On Hugging Face's Open LLM leaderboard, Falcon 40B was top-ranked a quarter ago. Now, it does not feature in the top 50. This churn outpaces the volatility of weekly music charts.
Even while we have talked about the excessive pace of innovation before (https://bit.ly/43hzHK6 or https://bit.ly/436RBOQ), one feels compelled to return to the subject for its massive implications across fields and themes. The dizzying pace of advancement hurts its current leaders, too.
The "instant copiability" enabled by GenAI, also a hallmark of the new era, means replications of any advance in weeks or even days. In the LLM arena, unique innovations rarely retain headline prominence for more than a news cycle before being outshone by the next jaw-dropping demo.
Yet many commentators still speak of "investing in AI" through a buy-and-hold lens, without accounting for the field's frenzied dynamics. Such strategies predicated on multi-year inertia carry substantial risk.
Consider the most touted beneficiaries: Nvidia and OpenAI. Some paint their traction thus far as self-reinforcing advantages, given customer lock-in to their operating-system-like platforms of CUDA and GPT.
With vertical integration as the theme of the era, both have all the tech giants' from disconnected fields entering their spaces. From Microsoft to Meta, purpose-built, AI-specific chip, aka Nvidia-defeating products, are planned by resourceful companies who rarely looked at chip-making before.
It is worse for OpenAI. Hundreds of LLMs are being built globally. The overpowering quality of alternatives like Anthropic's Claude reveals the precariousness of any single player's dominance.
In addition to intensifying direct competition, both face headwinds from adjacent innovations that erode the uniqueness of their capabilities. Advances like chiplets and CPU enhancements enabled by generative AI allow NVidia's chip competitors to close performance gaps. Even without competitive chip architectures, progress in optimizing and accelerating model training algorithms reduces the advantage of Nvidia's hardware lead.
OpenAI has to contend with regulatory scrutiny and technological countermeasures. As its capabilities raise concerns, regulators are increasing pressure globally. Meanwhile, web tools that help sites evade scraping by OpenAI's bots undermine their access to crucial training data. For both, maintaining advantage requires continuous invention with little scope of error.
There are scenarios in which leaders in some segments emerge as multi-decade, runaway winners by leading in innovation that keeps building on previous innovations, with each version only increasing the gap. However, this cannot be taken as given with any stock as a multi-year definite investment when we cannot even forecast where products will be a few months later. The best a long-term investor can do is keep the exit door open and hope this option is never needed.
