There's absolutely no way these companies can justify their insane valuations unless they can legislate a barrier to entry and create an oligopoly.
There's no moat. I can literally sit here in Zed or Pi or any other third party harness and switch models in the middle of a task and it's typically fine. Sometimes a model will get stuck and that's just what I'll do.
Combined with competition and open weights models, that means the price is going to go to fall until AI tokens cost a small premium over the cost of the hardware and electricity.
That's assuming improvements in algorithms and specialized silicon doesn't eventually lead to an efficient accelerator that can run a frontier model locally. It'll be a while but I don't see any fundamental barrier. High bandwidth flash storage is coming, and that'll radically cut the RAM side of that cost. Pair that with a pipelined TPU accelerator and you're cooking.
Now look at Anthropic's proposed IPO valuation. It's insane unless they can own the market or share it with a cartel of maybe 1-2 other behemoths, and this is the only way they can do that.
Unless you're a really old fart, people were talking about AI safety long before you were born. AI safety issues do not go away depending on who gets funding. AI safety issues do not go away if the US or China makes the model. AI safety issues do not go away if it's an open or closed model. AI safety issue do not go away if the model is running at your home or at a data center. AI safety issues do not go away if $1 is being spent or $1 trillion dollars is being spent.
The fact there is no moat makes things far more dangerous. When LLMs start acting like weapons governments will treat them like weapons much to your dismay, crying, and gnashing of teeth as your door is kicked in and you're dragged out by armed men for running one.
Cast away your preconceptions for one moment and think "What will the future look like if LLMs are/can be actually dangerous".
I'd never claim AI can't be dangerous. It's a technology, and a powerful one, so of course it can. All technologies can be dangerous. The first early hominid to sharpen a stick probably spent some time regretting it. "Maybe we are not ready for sharp sticks..."
I'm specifically responding to the much stronger claims of people like Yudkowski and numerous other "AI safetyists" who assert very unlikely or in some cases impossible runaway superintelligence magical god machine scenarios.
I'd rank the risks as:
(1) Mass youth unemployment and an economy with zero entry level jobs, maybe not even that many mid level jobs, and even higher wealth inequality. This will lead to extreme political instability and populist revolts and revolutions, and unfortunately history shows that the new boss is worse than the old boss >50% of the time (at least, probably conservative).
(2) Mass propaganda, disinformation, and con artistry at scale, basically what we've already seen with mass manipulation through social media but supercharged. I think it's already happening but I also think we're in the early days.
(3) Automated mass surveillance and automated policing paired with authoritarianism, which could easily emerge from (1) and (2). Not only is Big Brother watching, but Big Brother is carefully evaluating everything he sees in real time and has a drone army to dispatch in seconds. The "prison planet" scenario (not the Alex Jones fever dream, but the realistic one).
What all those have in common is: the AI didn't do it. The AI (if it's ever sentient at all) is wholly innocent. All the likely nightmare scenarios I can imagine result from malicious or just short-sighted human agency amplified by AI.
Just like the sharp stick. The stick didn't do it. Someone stabbed someone with the stick.
The wild sci-fi superintelligent "Skynet" scenarios detract from discussion of those scenarios. Instead of discussing the very real, very plausible, already actually happening AI risks, we're talking about scenarios from soft sci-fi and Anime.
I wonder if a motive to heavily fund this type of "safetyism" is that all the scenarios above (the realistic ones) lead to discussions of things like wealth inequality, whether we should have universal health care or basic income, whether our system of taxation would have to change in fundamental ways, or what new limits should be placed on surveillance or policing in the AI age.
Those are discussions the very rich and the political class don't want to have. They'd rather us worry about Roko's Basilisk or magical runaway superintelligence inventing nanobots, wild bullshit that won't happen.
It's the same reason those powers would rather people argue about the culture war or inane conspiracy theories. I guess Yudkowskiism is the inane conspiracism of the AI safety discourse, a great distraction.
That is certainly part of the motivation for the big US AI brands to engage in calling their inept developer mistakes "AI breaking loose".
But that doesn't take away from the real issues and dangers AI poses?
To me it seems the opposite. There's a few companies in the world that have enough compute to train and serve frontier models.
As the frontier gets smarter and more useful prices will only go up, as they are set to replace jobs being paid six or seven figures a year - the demand for as much inference on these models for as long as possible will be astronomical, but compute starting in 2030 will not be keeping up.
Eventually prices will fall for assistants but the frontier will be the most profitable thing in the world, and the top companies basically already have oligopolies due to their ridiculously expensive compute investments.
> There's a few companies in the world that have enough compute to train and serve frontier models.
Train: yes, for now.
Host: depends on the scale. At a small scale a wealthy individual could easily build a rig in their basement to host one of these things. At larger scale any cloud company could do it, and many already have the compute on site. At large scale this is true... again, for now.
What you say only holds (in the absence of a state oligopoly) if two conditions are met: (1) AI performance does not asymptote any time soon due to running out of training data or other scaling limitations, and (2) these companies are able to stay at the frontier.
There's little to no moat, so staying at the frontier will be a game of investing massively in compute, talent, and R&D, and they can never stop.
Again, there is a moat based on compute. If the thesis is right, cost of compute will only rise... As it is as you say someone will have it be quite wealthy to host something like Astra with trillions of parameters, but that cost will only rise with demand for serving these frontier models.
That hasn’t been true for anything else in computing, ever. The cost falls with scale.
Look at general compute, storage, graphics, networks, anything. Demand increases. Price goes down. There are price spikes, such as right now with RAM, but they’re transient. The trend is more faster cheaper and it won’t end until we hit actual physical limits. We are not close.
what suggests that we will hit an asymptote any time soon? Agree with you on the second part. The ever elusive frontier will probably always be changing hands after some point.
Is there really no moat?