Pinned post
Folks (see screenshot) have made the point that the labs only goal is to sell more tokens.
So I'll update what I said in the quote tweet to be more precise: if the labs' goal/long term approach is to sell/provide universal access to fixed-unit cost tokens, they will fail.
If they labs want to succeed, the goal of their DeployCo's and (their strategy writ large) needs to be to
1) develop a predictive model that understands which pockets of the economy, at any time, produce the highest returns to frontier intelligence and what those returns are, and
2) build the deployment meta-learner (a continual learning system) that allows them to realize those returns faster, more reliably, and in more valuable areas than anyone else.
Building/improving 1) tells the labs where to allocate capital (at the limit, compute) and how much to allocate, Building/improving 2) expands the pockets of economic complexity from which they can practically realize returns from allocating machine intelligence (areas of the economy that appear "API complete" to them)
Crucially, 2) provides the ground truth signals that allows 1) to become more predictive over time.
You could optimistically say that developing these capabilities will allow the labs to power positive sum intelligence auctions (to price tokens dynamically in a way that's win, win, win for everyone involved) across the economy. But for variety of reasons (technical, economic, political) I think this is unlikely to work.
The reality instead, is that continuous vertical integration over the (always evolving) distribution of the economy with the highest returns to intelligence—high and increasing task complexity—is the only way they can capture durable value. And these capabilities are what they need to transition to and power this vertically integrated business model.
The labs have realized that selling/provide universal access to frontier tokens AND trying to durably capture value from those tokens (which is required to attract the capital necessary to stay on the frontier in the first place) means continuously competing with your best customers: those that have high returns to frontier tokens.
They know this isn't a coherent strategy and that they need to change course. Selling your alpha to competitors means eroding your long term returns, so you will rationally stop or be forced, by the capital markets, to stop doing that.
(If the logic here is confusing, see @hypersoren's excellent piece below that provides a much clearer, analytical explanation for why you cannot capture durable value by selling tokens. I'm sure, or at least hope, that everyone important at the labs has read and/or understands this.
x.com/hypersoren/sta…)
Ok, so, now that they've decided that this is no longer tenable, and that they need to create and privately monetize machine intelligence via automated capital allocation and deployment capabilities, what does that mean for the rest of us?
Knowing this, the rational move for everyone else is to not fall for the trap of buying from (and in the process) leaking your scarce/singular context to your future competitors (the labs). This is true at the level of individuals, firms, and economies!
Cursor gave us a nice existence proof for what you don't want to happen to the rest of the economy. This supplier, to competitor, to existential threat dynamic was clearest to Cursor first, which is why you saw them post-train their own model first. Unfortunately, in their case, they realized it was already too late, and that the only economically rational decision was to sell to the lab overlords.
They sold because they knew that, in coding, the only only thing standing in the way of the labs getting access to the the scarce context necessary to produce a superior offering was capital. And capital was something the labs of course had in spades. Abundant capital (compute) allowed them to 1) scale pre-training and post training to the point where their coding products were immediately at parity with whatever Cursor could offer, and 2) subsidize tokens such that all of the new context/signal necessary to stay on the frontier/be more useful than Cursor was produced in their harnesses. "Harness / model co-optimization", etc.
Many smart people have extrapolated/generalized this as the likely end state of everything (not just coding), and come to the conclusion that they are better off joining the labs rather than trying to compete with them.
Luckily, lab access to the scarce context necessary to allocate capital / compete / effectively vertically integrate in other areas with high returns to intelligence seems to be more bottlenecked by humans/trust/time than compute/capital, which means that if everyone starts acting in their own best interest, we will make it out of this alive.
(see Dwarkesh's Data Black Hole, Tom Reed's Goodhart Singularity, @anjali_shriva and @joodalooped's RLWD pieces for intuition here)
The labs (not because they are bad people, but because capitalism/markets are a brutal natural selection machine) have of course been finding every way they can to surmount these diffusion bottlenecks, buy themselves more time, obviate these bottlenecks all together, etc. DeployCos, Thrive Holdings, calls to slow development by fearmongering about RSI, data retention under the guise of safety, regulatory capture, are some examples, but there are many more.
We will see their desperation accelerate as open-source catches up to the frontier and as this logic becomes increasingly legible to everyone. You'll see them using capital and aggressive political narratives to try to brute force/coerce these bottlenecks or obviate them. Again this is not because they are bad people, it's just what they need to do to survive.
In other words, the market will work as it should and reallocate capital from the labs to other economic actors, as long as we let it! We mustn't let lab narratives around RSI, existential risk etc. give them the capital, time or political ammunition they need to make the economy look like a central planner or the country look like one with a natural resource curse.
Luckily, particularly in the past few weeks, it seems savvy economic actors are already well aware of what's going on and are finally publicly calling the labs' bluff (with words and capital).
All of a sudden, everyone is post-training their own models, talking about the importance of an open eco-system and the existential need to own/control your proprietary learning loops (context, evals, selection infra)
CEOs like @satyanadella and @nikesharora are making it abundantly clear that this is existential (and that, of course, they are the ones who can help you!)
And, luckily, many have known this since the beginning. Firms like @formationbio architected themselves from day one (pre-ChatGPT) around proprietary learning loops that they do not rent from others OR rent to others. They instead are architected to monetize their search and learning in ways that cannot be distilled and commoditized. Structuring themselves this way is what will allow them to attract and deploy the capital necessary to provide durable value to the patients they hope to serve!
Even Lilly owns their own cluster, is hiring for 100s of AI related roles (yes, I've been monitoring their job board), and is partnering with the non-frontier lab AI entities that are incentive aligned to provide capabilities without becoming competition.
So we should be alright, if we don't let the psychosis swallow us.
So I'll update what I said in the quote tweet to be more precise: if the labs' goal/long term approach is to sell/provide universal access to fixed-unit cost tokens, they will fail.
If they labs want to succeed, the goal of their DeployCo's and (their strategy writ large) needs to be to
1) develop a predictive model that understands which pockets of the economy, at any time, produce the highest returns to frontier intelligence and what those returns are, and
2) build the deployment meta-learner (a continual learning system) that allows them to realize those returns faster, more reliably, and in more valuable areas than anyone else.
Building/improving 1) tells the labs where to allocate capital (at the limit, compute) and how much to allocate, Building/improving 2) expands the pockets of economic complexity from which they can practically realize returns from allocating machine intelligence (areas of the economy that appear "API complete" to them)
Crucially, 2) provides the ground truth signals that allows 1) to become more predictive over time.
You could optimistically say that developing these capabilities will allow the labs to power positive sum intelligence auctions (to price tokens dynamically in a way that's win, win, win for everyone involved) across the economy. But for variety of reasons (technical, economic, political) I think this is unlikely to work.
The reality instead, is that continuous vertical integration over the (always evolving) distribution of the economy with the highest returns to intelligence—high and increasing task complexity—is the only way they can capture durable value. And these capabilities are what they need to transition to and power this vertically integrated business model.
The labs have realized that selling/provide universal access to frontier tokens AND trying to durably capture value from those tokens (which is required to attract the capital necessary to stay on the frontier in the first place) means continuously competing with your best customers: those that have high returns to frontier tokens.
They know this isn't a coherent strategy and that they need to change course. Selling your alpha to competitors means eroding your long term returns, so you will rationally stop or be forced, by the capital markets, to stop doing that.
(If the logic here is confusing, see @hypersoren's excellent piece below that provides a much clearer, analytical explanation for why you cannot capture durable value by selling tokens. I'm sure, or at least hope, that everyone important at the labs has read and/or understands this.
x.com/hypersoren/sta…)
Ok, so, now that they've decided that this is no longer tenable, and that they need to create and privately monetize machine intelligence via automated capital allocation and deployment capabilities, what does that mean for the rest of us?
Knowing this, the rational move for everyone else is to not fall for the trap of buying from (and in the process) leaking your scarce/singular context to your future competitors (the labs). This is true at the level of individuals, firms, and economies!
Cursor gave us a nice existence proof for what you don't want to happen to the rest of the economy. This supplier, to competitor, to existential threat dynamic was clearest to Cursor first, which is why you saw them post-train their own model first. Unfortunately, in their case, they realized it was already too late, and that the only economically rational decision was to sell to the lab overlords.
They sold because they knew that, in coding, the only only thing standing in the way of the labs getting access to the the scarce context necessary to produce a superior offering was capital. And capital was something the labs of course had in spades. Abundant capital (compute) allowed them to 1) scale pre-training and post training to the point where their coding products were immediately at parity with whatever Cursor could offer, and 2) subsidize tokens such that all of the new context/signal necessary to stay on the frontier/be more useful than Cursor was produced in their harnesses. "Harness / model co-optimization", etc.
Many smart people have extrapolated/generalized this as the likely end state of everything (not just coding), and come to the conclusion that they are better off joining the labs rather than trying to compete with them.
Luckily, lab access to the scarce context necessary to allocate capital / compete / effectively vertically integrate in other areas with high returns to intelligence seems to be more bottlenecked by humans/trust/time than compute/capital, which means that if everyone starts acting in their own best interest, we will make it out of this alive.
(see Dwarkesh's Data Black Hole, Tom Reed's Goodhart Singularity, @anjali_shriva and @joodalooped's RLWD pieces for intuition here)
The labs (not because they are bad people, but because capitalism/markets are a brutal natural selection machine) have of course been finding every way they can to surmount these diffusion bottlenecks, buy themselves more time, obviate these bottlenecks all together, etc. DeployCos, Thrive Holdings, calls to slow development by fearmongering about RSI, data retention under the guise of safety, regulatory capture, are some examples, but there are many more.
We will see their desperation accelerate as open-source catches up to the frontier and as this logic becomes increasingly legible to everyone. You'll see them using capital and aggressive political narratives to try to brute force/coerce these bottlenecks or obviate them. Again this is not because they are bad people, it's just what they need to do to survive.
In other words, the market will work as it should and reallocate capital from the labs to other economic actors, as long as we let it! We mustn't let lab narratives around RSI, existential risk etc. give them the capital, time or political ammunition they need to make the economy look like a central planner or the country look like one with a natural resource curse.
Luckily, particularly in the past few weeks, it seems savvy economic actors are already well aware of what's going on and are finally publicly calling the labs' bluff (with words and capital).
All of a sudden, everyone is post-training their own models, talking about the importance of an open eco-system and the existential need to own/control your proprietary learning loops (context, evals, selection infra)
CEOs like @satyanadella and @nikesharora are making it abundantly clear that this is existential (and that, of course, they are the ones who can help you!)
And, luckily, many have known this since the beginning. Firms like @formationbio architected themselves from day one (pre-ChatGPT) around proprietary learning loops that they do not rent from others OR rent to others. They instead are architected to monetize their search and learning in ways that cannot be distilled and commoditized. Structuring themselves this way is what will allow them to attract and deploy the capital necessary to provide durable value to the patients they hope to serve!
Even Lilly owns their own cluster, is hiring for 100s of AI related roles (yes, I've been monitoring their job board), and is partnering with the non-frontier lab AI entities that are incentive aligned to provide capabilities without becoming competition.
So we should be alright, if we don't let the psychosis swallow us.
The goal of DeployCo is obviously to 1) turn each deployment into a training case for a meta learner that can automate the process of turning tokens into economic value and 2) as a search function to find which pockets of economic complexity produce the greatest (short and long term) returns on machine intelligence.View quoted post on X
