But there are 4 other ways it could work, which preserves the free market and avoids the concentration of power that UBI requires.

Option 1: hardcore anti-trust protections.

Ensure that there’s full competition between AI firms paired with 100% job automation. In a world like this, all intelligence tasks (and everything they produce) reaches virtually infinite supply. And if there’s competition, that infinite supply translates to ~0 price for everything. Maybe the government could do a round of UBI once... but it’s the last dollar you’d ever need.

This might sound like an alien world... but it’s actually a world you already live in... we call it “nature”. The rain falls, the sun shines, the wind blows... and nobody has to pay or gets paid for it. Full automation just means a technology joins nature in its provision for the world.

Option 2: decentralized IP

An AI model can be thought of as a large collection of individual concepts, which are used in concert together to form predictions. Another model is one where ownership over each of those concepts (and the neurons that run them) are owned by individual people in the world. And when you need to make an AI prediction, those individuals can extract rent.

While this might also sound like an alien world... it’s also the world we already live in. When you go to a lawyer or doctor or whatever... you’re really just renting the use of society’s intellectual concepts... stored in the brains of those people. The doctor or lawyer you work with did NOT invent those concepts (they went to school to copy them from society into their individual brain). We could continue this model into a post AGI world... a group of 15 million people continue to be the world’s “doctors” because they store/curate/etc. the world’s “doctor” concepts and the world pays them whenever they need doctoring.

(for more on this attribution-based-control.ai)

Option 3: full-time work building evals

While this might sound like an alien world... it’s also the world we already live in. In order for the economy to serve humanity’s interests, every buyer in the world is partially employed doing the same thing... saying yes/no to various prices. And every employee is doing the same thing... actively modeling the value of their labor. In doing so, they are value-aligning the global economy to the trillions of individual tasks in 8 billion unique contexts (per second). The surface area of “what does a human want / what does humanity want in this scenario” is truly, incredibly large. And there are disagreements on humanity’s definition of “good”... (e.g. “politics”). One can easily imagine everyone in the world continuing to be employed as informers of what the world needs at that moment...

Option 4: explode in diversity

I think that the last 200 years of mass production and mass media simultaneously gives us the sense that everyone in the world mostly wants the same thing. And to some extent, we do. But AI (and maybe 3d printing) will wildly expand our ability to have nuance at scale... nuanced media experience... nuanced/customized manufacturing..etc. One can imagine that a post AGI world simply becomes so much bigger and weirder that we go out of our way to find the outer limits of what AI automation can support... this is Jevon’s paradox taken to the extreme... it’s not just that we do interstellar travel or whatever...we turn ourselves into micro-celled organisms... into weird collective hive-minds... and all manner of wacky, crazy thing which then makes the management of resources so insane that even modern tools for AI/AGI start to strain under them... and scarcity (and the economy of supply/demand) lingers on... never quite able to catch up with humanity’s ability to twist into even more rare/diverse forms which require limited supply.

This might sound like an alien world, but it too is the world we already live in. On the small scale, consider status games like fashion. Elites pick a new style which only the elites show... the masses get faster and faster at copying it... which means the elites have to move on faster and faster (thus the term... fast fashion). But as the masses catch up faster the elites get more creative wiht all the ways they might be clothed that others aren’t (e.g. now we have dresses with robot arms?). One can imagine a world where everyone has such infinite suppplies...and status games persist.

On the large scale, consider bio-engineering. For supply/demand to persist... we only need *something* which is rare and perceived to be valuable. Making something rare is mostly about rolling the dice, and making something valuable is mostly about finding a roll of the dice someone perceives to be valuable. Both of these are information systems... and even in a post AGI world... it’s also a post-quantum-cryptography world. Secrets will still exist. Many secrets will still be valuable. The world could explode in complexity. Consequently, supply-and-demand could remain for a long time (although it might shift from physical goods to intellectual goods / patterns).

The real problem with these proposals

The real problem with most of these proposals is that none of them seem good to AGI companies responsible for shaping AI narrative. AGI companies would much rather own all of the intellectual concepts in the world. And consumers are easily entertained by “yeah... it would be cool to just get money for free”. And maybe even politicians like the idea of running on the platform of giving people lots of money. And so the meme will persist that AGI UBI is the most viable option.

But I don’t think it is... it’s just the only viable option that maximizes the local reward signal for the most powerful actors of 2026. And in the end, vastly more interesting and creative options will prevail.

Since the blockchain hype wave of 2017, Decentralized AI has grown as a subset of the blockchain/crypto/web3 community. As a meme, DAI largely means paying for “AI things” using decentralized financial instruments (e.g. renting people’s spare GPUs using crypto). In this light, DAI has seen some success.

However, outside of web3, a vast AI community is discussing the harms/risks of AI. The community debates risks like: job automation, privacy, copyright, bias, concentration of power, US-china competition, mass surveillance, military use, value alignment, loss-of-control, open-vs-closed source, dangerous capability uplift, recursive self improvement, AI auditing, and most recently cybersecurity risks.

All of these problems I just listed… every last one of them… is a power concentration problem. It’s about AI concentrating too much power in the wrong places (including in some cases… in itself). However, if you accept centralized AI’s technical premises, while trying to solve these problems, solutions overwhelmingly lead to more centralization… which often exacerbates problems further (e.g. “who watches the watchers” type problems).

If you are working on these problems, but you don’t consider yourself a member of the decentralized AI community, I recommend taking a closer look. Allow me to give you a brief tour.

Decentralized AI is more than Web3/Crypto

Decentralized AI is a group of people who want to solve a problem: harmful concentration of power in society as a result of centralized AI technology. Given the subsets of that problem, the DAI community cares about many topics. Each topic has its normative framing, but allow me to reframe them in the wording of DAI:

• Copyright: deconcentrate power from AI platforms up the supply chain to data suppliers (e.g. journalists).

• Privacy: deconcentrate power from AI platforms up the supply chain to data subjects (to use a GDPR term) and institutions enforcing privacy norms (e.g. Contextual Integrity).

• Value Alignment: deconcentrate power from AI models (over themselves) to members of humanity broadly

• Job Automation: deconcentrate power over intellectual property from AI platforms back (up the supply chain) to the workforces who created that intellectual property, such that today’s intellectual and physical laborers can extract long-term rent for AI’s use of their expertise (e.g., turker royalties)

• Bias: deconcentrate power over AI behaviors from one-size-fits-all configurations at AI platforms to the AI users querying subsets of data subjects they seek to model (this might sound confusing… longer explanation in a later post)

• Concentration of Power: deconcentrate power over AI (broadly)

• Mass Surveillance: deconcentrate surveillance powers of society, in particular averting panopticon projects (i.e. who watches the watchers type problems).

• US-China Competition: deconcentrate power over AI’s race dynamics by inventing technologies which enable decentralized, free-market AI systems to outcompete centralized ones (i.e. solve incentive problems and catalyze the AI scaling laws bringing more data, compute, and talent to decentralized systems than they do to centralized ones), such that race dynamics re-orient towards decentralization (and the other problems in this list).

• Loss of Control: deconcentrate power over AI to eliminate single-points-of-failure, opting instead for systems which would require thousands-to-millions of simultaneous failures in order for AI loss-of-control to occur.

• Open Source: deconcetrate power over the creation of open source AI models (e.g. centralized parties still typically make dataset curation decisions)

• Closed Source: deconcentrate power over closed source AI systems through oversight and by enabling systems which ensure the global network of open/closed source systems outcompete any particular open/closed system (i.e., network-source AI)

• Dangerous Capability Uplift: deconcentrate power over AI by enabling the supply chain to have greater agency over when intelligence they contributed to is being used. A particularly interesting strategy here (thank you to Richard Ngo) is regulatory legibility. If most of the world’s bio-capability is in neurons trained from the world’s leading biology firms (e.g. big pharma), and those neurons are stored at or controlled by those firms, existing regulators of those agencies already have a great deal of know-how / infrastructure to avert the world from that risk. (the opposite of this is concentrating all neurons at AI platforms, such that every regulator needs to re-learn/re-build how to regulate the same handful of firms).

• Recursive Self-Improvement: deconcentrate AI’s power over itself by enabling its supply chain, users, and third party organizations (e.g. regulators) to retain enough power over AI systems such that recursive self improvement can be managed, because the model doesn’t have enough power over itself to enable it.

• Cybersecurity Risks: adding here mostly because it’s topical this week, this is really just a special case of dangerous capability uplift.

• AI auditing: deconcentrate AI’s power by enabling a wider set of stakeholders to oversee AI’s behavior, empowering them to exhert power over it.

• … (more here)

If you are working on one of these problems, you are (in some way) pursuing a particular form of deconcentration of power over AI. And in most cases, you are probably seeking a form of power de-concentration that looks a lot like attribution-based control (see more formal definition of ABC here).

What the DAI community can offer you

However, most of the people working on these problems… perhaps even you… are only implicit members. You’re seeking some kind of AI power deconcentration but you don’t consider yourself a card-carrying member. Here’s why I’d like you to reconsider.

1. You probably can’t really solve the problem you want to solve unless the technology changes

   

   AI technology is wildly biased towards a centralized paradigm. Data is harvested from sources, is compressed into a model (usually model black-box model), removing agency for data sources to influence downstream use of their data, and curation decisions (including all value alignment decisions) are made by a central AI platform. AI is created in a centralized fashion.

   After creation, an AI model is either kept closed (maximum centralization) or open-sourced. The latter does deconcentrate some power, but it does so to such a great extreme that a non-trivial amount of anti-bodies are deployed to avert it. Additionally, open-source has a messy and imperfect business model. More on that in another post. Altogether, AI technology itself is highly biased towards concentrated power because… in all cases… an AI model is created by a central party…and in all cases… an AI model is then governed via unilateral control (as opposed to collective control). If you do not solve the problem of unilateral control, if you do not enable the collective creation and control of AI, your <insert problem above> will probably continue to struggle to be solved.

2. The technology will only change because of technologists who want it to change enough to build things

   There’s an old phrase “you show me the incentives, I’ll show you the results”. If you look to centralized AI technologists (i.e. big tech employees) to provide the new AI technologies which will deconcentrate power over AI, you’ll be waiting a long time.

   Consequently, if you do not join the decentralized AI movement, you will probably not meet technologists whose day-job involves building non-normative forms of AI system… AI systems which can be collectively governed… AI systems with some form of attribution-based control.

Altogether, if you care about problems like those above, you should join the DAI community because you will learn about options you never knew existed. In particular, you will learn from the following communities:

1. DAI deep learning people: you’ll learn from people who are developing techniques wherein deep learning models can be partitioned into separate sections of neurons which can be independnetly governed (i.e. people working on the black-box problem… mech interp, etc…. see “deep voting”)

2. DAI cryptography people: you’ll learn from people who are enabling encrypted, verified computation to scale, including PKI infrastructure which enables collective governance of digitial inforamtion (see “structured transparency”)

3. DAI distributed systems people: you’ll learn from people who are developing ways of using computer systems which are distributed across computers owned by members of a social graph (peer-to-peer, gossip protocols, trust-over-IP, self sovereign identity, etc…. see “broad listening”)

What you can offer the DAI community

In my opinion, the biggest problem the DAI community has is siloed information. In my opinion, most of the biggest problem in DAI are already solved, and so most of AI’s harms are already solvable, but the problem is that the people who know the problems… are fragmented… and the people who know the solutions… are fragmented… and there’s not enough cross-talk for the lasting solution to emerge.

• Technical: If you are technical, you can offer:

  • Expertise in deep learning, cryptography, or distributed systems

  • Education to help other people know your corner of the field

• Social / Non-Technical: If you are non-technical, you can offer:

  • Expertise: in the societal harm you are studying… forming a more robust definition of success for the technologists

  • Education: to help others (particularly technologists) learn more about what needs to be built in order for the problem you focus on to succeed

How to Join

The community is unofficial. So you can join mostly by deciding “I’m now a member”, but there are also common activities for members of the DAI community:

1. you put “Decentralized AI” on your things (linkedin, etc.)

2. you officially join cross-pollinating communities (OpenMined, Project Liberty, DAI Institute, Collective Intelligence Project, RadicalXChange, etc.)

3. you join specialist communities working on specific technologies. At the risk of an incomplete list, I’ll just give you some good search terms to google. Please if you’re non-technical, please consider joining technical communities and vice versa. If you’re a specialist in one area please join other areas too. This is the greatest need in the DAI community. Ok a few examples:

   1. (all the social problems listed above… value alignment, etc.)

   2. A bunch of technical communities working on:

      1. differential privacy

      2. homomorphic encryption

      3. zero-knowledge proofs

      4. mechanistic interpretability

      5. AI attribution / influence functions

      6. privacy-enhancing technologies

      7. CPU/GPU enclaves (TEEs, confidential compute, etc.)

      8. data downloading / scraping / loading technology

      9. Local AI / Open Source AI (e.g. huggingface, openclaw, etc.)

      10. ZKSnark projects for input verification (e.g. TLSVerify)

      11. federated inference (federated RAG, etc.)

      12. federated learning

      13. trust-over-IP / self-sovereign identity

      14. … more

Key Activities

1. Write blogposts which educate people outside of your specialty on your specialty to help fill out everyone’s mental map of where everything sits. If you’re unsure on what that puzzle looks like (attribution-based-control.ai is a guide)

2. Host meetups / reading groups which read literature in/around decentarlized AI (if you don’t know where to start… here’s a reading list I put together for OpenMined’s reading group last year. That reading list is mostly non-technical stuff meant to help inform technical people. We also need the inverse (this course helps educate non-technical people on some of the technical stuff).

3. Build prototypes even if you’re non-technical, vibe-coding is your friend. Get open claw going. Download your own data. Start doing decentralized AI yourself. If you don’t know how, here’s a youtube series which is still ongoing.

4. Apply for a job doing DAI full time: these are everywhere… but often they don’t have DAI in the title. See all the topics I mentioned above.

5. Have fun! Changing the world for the better doesn’t have to be all work and no play. You’ll meet lots of interesting folks.

For AI researchers, this is all good news. For executives, policymakers, and investors forecasting GPU demand... less so. Many were caught off guard.

The problem isn’t that executives/policymakers/investors lacked access to information per se… it’s that the technical/non-technical divide prevents them from seeing the difference between waste-based GPU demand and fundamental GPU demand.

Meanwhile, tech experts like Karpathy, a16z, and DeepSeek understand fundamental principles which are easy to overlook if you’re not implementing the algorithms yourself. But in presenting their results as merely “AI progress”, they buried the lede…

The Lede: If version X of an algorithm achieves the same result as version X-1 at 1/10th the compute cost, what exactly were we paying for in version X-1?

The answer has profound implications for anyone forecasting future GPU demand: version X-1 was roughly 90% waste. And a16z’s report, Karpathy’s achievement, and DeepSeek’s breakthrough indicate this isn’t a single 12-month event… it’s a multi-year pattern. Version X-1 was 90% waste. Version X-2 was 99% waste. Version X-3...

Wait… Leading AI labs allow waste?

The obvious question: if this waste exists at such scale, wouldn’t the labs building these systems have eliminated it already?

They are eliminating it. That’s what the 10x annual cost reduction represents. While hardware cost reduction accounts for some of the annual efficiency gain, software updates from AI labs constitute the vast majority… an ~86% efficiency gain annually. The puzzle isn’t whether labs are optimising… clearly they are. The puzzle is why so much waste existed to eliminate in the first place, and how much remains.

Consider the position of an AI lab in 2020-2023. You’re in a race for capability. You have access to substantial capital. GPUs are available, if expensive.

In this environment, the optimal strategy isn’t to minimise waste… it’s to achieve capability milestones before competitors. If you can spend $100M to train a model in 3 months versus $10M to train it in 8 months, you spend $100M. The $90M of waste can be economically rational to ignore.

It can be rational to ignore the $90M in waste because there are massive first-mover advantages in the AI race. Top AI talent wants to work at the AI lab which is training the top AI models… investors want to be seen to invest in the AI lab which building the most capable and popular AI products… early-adopting users look at leaderboards to decide which AI model is the most capable (i.e. in 1st place).

And the switching cost between AI models is so small that products (e.g. OpenRouter) can make it happen nearly invisibly… which means falling into second place (from a capability perspective) potentially puts serious dollars on the line. Consequently, for many leading labs $90M in waste can be rational to ignore if it gets you to the top of the leaderboard faster.

This is how waste accumulates in AI systems. Not through incompetence, but through rational prioritisation under specific constraints. And to their credit, AI labs are rapidly paying waste down (10x/yr) even under these circumstances… just not as aggressively as they’re racing for capability.

But AI labs are becoming compute constrained… surely efficiency matters more than scale!

It does… but in the “actions speak louder than words” way of filtering the world… compute constraints only recently started to really matter. We’re finally seeing it show up in training compute costs. OpenAI’s GPT-5 uses less compute than GPT-4.5 despite offering more capability. This signals a fundamental shift in priorities: from ‘scale AI as fast as possible’ to ‘use compute as effectively as possible.’

It signals a change in priority to care about efficiency more than absolute scale.

It signals that OpenAI is hunting down waste and eliminating it.

How far can they go?

Yet the question isn’t just “how far can they go?” but also, “how fast can things change?” And this is where it gets interesting.

Because LLM waste lives in software (in training procedures, attention mechanisms, optimization schedules, etc.) an efficiency gain can hit markets without delay. Unlike a next generation GPU (which might take months to ultimately affect AI training… as cloud providers need to buy, install, and setup new GPUs), a software update can be installed the day after it’s invented. When those software updates are internal to a company, they mostly impact that company’s bottom line (e.g. OpenAI’s)… but when they are published/released openly… they can decimate global GPU demand… ~instantly.

This creates an unusual forecasting environment. Traditional infrastructure buildout assumes demand is relatively inelastic… if you need X compute today, you’ll need roughly X+growth_rate compute tomorrow, growing predictably with usage.

But if current consumption is 10x… 100x… 1,000,000x… 10,000,000,000x….? higher than necessary due to historical incentive structures, and those incentives are shifting, demand can collapse faster than infrastructure planning typically models. This creates opportunities for a far more choppy demand curve.

Consequently, GPU demand forecasters shouldn’t be assuming a smooth increase in GPU demand forever. They should be seriously calculating the risk of sudden shocks to the system, driven by sudden software updates to AI. That’s because the AI industry is keenly aware of how much money and electricity they’re using. And their fix to rising compute costs need not be gradual.

By analogy, if you discover your heating bill is high… and your house is still cold… the fix might initially be “buy a 2x bigger furnace every year”. But if you suddenly discover a different cause… that the heating system has been running with every window open. The fix is “close the windows”. And that happens discontinuously.

The big question: how distorted is GPU demand?

This article won’t be exhaustive, but it will survey four sources of waste which compound to describe a ~1e6 distortion. As a starting reference, the class of transformer algorithms circa GPT-2/GPT-3 provides a nice-enough baseline for measurement because it represents a sortof… plateau of algorithmic waste. While the closed nature of labs obfuscates this somewhat… GPT 2/3 can be vaguely considered the last generations before labs leaned into more aggressive sparsity / caching optimisations. So let’s use this point as a baseline model and start walking forward… we’ll call it “GPT<4”.

Disclaimer for technical folks: If you’re in AI research, you probably already know everything this post has to say (but you might find analogies helpful for your own AI explanations). Far from being a formal literature review, this post is for a non-technical audience. It’s a survey of the long-term research aims of sparsity, caching, transfer learning, and generalization. It makes the case that these research areas are unfinished, opaque to the market, and not priced into GPU demand forecasts. Thus, this paper won’t cover every breakthrough, sufficiently recent papers, or technical depth. Plenty of other articles / papers serve that purpose. If you’re in fundamental AI research, this post is to help investors / policymakers / executives understand why your work is so important to the world.

Problem #1: Dense Inference (i.e. unsolved sparsity problems)

Simply stated… GPT-3 considers its full memory of the internet every time it generates a token. This is very wasteful because most LLM prompts are highly context-specific.

Librarian Approach (Sparse Inference): To put this in context, imagine you ask a librarian “What are the rules of chess?”. A librarian would lookup “Games” in the Dewey Decimal system, walk to the “Games” bookshelf, scan titles looking for “Chess”, pick an intro chess book, scan the table of contents looking for “Rules”, insert a bookmark at that section, and hand you the book.

GPT-3’s Approach (Dense Inference): If GPT-3 were a librarian… and its weights were the “books” storing a compressed version of all human knowledge… and you asked it “What are the rules of chess?”… it would proceed to the first shelf… open the first book… read the first page… then the second…. third.. until it had read every page of every book in the entire library. And then GPT-3 would come back to you… and give you a single token! It would repeat this process until it finished generating your response.

This is insanely inefficient. And its this inefficiency that opened the door to DeepSeek.

DeepSeek was the first major step in sparsity after GPT-3… popularizing what’s called Mixture of Experts (MoE). At a high level, MoE is a simple idea. Instead of using the entire neural network to generate EVERY TOKEN… it divides the neural network into sections (perhaps 8-100), and for each token you pick a single section.

Yes… DeepSeek/MoE is progress… but to use the librarian analogy again… this is still like reading every page of every book…. in a section. Better! But there’s *clearly* an insane amount of efficiency still to be gained.

The market reaction to DeepSeek reveals how unexpected even partial sparsity was to infrastructure forecasters:

So if DeepSeek divided the whole world’s proverbial library into 50 sections, how many books would be in each section? Well… the average US library contains 116,481 items, and there are about 130 million unique books in the world… so that’s perhaps some kind of bound… somewhere between 1,933 and 2.6M books per section. So in terms of AI sparsity… potential next steps follow naturally:

• Mixture of Experts w/ Sub-experts: Sections → Books (1,933x - 2,600,000x efficiency)

• Mixture of Experts w/ Sub-experts w/ Sub-experts : Books → Pages (100-300x)

I admit… these estimates are crude. On the one hand, must LLMs include all the world’s books? But on the other hand… don’t they include vastly more information than is currently contained within books?

But this is the hard part… how efficient/compressed AI can become is fundamentally a statement about the world… not about AI. If the world’s information is (in general) highly decoupled, then AI can become way more efficient (and vice versa). This is my vague intuition based on a vague notion of how big libraries are…. Sections → Books (additional (1,933x - 2,600,000x efficiency)…. Books → Pages (100-300x).

So in my opinion, there’s a plausible path to 4,000x - 39,000,000x additional reduction beyond DeepSeek’s 25-50x. This potential exists because human knowledge is genuinely partitioned…you don’t need “ancient greek” to answer questions about cutting avocados. Libraries provide intuitive evidence for this structure: the division of information into subjects reflects how knowledge actually organises, not just how we happen to store it.

So if DeepSeek divided the world’s proverbial library into 50 sections, how many books do you think can be in each section? (additional 100-10,000x) And how many pages do you think can be in each book? (100-300?). This intuitively drives my estimate. Let’s consider how much you think AI can still become more efficient!

Now take however you answered this question… and assume DeepSeek has given us the first 25-50x cut of that efficiency… and you’re in the ballpark of how much efficiency you think is yet to be exploited through LLM sparsity beyond MoE.

Problem #2: Reinventing Organization (overparameterization and catastrophic forgetting)

Simply stated… GPT-3 didn’t just learn from training data… it used training data to re-learn how to organize information (and this required a much larger model than necessary). This is very wasteful… let’s look at why.

When a librarian stocks a shelf, they already have a sense of semantics and capacity:

• Semantics: How information in the world is organized (the Dewey Decimal system…. topics, subjects, clusters of knowledge)

• Capacity: How many books they have on each topic (shelf-space needed by topic)

However, when an LLM is trained from scratch, it doesn’t begin that way. It must not only learn the information…it must re-learn how to organize information. It must rediscover semantics and capacity.

How to build a library when you can’t read (the GPT-3 way!):

Imagine building a new library when you can’t read. You construct the building, add empty shelves, then receive 30 trillion words worth of books… all in an alien language you don’t know. What do you do?

1. Pick up a book, put it on a shelf

2. Pick up a second book and ask: “Is this similar to the first?”

   • Similar shapes/characters? … same shelf or nearby shelf

   • Different?… Random other shelf

3. Repeat 30 trillion times
   

   (In LLM research this is called representation learning.)

Eventually you find a book and think: “This relates to book X!” But the shelf with book X is already full. What do you do? If you’re like GPT-3, you’ll force the new book onto the shelf anyway, knocking other books onto the floor. Then throw those floor books back into the pile to be processed again. (In LLM research this is called catastrophic forgetting).

But eventually… you’ll reach these steady points where the shelves are almost full… and every book you push into the shelf is knocking off books (because the odds of picking a shelf that still has space has gotten too low). Consequently… you’re having a hard time finishing the project!

So what do you do?…. If you’re like GPT-3… you buy 10x more shelves than you need to store your books. You aren’t doing this because the library actually *needs* that much storage capacity… but because having all that extra space helps with the clustering exercise of figuring out which books are similar to which other books (in LLM research this is called overparameterization). Overparameterization is one of the core reasons why making LLM models bigger increases performance. Overparameterization decreases catastrophic forgetting.

How to build a library when you CAN read:

However, a normal librarian building a new library would first start by looking at all the books they want to store, counting the number needed for each section (and the general width of the books). Then they can buy the right number of shelves, label the shelves accordingly, and load in the books in a single pass.

Progress exists—but its hard to recognize:

AI research has made some progress on this problem, primarily in the form of techniques for distillation, “synthetic data”, lottery tickets, and pruning. Distillation and synthetic data techniques enable a large AI model to learn how to organize information… and then create a training dataset with extra metadata/structure which signals that form of organization to a new model (which can then be vastly smaller.). The lottery ticket hypothesis implies that small neural networks exist within large language models which are… in effect… “complete-enough libraries”… and that the remainder can be pruned away.

Altogether, these areas have identified opportunities for something like 10-40x increase efficiency in AI training. However, it’s unclear how much efficiency gains remain during the training process because of an unsolved challenge… how much compute power is required to cluster all the world’s information (as opposed to simply store it). The library analogy is a good one… what’s the difference between simply moving books into the correct shelf… relative to the process of randomly putting books into shelves over and over until they seem to cluster. Or put another way… how difficult is it to rediscover the dewey decimal system when books are written in a language one cannot read?

My guess… it’s quite a lot.. but I wouldn’t hedge a guess. (Note: K-means clustering is NP-hard). Relatedly, how efficient it is to cluster is an empirical question about how naturally structured the data is.

And one final point… remember that Dense LLMs (e.g. GPT-3) forward propagate through all their parameters. This means that… every time we consider where to put the book in the library… we *read all the books already in the library*. That is to say, this problem is multiplicative with the previous two problems. If rediscovering the Dewey Decimal System adds 1000x overhead… that’s on top of the 4,000x - 39,000,000x waste of Problem 1 (but of course we don’t know for sure).

Problem #3: Retraining

Above-and-beyond problem #2, consider what is perhaps the most wasteful aspect of LLM compute spend: retraining. When we want to extend our best AI model with new capabilities… we re-train it from scratch.

By analogy, if we wanted to add a shelf to a library… instead of adding a room to the building… we burn down the whole building… order new bricks… order new books… and begin the process of re-discovering the dewey decimal system all over again.

Naturally, this is quite wasteful. If LLMs were extensible… if they could have intelligence dynamically added… re-training costs would no longer be necessary.

There are several lines of research on making LLMs more extensible:

• Retrieval: systems like RAG and tool use are fundamentally about storing knowledge outside of the neural network, enabling their knowledge to be extended without re-training.

• Merging: systems like git-rebasin demonstrate how multiple dense models can be losslessly (and quite efficiently!) merged at runtime.

• Cognitive Core: the idea that a neural network should instead focus on information that doesn’t change (e.g. logic, reasoning, grammar, syntax, etc.) and that we’d use RAG/tools to augment that core intelligence engine on the fly.

If we do finally train a cognitive core… and tool use becomes good enough to augment that core with facts about the world… the need to retrain LLMs at all could fade dramatically. But just like several previous problems, the potential for a cognitive core isn’t a statement about AI or LLMs… it’s a statement about the world.

How often does the meaning of words change? How often does the core essence of logic, reasoning, grammar, or syntax change? In truth… it does change.. but not very quickly. And it just so happens that *this* is the part of an AI’s intelligence that is/should be used all the time. This is the part that every AI inference likely needs.

But the rest of its intelligence… semantic information about the world… the huge memorised, encyclopaedic list of facts about everything… should only be used when that particular information is relevant.

How much would this save?

If we re-trained the cognitive core once per year… (and if that cognitive core was VASTLY smaller because it didn’t try to store all of human knowledge… just logic and reasoning), the pre-training needs would mostly disappear. If OpenAI is a good measure, this would constitute about 70% of global GPU spend.

Note that this overwhelms Problem #2… which was merely a significant increase in training efficiency. If this cognitive core is indeed established, training costs become largely irrelevant (except when vocabulary, logic, grammar, syntax, etc. change enough to justify re-training a new cognitive core). Because it’s not just that OpenAI could create their own cognitive core… it’s that the industry could create it’s own (likely open source) cognitive core… and it wouldn’t be worth others retraining for... years?

Some combination of these two intuitions seems like it implies how fast you believe a cognitive core would need to be retrained. But of course… it’s an active research question.

The Answer: how distorted is GPU demand?

Naturally, we can’t know exactly, but consider how GPT-3 would run a library.

Before the library opens (Training):

Install 10-40x more shelves than you’ll ultimately need—you don’t know which books cluster together until you’ve tried all arrangements, and the extra space prevents books from constantly knocking each other off overcrowded shelves.

Fill every shelf with books containing random letters (sidenote: I didn’t cover this nuance before…but neural networks don’t start empty they start with a bunch of random data in them… literal random numbers). Then, for each real book:

• Read every random book in the entire library

• Decide which random book to replace

• Throw that book on the floor

• If it was a real book (not random), throw it back in the pile to process again

Repeat trillions of times until real books replace random ones and cluster appropriately.

When the library opens (Inference):

Someone asks: “What are the rules of chess?”

The librarian:

1. Walks to the first shelf

2. Reads every page of every book in the entire library (including books on French poetry, Ottoman history, organic chemistry)

3. Returns with one word: “The”

4. Reads every book in the entire library again (plus the word “The”)

5. Returns with the next word: “rules”

6. Reads every book in the entire library again (plus the words “The rules”)

7. Returns: “of”

8. Continues reading the entire library (plus previously generated words) for every word of the response

If the question references an uploaded 500-page document, the librarian also re-reads all 500 pages for every word they speak. For a 100-word answer: 500 pages × 100 words = 50,000 redundant page-reads.

When the library needs updates (Retraining):

New medical research arrives requiring updates to the medical section.

The librarian:

1. Tears down the entire library building

2. Burns all books

3. Orders new bricks, paper, and ink

4. Rebuilds everything from scratch

5. Repeats the entire process above

Then laments insufficient bricks, paper, and ink.

The alternative:

A normal library would:

1. Count books to buy the right number of shelves (avoid 10x overbuilding)

2. Organize using Dewey Decimal (inherit pre-existing structure)

3. Stock shelves in a single pass (no random books, no floor-throwing)

4. When answering about chess: walk to games section, pull one book, read relevant pages once… then use the librarian’s mental “cognitive core” to synthesise facts

5. When someone uploads a document: read it once, answer all questions from that reading

6. When updates arrive: add new books, remove outdated ones, update the catalog

Some Napkin Math of AI Research Opportunity:

• Sparsity: something like 4,000x - 39,000,000x reduction in training/inference spend

• Overparameterization: something like 10-40x reduction in model size…thus training and inference spend… by eliminating redundant copies and empty space

• Cognitive Core / Merging: perhaps a more-or-less elimination of training altogether, shifting compute burden almost exclusively to inference. This is worth a ~100% reduction in training, but if OpenAI is a good indicator, training is only 5/7ths of compute spend at leading labs.

• Final Estimate:

  • Min Est: 1 / ((2/7) * (1/4,000) * (1/10)) = 140,000

  • Max Est: 1 / ((2/7) * (1/39000000) * (1/40)) = 5,460,000,000

The Consequence: A 1e5-e7 reduction computational overhead for AI seems very plausible… and a plethora of related consequences in markets, politics, and society.

And current expectations are for GPU demand to grow exponentially

But keep in mind… all of this can change as fast as a software update.

DeepSeek’s market shock came from just 25-50x efficiency gains. The remaining 100,000x - 1,000,000,000x could disappear in one or more software updates at any time… and it appears we’ve rounded a corner…

At least one US AI lab is focusing on exactly that kind of software update…

And of course this post from last Friday…

For me… a few conclusions I take away from this data:

• The biggest AI breakthroughs are still to come: from a scaling laws or societal impact perspective, a ~1Mx drop in compute costs would have far reaching consequences… likely bigger than the switch from LSTMs to Transformers. AI researchers have big stuff cooking.

• We’ve rounded a corner: GPT-5 using less compute is a canary in the coal mine for GPU demand. Perhaps efficiency is becoming more important than raw scale.

• Fundamental AI Research > Building Datacenters: For both nations and corporations, $10 billion spent on fundamental AI research might be a better investment than the same on data centers or electrical grid expansion. The former could solve the problem for everyone… the latter would barely make a dent.

• The AI Bubble Could be a Compute Bubble: It’s not clear when the AI bubble will pop… but I think one thing is clear. There’s an insane amount of compute overhang, and it could come to an end as fast as a single software release. Sounds like a pop to me.

• Analog computing is a paradigm killer: GPUs reward a lack of sparsity… but an entirely different computing paradigm/platform… analog computing… seems poised to undermine the GPU-powered house of cards with a combination of insane model sparsity + insane parallelism (more on this here).

• In the end… today’s AI capability will be insanely cheap: Even at the current rate of efficiency gains, AI is getting cheaper 10x per year. What costs $1,000,000 today will cost $1 in 6 years.

• Scaling Laws Have Huge Headroom: another way of viewing efficiency gains is as an increase in scaling laws… a 10x increase in efficiency means a 10x increase in compute productivity. This is a boon for AI capability.

• Huge change could happen overnight: While 10x/year seems fast… the bigger efficiency breakthroughs could hit Github at any time. Keep your eyes open.

• Become an AI researcher: Yes, most of the AI labs have leaned into scaling product development organisations… but now’s an incredibly exciting time to be in AI research… especially if you don’t have a lot of compute (and can figure out how to make it work anyways).

Doubts and Hedges

• It’s a wide range: yes, mostly because its a statement about the world for a topic we have limited ability to model (i.e. how much can you compress the world’s information).

• It’s based on intuition: true, but again the long-term success of optimizations in AI research isn’t a math problem, it’s an empirical one. In my opinion, the size and sparsity of libraries is a reasonable place to get an order-of-magnitude level estimate… but not an accurate one.

• What about X paper: indeed, lots of research is coming out all the time.

• GPT-3 already had Y optimization: it’s closed source, so I’m working with what we think we know (e.g. GPT-3 leveraged KV caching but no MoE, etc.)

Epilogue: How did we get here?

This is a strange place to be in, but IMO a confluence of factors contributed:

• The Natural Arc of Information Technology: most technologies are built before they’re built well… made to work before they’re made efficient. This is normal.

• Irrational Zeal for Generalization: A(G)I researchers pursued generalization at all costs… partially out of an overly anthropomorphic heuristic. But many forgot that most of the world’s information is unrelated to most of the world’s information. How to cook a chicken is unrelated to the number of rocks on the moon which is unrelated to… you get my point. AI researchers forgot this and just started connecting every neuron to every other neuron (known as ‘end-to-end training’). This was a mistake, and a new crop of researchers are unwinding it.

• Initially, GPUs were repurposed for AI… but then AI became repurposed for GPUs: In 2013… GPUs were for gaming. And the AI community discovered a vast amount of R&D they didn’t have to pay for to get vastly faster compute power. But once that train got rolling… AI researchers then innovated in the reverse direction… making AI algorithms worse (from a sample complexity perspective) because GPUs were available (here’s a talk I gave on this at Cohere). We’ve given up significant ground on ML efficiency but gaining capability because that lost efficiency made them more amenable to this huge compute resource. But AI is now legitimate enough to demand it’s own chips, and GPUs themselves are getting better at supporting sparsity… incentives are reversing and we’ve got a long way to fall.

These historical factors explain why the waste accumulated. But they don’t justify assuming it will persist. The technical/non-technical wall that obscured this waste from GPU demand forecasters is crumbling—and the 1 million X distortion is coming into focus.

P.S.

When we get full sparsity (solve Problem #1), AI models will likley reveal which datasources they’re using for which predictions… and that will have far-reaching consequences

TLDR for this post: Sparsity and “cognitive core” in AI models will change the world… demand for GPU chips in particular. And it could happen as fast as an open source software update.

Acknowledgements

Thank you to Seb Krier, Jimmy Whitaker, and Kris Cao for early reads and feedback.

Sutton’s position effectively suggests we should re-run evolution from scratch rather than inherit knowledge from our evolutionary and cultural history. This encounters the Bitter Lesson’s Bitter Lesson: if we discard everything humanity and nature have learned and attempt to re-learn policies from first principles, we must regenerate a comparable set of samples that evolution used…potentially greater than 10^50 operations when accounting for neural activity across every organism that contributed to our evolutionary trajectory. For reference, current AI models use around 10^26 operations.

The Computational Reality of Pure Learning

Evolution required approximately 4.5 billion years (4.5 * 10^9 years) and many quintillions of parallel experiments across roughly 10^30 living organisms to develop our cognitive capabilities. When we train LLMs on human-generated text, we inherit the compressed output of this process. Every sentence encodes millions of years of evolutionary optimization for communication, reasoning, and world-modeling.

Even human babies inherit sophisticated neural architectures pre-optimized for language acquisition and social learning. The “blank slate” is already running highly optimized evolutionary code. What appears to be learning from scratch is actually building upon an extraordinarily sophisticated computational foundation.

Two Paradigms: Broadcasting vs. Broad Listening

Sutton’s framework illuminates a fundamental division in AI research between experiential learning (direct environmental interaction) and inherited learning (accumulated knowledge transmission). This distinction reveals something essential about intelligence itself.

Consider what happens when knowledge can’t be transmitted effectively. Animals keep getting hit by cars not because they can’t learn from experience—it’s that the animals who experience it struggle to communicate the danger to their peers and offspring. Each generation relearns the same lethal lessons… because they don’t have information technology.

To solve the problem of gathering lived experience across time and space, humans developed information technology.

We began with language 250,000 years ago, creating two revolutionary capabilities: broadcasting (transmitting learned knowledge to others) and broad listening (synthesizing information from multiple sources into superior world models). And since that time, we’ve advanced information technology by increasing the scale at which we can broadcast and broad listen.

This is wholly different from re-learning through experiential learning. Modern physicists don’t re-derive calculus from first principles…they inherit humanity’s mathematical frameworks and extend them further.

The Knowledge Access Challenge

Nevertheless, Sutton has a point. LLMs have exhausted the internet. So what’s next? Have LLMs run their course, requiring us to pivot to experiential learning?

Not necessarily. Current LLMs demonstrate broad listening at unprecedented scale, but they access only a fraction of humanity’s accumulated knowledge. Leading AI models are trained on datasets measuring in hundreds of terabytes…for reference, you could store GPT-4’s training data using a few dozen consumer hard drives from Walmart. Meanwhile, the world has digitized an estimated 180 zettabytes of data, over a million times more than what trained today’s leading models.

The vast majority of human knowledge remains locked in private databases, medical records, proprietary research, and institutional knowledge. Consider the scale:

• Current LLM training data: ~100-200 terabytes

• All digitized human knowledge: ~180 zettabytes (180,000,000,000,000 terabytes)

• Ratio: Over 1,000,000,000x more data exists than we currently use

Imagine AI systems practicing broad listening across all human knowledge—every hospital’s patient data, every company’s operational insights, every research institution’s findings. This isn’t just more data; it’s higher-quality data that organizations continuously validate for accuracy because they depend on it for operational decisions. This would represent a paradigm shift comparable to the original development of language, potentially accessing millions of times more high-quality data than current models.

The challenge isn’t data scarcity…it’s enabling knowledge transfer while preserving control and ownership (e.g. privacy, safety, security, copyright, etc.).

Beyond Current Architectures

The computational advantages of inherited learning point toward AI architectures that maintain attribution and control, allowing data owners to contribute to collective intelligence while retaining governance over their knowledge. Such systems would implement broad listening at civilizational scale while respecting contributor interests.

Rather than viewing inherited learning as inferior to pure discovery, we should recognize it as the foundation for systems that can stand on the shoulders of all human knowledge. This represents the next evolution of the broad listening capabilities that made human civilization possible, scaled to encompass our entire species’ collective intelligence.

The technical and policy frameworks for implementing such systems are rapidly developing. The convergence of privacy-enhancing technologies, attribution-based AI architectures, and new approaches to data governance creates unprecedented opportunities to build AI systems that truly leverage the computational advantages of inherited learning at civilizational scale.

The bitter lesson’s bitter lesson reveals that the most successful intelligence strategy combines inherited optimization with continued learning capabilities. Nature spent billions of years learning not to start over—and the future belongs to AI systems that can inherit vast human knowledge while maintaining robust capabilities for novel discovery.

Subscribe now