BTB

A few Macro colleagues have reach out asking where I stand on the AI bubble debate. I thought to share my thoughts here as focusing on whether AI is a bubble or not is the wrong way to frame the technological shift the world is experiencing. Don't be a maximalist. In the nuance is where I find the value. Long post 👀 There are many developments in the AI ecosystem that are happening at the same time, one can't just group them all together and categorize them as a bubble or not. Infra, models, tooling, agents, apps all sit at different points in the S-curve. Categorizing everything as a bubble, is just underestimating the enormous and transformational value of AI technology - this is the new paradigm so get used to. First, one needs to acknowledge that the majority of investors (macro, equity, L/S, retail, FI RV, etc) know very little about the inner workings of LLM architecture. The majority is not following the leading edge of research, so have no anchoring insights of where things are potentially going and why the research community is focused on those topics. From this perspective, when investors question the technology because LLM can't count how many 'r' are in 'strawberry' (issue with early models) or hype the technology thinking it will replace a large number of jobs in the software industry overnight, it just tells you that the majority don't know what LLMs are and what are not. Second, you don't need to be a machine learning expert to realize the world is short of compute. But is helpful to step away from the AI narrative to understand why is this the case. As Jensen argues, we are moving to a new paradigm of accelerated computing, which goes beyond the AI ecosystem. Think of scientific computing and simulations, database acceleration, real time analytics, digital twin economy, edge computing transformation, etc. When Jensen talks about the capex needed going forward, he is thinking about this paradigm shift, he is not thinking we need more compute so more @sama can be meme via Sora2. Third, while many out there are calling US equity markets are in a bubble, the reality is that the AI ecosystem is moving in tandem with AI adoption. For the bulls, we are very early on adoption. For the bears, there will be mini-cycles where adoption feels frozen. In my view, instead of focusing on grouping the whole ecosystem into a bubble or not, focus on trading these adoptions cycles. 2 scatter charts below showing a linear relationship between adoption rate and performance of AI segments. Adopters are trading in tandem. More speculative stuff like bitcoin miners have moved ahead of adoption (this is due to the wave of deals we seen since Sept 8-9) Fourth, in my view adoption in aggregate will happen slower than people hope. Note, I was one of the early macro people that were bullish adoption. But by developing my own AI ecosystem, I now have a much more nuanced view -> Consumer adoption is happening fast, hence the response from @OpenAI to bring to market new monetization apps. Sama knows that to get to AGI, he needs to make $$ and the way to do this is via consumer monetization. But enterprise adoption is happening much slower than I anticipated, and this is because the technology is still immature. In my business, I need deterministic answers from the models, calibrating the model to do this consistently has been and still pretty challenging (even temperature at 0 won't guarantee it) - but we are making progress. At a dinner this summer with the great @wolfejosh I learned about the frictional frontier. He wrote about it in his investor letter - a must read if you want to understand many of the pieces that are needed to fall into place to facilitate complex system adoption (i.e. enterprise at scale) Fifth, I think we are in a once in a lifetime technological acceleration. As an investor and a computer science engineer, the AI revolution is a gift. And while I'm very bullish on the future and how AI can transform my business and the world economy, I'm humble enough to understand there will be AI winters along the way. @dylan522p highlights that if the models don't improve we may run into some trouble. This is something I'm very focused as I believe the "wow moment" of the LLM architecture maybe behind. Model improvement is unlikely to come via cognitive quantum leaps, rather it will come via better orchestration, tool calling, efficient algorithmic improvements, etc. (I say this humbly because I don't want to bet against the AI research community that are tireless working on advancing the ecosystem). Note, all these are very positive developments which the AI research community cheers pretty much every other day, however for the majority of investors, that don't know what an LLM can and can't do, these improvements feel irrelevant because their expectations are on that next "wow moment".