In this note, we return to the artificial intelligence gold rush that has – amid sporadic bouts of volatility – continued to fuel global equity market gains through the first half of this year. Having previously taken a look at the challenging business economics demanded by egregious valuations, here we consider the potential of the technology itself, asking whether it is likely that we are really at the dawn of a new technological and economic era that, in the words of Demis Hassabis, the Nobel prize-winning CEO of Google’s DeepMind Technologies, will be “ten times bigger than the Industrial Revolution, and maybe ten times faster”.
The magnitude of that claim is difficult to convey. The industrial revolution was not a mere efficiency gain for manual labour, it prompted nation-building waves of urbanisation while facilitating the economic and cultural integration of newly dynamic cities, it spelt the end of feudal rent extraction and the birth of the moneyed middle classes, and it promoted education, particularly in the sciences, as a means to a highly productive end. To expect ‘generative AI’ to deliver anything like this scale of change – never mind a multiple of it – seems, to us, extravagant.
Yet sceptical as we are of the transformative potential of AI – or at least the large language models (LLMs) that currently profess to this status – over anything approaching a reasonable investment horizon, this is ultimately subsidiary to our core argument. For investors, the debate is not really about the possibilities engendered by the new technology. It is about the tendency of market participants to extrapolate technological advances into the promise of abundant future profits and to behave irrationally in pursuit of a share of those imagined spoils. What investors must understand is that, at this scale of investment, and at prevailing valuations, anything short of the mind-bending and imminent economic and social revolution implied by Hassabis is likely to leave those funding the technological roll out nursing heavy losses.
We begin by contesting whether LLMs are either artificial or intelligent. Hassabis and others aspire to the creation of ‘artificial general intelligence’ (AGI), a system that can understand and apply knowledge across all cognitive tasks, beyond those on which it has been trained. The concept of AGI is often simplified as a single tool capable of surpassing human expertise on any subject, at any time, in any situation and any environment. Certainly, this could be revolutionary, opening a path to John Maynard Keynes’ ‘leisure society’. Unfortunately, our reading leads us to believe LLMs can never attain this status.
LLMs are ‘trained’ by ingesting the contents of the world wide web. The patterns learned from this vast depository of the written word effectively inform an enormous probability tree that enables the LLM to predict a sequence of words, one after the other, that together make up a plausible response to the initial prompt. There are at least three problems with this approach. First, no quality filter is applied to the training data (Wikipedia and Reddit are, apparently, the two largest training resources for LLMs). This ties into the second issue; that LLMs are optimised for plausibility, not accuracy or truthfulness. Thirdly, LLM developers have long-since exhausted all available training data, raising questions over the extent to which their models can continue to improve.
As we understand it, these issues are intrinsic to LLMs. Pattern recognition – and the prioritisation of plausibility over accuracy that stems from it – is what enables LLMs to transcend a fixed bank of known answers to known questions and respond to the infinite number of possible user prompts. However, this means so-called hallucinations (in which LLMs appear to invent non-existent sources, historical events, case law or even ‘must read’ novels) are a feature, not a bug. It means that essentially identical prompts can provoke differing answers. It means that models can be easily misled, sometimes to the extent that supposed guardrails are easily breached. Perhaps most importantly for those pinning their hopes on LLMs as the route to AGI, it means that iterative refinement of the models and the introduction of ‘reasoning’ processes cannot eliminate their inherent flaws.
In our view, the potential of a technology that provides plausible rather than accurate answers, that is given to fabrication and that will earnestly provide contradictory answers to identical questions is necessarily limited. Would you trust such a technology with the details to your bank account, accepting mistakes are inevitable and of unpredictable severity? Is a large company – perhaps one that operates under stringent regulatory or legal constraints – likely to hand control to a confident yet potentially reckless technology? Optimists claim it is only a matter of time before these issues are resolved and the technology can be trusted. It bears repeating that the limitations of LLMs are inherent: they can be reduced, but they cannot be eliminated.
That is not to say that LLMs cannot be useful. Many of us will have turned to them in a personal capacity, perhaps seeking inspiration when looking at the contents of our fridges or when traveling in a country where the language is unfamiliar. Commercially, several of the investment management firms with whom we engage have used AI to improve the ease with which their libraries of proprietary research and analytical inputs can be interrogated. A few years ago, this would have been called “big data” and the marginal efficiency gains it promised would have been impressive. Several hundred billion dollars’ worth of investment later, the gap between this and “ten industrial revolutions” is glaring. Efficiency is lubricant for corporate profitability, but it is clear that the gains need to outweigh the costs. With numerous companies cutting back their AI budgets as measurable gains remain conspicuous by their absence, the scale of investment that has already flooded into the technology must surely be called into question in due course.
The latest wave of excitement has been driven by the coding abilities of recent models from Anthropic and OpenAI, and the ability of such models to manage increasingly time-consuming tasks with minimal intervention. Users with sufficient budget can even establish multiple ‘agents’, powered by generative AI models from competing providers, to monitor, check and correct each other’s output. This is undeniably impressive technology. Yet we still struggle to view it as revolutionary. Certainly, new ideas can come to fruition more quickly, allowing new ‘apps’ to be developed at breakneck speed. No doubt some of these apps will enhance productivity – some may even offer something genuinely new and useful. However, we suspect the vast majority will disappear without a trace. Furthermore, the code generated by these models is likely to contain vulnerabilities in security and adaptability that present significant hurdles to large scale commercial deployment. Notwithstanding these challenges, 18 years after the launch of Apple’s ‘App Store’ and the proliferation of ‘software as a service’ business models that followed it, there are legitimate questions over just how economically impactful code-based change can be. Convenient? Certainly. Profitable? In places. But a revolution? We suggest not.
In the course of our research, we have been presented with myriad use cases for AI. For each, it strikes us that it is the consumer, not the producer, that stands to benefit. There is a self-evident limit to the amount a consumer would pay for an AI agent capable of scouring the internet for the best available deal on a given product, for example. Even assuming vendors do not respond by blocking agents, adopting dynamic pricing or altering their listings to confuse agents incapable of true discernment, is this amount enough to justify trillions of dollars of investment in the development of the agent and the infrastructure that underpins it?
Elsewhere, advances in the field of protein folding – the medical research endeavour critical to both understanding the causes of many diseases and developing targeted drugs with which to treat them (and, not coincidentally, the area for which Demis Hassabis was awarded his Nobel prize for chemistry) – should both increase the frequency with which new treatments are discovered and lower the cost of their production and the price developers can hope to charge for them. Though the benefits to humanity could be significant, the benefits to investors are much harder to identify.
And so we return to the core of our conviction. It is not necessary to think AI a modern Mechanical Turk to have serious misgivings over the possible returns for investors. Though we suspect the technology is not as revolutionary as its proponents would have us believe, the strength of our conviction lies not in our scepticism over the practical application of the technology, but in the eye-watering valuations and growth forecasts investors have attached to companies operating across the breadth of the AI supply chain. We believe sky high expectations must at some point collide with the unforgiving business economics of the LLM providers and the revenues they can realistically aspire to.
For investors, the debate is not really about the possibilities engendered by the new technology. Rather it is about the tendency of market participants to confuse potential for certainty, to extrapolate the recent past into the long-term future and to believe the risk of missing the upside is greater than the risk of bearing the downside. We believe these exuberant behaviours are driving markets today, as they did during the dot-com bubble twenty-six years ago.
In such an environment, where even an impact equivalent to ‘nine industrial revolutions’ might constitute a disappointment – and where our research leads us to believe that even one is highly unlikely – we consider it our duty to Heronsgate clients to remain disciplined, to rely on rigorous analysis rather than hope or conjecture, and to communicate our conclusions clearly. We believe there is significant risk embedded in equity markets trading on distant hopes for a nascent – and thus far staggeringly loss-making – technology, and strongly urge long term investors to tread cautiously amidst the euphoria.
