←
illiquid9,607 posts
illiquid avatar

illiquid

@lefttailguy

routers eat the world; consilient software observer; crafting an alchemist's paradise. economic meta-learner. DMs open.

◷ Joined September 2020▧ 868 Photos and videos
illiquid avatar
illiquid@lefttailguy·Jun 24

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/status/2056…) 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.

Post attachment

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.

illiquid avatar
illiquid@lefttailguy·22m

New Made in China label just dropped

SITUATION DETECTED: The White House and Treasury are discussing a proposal that would require US companies to publicly disclose the use of Chinese open-weight AI models on their websites, aiming to discourage their use through public pressure, per Reuters.

illiquid avatar
illiquid@lefttailguy·Nov 17

.@AntithesisHQ for everything. I’m incredibly bullish on the team + platform.

Every piece of critical infrastructure or network (grid, telecom, finserv, etc.) will need a digital replica. A simulation suite. A deterministic testing suite. Etc. The tokens will continue to burn.

illiquid avatar
illiquid@lefttailguy·Nov 16

Peter Thiel has done a lot of good for Silicon Valley and the world. But one of the ways he (unintentionally) fell short is by making some of the smartest, most well-capitalized people in the world allergic to ideas of convergence, 1 to n, etc. In SV, it's incredibly low status to think in terms of convergence, and to get excited about solving highly tractable problems where capital + disciplined execution are really the only barriers. This shows up in how people allocate their philanthropic dollars and how they spend their time. Of course, radically new paradigms are extremely important to progress, but I think people in SV underestimate how much the process of diffusing existing paradigms / executing against tractable objectives can teach you. For example, Zipline perfected it's drone delivery systems by delivering medicine to hospitals in Africa.

Tech people think too much about forests and not enough about trees Loving “humanity” while not caring about humans. Building for “future generations” while not raising children But the optimal amount to sacrifice rationality for sentiment is not zero Sometimes the best way to make the world better isn’t through future hypotheticals or abstract aggregates but solving specific problems, for real people, in the present

illiquid avatar
illiquid@lefttailguy·Nov 16

Meta continues to feel very prescient as this machine that invents primitives for adding useful economic information to the general purpose technology. Ways to make our bids and asks legible through interactions that feel “mindless” / frictionless / natural.

illiquid avatar
illiquid@lefttailguy·Nov 16

Does AI contribute to its own growth more than say, the facebook newsfeed contributed to its own growth? At present, I'd say it does not.

And why should we think today's AI is so vastly more easily growable that normal tech?

illiquid avatar
illiquid@lefttailguy·Nov 16

Kinks in the singularity: the origin of language, the origin of christianity, the origin of the joint stock corporation, the origin of digital computation, the origin of deep learning, the origin of ____ (something secret that I know about but won't tell you).

illiquid avatar
illiquid@lefttailguy·Nov 15

Unlike people who think the frontier labs will interalize all the gains, I think that the quest to make machine intelligence useful is likely to look like a decentralized search. Much like evolution. Scaling that search requires reliable / shared / trustless primitives for parallel experimentation, interaction, transaction, etc. All the tech trees that can be used will be used. Crypto, AR/VR, HUMANS etc. machine intelligence increases the option value of the rest of the technology stack.

In the coming machine intelligence biosphere, companies like Cloudflare and Stripe provide the fundamental primitives--the oxygen, carbon, sunlight, etc.--that allow swarm intelligence to flourish. And, importantly, these primitives become more and more useful as intelligence becomes more and more diverse.

Post attachment
illiquid avatar
illiquid@lefttailguy·Nov 15

"1) it may be obvious at this point, but it is worth noticing that economic value from ai accrues to frontier labs and not to application ai companies" That's wrong. The plurality of economic value from AI has accrued to the infrastructure layer (Nvidia, Micron, Siemens Energy, GE Vernova, TSMC, etc.) Nvidia alone has added ~$3.6 trillion in market cap since the end of 2022. And, you're grossly under-accounting for "application layer" success here: Palantir, AppLovin, Meta (even pre-frontier lab status), etc. have collectively accrued ~$2 trillion since ChatGPT came out in November 2022. Cognition, Cursor, Stripe, Ramp, count too. Lastly, there is a (substantial) possibility that we are in a temporary equilibrium re: value capture. I don't doubt that the frontier labs have created the plurality of the value, but I still have lots of doubts that they'll capture it in the fullness of time.

some quick thoughts on ai & value creation 1) it may be obvious at this point, but it is worth noticing that economic value from ai accrues to frontier labs and not to application ai companies 2) frontier labs have created ~$2.5tn of market value since 2023 while application ai companies have created ~$360bn and this includes companies like elevenlabs 3) in addition, openai and anthropic have ~$130bn of revenue while application companies have ~$13bn in total revenue; > $115bn of frontier revenue doesn't even pass through application ai companies 4) i think that this emphasizes the fact that most value from ai does not come from productitization because if it did we would expect to see more revenue at application companies 5) instead, the value seems to come from the model itself and then from integrating those models effectively into existing organizations; but, the later is not a product feature movement 6) there have been signs of this since at least 2023, when it was so easy for large companies to introduce ai variants of their existing products, it was just an api call 7) but, i think we can more definitively say today that the model is the product and the scaffolding that you build around them is much less important on the whole 8) and, this is very good for the frontier labs, which may become less organizationally nimble, but which will retain a very strong advantage in model training itself 9) due to high capital requirements for the necessary compute, an existing lead in research agents, etc... 10) and, frontier labs will therefore be likely to both create and capture most of the concentrated value with respect to ai, across a very wide range of verticals

illiquid avatar
illiquid@lefttailguy·Nov 15

The success of institutions will increasingly be a function of its ability to identify, produce, optimally coordinate / allocate Comprehensive Designers. Many of our most powerful institutions have almost the exact opposite objective function at present. This is probably the largest sovereignty gap that currently exists.

Post attachment
illiquid avatar
illiquid@lefttailguy·Nov 15

BendingSpoon's return on incremental invested capital is the right economic KPI to track to ascertain whether we are in a software only singularity. If we see that Bending Spoons can deploy progressively more capital without diminishing returns, we have our answer.

illiquid avatar
illiquid@lefttailguy·Nov 13

In the coming machine intelligence biosphere, companies like Cloudflare and Stripe provide the fundamental primitives--the oxygen, carbon, sunlight, etc.--that allow swarm intelligence to flourish. And, importantly, these primitives become more and more useful as intelligence becomes more and more diverse.

Post attachment

Over time, more useful capabilities will be possible on increasingly smaller models while tps rises. Cloudflare is going to capture value on tasks that must be done quickly and in a responsive environment with low latency.

illiquid avatar
illiquid@lefttailguy·Nov 13

I wonder if OpenAI ever becomes a health insurer (with the primary motivation being access to linked longitudinal clinical / biometric data) .

Post attachment
illiquid avatar
illiquid@lefttailguy·Nov 13

when this hits 100%, will we have an AI elon musk? If not, we've got a long way to go.