Goldman Sachs: AI Is overhyped, expensive, and unreliable
404media.co
404media.co
(88 points, 12 days ago, 157 comments) https://news.ycombinator.com/item?id=40837081
(89 points, 1 week ago, 47 comments) https://news.ycombinator.com/item?id=40885632
But it's only overhyped if AI stays that way.
The hype isn't coming from where AI is today, but where it will be in 2030 and 2040.
What excites the true believers is the slope, not the intercept.
The nearly $1T in investment hype since Q42022 was chasing the possibility that revolutionary commercialization of generative AI might arrive any day. That no longer seems likely.
As you note, there seem to be at least a few more profound discoveries that need to be made and matured, and so that revolution is still a long way out (from the perspective of these investors), even if it looks meaningfully more possible than it did two years ago.
You are free to disagree, but don't pretend like you don't understand what the AI crowd is saying.
March 2021 yeah, well known all over the world, everyone talks about it but not much has really changed in a while. Some fear COVID might mutate into something deadlier, and some think generative AI will turn into something completely different.
The quick pace was GPT-2 and GPT-3, those happened 2019 and 2020, today it is a crawl in comparison, GPT-3 was like covid in march-2020, at the time most were barely aware of it but today it is mainstream.
The revolution you are talking about has already happened, people all over the world interact with their computer using natural language, so your prediction here is a bit late.
Not much unexpected will happen now unless a new breakthrough is made. Similarly covid in march 2021 were well understood and not much would happen unless it mutated into a much deadlier strain, which of course is unlikely to happen. Covid in march 2021 was still a massive thing, as is AI today, but just that it is a massive well understood thing not something that will completely change the world in the future.
By the second week of March I was well aware of COVID. By the third I was locked in my apartment.
In January someone at the airport in Bangkok thought it was important to scan my body as I arrived, but I had no idea why
The slope changed between 1990 and 2000.
There's a pattern somewhere, here.
Arguing by example is stupid. Plenty of other examples have been proposed elsewhere in this thread. My only contribution is that I think LLMs are not going to yield anything more than incremental efficiency gains. If the only result after 7 years is that NVidia became a new $1T company and no one else then I think it should be becoming more obvious that this is a gold rush situation and not an iPhone situation.
[1] - https://en.wikipedia.org/wiki/Aviation_in_World_War_I#cite_n...
What a coincidence that we have people trying to ban AI at a domestic and international level
> What do we have? Better autocomplete? Worse search results?
It's ridiculous to minimize LLMs as better autocomplete.
Also I get it, for most people remembering how things were more than a couple of years ago is hard. Can you even remember when you couldn't talk to computers?
Software mainly seems to enable new business models, of which some tangentially improve our quality of life, others are actively harmful. Contrasting this to advances in technologies like solar panels or electric cars almost makes me wish I had chosen a different career path.
It's more that by 2040 AI/robots will be considerably smarter than humans and we can kick back while they build us palaces and yachts. Not that that will definitely happen but it's not a business as usual scenario.
> What excites the true believers is the slope, not the intercept.
The phrase "true believers" jumps out to me here. It sure makes AI (well, ML) sound more like a religion than a field of technological research.
Yes, because what OpenAI releases today helps us improve the prediction of what it may release in a decade or so.
It does neither. What OAI or others release next does, in the very best case, help set a floor. And if it's a blind alley to a local maximum, even that isn't super-helpful.
Each release of an LLM is largely a black box. We don't know exactly how they work, what they learned, why they learned it, or what their limitations are.
How exactly does that help us predict what could exist in a decade or two? And how do we align that predictive ability with the fact that so many people working in the industry would not have been able to predict where we are today a decade or two in the past?
Scott McNealy circa 2002[1][2] had some relevant food for thought:
>> But two years ago we were selling at 10 times revenues when we were at $64. At 10 times revenues, to give you a 10-year payback, I have to pay you 100% of revenues for 10 straight years in dividends. That assumes I can get that by my shareholders. That assumes I have zero cost of goods sold, which is very hard for a computer company. That assumes zero expenses, which is really hard with 39,000 employees. That assumes I pay no taxes, which is very hard. And that assumes you pay no taxes on your dividends, which is kind of illegal. And that assumes with zero R&D for the next 10 years, I can maintain the current revenue run rate. Now, having done that, would any of you like to buy my stock at $64? Do you realize how ridiculous those basic assumptions are? You don't need any transparency. You don't need any footnotes. What were you thinking?
>> I was thinking it was at $64, what do I do? I'm here to represent the shareholders. Do I stand up and say, "Sell"? I'd get sued if I said that. Do I stand up and say, "Buy"? Then they say you're [Enron Chairman] Ken Lay. So you just sit there and go, "I'm going to be a bum for the next two years. I'm just going to keep my mouth shut, and I'm not going to predict anything." And that's what I did.
Yeah but the investors they're talking to aren't thinking about 2030-2040. They are thinking about 2025. If they don't get their 2000% profits they are imagining now they will declare it a FLOP!!!! and drop it. Doing harm to the overall development of a technology that does have its merits but is just propped up to a position it's not ready for by a long shot.
With this kind of pressure a slow slope is not enough to keep investors happy.
These guys are all hitting themselves in the head every day that they didn't buy bitcoins when they cost $0.01 and they are constantly telling themselves the next big thing will make them rich beyond their wildest dreams. They are ruining good tech this way. Metaverse was pretty decent too, it was just not for everyone yet (and it won't be for a loooong time), but it does have its niche uses.
I'm among the AI true believers but the change above made me have second thoughts.
Now crypto is used primarily for crime and speculation, which is exactly what we'll be using GenAI for in a few years, especially as it poisons its own source of training data.
I’d argue that transformer-driven AI could follow the exact same trajectory: hype cycle driven by people imagining legitimately transformative use cases, which is ultimately popped when we realize the actual here-and-now implementations don’t work that way.
The idea of transferring value over the internet using one common, frictionless, ownerless system isn’t dead; it can’t be, because it’s simply too good of an idea. But nothing actually offers that. Actual artificial intelligence is, likewise, too good of an idea to ever be “dispelled” or dismantled. The only room for doubt is whether the technical designs we have today are actually AI… or just some wannabe scam coins that LOOK like AI.
GenAI is used by people for all sorts of purposes. Anecdotally, it has largely replaced my own usage of traditional search engines.
This strikes me as a false equivalence the same way comparisons to the dot com bubble do.
Depending on the search, I either don’t care, or I use something like perplexity, which includes its references.
I don't think the current generation of AI will go away -- there are too many use cases for it. I think it will take some time as people and companies experiment with it, find what works and what doesn't. Then, some use cases will disappear but others will remain -- like how autocorrect came from previous generation of AI tech.
---
With the code autocomplete it works best when there is repetitive logic such as doing the same thing for each variable in a class.
With interactive story telling it can be good, but can often say things that don't make sense or not pick up on the subtext of what you are saying. It can still be entertaining. -- I still use other forms of entertainment.
On the TTS side, I've been using it for reading articles and stories for a long time. The current gen models are very good quality wise, but need work on accuracy and artefacts in generation.
This is a fallacy that comes from not doing discounting. When you apply discounting with any realistic discount rate, current investments don't pay off if AI becomes something in 2040.
Fro investors eventually is not a good enough. They move money into some other asset and wait until the time comes.
I think a big problem is the amount of companies rushing to add AI to it's products when it's not really that good at the moment or not really that useful so there is somewhat of a backlash against it and people hearing the headlines about AI changing the world, seeing that AI has been added to a product they use, testing it and seeing it's useless and then wondering what all the fuss is about.
I stay up to date with AI news, I follow it quite closely and I interact with chatbots somewhat, however I've yet to find it useful in any real use cases for me, since I don't code and the answers it gives for everything else are usually fairly generic nonsense, but it's amazing just that they can do that and the improvement is there. Image generation is pretty crazy too, video is getting there and audio for sure is going to be done, so I wonder what happens to the music industry? Instead of having Spotify you can generate your own music on a whim. As a music lover, that is going to be very odd.
but the first half vs the rest of the list did not receive the same proportional funds, at least from the tech phase from what we can see.
the core difference i can see now is that the sota is captivating the crowd enough, whether from demos alone or from personal tests.
Do you have information they don't? Can you show that LLM tools are worthy of the hype (investment in particular), are consistent, and reliable? That's certainly not my experience.
It's worth saying the emperor has no clothes, especially if we want to avoid the bubble pop, which admittedly GS has an interest in.
Everything is going to need to be refactored for a world where bad actors have truly unlimited time and attention to invest in identifying privacy and security vulnerabilities.
I did not imagine I would ever agree with Goldman Sachs on anything in my life.
Now, I do think it has its uses, but it's once again way overhyped like all the hypes that came before it. As always there is a certain use to it but it's way overblown.
Some banks that develop their own applications will start using it heavily when they will realize that their in-house trading platform can execute orders 0.0001 seconds faster (or something insanely small).
I strongly believe that the only thing that will hold AI/LLMs back is regulation, and nothing else.
Amusingly, Microsoft had an autosummarize feature in Word as far back as, like, 1997. The fact that this appears to be the rare feature that actually got removed from Word could be an indication that this use case might not be as compelling to users as one might think.
I think the chat feature is also really helpful "tell me more about this aspect" kind of thing. That really helps make the summary more tailored to your needs.
Turns out there were some problems that crypto currencies hadn't solved yet, and those still aren't solved. At the time people expected the tech to advance to solve all those things, but they failed to do so.
Same thing applies to current day AI, in order to revolutionize the world like the hype expects they need more things that are completely unknown whether we are actually able to solve or not.
Until those things are solved generative AI is just new search, translation and some funny pictures. Useful but not that world changing. Internet 1995 you could already connect and talk to people from all over the world in social networks, do banking, shopping etc, all the technical problems were already solved, people just hadn't caught up to it yet.
Internet just needed social changes to happen, generative AI still requires technological breakthroughs.
Just look at how popular TikTok is (it wouldn't be without its powerful recommendation model), or the fact that the majority of the world's population has to rely on translation models and subtitles to translate the vast amount of English text online. Bitcoin is never going to "revolutionize" banking, it's really just a downgrade, no privacy, no fraud protection.
This TikTok ads team job post[0] does, however, mention "AI-powered smart video generation (we are also exploring AIGC)" (AIGC = AI-generated content) which implies to me that they are looking to genAI to make new content.
[0] https://careers.tiktok.com/position/7189645714418141499/deta...
I’ve still yet to discover a use for crypto other than waiting for it to increase in value and sell to the next person
I am not saying bitcoin is more important overall, I'm just saying that it is fair to put the two on similar levels today, I believe generative AI will ultimately have important use cases, but as they are today that is far from a given.
It is possible that companies will spend hundreds of billions of dollars on electricity and hardware just to come up with stuff that are barely better than we have today, and then it turned out to be a massive big waste poured into overhyped technology.
Could also be that it gets good enough to do a lot of useful stuff worth all that effort, but you can't count on that happening. The main difference is that we today know crypto currencies was a dead end, we don't know that about transformer based generative AI so it makes sense to make that bet, but it could still be a massively overhyped dead end.
AI is more comparable to GUI or internet or cell service than it is to cryptocurrency. Blockchain, if you must, but I don’t think anyone could saw AI has to more relevance than blockchain.
The same isn't true for generative AI today, in order for it to revolutionize the world similarly to how the internet did it still need some more unknown pieces, which you are expecting to happen but it might not. The internet however was already completely solved when it was hyped, very different.
The difference between internet and AI is one is mature and the other is nascent. That’s it.
Because Goldman Sachs have been using AI for well over a decade. They're saying one (popular right now) thing, and very much doing another (quietly).
Their employees are capitalizing on this latest AI boom [1] to make their efficiency go way up, radically changing their processes... And a lot of people who were slow to grasp what AI can do for their workflow are being made redundant [2].
I can only imagine what those people laid off are thinking of this statement.
1 - https://www.wsj.com/articles/goldman-sachs-deploys-its-first...
I came to this conclusion a while ago :-P
If GPT 5 doesn’t meet high expectations, another winter could soon arrive.
An AI winter where AI will be used everywhere.
Right in the preamble
..despite these concerns and constraints, we still see room for the AI theme to run, either because AI starts to deliver on its promise, or because bubbles take a long time to burst.
So many HNers love to be contrarian about anything resembling a hype train, and feel validated when a report like this comes out. I think the reality is that people moving forward with integrating AI into various aspects of their business and reaping the rewards aren't here making snarky "I told you so" comments, which seems to be every other comment on this and the other threads about the same article.
i see people constantly ragging on LLMs as being infringing and bad and wanting to see them fail
If something is unreliable, it is, by definition inconsistently unreliable.
I tried copilot once and it was pretty keen on sneaking itty bitty bugs in to my programs very frequently.
It is fine for certain boiler plates, but it wasn’t “great” for anything that templates weren’t already great for.
Not chatting with and asking to write something open ended that even you don’t know how to do.
Digital coding assistants are here to stay, but not at the $10/month price point. Maybe $10/year.
Unreliable for what? There's so many different ways which people use these services. The way I use them, it's not unreliable. Because I don't use these services in a way which would be unreliable.
Instead, it seems the broader conversation that the huge investment in AI might not pay off right away is very plausible. I pay $20 per month for one of these services. And that $20 service is maybe the most expensive to implement of anything I have paid $20 monthly by a long shot.
And this is important for the audience of Goldman Sachs. Maybe it's not so important for the typical reader of Hacker News. Who reads Goldman Sachs papers?
>questions whether generative AI will ever become the transformative technology that Silicon Valley and large portions of the stock market are currently betting on...
Who really cares if the AI is generative or not? AI is pretty much bound to be a transformative technology.
>higher productivity (which necessarily means automation, layoffs, lower labor costs...
Nah. It can also mean more stuff produced. And probably will.
That said current asset prices may well be a bit inflated. And some startup could come up with a better algo for self improving AGI rendering the other companies not worth much. Bit like how Google came along and rendered the other search companies not worth much.
What a joke.
I'd like to see the quants that are using ChatGPT for HFT. Apples and oranges.