bubble is if people are reasoning incorrectly about the thing. Or at least, it's a bit reductive. There are risks associated with AI investment. The important people at FAANG/AI companies are the ones who stand to gain from investments in AI. Therefore it is their job to downplay and minimize the apparent risks in order to maximize potential investment.

Of course at a basic level, if AI is indeed a "bubble", then the investors did not reason correctly. But this situation is more like poker than chess, and you cannot expect that decisions that appear rational are in fact completely accurate.

Re: How the AI Bubble Will Pop

#334

Post by triceratops »

_heimdall wrote: post #278
jennyholzer wrote: post #71

Earlier quoted context omitted.

name the podcasts

This is an odd anecdote to ask "show your work." I don't catalog shows and episodes where any particular topic comes up, and I follow over 100 podcasts so I don't have a specific list you can fact check me on. Personally I could care less if that means you choose not to believe that I hear the Taiwan risk come up often enough.

> I follow over 100 podcasts

How? Do you read summaries? Listen at 3x speed 5 hours a day?

Re: How the AI Bubble Will Pop

#335

Post by fred_is_fred »

ninetyninenine wrote: post #211
driverdan wrote: post #188

I keep seeing articles like this but does anyone actually think we're not in a bubble? From what I've seen these companies acknowledge it's a bubble and that they're overspending without a way to make the money back. They're doing it because they have the money and feel it's worth the risk in case it pays off. If they don't spend, another company does, and it hits big they will be left behind. This is at least insura…

HN isn't always right. There was massive pushback against self driving and practically everyone was saying it would fail and is a bubble. The level of confidence people had about this opinion was through the roof. Like people who didn't know anything would say it with such utter confidence it would piss me off a bit. Like how do you know? Well they didn't and they were utterly wrong. Waymo showed it's not a bubble. A…

What was promised with self-driving and what we have are orders of magnitude off. We were promised fleets of autonomous taxis - no need to even own a car anymore. We were told truck drivers would be replaced en-masse and cargo would drive 24x7 by drivers who never needed breaks. We were told downtown parking lots would disappear since the car would drop you off and drive to an offsite lot and wait for you. In short a complete blow up of the economy with millions of jobs in shipped lost and hundreds of billions of spend on new autonomous vehicles.

None of that happened. After 10 years we got self-driving cabs in 5 cities with mostly good weather. Cool, yes? Blowing up the entire economy and fundamentally changing society? No.

Re: How the AI Bubble Will Pop

#336

Post by danaris »

throwaway-0001 wrote: post #308
danaris wrote: post #289

Earlier quoted context omitted.

That "if" is doing literally all the work in that post. Claude is not, in fact, smarter than the average person. It's not smarter than any person. It does not think. It produces statistically likely text.

Well, I disagree completely. I think you have no clue how’s the average person or below. Look at instagram or any social media ads, they are mostly scams, AI can figure out but most people don’t. Just an example.

I don't have to know how smart the average person is, because I know that an LLM doesn't think, isn't conscious, and thus isn't "smart" at all.

Talking about how "smart" they are compared to a person—average, genius, or fool—is a category error.

Re: How the AI Bubble Will Pop

#337

Post by dimava »

Gerardo1 wrote: post #283
steveklabnik wrote: post #250

Earlier quoted context omitted.

Many of the companies (including OpenAI) have even claimed the opposite. Inference is profitable; it's R&D and training that's not.

It's not reasonable to claim inference is profitable when they've also never released those numbers. Also the price they charge for inference is not indicative of the price they're paying to provide inference. Also, at least in openAI's case, they are getting a fantastic deal on compute from Microsoft, so even if the price they charge is reflective of the price they pay, it's still not reflective of a market rate.

DeepSeek on GPUs is like 5x cheaper then GPT

And TPUs are like 5x cheaper then GPUs, per token

Inference is very much profitable

Re: How the AI Bubble Will Pop

#338

Post by Workaccount2 »

manishsharan wrote: post #253
fabian2k wrote: post #74

Earlier quoted context omitted.

AI can be quite impressive if the conditions are right for it. But it still fails at so many common things for me that I'm not sure if it's actually saving me time overall. I just tried earlier today to get Copilot to make a simple refactor across ~30-40 files. Essentially changing one constructor parameter in all derived classes from a common base class and adding an import statement. In the end it managed ~80% of t…

This is exactly my experience. We wanted to modernize a java codebase by removing java JNDI global variables. This is a simple though tedious task. And we tried Claude Code and Gemini. Both of these results were hilarious.

LLMs are awful at tedious tasks. Usually because it involves massive context.

You will have much more success if you can compartmentalize and use new LLM instances as often as possible.

Re: How the AI Bubble Will Pop

#339

Post by torginus »

reilly3000 wrote: post #213

I don’t like that he cited $12B in consumer spending as the benchmark for demand. Clearly enterprise spending has and will continue to dwarf consumer outlays, to the tune of $100b+ in 2025 on inference alone, and another $150b on AI related services. I see almost no scenario where the value of this hardware will go away. Even if the demand for inference somehow declines, the applications that can benefit from hardwar…

Also, as others have pointed out, if the next Pixel phone or iPhone has 'AI' as a bullet point feature, then people buying and iPhone will count as 'consumer AI spend', that's why they're forcing AI into everything, so they can show that people are using AI, while most people are ambivalent or hostile towards AI features.

Re: How the AI Bubble Will Pop

#340

Post by thisisit »

aurareturn wrote: post #11
hemloc_io wrote: post #7

The most frustrating thing to me about this most recent rash of biz guy doubting the future of AI articles is the required mention that AI, specifically an LLM based approach to AGI, is important even if the numbers don't make sense today. Why is that the case? There's plenty of people in the field who have made convincing arguments that it's a dead end and fundamentally we'll need to do something else to achieve AGI…

Where's the business value? Right now it doesn't really exist, adoption is low to nonexistent outside of programming and even in programming it's inconclusive as to how much better/worse it makes programmers. I have a friend who works at PwC doing M&A. This friend told me she can't work without ChatGPT anymore. PwC has an internal AI chat implementation. Where does this notion that LLMs have no value outside of progr…

> This friend told me she can't work without ChatGPT anymore

I am curious what kind of work is she using ChatGPT such that she cannot do without it?

> ChatGPT released data showing that programming is just a tiny fraction of queries people do

People are using it as search engine, getting dating advice and everything under the sun. That doesn't mean there is business value - so to speak. If these people had to pay say $20 a month for this access, are they willing to do so?

The poster's point was that coding is an area which is paying consistently for LLM models so much that every model has a coding specific version. But we don't see same sort of specialized models for other areas and the adoption is low to nonexistent.

Post reply on HN