How are they able to sustain keeping this free? Or is the idea that it's a demo for now and they try to charge for this later?
Copilot+ runs locally, but has the quality of 2 years ago, it's basically useless for anything but shitposts and spam.
The cloud AI tools all burn hideous amounts of money, all ran at a loss.
AGI is a red herring, and will not happen. The architecture of generative-AI systems simply doesn't permit the required logic and reasoning capability.
Even the "Actually, Indians" concept of outsourcing the tertiary sector to the developing world by way of having low-skill workers clean up AI generated trash is unviable. (It both doesn't work, and is politically doomed.)
What's going on here is that tech companies are tearing up everything to pump their stock prices after the covid-tech-boom and ZIRP ended. Burn down their core products to keep the bubble going just a bit longer.
For the other companies in your list and more specifically the ones that are developing LLM's they are investing billions of dollars in the hope that they will make all of those back and then some in the future. Meta for example said they invested $10bn into AI just last quarter alone.
At this point I see training costs as basic research and R&D, partially and not entirely oriented at -making training scale- and cost optimization for next generations of foundational and fine tuned models. For example Falcon is literally basic research funding.
That’s relevant because if they stopped with say Claude Opus and 4o and did no more training they would be handsomely profitable indefinitely because the product is that useful. However it’s an arms race and the limit of effectiveness hasn’t been reached as training costs fall dramatically. So it’s not the right time to stop because whoever stops first loses everything to whoever doesn’t stop. More they’re feeding off each other in a virtuous cycle.
Once diminishing returns kill the race whoever is left in the race will have a handsome business indefinitely as the moat to build such a return diminished model is probably enormous. But if inference optimizations keep going as they’re going it’ll be crazy cheap to operate.
The second layer market of tools that constrain, optimize, and effectively apply the models in effective ways will be the first to really turn a profit. Many hype wagon AI companies are already being snapped up by larger companies to bootstrap their internal work.
> The bigger AI companies like OpenAI and Anthropic are not losing money in their monetized APIs.
[CITATION NEEDED]
Inference is more expensive than claimed, used extensively as a 'slot machine' with users trained to just keep re-generating until they get something useful, and only keeps getting more expensive as model quality has to go up.
And in practice, training the model is far less one-off than claimed. Current tools are not sufficient.
> Every place I’ve worked is absolutely reducing costs and doing new / more business as a result of their LLM use.
Unless you are working in SEO, Marketing, or spam, I don't believe you.
LLMs aren't reliable enough to replace actual human labour. While it's true many companies are fooling themselves into believing they're reducing costs, in practice other staff is picking up the slack. This is unsustainable unless your company has massively overhired.
Things like "AI generated software tests" are a farce. The consequences aren't immediate, but will show up long term.
I don’t feel like you really have much experience using LLMs in business. However an example of where they’re very powerful is in summarization. For instance we have a pretty complex customer support model for our fraud and other cases with various disparate data sets including prior cases, related possible fraudsters identified via our fraud models, etc. We built a copilot LLM multi agent system that has access to various functions as sub agents that are prompted and context aware of how to summarize their specified data sets. They also have the ability to render widgets on demand or if their context implies it’s relevant. This allows quite a lot of complex high cognitive load information to be distilled rapidly and the investigators to interrogate the copilot on a case. As the copilot develops “answers” as a summary it dynamically renders an appropriate contextual dashboard with the relevant visualization.
By structuring the application as a multi agent model we can constrain the LLM to pretty well specified tasks with fine tunings and very specific contexts for their specific task. This almost entirely eliminates hallucination and forgetfulness. Even if it were to do so the actual ground truth is visualized for the investigator.
Prior systems either dumped massive amounts of cognitive load in the investigators face or took man years of effort to create a specific workflow, and in an adversarial dynamic space like fraud you need a much more dynamic approach to different types of new attacks.
We aren’t replacing anyone. That’s not our goal. In fact we grew our investigator footprint because both our precision and recall have grown dramatically making our losses much less. We hire more skilled investigators and greater number to address more suspected cases faster and better.
Listen. When John Henry battled the steam drill he did win, but it killed him. Go to any modern bore site and you won’t see less people working on the tunnel but more people - people who aren’t there for their strong back and ability to swing a pick but because they’re highly trained experts. They’re just building more complex tunnels that don’t collapse and don’t lose dozens of workers per dig.
This form of automation is no different in my experience so far.
So, if all you can see is SEO and grift, it might be a lack of imagination and experience on your part and some magical AI thinking sprinkled in. All your points about LLMs failures are true but they also all have solutions that don’t require slot machines as you say or imply it’s all a scam. They’re a tool like any other and they require handling in specific ways to be most effective. Even if chatgpt is a pretty unconstrained interface and that leads to issues doesn’t mean that’s the only way to use the tech.
Use of LLMs to generate software is dumb. Although a protio, LLMs are actually pretty remarkable at generating Cucumber tests as Gherkin is a natural language grammar that plays into their native strength better than producing computer language grammars. This is useful if say you have business people or whatever writing effectiveness or whatever testing where they can provide a specification of policy and a well prompted LLM can generate pretty exhaustive cucumber tests (which can be pretty redundant and formulaic when asserting positive and negative cases exhaustively) which can then be revised by hand as needed. Since they’re natural language as well the business people tend to be pretty good at debugging the tests up front and with a large set of cucumber tests written by hand you’ll see tons of errors anyways. The LLM tests tend to be much much higher quality than the human written ones.
But to say something useful, let me try to elaborate my general criticism here:
> Prior systems either dumped massive amounts of cognitive load in the investigators face or took man years of effort to create a specific workflow, and in an adversarial dynamic space like fraud you need a much more dynamic approach to different types of new attacks.
This begs a question: Why didn't a computer system to summarize this data already exist? Or rather, what stopped the prior systems from doing this work? (And I'll consider conventional machine learning; classifiers and the like, as traditional computer systems here)
And there's generally two options here:
1. Conventional computer systems absolutely could do this work, but they just haven't been built. (Say, because nobody signed off on the R&D but would sign off on AI hype R&D)
2. The LLM system is doing a task the conventional computer system cannot do.
Number one's problem is simple: It's just inefficient and wasteful. Number two is a red flag: There's very little overlap between the things a conventional computer system cannot do, and the things you can trust an LLM to do reliably.
As you describe this system, selecting which data is relevant for fraud investigation is a very traditional classification task. Using normal machine learning for that is basically industry standard.
So what's the LLM actually doing? Subtract the hard logic of normal software, and the classification of machine learning, and the answer is generally: A complex nuanced reasoning task.
But that's precisely what LLMs are not to be trusted for, because they are incapable of that kind of reasoning.
> Listen. When John Henry battled the steam drill he did win, but it killed him.
You're missing the point I was making with that remark. It's not about firing people or not.
It's that these systems are dangerous to evaluate from a high level. It's very easy to miss externalities that'll tip the entire endeavour into a net-negative. You need the investigation of what exactly the AI systems are doing, on a specific detailed level.
E.g.:
> This is useful if say you have business people or whatever writing effectiveness or whatever testing where they can provide a specification of policy and a well prompted LLM can generate pretty exhaustive cucumber tests (which can be pretty redundant and formulaic when asserting positive and negative cases exhaustively) which can then be revised by hand as needed.
"A specification that has been prompted into sufficient detail" is just a program. You're describing the most inefficient declarative programming stack on the planet.
Granted, the programming stack to actually declare business rules this way isn't very good, but using AI here is just an error-prone transpiler.
It's very easy to "looks good to me" these tests and claim the project a success, yet miss subtle errors in the generated tests. I remain skeptical about how well these tests will hold up in the longer term.
The LLM isn’t used for reasoning at all. The human does all the reasoning. The LLMs task is summarization and semantic analysis of relevance which LLMs are fantastic about especially in a well managed and fine tuned environment with guard rails and context scoping. It’s a true copilot scenario and the LLM takes direction from the human and answers questions only. All decisions are investigator driven. This is the right relationship. The LLM coupled with IR tools does information retrieval and summarization and the human makes decisions and reasons.
Press X to doubt.
I highly doubt if ChatGPT API is losing money. Yes, I've read claims saying so. No, I haven't seen a credible one. And it's getting cheaper and cheaper, currently even faster than Moore's law (gpt-4o is 6x cheaper than gpt-4).
I am probably more privacy oriented than the next guy, and I’m just not really seeing the pitch here.
EDIT: I mean, you don't need to look far to see this effect; see e.g.: https://news.ycombinator.com/item?id=40447474.
Also, if DALL-E is anything like Stable Diffusion, it's effectively free to run. For comparison, my 5+ years old machine that I recently put a 4070 Ti in, can generate images with SD all day long without breaking sweat, and does it faster than even paid on-line services I've interacted with so far. That same machine chokes on any LLM above 8-10 billion parameters; it can handle that much with full GPU offload, but try anything larger and it becomes just an elaborate space heater.
(I tried running 8x7b Mixtral on it once, and managed to OOM and crash the system :/).
They say it’s all local. Maybe that’s true. Maybe for now. Will it be forever? Is the recall local, but insights are collected? Time will tell. I certainly won’t be signing up to be part of the experiment.
LLMs need lots of input
If that was true, there would be a free laptop by now.
(the above is technically parody and not precisely serious)