The distinction is adjusted vs non-adjusted basis. Serving inference is profitable, but serving inference for the last released model doesn't cover the cost of the model in training.
Dario explained it very clearly in an older interview. If every model is bigger and more expensive to train, each model can individually be profitable, while the company as a whole still shows a loss at any given point in time. This is why he likes the idea of regulation so much, it'll let them take a breather and recoup.
> Dario explained it very clearly in an older interview. If every model is bigger and more expensive to train, each model can individually be profitable, while the company as a whole still shows a loss at any given point in time.
“We’re profitable as long as you ignore our ever increasing cost of goods sold, we promise it’s not a pyramid scheme” - ~~Charles Ponzi~~ Dario
But the reason the cost is ever increasing is the steeply ascending quality. There's no (non-doom) scenario where this game does not end with profitable cheap tokens. Wherever it'll stop scaling, it'll be profitable to serve tokens at that scale. And of course if it doesn't stop scaling, it's the singularity and all this thought of profits is quite secondary.
An industry cannot fail because it's too successful.
edit: Other explanation. The industry has found a moneyprinter. It prints money, and several companies are building them and printing money. However, each one of those companies is cashflow negative, because they're all trying their desperate best to build an even larger moneyprinter. Why? If they don't build it and the next company over does build it, and it works, then company 1 is differentially outcompeted. what's more, the more expensive each moneyprinter is, once a company gets strongly ahead the second company won't get the financing to keep up with that one. ... But it's still moneyprinters. This is not an economic race that ends up with "and thus everyone stopped printing money."
> But the reason the cost is ever increasing is the steeply ascending quality.
Sure, if you say so.
> There's no (non-doom) scenario where this game does not end with profitable cheap tokens.
I doubt it.
> Wherever it'll stop scaling, it'll be profitable to serve tokens at that scale.
It’s much more likely, imo, that local AI models will completely destroy the entire business model of inference. For any entity that would be spending enough to replace multiple employees (allegedly), the cost of the hardware to do so is completely affordable. This has the major benefit of eliminating per token costs, and practically guaranteeing that some “AI company” can’t steal your company’s secrets.
At an individual person level, this hardware will take much longer to obtain, but it will happen eventually.
> And of course if it doesn't stop scaling, it's the singularity and all this thought of profits is quite secondary.
I still believe, strongly, that LLMs are the wrong path to AGI. There’s trillions of dollars being put into turning LLMs into AGI; however, there has no zero results showing that LLMs are capable of “learning” anything. “Skills” and expanded context windows don’t count. The “model” needs to be able to modify itself, in real time, for that to be possible.
We’re in a local maxima that strongly refuses to accept that it’s a local maxima. (Diminishing returns for the amount of money poured into “training” is beyond obvious to anyone paying attention).
> An industry cannot fail because it's too successful.
Sure it can. The fossil fuel industry is wildly successful.
It’s also boiling the planet alive, but that’s okay, it’s generating shareholder value so everything’s fine.
Edit for your edit:
> edit: Other explanation. The industry has found a moneyprinter. It prints money, and several companies are building them and printing money. However, each one of those companies is cashflow negative, because they're all trying their desperate best to build an even larger moneyprinter. Why? If they don't build it and the next company over does build it, and it works, then company 1 is differentially outcompeted. what's more, the more expensive each moneyprinter is, once a company gets strongly ahead the second company won't get the financing to keep up with that one. ... But it's still moneyprinters. This is not an economic race that ends up with "and thus everyone stopped printing money."
By your own definition it is impossible to achieve profitability without an ever increasing amount of money flowing into it.
There was a business about moving stamps around that needed similar levels of cash infusion to sustain itself. Eventually its bubble popped too.
> It’s much more likely, imo, that local AI models will completely destroy the entire business model of inference.
You're massively underestimating how much cheaper it is to serve models in a datacenter than at home.
> By your own definition it is impossible to achieve profitability without an ever increasing amount of money flowing into it.
No, what? No. This is just a total misreading of what I'm saying. It's extremely easy to achieve profitability. It just requires your business to stop chasing more profitability. The problem is just that then the other businesses who are still chasing more profitability will eat its lunch. It's a market shape that easily supports n-1 bankruptcies but never n bankruptcies.
it doesn't seem that complicated to me. what do you think the cost is for AWS to serve inference it sells to Anthropic for "$1.00"? I think it's about $0.11. Then, ask yourself, did Amazon sell it to Anthropic at all? No, they traded $0.11 real dollars for $1.00 of Anthropic shares (the CLN that created the real loss Anthropic is talking about), which are now worth $10.00 (which is why people talk about this as a paper loss). So... IS it profitable? It's profitable, but Amazon didn't JUST sell inference when they turned $0.11 into $10.00...
It's really quite interesting how many HNers hate AI; which is odd, considering the zeitgeist of this community. It may even be more intense than the Apple-hatred.
Hate marketing BS and wealth extraction? Absolutely.
I see OpenAI, Anthropic, etc (the whole LLM-hype/boom) leaning far more towards being the latter than the former, in which context the dislike makes perfect sense.
I don’t hate it. I just hate the cult-like enthusiasm for it and the associated false claims. Again, your isolated anecdote does not trump the carefully compiled data presented in this post.
We get way too hung up on data; especially when it reinforces our personal bias.
Some time ago, I had a discussion here, where I mentioned that the average commute time from the suburban community in which I live, to downtown Manhattan, was generally 90 minutes to two and a half hours.
This was based on purely anecdotal data. That data includes the daily news on the radio and TV, which considers commutes from suburbs, important enough to dedicate significant airtime.
It was attacked, because some organization published some statistics, that said the average suburban commute time was 27 minutes. I was told that I was exaggerating, and being overly negative.
Here's a math word problem:
Harry lives 25 miles, as the crow flies, from his destination. However, the roads make it 40 miles.
Of those roads, 10 miles are suburban side roads that Harry needs to take, to get to the highway. The average speed is 25 MpH.
Once Harry hits the highway, though, the real fun begins. It's stop-and-go traffic (20MpH) for 28 miles. Once he gets to the city, he needs to navigate another 2 miles, through the most God-awful traffic on Earth, to where he needs to drive around for up to a half hour, looking for a parking space.
How does Harry make the drive in 27 minutes?
A fairly significant number of folks just took that ridiculous statistic, and were quite comfortable, calling me a liar, based on it. It was kind of jaw-dropping; considering the types of folks that we have, here.
That sounds a bit more "cultish" than we might be comfortable acknowledging.
> We get way too hung up on data; especially when it reinforces our personal bias.
While I don't disagree that we generally get too hung up on data, "It's a problem when the data matches the hypothesis" is an odd blanket interpretation of the scientific method.
I'm confused, your rebuttal to statistics is another anecdote? Was the quoted statistics limited to car commute or includes public transport? Your commute may be long, but it may not represent the average as well. Maybe it's a sign to switch away from cars instead?
Anecdotes are poor because it is highly context dependent. You had to make a long description of Harry to make your point, and yet it has not, to my mind, disproved the statistics, though I would like the source to the stats myself if you could provide it, thanks.
The commute only takes 27 minutes if we ignore that everybody else has an LLM- erh, car I mean.
Think you’re clever because your LLM contacts 100 businesses? Imagine 100000 businesses doing the same thing. You’ve just eradicated human support, because nobody’s going to sit through this BS bot future. Good luck when you get auto-flagged for some reason and there’s not a single human to revert or care.
"Unproven" is in the eye of the beholder, but to me it certainly looks like the sudden take-off in arithmetic ability (to name one example) at approx 10^22 training FLOPs across all LLMs is emergent. (ref. https://hai.stanford.edu/news/examining-emergent-abilities-l...)
Many, many people here have already been sucked into the TESCREAL orbit, and have a quasi-religious belief that AI will be the most important thing to ever happen. Anything that challenges that stance is anathema.
In practice, hallucinations have only mitigated within verifiable domains. The vast, vast majority of human activity is unverifiable, or at least significantly less verifiable than mathematics.
So you're mad at being consistently incorrect and being held to account with what is essentially, "believe my assessments that I have to continuously caveat as they become invalidated, over what you see and hear."
System 1/2 is the pop version, but the fast/slow distinction is prevalent in both RL and computational cognitive science, sometimes going under different names: procedural/deliberative, model-free/model-based, automatic/controlled, associative/rule-based, autonomous/algorithmic, etc.
It's indeed a terrible idea - as in, it's great. You get benefits of cross-marketing: you ride on a popularity of a well-known book, and as you also drive more sales of it, even if you don't have a deal and don't benefit from that directly, you strengthen the loop and solidify your brand.
This choice doesn't really constrain what the organization can do, either. Pop-sci books have plenty of wiggle room in interpretation, and afford a lot of "you're holding it wrong" dismissals of criticism, that with a bit of clever copywriting, the organization can do absolutely anything and still claim it's embodying the framework/theory of the book.
Or you could just accept that other people disagree with you. It’s perfectly valid to have the opinion that AI models are not “thinking”. Instead, many AI enthusiasts seem interested in browbeating others and policing discussion.
Can you explain to me how a Millennium Problem gets solved without any thinking or intelligence?
If someone's opinion is illogical and clearly incorrect (not based in facts or in reality) it is not valid and doesn't not have to be given any serious consideration in a disagreement. Personal opinions are subordinate to facts and reality.
there is no proof that symbolic manipulation leads to an internal phenomenology.
in english: doing advanced mathmatics does not create thoughts or feelings.
or awareness of one's own existence.
it does not create a 'self' able to deduce a priori the existence of oneself or the external world.
or qualia.
again in english: there is nothing it is like to be an llm.
someone who thinks llms are conscious is probably a panpsychist. i don't think they have the features of consciousness defined even by functionalism.
here is one fairly famous thought experiment. mary lives in a black and white room. but she knows everything about color. in fact she predicts it so accurately nobody can tell the difference. she steps outside into the world of color. in the world where llms are conscious, she already knew color, so she experienced color, so there is nothing new it is like to see it.
You are making several very strong, unqualified, absolute statements which are really only your opinions. You are also apparently unaware of the large body of top level philosophical writing on this subject going back several generations which has established the current consensus (among those who are aware of it) that these questions cannot be answered in such absolute terms.
I think you need to read more about this subject before presenting your unsupported intuitions as fact.
First off, there is no accepted definition or test for consciousness. The concept is something we cannot prove even exists (outside of magical or imprecise, human-centric definitions).
Second, there is no common definition of thinking that requires consciousness as a prerequisite (that seems to be a false/invalid statement you're making to try to gloss over accepted definitions for thinking/intelligence that describe/fit AI agents well).
And third, if we can't prove consciousness even in animal species close to humans (or even other humans) in cognitive abilities, then you cannot in good faith (being honest about what we know or don't know empirically) say that an LLM Is not conscious (and the same goes for an LLM thinking or having intelligence).
It is more likely that humans made up a term (consciousness) to try to separate human intellect from that of other animals (for veiled religious and mystical reasons, along with concepts like "the soul") and now people are falling back into that flawed definition in new quasi-religious debates around AI (like the debate around evolution previously) that challenge our deeply held (but flawed) beliefs/opinions around our understanding of our own abilities and our incessant need to feel special/supreme in the universe so we don't have to come to terms with hard realities and nuanced understandings about the reality in which we live.
reply