Season 6, Post 12: California Dreaming: Notes from the Valley

Anyone vaguely familiar with how the Large Language Models that drive products such as Chat-GPT work will know that they have a tendency to hallucinate. After having spent most of last week in California, your author sometimes wondered if he were hallucinating. Jetlag notwithstanding, the fervor around all things AI verged almost on the hyperbolic.

Land at San Francisco airport and drive to the city. Almost every billboard along the way will tell you how AI can ‘10x’ or ‘turbo-charge’ your business. Sit in a bar and sip beer. Listen to the conversation around you and nearly everyone is talking about how they’re working on an AI project. If not, then their friend is. At times, it almost felt like 1999 and the moment when the Internet went mainstream.

Nobody said “this time it’s different” in as many words to your author, but several came close. During his time, your author met with 13 companies ranging from over $500bn in size to a Series-B funded start-up. At 5 of these businesses, he met with C-level management. He was told by one senior executive that AI is “profoundly different.” Another said “levels of interest are unbelievable… as barriers to resistance melt.” AI, he was reassured by a third business leader “will make us quicker.”

Fancy words for sure. If anyone is going to fashion the tools that will enable AI to become mainstream, then it will be those in the Valley. San Francisco might have one of the most appalling homelessness problems among major US cities, but it attracts over 40% of the country’s VC funding, more than three times the level of the next closest state (New York).

Why are we still in an early innings with AI? Your author was reminded that pre-trained models can already pass Bar exams. An average law student would need 400 hours to achieve a similar feat. By contrast, a teenager could learn to drive in a tenth of this time, but no autonomous car can come close. Indeed, your author heard how almost all the self-driving cars that had previously operated on a trial basis in California had been retired owing to safety concerns. If driving successfully is about prediction, contextual learning, vision and more, LLMs lack full functionality today. This implies the need for more training of models and further inferencing. It also means more GPUs and more silicon.

Believe such a world view and “no way is this a hype a cycle”, as one Chief Executive put it to us. NVIDIA is, we were told, training models “to do things never done before.” No surprise then that AI is now part of every business (and investor) conversation. Few, of course, were willing to share their views on valuation, although one highly experienced person with whom we spoke noted that “the market temporarily overvalues things it does not understand.” We will not opine here on this topic. It was, however, reassuring to witness innovation at work on this trip in areas other than just AI. Your author was particularly impressed with progress being made in robotic-assisted surgery and lab-grown food but will save these topics for another day.

19 March 2024

The above does not constitute investment advice and is the sole opinion of the author at the time of publication. Past performance is no guide to future performance and the value of investments and income from them can fall as well as rise.

Click here to view all Blog posts.

Alex Gunz, Fund Manager