Artificial Intelligence – Lifecycle of a bubble

The 4^th AI bubble is in full swing.

• I will be the first to admit that large language models (LLMs) can do far more than I expected but the problem is that expectations are now that super-intelligence is just around the corner which has triggered all the usual irrational behaviour that precedes a large correction as reality re-asserts itself.
• RFM research has concluded that LLMs are really good at understanding and conversing with humans and at storing large amounts of unstructured or structured data in a way that is very easy to recall and analyse (see here).
• However, contrary to the general opinion, RFM concludes that because LLMs are based on the same fundamentals as all other deep learning systems that have gone before them, they will be subject to the same limitations.
• It is these limitations that mean that generative AI is unable to reason or understand causality but instead can weave a convincing illusion.
• This illusion is so good that even participants in the field have been taken in which in my opinion greatly increases the scale of the bubble that is currently being inflated due to the argument to authority.
• There are signs of this everywhere with everyone now announcing an intention to develop generative AI or throwing money at start-ups with very few questions being asked.
• The latest example is from Salesforce which has suddenly increased the size of the generative AI fund that it launched in March 2023 from $250m to $500m.
• The companies that it is investing in like Cohere, Anthropic and so on have no revenues but, because they are working on the latest hot area, their valuations will be beyond anything that could ever be rationally justified.
• This is the approach that SoftBank used in the past and which got it into so much hot water and is a sign of a bubble in full inflation mode.
• Currently, extravagant promises are being made by these companies and because they have no revenues, there is no way to have any certainty of what revenues they will be able to generate or to do any meaningful due diligence.
• Generative AIs are massive black boxes with no visibility as to how they work meaning that investing in this area is even more risky as it is almost impossible to know what one is buying.
• Consequently, under the well-established doctrines of rational investing one should pay less for these companies given the higher risk being assumed not more because they are trendy.
• However, while capital for this theme is abundant, there will be an increasing number of companies appearing to soak up the available funds regardless of whether they have a viable business case.
• The turning point is likely to occur sometime in the next 24 months when the start-ups that have soaked up all the money will need to seek their next rounds.
• This typically happens after start-ups have met some of their targets and need more money for the next stage of their development.
• The meeting of targets enables the start-up to increase its valuation which is why they don’t raise all the money they need in one go.
• The problem here is that almost all of this latest batch of generative AI start-ups will fail to hit their targets or keep their promises.
• This means that when it comes to refinancing, it will be mostly refinancings at lower valuations or highly dilutive consolidation.
• This is exactly what happened with autonomous driving and SPACs.
• This will be what precedes the general awareness that the generative AI theme has been overhyped and that reality is in fact a good deal lower than expectations.
• This has happened many times in the past with other new technologies and generative AI has all the hallmarks that indicate that it will repeat the experience.
• I am not saying that generative AI has no use nor am I saying that it won’t cause a lot of disruption and make a lot of people a lot of money in the process.
• All I am arguing is that expectations and reality have become wildly out of touch with one another and that there will be a painful reset when reality returns as it always does.
• Nvidia is undoubtedly one of the winners of the current craze but even with much higher expectations, its multiples are so high that it is impossible to justify the share price using fundamentals.
• Instead, I would be looking at some of the knock-on effects of generative AI such as a large increase in the amount of inference (running AI models) happening on devices rather than in the cloud.
• Other areas would also include a new industry being born focusing on prompt engineering as well as products and services being offered to help companies create LLMs for their own internal use.
• I continue to think that corporate data companies like Oracle or Snowflake could be on the receiving end of this disruption although they have time to take remedial action.
• For now, the mania continues unrestricted.