Blog

The Problem with Funds

Published 24 April 2026

At the most basic level, collective investing works because it allows specialisation and economies of scale. It would be time-consuming and inefficient for every person to manage their own investments, and it makes sense to delegate one's investment choices to a specialist who has a natural interest in asset pricing. They can devote their time to following the news and becoming an expert in the field, freeing up the time and mental energy of the rest of the group to focus on their own specialties.

Staying informed can also be expensive; research houses and software subscriptions usually charge per user, not to mention attending conferences or conducting site visits, which is where economies of scale can help.

These benefits are not unique to investing, they are almost ubiquitous across capitalist based economies! In almost every other area of business, people form equity-based limited liability companies to harvest these gains while conducting whatever business they happen to be in. It's understandable that funds have become the primary vehicle for investment (people are used to separation of client money from the institution holding it, from the way banks operate) but we don't think it's the right approach.

The fund structure suffers from two major disadvantages:

• Fund managers charge an annual fee, and when someone is charging you for something, they feel pressure to justify it. This, along with almost every bias humans are born with, will encourage over-trading. Sometimes trading frequently is the correct thing to do, but many times the correct thing to do is nothing, and you don't want someone managing your money who feels they, to quote the great Charlie Munger, "can't sit on their ass" even if that's the best thing for their investors.
• Fund managers are in the volume game. It's simple maths to note that if a manager charges 10% of returns, they are better incentivised to make 10% on $10 billion than 25% on $1 billion. A manager maximising their own profit will then spend much of their time courting new clients, rather than focusing on delivering returns for their existing investors.

We want to stress that this is not a criticism of fund managers; no doubt there are many of them who do their best to put their investors first in spite of these misalignments. Our criticism is of the structure itself and the system of incentives it creates.

There is a better approach that is very straightforward: structure the investment vehicle as a normal company and issue investors with ordinary shares at Net Asset Value (NAV). Day-to-day expenses can then be paid directly from the firm's funds, and instead of an annual fee, senior staff can be rewarded by stock options paid based on NAV performance. This ensures that the team is incentivised to focus on exactly what investors want them focused on - increasing shareholder value. When it comes time to divest for whatever reason, an investor can do so at NAV via a redemption. Alternatively, one could create an internal market for the shares or even list them on exchange.

This is how we want to structure Gradient Ascent. We want to feel incentives are as closely aligned as possible between the team and our investors. Of course, there are still some reasons to scale the business. A wider investment base means the fixed costs get spread across more investment and/or would allow us to increase the team size if we feel it is beneficial to do so. The point is that these are benefits for both investors and the management team.

This is the logical and obvious thing to do, but we don't need to rely solely on abstract thinking. Warren Buffet and Charlie Munger lived this approach very successfully for many years, and at any opportunity advocated it as preferential to the fund approach. We are convinced that this is a better structure for investors and money managers who are focused on a sufficiently long time horizon and hope that you will be too.

Topics of Interest

Ongoing publication · By Jamie Cassidy

This is an eclectic collection of random thoughts, half-baked ideas, and things that I feel might be very important but don't know enough about yet. If you disagree with me, or just feel the urge to educate me because my words betray a naïve understanding, please let me know!

Lab Moats

I am mildly concerned that the labs will struggle to create a lasting moat. I believe they can keep setting the frontier, but I wonder: will they become victims of their own success? Let's use Software Engineering as a timely example. Right now, it's worth it for many firms to pay a premium for Claude because it can do tasks that other models are simply incapable of doing. However, at the current rate of progress, it seems quite likely that in the near future all the major models — including some open-source ones — will be capable of doing all software engineering tasks in under a day with perfect accuracy.

At that point, what are the economics behind paying for the frontier model? With open source, I can already replace each team with a prompt engineer who is producing the work of 7 or more software devs. They spend half of their time working with the agents to review the work and set the next task, and the other half in meetings, or piloting sessions with users. It's a heavy lift for the frontier to replace this seventh person completely, but even if it does, the bulk of the total cost savings have already been made, so the remaining marginal value that the frontier replaces with each additional upgrade is shrinking.

There is a lot of concern in markets that compute will become commoditised and all the value will accrue to the frontier models and the consumers. However, I can see the opposite occurring, where the models become commoditised (because they all achieve perfect results for many human tasks), so instead the value accrues at the bottleneck – which is compute. In fact, perhaps one way the frontier labs could command margin is through efficiency – delivering the same output for fewer watts of datacentre capacity.

The other problem, especially for OpenAI and Anthropic, is distribution. For now (while they have brand loyalty and fully differentiated models) it's okay, because the friction of an added service is worth the effort, but Microsoft and Google make it remarkably easy to default to them as full-service solutions. A small example is note–taking. When using Google Meets, until recently, people had been either using separate note–taking software, or even recording meetings on their phones and uploading the audio file to Claude. Now though, Meets has an inbuilt note–taking feature. Even with a preference for the better quality notes from Claude, it's difficult to justify the hassle. I struggle to see how the standalone labs will stay relevant to businesses against the hyperscalers, given the large distributional disadvantage.

AI Winners Outside of Tech

I'm keen to understand who the big winners will be outside of technology names. Here, I'm thinking of companies with high wage bills for white collar workers and also high degrees of regulatory capture. One thought I've had is that investment banks might fit the bill quite nicely. In particular, those without loan books might have a strong advantage. (Those with loan books would suffer in the event of large increases in unemployment, especially if the job losses are concentrated among well-paid white-collar workers who previously would have had very high credit scores, and therefore whose loans are priced as extremely safe). This is just one idea among potentially thousands of non-tech AI winners.

Deflation and Employment Loss

Many (if not most) new technologies are deflationary in the specific, and expansionary in the general. We all know that technology has been driving economic growth for centuries, but it is important to acknowledge that in the specific, technology is deflationary. The motor car lowered the price of transportation and killed the horse industry, but it created manufacturing jobs. Over time, the lower cost of transport facilitated the growth of many other industries, including novel ones, like tourism.

I suspect that the way our economies work is that lots of small disruptions are happening all the time, creating isolated deflation and improving overall productivity via the process of creative destruction. Even if the gains come at a lag to the deflationary unemployment, at any given time the gains from previous disruptions outweigh the losses from new ones. I'm just speculating here (I'm sure there are studies one could find to explore this time lag effect, but I have not yet done so) Nonetheless, if we follow the thought, there is the inevitable question about how this process reacts to a shock that disrupts a significant amount of the economy in one go. I'm sure the economy will adapt in time to the productivity gains, but in the short-term, could the deflationary effects be large enough to affect the whole price level? And — at the same time — could the job losses be sufficiently large to impact the overall employment level across the whole economy? I don't have the answers, but I think it is a serious question that has widespread implications for general human welfare, and for investing.

The Looming Capital Shortage for AI Infrastructure

Published 1 April 2026 · By Jamie Cassidy

The release of ChatGPT in November of 2022 made clear to anyone watching that a Rubicon had been crossed. This model appeared to have developed a form of emergent intelligence, in a way that was both magical and jarring for anyone using it for the first time. It was based on LLM infrastructure, which takes massive amounts of data and compute to develop, and perhaps more significantly, a very large amount of compute every time it is used. This sparked an enormous wave of investment in compute capacity in 2023, and investment has doubled every year since, a trend that will continue in 2026. Even at this level of investment, capacity cannot keep pace with demand, as models get bigger and more useful across a variety of human endeavours.

Exceptionally high growth rates are not unusual for a new technology, but it is trivial to know that nothing can sustain that kind of growth forever, so the nuanced question becomes: how far can AI demand scale before its growth normalises? This is an impossible question to answer precisely, and the only sensible way to get in the right ballpark is to consider what we in markets call the TAM (Total Addressable Market). That is to say, how much money currently gets spent on productive activities that this technology is capable of disrupting, bearing in mind that every time the technology advances, the TAM expands. Right now, I would describe frontier models as having "Task Specific Functional Intelligence," meaning they are capable of completing some sets of tasks better (or at least to the same standard, more cheaply) than humans.

The corresponding maths is in the appendix below, but the net result of my calculations is that even in the current paradigm, demand for AI compute growth of 50% a year seems realistic before the limits of scale start to kick in; that's much lower than the 100% a year we've seen to this point, but is still very substantial given the scale we're already at.

Note: these numbers are based on steady progress within the current paradigm. We will never make investments based on guesses about when AGI or other forms of step-change improvement arrive. Personally, I find quite compelling the arguments put forward by Francois Chollet that general intelligence will not emerge by iterating and scaling within the current LLM-only framework. We are looking for investments that will win if the progress is slower than insiders expect, and win bigger if there is some dramatic breakthrough.

The Build-Out

Up to and including 2026, the build-out of datacentres has been constrained by available supply of semiconductor equipment (GPUs, memory, networking cables etc). However, all that is about to change as a new constraint becomes the bottleneck: money. In my view, there is no way that the capital required to keep up with 50% demand growth will be invested in datacentres over the next several years.

To understand why, let's talk about who's been paying for all this so far: Amazon, Google, Microsoft and Meta, otherwise known as the hyperscalers. Together, these companies will account for roughly 85% of total AI infrastructure investments in 2026. That's because the CEOs of these companies have been uniquely placed to drive this investment. They're close enough to the technology to understand its potential, and have previous experience in physical infrastructure, as well as access to 100s of billions of profits from their existing businesses. They have invested more and more of their available cash (free cash flow after shareholder returns) on AI-related capex over the last 4 years; given those investments are showing great returns so far, there's every reason to expect them to continue investing. There's just one problem: in 2026 they will increase to the point where they are using all of this available cash, and in some cases more, and once you get to 100% there's no way to keep increasing.

Let's look at Google as an example: to double their capex in 2026 while keeping their shareholder return policies in place, they have had to turn cashflow negative (see table below). This is not a problem for Google, given the size of the company: they have very low debt and $125B of cash and equivalents on hand, but it shows that this is almost certainly the last year in which they can grow their capex substantially faster than their earnings. Perhaps they could cancel their buyback programme (which shareholders would hate!) and get to 50-60% in 2027, but even if they do this, they hit a hard limit in 2028.

It's a similar story with the other companies. If you believe (as I do) that they will be rewarded for these investments, then their earnings will continue to grow at healthy rate of 20-30%, and they can continue to reinvest in capex at that rate, but the days of 100% capex increases are behind us, and there's no way they can keep up with the required pace of 50% per annum over the next 4-5 years. If we assume they average 25% capex growth over the period, that will leave a cumulative investment shortfall of $1.7T relative to demand. Others will need to step up to fill the gap.

There are others in the space, most notably the labs OpenAI and Anthropic, who will naturally move towards building datacentres in time. However, despite having just raised $130B between them and both planning IPOs in the near future, I still feel like while they are loss-making operationally, capital will be too scarce to invest heavily in capex. Once they have their profitability flywheels up and running I'm sure that they will get involved on the physical infrastructure side, but that counts out any serious contribution during the next 3-4 years (even by their own estimates). Their capital will be better spent on top staff and research.

That leaves the neoclouds: specialised companies that build and operate clouds of GPU datacentres. Their business model is to lease GPUs or tokens in various forms to hyperscalers, labs or end users. These include Oracle (an older company which has repositioned itself for the AI era) as well as newer firms such as CoreWeave, Nebius and a host of private companies. These are growth companies, whose mission is to scale up fast to fill the demand gap that I've described, but to do so they need to raise capital through equity and debt, both of which will be easy or difficult in proportion to how public market investors are sufficiently convinced that their investments in these extremely expensive datacentres will pay off.

The Sentiment Gap

To say that investors remain unconvinced would be a massive understatement. I find myself caught between two worlds, one where I listen to podcasts and read blogs from AI insiders, and the other where I track the day to day movements in stock market prices. Not only do they have contradictory views on the path forward, but the two camps seem either oblivious to the existence of the other, or so entirely dismissive of the others' views as to not even be worth mentioning.

I recently listened to an interesting but ridiculous conversation between Dwarkesh and Dylan Patel (founder of Semi-Analysis). It was interesting because the details were very well laid out and it's clear that Dylan is an expert in his field; it was ridiculous because the premise was so farfetched. They were discussing the bottlenecks which would prevent the supply chain from delivering over 200GW per year by 2030. Dwarkesh, I promise you that achieving 200GW in 2030 is not something any of us need to worry about. This scenario, which was discussed in minute detail, just assumed the $15-20 trillion PER YEAR (appendix 2) in required funding would be available; they didn't so much as mention it as a constraint.

Let's compare that to what we are seeing in capital markets on a daily basis:

• NVIDIA is trading at roughly 23x reasonable estimates of 2026 earnings (>$8); this puts it below blue-chip stocks whose earnings grow consistently with GDP, implying that risks to downside for future earnings exceed upside potential. In short, the market is pricing more of a discount to downside risk than a premium for future growth for future NVIDIA earnings. Check out our Pricer page for more on this.
• At Google's last earnings release, their profit and revenue numbers exceeded even the most optimistic expectations, but when they announced a capex number far above expectations, the stock sold off. The implication here is that Google made more money than anyone was expecting, driven partially by their investments in AI to that point. Then, they told the market that they were doubling down on these bets that are working better than everyone expected, and the market absolutely hated it, it hated it more than it liked the good numbers.
• The neoclouds—especially the ones more aggressively financing fast expansion with debt—have seen their stocks crushed in the past 6 months; Oracle and CoreWeave, for example, are down 47% and 32% respectively.
• Memory names continue to be priced on extremely low multipliers. They have rallied a lot, but nowhere near as much as they have raised their prices and increased their volumes, so the stock prices are significantly lagging the jump in earnings. This is because the market believes this spike in memory demand to be temporary and that at some point the cycle will revert, leaving these companies with a glut of supply. It is very likely that the current, very acute, supply-demand imbalance will ease somewhat over the next couple of years, but that is not what's being priced: what's being priced is essentially that AI demand is a fad.

Investor Sentiment Matters

Try as CEOs might to detach themselves from short-term share price movements, the reality is that the price of a company's stock does matter, regardless of how big they are and how much cash they have on hand. It matters for staff morale, for how your company is viewed by your customers and suppliers, and for the CEO's own confidence in their decisions and the direction that they are leading the company in.

It matters even more for growth stocks, who are still reliant on capital markets to fund future investment. If your shares are trading at all time highs, it's much easier to raise capital to pay for future investment; with a lower market cap, raising capital for future investment is both very difficult to do and (if done below the price they bought it for) very unpopular with shareholders. Meanwhile, in the debt markets, Oracle already had to cancel one of the datacentres it was building for OpenAI because it couldn't source the financing.

So the feedback the market is giving to neocloud CEOs is clear: be conservative with your investment approach. To the extent that they listen to this feedback, the net result will be that they grow their capacity much less quickly than they would have done in a more supportive market.

Net Result of Supply Analysis

The net result of all of this will be that across the hyperscalers, labs and neoclouds, supply will be unable to scale as fast as demand requires. In fact, if you were to trust the implications of NVIDIA's share price, it looks like capex growth to drive this supply is going to stop abruptly in 2027 and flatline from there. Thankfully, I don't think that will happen; I think the hyperscalers will continue to grow their annual investment at 25%, while the neoclouds and labs will grow their investment at 50% (albeit from a much smaller level). Over time this will lead to a large and growing supply deficit, to say nothing of the existing deficit—as GPU prices and availability shows, demand is already well above supply.

Perhaps a small shortage of supply would not be a bad thing; it's important that there is some demand constraint to ensure the supply chain is sustainable (i.e. profitable) and it's sensible to have an ecosystem where model efficiency is rewarded. However, a capacity gap of 30% or higher (which is where we are headed unless something changes fast) will prevent us from getting the best out of technology.

Solutions

Given the current state of the market, it seems almost impossible that this shortfall can be fully avoided. However, I think the gap can be partially filled if:

• The profit motive is more heavily incentivised. Neoclouds are currently able to sign five-year GPU rental deals with a total margin of roughly 25% over the lifetime of the project. Much of this margin is eaten up by financing costs, but it roughly allows the neocloud to make a small profit on the deal, plus whatever residual value the neocloud can derive from these 5-year-old chips. So, if you are not convinced that these chips will hold their value, this is a pretty unattractive deal. If the deal margin drifted closer to 35%, then a lot more investors will be interested (because the deal justifies itself in hard cash terms, and the residual value of the chips becomes something of a free bet).
• The finance community directs more money towards AI related investments. As a society, we want to move our accumulated resources (savings) towards the most productive investments. It's our job as financial professionals to convince the sceptics.

Takeaways

The pace of the AI infrastructure has doubled every year since 2022. There have been many bottlenecks along the way, but capital has not been one of them as the HyperScalers have paid the bill so far. Even with this annual doubling, demand for compute continues to grow faster than supply.

Their pockets are deep, but not infinitely so, and we are getting close to limits. Hyperscaler investment growth will begin to slow down in 2027 and slow dramatically in 2028. If demand growth continues at the current pace, this investment will need to be replaced by the market, or the demand shortfall will start to grow dramatically.

This can happen, but market sentiment towards AI investments will need to change dramatically before it does, and even if that sentiment reversal occurs quickly, it will still take the market time to fully adapt given the scale of the funding shortage.

Notes

1. From the view in Q1 2026, I think we can conservatively say that by 2030 AI will be doing at the very least 50% of tasks previously completed by software engineers, and potentially 30% of all tasks currently done by remote workers. The global economy spends $2T on salaries for software engineers and north of $30T for remote workers, implying a TAM for AI between $600B and $9T. Given that this will by no means be the only use for AI (come on digital personal assistant!) and that demand for this kind of work will increase dramatically as the price goes down, I ultimately feel that even the higher end of this range will be conservative, but let's use $6T (there's nothing I hate more than false precision).
   
   If we assume half that goes to consumer surplus ($3T) and 30% of the remainder goes to lab margin, that leaves $2T for datacentre owners. Given the high depreciation and capital scarcity, I assume they require 25% return on the cost of build, so AI capex of up to $8T up to 2029 (datacentres take a year to build, so in 2030 only the ones paid for up to and including 2029 will be in use). Starting with the $1T figure which we know will be spent in 2026, this corresponds pretty well to a 50% a year growth rate over the period.
   
   (I've excluded pre-2025 investment because it was not huge in the context and GPUs age fast!)
2. The claim that 200GW p.a. in 2030 would require $15-$20 trillion p.a. is based off a current cost per GW of $50B for projects in 2026 and costs have been increasing by roughly 15-20% a year. Even if cost increases slow to 10% over the coming years, that implies nearly $75B per GW in 2030 or $15 trillion for 200GW, while if they continue to grow at 20% it will be above $100B per GW, which implies the top end of the quoted range.

AI: General Thoughts and Timelines

Published 20 April 2026 · By Jamie Cassidy

Introduction

The goals of this post are to clarify:

• My thinking about AI in what I see as the current paradigm
• The requirements for AGI as I define it
• The additional barriers that will need to be overcome to achieve Dario's much discussed "country of geniuses in a datacentre"

I want to caveat the piece by noting that these thoughts are just the framework I use for my own understanding. I am constantly updating them as I learn more and the frontier changes. As a result, I have wider and less aggressive timelines than one might expect from a person starting a company to invest in AI.

While I have a lot of uncertainty about the path, I am very confident in the eventual destination: artificial intelligence will far surpass its creators. In fact, my most strongly held view on this topic is that there is nothing special about humans, or organic intelligence in general.

Everything we do – every instinct we have or creative thought we experience – is the result of some set of algorithms that are replicable in silicon. This appears to me to be a clear and obvious conclusion of the rejection of the paranormal. It then follows that artificial intelligence, which is not limited by brain size and which is evolving much more quickly than natural selection allows for organic life, will continue to grow in complexity and understanding long after it has left human level intelligence behind.

Current Paradigm: Task-Oriented Functional Intelligence

Before getting into AI, I want to highlight some general concepts which will be useful to contextualize the discussion.

Let's define a task as a piece of work to be done. A task can be easy or hard, long or short, but it is well defined, and the result of its completion is a concrete output or impact on the world.

Let's define a goal in this context as the object of a person's ambition. There are hard goals and easy goals, but the crucial difference is that a goal is a much more abstract concept than a task. Specifically, the course of action required to achieve a goal is much less clear. Therefore, the completion of a goal requires a translation layer, where one must break down the goal into tasks which one believes – assuming they are collectively completed – will achieve the goal.

I raise these concepts because I view the current paradigm of AI to be task-oriented. Over the last several years, the complexity of the tasks that the frontier models can do has expanded extraordinarily, but these tools still work best when given a well-defined task, and in fields where they have received adequate training and been provided with all the appropriate contextual information. They are not, at least in my experience, reliably capable of achieving even quite basic goals. Certainly the concept of reasoning is helpful here, and a step in the right direction, giving models a framework to complete the translation layer I've described above. I also think the expansion of the agentic approach (which could be argued as a paradigm shift in itself) will help. However, for many goals the translation to tasks requires a much broader context than they have access to, and is heavily dependent in most cases on the dreaded 'common sense.'

Now let's tackle the nature of intelligence. Perhaps my favourite part of discussing and thinking about AI is that it drives me to a deeper understanding of intelligence itself. Every time the frontier moves forward I find myself questioning my previous understanding of the concept of intelligence. Let's define a couple of terms on the topic, which can serve as building blocks, both for this discussion of the current paradigm, and later on when we turn our attention to the popular but nebulous concept of general intelligence.

Understanding Intelligence: a system which understands the reasons behind its own actions.

Functional Intelligence: a level of intelligence which allows the system to produce results similar to those of an understanding intelligence under a given set of circumstances, without necessarily having that understanding.

My favourite example of what I'm calling functional intelligence is the "living dead ant" experiment from the 1950s. Ants can be observed to carry dead ants out of their nests. This is quite sophisticated behaviour for an animal with a relatively small brain, so E. O. Wilson decided to see if he could figure out how this came to be. He realised that the dead ants were releasing a particular fatty acid, and that if you put that acid on a live ant, it would be ejected from the nest, despite struggling actively and generally showing clear signs of being alive! To me this is a nice example of an animal acting with functional rather than understanding intelligence.

As someone interested in behavioural economics, I can attest to there being many examples which show that – while humans are capable of understanding intelligence – many of our actions are guided by less sophisticated heuristics. There are lots of published papers about this, although my favourite example is actually from personal experience. I know many, quite capable people, who continue to fill their cars with fuel in round monetary amounts. I believe this habit developed from a time when people paid for fuel in cash, had a lot less money and had limited access to credit. There were reasons for an understanding intelligence to form this habit, but now that it has been established, humans who have adequate money and access to credit are partially refuelling their cars for no sensible reason – acting much more like a functional intelligence.

In terms of how this relates to the current frontier and paradigm, I believe it is clear we are at the functional intelligence stage with current models. They might be a lot more sophisticated than the ants, but fundamentally, they are still doing the same thing. The difference between knowing something as a standalone fact and understanding it in context, is that the latter requires the processing of the new information with regard to all previously held views. One consequence of this is that one should expect much more consistent conclusions from an understanding intelligence. I think the blatant guessing (not hallucinations), thank you Scott Alexander, which plagues current model outputs exposes this reality most clearly.

I'm nowhere near as confident as LeCun or Chollet seem to be that LLMs are fundamentally limited and that understanding intelligence will not emerge from this paradigm, but I find their arguments compelling enough to be extremely plausible. I suspect that we will never find out, because it seems that the next paradigm shift will come long before we reach the limits of this one.

There are so many different paradigms being explored by very clever and motivated people, that at least one of them is likely to bear fruit before this one runs out of room. Nevertheless, I think it is an interesting thought experiment to think through the limits of this one, and – because I believe progress within a given paradigm can be predicted with high confidence – I feel it is an excellent way for an interested investor to try to put a floor on AI progress over the coming months and years.

I expect that task length and complexity will continue to expand. Probably more impactful, I think the breadth of digital tasks for which models are useful will continue to expand. There are some intrinsic attributes to software engineering (verifiable, natural overlap of skills with AI researchers) that make it a natural area of strength for models in this paradigm, but I think that most of the reason current models are exceptionally good at SWE is because Anthropic has made such a deliberate push in this field. I believe similar pushes in similarly well-suited fields, of which there are many, would yield similar results.

Economic Implications

I expect that within this paradigm something like 50-80% of tasks previously completed by knowledge workers will be completed by AI. There will, of course, be some (probably a lot) of new activity that was either not achievable or not economical before this technology, but let's ignore that for the purposes of establishing a floor to economic impact within the current paradigm.

I like to think of a given company of reasonable size as organised in the following way:

• Leadership: Sets company goals (& oversees senior management)
• Senior Management: Sets department/regional goals (& oversees management)
• Management: Translates goals into tasks, allocates tasks to staff (& oversees staff)
• Front Line Staff: Completes tasks

To simplify let's use a 7:1 staff/manager ratio which implies around 85% of employees are front line staff regardless, of organisation size. Another simplifying assumption would be to say that front line workers are task-oriented, and managers are goal-oriented. This is clearly wrong but, if anything, this is an underestimate of how much of work is made up of tasks, because I would be confident management staff do more task-oriented work than front line staff do goal-oriented work.

Of course, management and leadership get paid a lot more per person, and in fact take home roughly half of total staff compensation in most firms, despite being so heavily outnumbered, though this varies widely by industry. It might be prudent to use a slightly lower number for front line worker compensation, but I think given that AI will make goal-oriented employees more productive by doing most of their tasks for them, cutting staff costs by 50% is a reasonable estimate.

Global GDP is roughly $125 trillion. Just over half of this, $65 trillion goes to labour, and of that roughly half goes to knowledge workers. So I believe AI in the current paradigm has a current TAM in excess of $20 trillion, on the basis that it is capable of replacing 75% of task-oriented knowledge work. I make no assumptions here about whether the staff doing these tasks currently will be replaced or will move up the value chain, though I suspect it will be a combination of the two.

Another area for potential AI gains within the current paradigm is completing tasks that no human can currently do. As previously discussed, tasks do not require a general understanding of the world, only the requisite knowledge and skill for that task. It seems very likely that AI within this paradigm will accumulate a combination of knowledge and skills that humans do not currently have, and potentially could never achieve. We've seen this already with AlphaFold (which of course is not an LLM but I would still consider it in the current paradigm), and there are now many other companies pursuing this line of inquiry in various divisions of science, but in particular in biochemistry.

Finally, we will have industries which would have theoretically been possible before now, but weren't feasible or economically viable before AI made them so. I don't know what these are yet, but I do know that a single person startup is now a realistic thing, and that hiring the necessary people was a major blocker for startups to this point. This is only one way in which AI will grease the wheels of industry. Tourism on a mass scale, and especially day-trippers, would have been completely impossible to imagine in the horse–and–carriage era, and then going for a day trip became an obvious and logical thing to do once one had a car. I don't think there is any question that AI will have its own equivalents.

In short, I believe the economic impact of the models we have now – and will have in the future – will be massive, even with no assumed paradigm shifts. That is why I am so bullish on AI related investments, quite aside from my views on AGI and related timelines.

AGI: Thoughts and Timelines

Let's start with a definition: AGI is a system which operates as an understanding intelligence in all fields of human endeavour, and has the ability and algorithmic flexibility to teach itself about new fields, such that it can operate at a level of understanding intelligence in those fields too.

I think it is fine for an AGI to use heuristics in some cases (doing so has clear efficiency advantages) but the system must then regularly reassess the validity of these heuristics, using its understanding intelligence, and change them when appropriate. Also, let me clarify that in this context I'm referring to a particular instance of the model or agent. A model then parallels better to species, or some class of individuals.

I also find it useful to think about the attributes I would expect from AGI:

• Consistency of views. If it's easy to make the thing contradict itself by framing the same issue in different ways, it's clearly not a general intelligence. To be generally intelligent, I think the agent must have a cohesive understanding of the world and universe it occupies. Perhaps a world model is just knowledge of a sufficiently large list of facts, but I suspect there's more to this. I think it's about having that knowledge, yes, but also analysing these facts internally, working out how they fit together, and questioning where they appear to contradict each other, rather than simply having each fact sit in isolation in static memory.
• Learning efficiency similar to or beyond human level.
• Long term memory (okay for this to be compressed/summarised, although ideally raw text data is saved somewhere and agent can retrieve when required)
• 2^nd Order Knowledge: having a sense of how confident you are about a particular opinion or how good you are at a skill. This is sorely lacking in the behaviour of existing models, whether by accident or design.
• Latency requirements to keep up with real world inputs at human level. For example, follow along in a group setting, synthesising the plethora of available information, from the topic under discussion to body language.
• Continuous learning: this isn't just about acquiring knowledge over time, but also synthesising that knowledge, updating its views on all topics related to this new information and seeking to draw new conclusions where possible.

From where I'm sitting, this feels more than a couple of years away. One argument made by those who believe AGI is closer or even here is that those who claim it is not are constantly moving the goalposts, but I feel this is the wrong way of looking at it. The reality is that the goalposts remain in the same place, but as we go through the journey towards them, we learn that they are further away than we first thought. So when we were 50 steps away, we thought we were 10 steps away. And now, 20 steps into the journey we realise we are at least 20 more steps away.

One interesting thing is that because there are serious people at either end of the spectrum (Chollet to Dario), you can reasonably hold any position in between these two without being accused of disagreeing with an expert; any view disagrees with either one or both to various degrees.

I quite like the idea that a generally intelligent agent will be composed of several different models, in a similar way to how the human brain is composed of different parts. Perhaps they use an LLM for language and thinking, and also have access to Deep Reinforcement Learning models for when they are developing skills for repeatable tasks. The concept of Mixture of Experts already exists within LLMs; why not apply the same architecture to decide which type of model to use, rather than just which weights to turn on and off?

Turning all of this into a set of falsifiable statements (to the degree probabilistic statements are truly falsifiable!), here is my view:

• LLM only AGI – Probability ~ 5%. AGI is achieved via LLMs only, emerging from combined progress of scaling, RL and thinking; if so perhaps it's reasonable to believe it could happen as soon as 2 years from now. I don't think so but I won't exclude it.
• Mixture of Models – Probability ~ 60%. An agent armed with multiple components, all of which exist today (LLM, World Model, DRL) and the ability to sensibly choose how and where to use each one can achieve AGI. If so, perhaps 4–10 years seems realistic.
• Unknown Breakthroughs Required – Probability ~ 35%. Maybe we will need an entire new paradigm, or even several paradigm shifts, before AGI is achieved. This would of course take longer. However, given the rate of progress to this point, the number of smart people focused on this problem, and the level of productive value AI will bring over the coming years, I still struggle to see this being more than 25 years away. My range here is then something like 8–25 years.

Economic Implications

If I'm right about AGI being multiple steps from the current paradigm, I expect, alongside the progress within this paradigm, to see lots of interim steps along the way. I'm not sure what these will look like, but I am excited to find out what they will be and all the ways that humans will find to make them useful, and perhaps even ways that they will find to make humans more useful!

A practical consideration I want to mention while on the topic of AGI, is that there is a big difference between a model capable of AGI, and an agent that is AGI in practical terms. As I've described above, this agent will need a constant stream of information about the world, and to synthesise that information to constantly refine its knowledge base and internal world models. It also needs a cohesive and near-complete memory. This amounts to the agent having near-constant access to compute and memory that is not directly related to any economically useful activity it is engaged in at a given time.

Takeaways

• I think the average person is heavily underestimating the power of AI to transform the global economy within the current paradigm, especially when you consider the power of agents to extract more from current model capability.
• I suspect at least some within the AI industry are extrapolating the current rate of progress (within this paradigm) to imply future progress outside the paradigm occurring at the same rate, which seems like a mistake. I think this might explain why there are serious people on both sides of the long and short timelines argument. A specific example is that people are not giving enough credence to the gap between task completion and goal completion.
• Personally, I feel like there are a lot of difficult steps between current capabilities and what would be required for me to consider a frontier agent a full colleague in my own workplace.
• Capability is one thing, and capacity is quite another. I think that the level of continuous compute required for an AGI-level agent would be so expensive that it would make wide adoption of such agents within existing businesses uneconomical for quite some time. It might make more sense for companies to keep their compute for task–based work, with humans handling coordination and task specification, not because the AI capability isn't there, just because humans can outcompete on price (and availability at scale).
• I think models will progress to super-human levels (ASI) for task-oriented work in several fields before getting close to AGI, which could exacerbate the same impact as the previous point. This is in no way to underplay the transformational power of AGI agents. I have no clue what crazy explosion in economic activity they could cause. I simply want to point out that the time delay between AGI and the point when humans are obsolete for all knowledge work, including co-ordination and goal setting, might be quite long.