Precisely. Hallucinations were improperly named. A better term is "confabulation," which is telling an untruth without the intent to deceive. Sadly, we can't get an entire industry to rename the LLM behavior we call hallucination, so I think we're stuck with it.
You are implicitly anthropomorphizing LLMs by implying that they (can) have intent in the first place. They have no intent, so can't lie or make a claim or confabulate. They are just a very complex black box that takes input and spits output. Searle's Chinese Room metaphor applies here.
They're selling a commodity product in a Red Queen's Race. OpenAI, Anthropic, and whatever others get spun up while the gold rush is on are each building costly models at a vast cost to try to get a bit ahead of each other. The economics of this are ugly. The business of building massive LLMs looks more like air transport than the dot com gold rush in the '90s. There will probably be much useful stuff in the wreckage when it's all over, but I've not seen much to inspire confidence that the useful stuff will include profitable companies. This looks like it has all the makings of another WeWork, only this time with an epochal AI winter as the aftermath.
It's unfathomable to me that any of these companies have any kind of moat.
People here like to slag on google for being late to the chatbot party, but they've been using ML the entire time and integrating it into various products for ages. I kind of wonder if the only reason they were "behind" was the lawyers were less brazen about the copyright situation.
I agree with the commodity take, and I personally bet on Google eating everybody else's lunch eventually, because there's a lot of other business behind them and they can afford to undercut competitors. They aren't a one trick pony.
Google produces the most "safe" ( e.g. functionally useless ) models available.
The market demands models that don't fail constantly with HAL-like "Sorry, Dave, I cannot do that for you" moralizing responses.
An API that is used for mission-critical purposes and that randomly fails with "HTTP/1.1 406 You Are A Terrorist And/Or Hitler" is a BUG, and the market will coalesce around models that don't have this bug.
this whole thing produced some useful tools, if nothing else. I'll be able to use a few of them to keep warm during the next AI winter -- much like every other AI winter.
I imagine that OpenAI's operating revenue highly exceeds their operating costs. They spend virtually all their money on training and R&D salaries, not inference and operating salaries.
There are three likely outcomes:
1. AI plateaus. OpenAI slashes the R&D budget to become profitable with revenue in double digit billions and profit in single digit billions. Valuation likely similar to today's.
2. AI doesn't plateau. OpenAI makes a killing. (Hopefully metaphorically, not literally)
3. Scenario 1 or 2, but it's a company other than OpenAI that wins.
Bold of you to assume AI plateauing would be somehow intrinsically obvious to everyone, or even to the minds at OpenAI. Tunnel vision is common in tech, particularly when your salary (or funding) depends upon it.
> I imagine that OpenAI's operating revenue highly exceeds their operating costs.
The Information estimates that OpenAI is spending $4 billion just to run ChatGPT and their APIs, along with $3 billion in training and $1.5 billion in salaries.
Don’t agree. Inference costs have been in a race to the bottom for a while and it’s likely imo that OpenAI is gross profit negative on inference right now
If AI were to plateau, OpenAI would be one of many providers without a clear edge, they'd lose market share. Companies might even start competing on price. Imo, it's not clear any of the software providers would really do that well in an "AI is commodity" scenario, HW companies might though.
I'm running Phi3, Llama 3.2 and Mistral Nemo locally and they're decent enough for many things.
That's delusion, not imagination. Inference is far and away more expensive than R&D and training. It dominates the cost at these companies because it represents most of the unit economics. If OpenAI is not profitable, it's because each marginal customer costs more than they bring in. That's especially true for heavy users since OpenAI charges a fixed amount per customer per month.
Couple that with a lack of pricing power thanks to all the other similar products in the market.
How will they continue to operate into the future with those kinds of losses? More rounds of funding? Loans? Charge more? Most companies with losses that big would be bankrupt, e.g. GM in 2009.
Why do you think cost would expand linearly with revenue? That's typically not how technology companies work. Compute gets cheaper over time and R&D expenses eventually cap out.
We're in the "add billions in dollars of datacenters" and "spin up nuclear power plants" level of infrastructure needed for their growth, so if anything, I might be underestimating the costs if they grow as much as Sam claims they can.
Didn't Sam ask TSMC to spend $7 trillion on new fabs? By comparison, $18 billion/yr spend seems very small.
> Commodities can be massive businesses with competitive moats. Oil is a commodity BP and Exxon do just fine financially speaking.
They invested in exploration and now they control those oilfields. They built refineries and have the systems and experienced people to operate them. Meta can't release a LLamaOilfieldAndRefinery which I can operate by just spending a few thousand on gpu's.
Oil is a special commodity in the sense that it's as important to civilization as water is to humans. It's also special in the other sense that we have geopolitical strife over who gets to extract it, sell it and to which market...
I don't see how a chatbot meets any of those criteria
> I don't see how a chatbot meets any of those criteria
Calling these things "a chatbot" is likely limiting your vision: some of the stuff people build by fine-tuning LLMs, such as the ones OpenAI offer, use them to generate database queries matching their customers' database schemas.
"Chatbot" is simply a convenient UI for an LLM in the same way that a web browser is a conventional UI for email. (And in this analogy, anyone calling an LLM "autocomplete on steroids" would be making the same mistake as someone saying "Wikipedia is just TCP on steroids").
I expect LLMs to continue to be extremely generic in the same way that web browsers are (market history shows periods where one browser dominates the market despite open source) or like spreadsheets (where Microsoft Office is, or was last I looked, dominant despite free offerings being good enough for most people).
Society needs intelligence. We started using mechanical aids because it became impractical to perform census work by hand as the population increased, AI is a continuation of this process: we need it, it's a commodity, there may be a market opportunity despite free and/or open source competition, and (like Netscape, like Internet Explorer) there's no guarantee that the winner in one year will still be leading the next year or even existing a decade later.
>This looks like it has all the makings of another WeWork, only this time with an epochal AI winter as the aftermath.
I'm as cynical as it gets on this forum, as evidenced by my comment history.
But you're comparing these AI companies to WeWork? Really?
WeWork was a real estate company operating in a historically favourable environment (0% interest rates) pretending to be a tech company. They literally rented office space. What do they have in common?
I notice there's a new generation of "grey beard" programmers constantly talking about how "useless" AI is for programming, and they can do everything faster. Meanwhile, there are tons of us out there who are paying $20, $30, $50 per month and upwards for these tool as they are, and wouldn't want to go without. Ever. And we have no idea where it's going to go. Maybe you're missing something?
> Meanwhile, there are tons of us out there who are paying $20, $30, $50 per month and upwards for these tool as they are, and wouldn't want to go without.
I'm paying for Claude, which I find super helpful (although mostly for side-projecty stuff rather than dayjob). I'd definitely pay double what it costs today, particularly if I went back to shorter-term environments where I think it shines.
If you can't hook it up to your codebase/you write in a language where it's not great (it doesn't seem good at Elisp at all, at all) then I could see how people might not find the value.
Nonetheless, despite finding these tools useful, I too am sceptical about whether or not there's a valuable business there.
For context, I said this about Uber, WeWork and a bunch of other startups that never really monetised. Note that I also said that about Facebook, where I was completely wrong.
AI is a real threat to programming compensation. It's already kind of happening as junior dev work has become much harder to find (and consequently people will take jobs paying less), and senior devs are starting to lean on LLM's for basic code.
If you are paying attention, this is a pretty terrifying prospect if you are building a $200k+/yr life on the basis that SWE will pay like this forever. A machine is coming along that genuinely might be 80% of your skill set in a few years, makes it very difficult to negotiate generous packages with the remaining 20%.
> A machine is coming along that genuinely might be 80% of your skill set in a few years
Something similar happened when COBOL came out. Same with website builders. The big difference between most industries versus software is we aren't even close to satisfying the world's desire for software yet, so increases in productivity just gives us bigger leverage.
From a greybeard, I'll say this: This is nothing new. I don't see that this will disadvantage anybody other than those with a failure to adapt. This is about on par with IDEs that can stub out tests and method signatures on the scale of things to worry about.
> A machine is coming along that genuinely might be 80% of your skill set in a few years
Is this supposed to be scary? This scenario is absolute stonks if you're a developer worth your salt. A machine makes me 5 times more productive, and I don't even have to commoditize my complement because they're taking care of that themselves?
This assumes companies have finite backlogs. AI will raise productivity, but to date no company has said 'our software is done, send all the programmers home.' Instead companies expand what they want from their software.
The rate of change could make the short term bumpy as companies try to play around with less dev work. Eventually though competition will push companies to raise their productivity to the new baseline (programmers + AI).
One thing that is a danger is if devs ignore what's happening in AI. I remember when Google first came out having to learn the art of querying Google to get what I was looking for. AI/LLMs looks to be the same - ignore learning how to leverage them at your own peril.
This 100% does happen, all the time, mostly in SMB.
Hire a team to build a project, when it's finally satisfied most of your requirements, you progressively cut staff until only Jim is left, and Jim spends the next two decades maintaining the system, growing out his hair and beard, piling kludge on top of kludge, and drinking heavily until retirement.
The HN bubble, focused so intently on BigTech and FAANG, is woefully unaware of how things work basically everywhere else.
According to Fidelity, Twitter has gone from a 44 billion dollar company to less than 10 in 2 years. Clearly we should not be using it as an example of how to run a technology business.
I think we're misidentifying AI as the catalyst when it's more like the accepted excuse for diminishing compensation for software developers. Ownership wanted to drive down costs as salaries were getting out of control. This includes not just wages but work/life balance and other perks.
Personally I think its cyclical as software is such a key component of communication and automation - and so we'll see future growth periods but probably not to the same extent as the last decade where seemingly every undergrad was compelled to study Computer Science.
I don’t think that software salaries are coming down yet as a result of Copilot/ChatGPT, they’ve come down because of overexpansion during COVID and a big rise in interest rates that makes raising funding and investment much more challenging for most technology companies and forces them to become profitable sooner rather than later (with some not able to make that transition and therefore failing).
Junior positions being hard to find are the result of the post Covid market correction and the relative expense of borrowing money in a high interest rate environment, not of chatbots replacing juniors. Executives might be saying it's AI but they're wrong.
AI is a polarising. And the danger with such things is once you’ve made up your mind, it’s hard to look at things from the other side regardless of whether you are for or against it. And arguments supporting either POV starts to look like shilling.
What they have in common is that they're selling a (now) commodity product for margins that will be driven further down over time as other companies catch up and the huge leaps stop happening. With WeWork it was other old-guard real-estate managers realizing they can do what WeWork does for cheaper.
Comparing to airlines was fine, but you take issue with a comparison to real-estate? WeWork was also beloved by their customers and had leadership who were a bit off the rails.
The fact that switching models is changing one string in AWS bedrock means that nobody is going to be able to charge a significant premium.
I'm genuinely curious what makes the current crop of tools so compelling/productive to you personally? Been doing this for twenty years so I guess I'm a greybeard.
I've always found the time-consuming part of this job to be understanding the context of the change rather than making the change itself. Essentially, trying to understand the existing code and business requirements and how they all fit together. I can definitely see the potential for AI to help make this easier but I haven't found the current tooling to be any good at this.
>But you're comparing these AI companies to WeWork? Really?
It's the first tool that came to hand: I'm sure there are better historical comparisons if I bothered to look. For professional reasons, I was well acquainted with WeWork when they were a big deal. It was clear to many of us in advance of their collapse that WeWork was heading for a hard crash based on their lease commitments and other public data. In this case, public data, such as for OpenAI and Anthropic, strongly suggests that, like WeWork, the economics of the businesses don't make sense. There are some fundamentals that no amount of innovation can overcome. Committing to leases you can't conceivably cover is one of them. Spending $2.35 for every $1 of revenue is clearly another, absent some breakthrough.
WeWork is not a perfect example. But if OpenAI flames out, it will be mentioned in the same breath as WeWork. The reasons are not the same, but they do sort of rhyme.
That suggests that there are scenarios under which we survive. I'm not sure we'd like any of them, though "benign neglect" might be the best of a bad lot.
They are kind of like pacifiers for adults, aren't they? The term "fondleslab" captures that; I believe I first read that moniker years ago in some article on the Register.
It depends on the company. It's kind of like a menu item that says "Market Price". You know it's not going to be cheap. You don't know until you ask if the price is less than the value offered.
Everything can be taken from a man but one thing: the last of the human freedoms—to choose one’s attitude in any given set of circumstances, to choose one’s own way.― Viktor E. Frankl, Man's Search for Meaning
It seems plausible that the immune system requires stressors to function well, and the absence of stressors may reduce immune system "strength". Selye's general adaptation syndrome seems applicable here as a first approximation. Muscles atrophy from disuse. Why not also immune systems?
The immune system is never dormant. It is constantly engaged. We don't live in a sterile environment, don't eat perfect sterile food, don't breathe perfect air, etc. There are multiple aspects to our immune system, and a lot of them are being stressed daily.
