It’s not just that. Oftentimes contracts stipulate that the client’s data can’t be transferred across certain boundaries. If you have signed such an agreement, even sending the data to a service on the same cloud provider but in a different region could be a huge compliance violation.
The old saying goes, the market can remain irrational longer than you can remain solvent.
I’m not necessarily expecting a crash any time soon. (But we average a major correction, what? every 8 years? So if you keep predicting one long enough you will eventually have been right all along.) But I do feel comfortable saying OpenAI and Anthropic are overpriced. For more or less the same reason Cisco was overpriced in the late ‘90s. It’s not that what they were making wasn’t valuable; it’s that we got out over our skis a bit over how much of it the world could actually manage to consume in the immediate future.
I can’t imagine them actually being bullish about exponential growth, when both seem instead to be stagnating. I’m more inclined to believe they’re just maintaining a level of hype in public because that’s what you do.
They’ve claimed a big revenue run rate for this quarter. But it’s non-GAAP, so you kind of have to assume shenanigans. Earlier this year they were telling a court their revenue was like 1/4 of what they had told the public. I consider the number they came up with when they had to worry about committing perjury to be more trustworthy (because I’m a pill), so that would also indicate shenanigans. My guess is they are inflating that revenue run rate figure by booking token pre-payments from enterprise contracts now instead of spreading it over time as GAAP would mandate. And at the same time their big enterprise clients are talking about scaling back their usage.
So we’ve got a combination of signs that they’ve been inflating their revenue growth, and signs that their customers are losing their appetite for contributing to that revenue growth. I suppose it’s not a slam dunk, but it feels to me like as strong an indicator as one could hope for a private blitzscaler startup like this.
"Although the company has generated substantial revenue since entering the commercial market—exceeding $5 billion to date—it has nonetheless had to raise more than $60 billion in outside capital to fund its operations".
Oh, to be clear, I'm not saying there is evidence they're all a-okay. I just hadn't seen any evidence that they were stalling out. (I have for OpenAI.)
I could have paid cash for my car, but that would have been a bad move. I wouldn’t have had any liquid assets left over for getting me through a rainy day. The interest I paid on the loan was an acceptable price for reducing my overall risk exposure.
Even if Alphabet has $80B sitting in the bank, they could quite reasonably arrive at a comparable decision.
This is a specious argument. I have not studied the case law, but I would guess that the reasons why courts decide in favor of gun manufacturers generally don’t apply to AI. Becauee the guns in question are not able to autonomously shoot people, and because they generally work as advertised.
A more accurate analogy would be Tesla and Autopilot. And they are being held liable in courts. They are being held responsible for autonomous behaviors that are not fully under the control of the operator, and they are being held responsible for misleading operators about the capabilities of the product.
Boeing got in trouble for MCAS, with a comparable legal basis.
I suspect that letting agents spin away unattended for long stretches of time will become less and less popular as more and more companies blow their token budgets and start requiring some answers to difficult questions before agreeing to further loose the purse strings.
I've been enjoying journalist Ed Zitron's recent diatribes about how impossible it is to find a business leader who had a plan for measuring their ROI from adopting AI coding.
What he says he's consistently hearing from them mirrors what I saw at my own employer: they thought they had ROI metrics, but they actually only had usage metrics such as "lines of code committed" or "number of pull requests". The only way those could possibly work as an ROI measure is if your business charges customers by the line of code.
Measuring productivity of developers isn’t really in line with what needs to happen, either. A team can be incredibly productive and still generate negative 100% ROI if what they are building so industriously is stuff that nobody wants to buy.
Which reflects another thing I’ve seen at work. A lot of what AI coding has enabled is diving headfirst into quagmires. Our costs have spiked - not just because of the token spend, also because we gotta pay the cloud platform to run all these new services, operators to operate them, marketers to market them, etc. - but revenue hasn’t budged.
But at least pre AI, most managers presumably subjectively measured devs on relevant performance. Using systems where employees who burn the most tokens ($) per week ‘win’ is crazy - just ask the AI to spin up a subagents to implement every conceivable approach to a task, then spin up n agent judge to pick the winner, and repeat. You've immediately got 50x or whatever your previous usage from that alone.
That's been the dream for the 40 years I've been paying attention. And in that time, I've seen plenty of incremental changes but never the kind of sudden sea change that the hype machine anticipates.
The perennial reality is that automation is inherently inflexible, so there's only so much of it that you can do before you've committed a huge strategic blunder by making your business resistant to change and severely curtailing its ability to cope with situations that don't cleanly fit the mold. So then we need to hack in ways to deal with the exceptions, but, since they're hacked in, they're often painful and time consuming. Sometimes so much so that after the new process stabilizes it turns out to be even more cumbersome and require more manual effort than the system it replaced.
When anyone other than a technologist suggests doing that kind of thing, we call it "bureaucracy", and we hate it. I think maybe what we have trouble seeing is that there's actually a pretty fundamental difference between automating purely technical processes like server deployment, and automating processes that are fundamentally about mediating human interactions.
I think it might be even worse than that. It seems to be a multiplier for the Dunning-Kruger effect. Possibly because being trained to exhibit positive demeanor means that it will always tell you you're the best, no matter what.
Not sure why I saw this getting downvoted. This is exactly what is happening to that person, they think they are way more capable then they really are that's why they're taking on hard challenges and high complexity tasks when they're not ready for it.
And I do think it's important to be aware of this. When I've got a good idea, the agent says, "Good idea!" When I've got a bad idea, the coding agent also says, "Good idea!" And I'm most inclined to just go with what when I'm outside my realm of expertise, because that's the time when the coding agent promises to save me the most time. Because developing a proper understanding of things like the problem space and tech stack is often the most time-consuming part of the job.
I'm getting the impression that LLMs are just not very good at "reasoning" about time. I have definitely had success getting a coding agent to produce decent concurrent code, but I had to basically lead it by the nose, and I strongly suspect that in most cases it would have taken less time to just do it the old fashioned way.
I've had good luck having it translate TLA+ specs to programming languages. The specs are written by me and my fingers, and I've done most of the interesting concurrency reasoning beforehand.
I'm pretty sure it still saves me time, and if nothing else it's an excuse to write TLA+, and that's fun.
reply