The AI dooooooom thread

[crazy canuck]

Where is their revenue going to come from? Companies will not have shed millions in salary to replace by spending even more for AI.

The private person market doesn't exist.

That is the question that has not been answered

[Sheilbh]

Where is their revenue going to come from? Companies will not have shed millions in salary to replace by spending even more for AI.

The private person market doesn't exist.

The world economy's about $120 trillion - somewhere between 50-55% of that goes to labour. That's the goal. Replace the share that goes to labour to a share that goes to capital (how the world economy functions without paid workers is less clear - but I think this also goes slightly to the slight incestuous elment of it all). I also find the cost point a bit uncertain - I think that relies on the cost of AI being stickier than the cost of employing human beings and I'm not sure about that (for example I'd suspect you'll have a big impact with, for example, the company that fixes the memory issue - but then again, maybe it can't be fixed).

Obviously there are AI resistant jobs but even shifting just a bit of that is significant.

[crazy canuck]

That's utter nonsense.  The revenue these AI companies need to pay for the infrastructure bill out is going to be a lot more than shifting labour markets a little bit.

And by the way labour costs have gone up because companies need to employee senior people who
Know what they are doing to work with and fix the errors AI makes.

[Sheilbh]

Where's the nonsense? I've described their goal and you've said it's not possible/going to happen. The answer to GF's question is there absolutely is a private person market. It's the share of the economy going to labour and they're aiming for it.

OpenAI's mission statement was about ensuring "artificial general intelligence" is to the benefit of humanity and their definition of artificial general intelligence is "highly autonomous systems that outperform humans at most economically valuable work". The replacement of labour by capital is one that's played out many times before. I think this is different because of the nature of the replacement. The global share of the economy that goes to labour is around 50-55% of $120 trillion - that's the prize they've got their eye on and a little bit of re-arranging on that global reality (especially with techs trend towards monopoly) is a very big prize indeed. But I think their goal is bigger - this isn't a spinning jenny but aiming for "most economically valuable work". So their goal is 25%+1 of $120 trillion which is not insignificant (but even 1-2% of that is big) - and again I'm less confident that the cost of infrastructure is going to be less sticky than the cost of human labour.

Whether it's going to happen etc is a separate issue. I have no idea but I lean far less confident in any direction.

[crazy canuck]

The goal is as realistic as the plan to colonize Mars.

Every age has its suckers, it's just that this age has more than usual.

[The Minsky Moment]

The world economy's about $120 trillion - somewhere between 50-55% of that goes to labour. That's the goal. Replace the share that goes to labour to a share that goes to capital (how the world economy functions without paid workers is less clear - but I think this also goes slightly to the slight incestuous elment of it all).

Put aside the question of whether that makes sense, and let's assume that happens. $30 trillion that now goes to compensate labor now goes to "capital" because AI has replaced the workers.  I agree that's the assumption baked into the AI bull case and what sustains the xAI valuation in the SpaceX IPO.  4X annual revenue up until 2030 = $30 trillion revenue, which validates the SpaceX valuation and Elon's current net worth, and which allows Anthropic and OpenAI to emerge as the leading global companies as opposed to setting records for the largest cases ever to hit the Delaware Bankruptcy courts.

Let's assume that the capability of AI is sufficient to achieve that level of rapid substitution, that political backlash doesn't crush it, that the data centers stuck in the imagination stage get built to provide the "compute," that the massive energy generation capacity required to make those data centers run suddenly materializes, complete with oceanic capacities of water for cooling.  Everything comes together like a George Peppard plan c. 1983 and all the skeptics end up fools to be pitied.

There is still one massive unwarranted assumption left.  That much of the cost "savings" from redundant labor gets channeled into revenue for the AI companies.  But absent a global monopoly that seems really unlikely.  None of these companies has a decisive secret sauce.  Whatever the leading cutting-edge model is in time T, becomes the model that all the followers are using in T+6 months. Unless that changes fundamentally, the industry can't control pricing and AI is doomed to be a utility.  There's always going to be a DeepSeek offering 98% of function for a fraction the price.

[Sheilbh]

So I think there's a few sides to that.

On the first section in terms of the bankruptcy there's two other key elements in relation to these companies in particular as opposed to "AI" in general. One is that from my understanding AI basically is the growth in the US economy right now. The other is that I think there is a sense - which may well be wrong but I think is strongly held - that this is an essential frontier technology to have "sovereign" versions of and that China is significantly investing and implementing AI in its economy (as on everything I think we need to incorporate China a lot more into the way we in the West think about the world). Those two factors, to me, point to a too-big-to-fail, Federal government stake/ultimate backstop that is implicit in this companies. It'd be an absolute shitshow politically but if they're the one bit of your economy that's growing and they're producing technology that people think is essential to have with a bloody big American flag on, then I'm not sure "failure" is going to be an option (there may be individual losers - from what I can see the Federal government has zero interest in Grok but are desperate to keep Anthropic).

In the context of AI companies my understanding is that at the token level they're making money. The actual products are profitable with a decent margin - not as good as pure software (also moving to rely on compute/everything-as-a-service) but solid. The big costs and money sinks are in building and training the new models because they're not at that point of replacing "most economically valuable work". So I think the question is what happens when/if they reach that point because it seems to me once they're "good enough" they don't necessarily need to carry on developing - certainly not at the same pace, certainly not at the same cost. This is why I end up more on the side of thinking this needs to be under democratic control through sovereign, state backed or owned mechanisms (that could also say, for example, compensate human ceativity, respect IP rights and decide that trying to replace "most economically valuable work" is a less good use of these resources than AlphaFold even if it's a better market proposition).

I totally agree on the utilities point - I've seen some report that stickiness is increasing and it's very like SaaS where they're becoming more embedded, people have preferences and moving at an enterprise level is becoming more difficult. I think that's the direction of travel and as I say where I think they're planning to end up is not a million miles away from Microsoft, Oracle, Salesforce etc. I think there is a geopolitical complication for other challengers in that I suspect this will be an area where you can't use DeepSeek if you ever want a US government contract, say. I could be wrong but I suspect there'll be consolidation and national champions and it won't be a global market. At this point it'll be whether you're on the American or Chinese version (though Mistral is solid). But I think at the minute their path is to become essential, difficult to move away from enterprise products (less sexy than existential risk).

I think the other slightly wider point is that I think there's a difference between these specific companies right now and AI. The examples people use of bubbles in this case are railways and the internet - not the South Sea Bubble - because there is an underlying technological change. The railways and the internet continued, expanded and became more important following those bubbles bursting. My view from the start has basically been the binary of either this is going to replace significant portions of the labour market with enormous social, economic and political consequences or it's one of the biggest misallocations of capital in world history which will fail producing enormous social, economic and political consequences - my views have now evolved to admit the thrilling possibility that it could be both. These companies could fail, there could be a crash - but the underlying technology and its "promise" (as currently and rather unpleasantly imagined by Silicon Valley) have had significant breakthroughs and will continue to develop. I don't think the end of a bubble means the end of AI (particularly not in China where I suspect it would accelerate).

[The Minsky Moment]

On the first section in terms of the bankruptcy there's two other key elements in relation to these companies in particular as opposed to "AI" in general. One is that from my understanding AI basically is the growth in the US economy right now. The other is that I think there is a sense - which may well be wrong but I think is strongly held - that this is an essential frontier technology to have "sovereign" versions of and that China is significantly investing and implementing AI in its economy (as on everything I think we need to incorporate China a lot more into the way we in the West think about the world). Those two factors, to me, point to a too-big-to-fail, Federal government stake/ultimate backstop that is implicit in this companies.

It isn't too big to fail in the economic contagion sense.  The economy needs an operating banking system to function. It doesn't strictly need a wall of money being thrown at RAM and GPUs.  A fall in investment will hurt the economy but that's always true.  I don't think the government can step in and become the investor of last resort without transforming into the PRC.

Which bring in the strategic element. If the competition boils down to who mobilizes the largest mass of "compute" then China wins, it's just a question of time, since the key physical bottleneck is power generation. But I think the whole premise is full of questions to begin with.

In the context of AI companies my understanding is that at the token level they're making money.

That's not a very impressive achievement.  It's basically saying - ignore the parts of the business that cost enormous of gobs of money, and then look how profitable it is.  Just talking about that reminds me of the crazy alternative measures that got stuck in prospectuses at the tail end of the 90s dot com boom.

So I think the question is what happens when/if they reach that point because it seems to me once they're "good enough" they don't necessarily need to carry on developing - certainly not at the same pace, certainly not at the same cost.

The two possibilities are: (1) They never reach that point because the business case always relies on continual optimization or (2) it turns into a cheap utility with no pricing power.

My view from the start has basically been the binary of either this is going to replace significant portions of the labour market with enormous social, economic and political consequences or it's one of the biggest misallocations of capital in world history which will fail producing enormous social, economic and political consequences - my views have now evolved to admit the thrilling possibility that it could be both.

That's been my take all along. The dot com boom optimists weren't entirely wrong, they were just way too aggressive in their timetable.  If humanity still exists in 30 years, AI is going to be big. But that doesn't mean 2-3 trillion in investment now can be financially validated in the 2020s.

[Sheilbh]

So I suppose on that final point - I find the technology and the questions around that more interesting (and alarming in the wrong hands/for the wrong purposes such as replacing "most economically viable work") than these companies or if the investment levels are financially justifiable right now. As I say my default is there's enormous political, economic and social challenges coming either way - what I think is the bigger question is what this underlying technology is/will be capable of. And I think those two are often conflated.

I think on the transforming into the PRC - I think we're all on that journey to some extent or other. See the US' acquiring a stake in Intel various movements on industrial policy and protectionism in the more rationally acting bits of the world. I think there'd absolutely be a a stake acquired in any bail-out, but I think more broadly we're coming into an age of "new state capitalism" and I think this is an area that needs champions. My concern is, like Thierry Breton's, that Europe is at this point heading the position it is with social media and with the internet (and energy) of picking between the US or China and pretending that regulatory superpowers means anything in that world. Which is why I hope Mistral does well.

I have to be honest I'm always slightly baffled on the concern around power because from my understanding the US hasn't seen an increase in electricity generation for twenty years and the grid is not great. Even aside from AI, like everyone else, the US needs vastly, vastly more electricity for energy transition and a grid that can handle it. It's a problem with AI but it seems like it's a bigger problem that needs fixing through more generation and better grid (which would also help the AI industry). I slightly worry this will negatively polarise in an unhelpful way where new electricity and new grid becomes AI becomes bad because of the growing opposition - when new electricity and new grid is essential for so many other things and basically, in my view, the 21st century.

[crazy canuck]

On the first section in terms of the bankruptcy there's two other key elements in relation to these companies in particular as opposed to "AI" in general. One is that from my understanding AI basically is the growth in the US economy right now. The other is that I think there is a sense - which may well be wrong but I think is strongly held - that this is an essential frontier technology to have "sovereign" versions of and that China is significantly investing and implementing AI in its economy (as on everything I think we need to incorporate China a lot more into the way we in the West think about the world). Those two factors, to me, point to a too-big-to-fail, Federal government stake/ultimate backstop that is implicit in this companies.

It isn't too big to fail in the economic contagion sense.  The economy needs an operating banking system to function. It doesn't strictly need a wall of money being thrown at RAM and GPUs.  A fall in investment will hurt the economy but that's always true.  I don't think the government can step in and become the investor of last resort without transforming into the PRC.

Which bring in the strategic element. If the competition boils down to who mobilizes the largest mass of "compute" then China wins, it's just a question of time, since the key physical bottleneck is power generation. But I think the whole premise is full of questions to begin with.

In the context of AI companies my understanding is that at the token level they're making money.

That's not a very impressive achievement.  It's basically saying - ignore the parts of the business that cost enormous of gobs of money, and then look how profitable it is.  Just talking about that reminds me of the crazy alternative measures that got stuck in prospectuses at the tail end of the 90s dot com boom.

So I think the question is what happens when/if they reach that point because it seems to me once they're "good enough" they don't necessarily need to carry on developing - certainly not at the same pace, certainly not at the same cost.

The two possibilities are: (1) They never reach that point because the business case always relies on continual optimization or (2) it turns into a cheap utility with no pricing power.

My view from the start has basically been the binary of either this is going to replace significant portions of the labour market with enormous social, economic and political consequences or it's one of the biggest misallocations of capital in world history which will fail producing enormous social, economic and political consequences - my views have now evolved to admit the thrilling possibility that it could be both.

That's been my take all along. The dot com boom optimists weren't entirely wrong, they were just way too aggressive in their timetable.  If humanity still exists in 30 years, AI is going to be big. But that doesn't mean 2-3 trillion in investment now can be financially validated in the 2020s.

On that last point, it depends on what you mean by AI, if you mean generative AI. I don't think it will be around in 30 years. If you mean other small scale applications that actually aid the user to carry out a task-certainly.