OpenAI is less good as an investment than their revenue curve would suggest because the US government is likely to cap the size of future training runs (to redu

OpenAI is less good as an investment than their revenue curve would suggest because the US government is likely to cap the size of future training runs (to reduce AI extinction risk) and there are plenty of open-source offerings suitable for the work of fine tuning and otherwise exploiting the economic potential of models of the size of GPT-4, so it is hard to see how OpenAI would be able to construct much of a moat by which to protect their prospective future profit margins. The idea that AI research must be regulated to prevent harms, including human extinction, is ridiculed on this site and in Silicon Valley, but policy is decided by representatives for the entire US voting population, not just the technologists of Silicon Valley, and polls show widespread skepticism in the broader population about the wisdom of continuing to create ever more capable AIs. Also, the people lobbying for a cap are disciplined, smart and well-funded. Companies that focus on applying the models we have now will probably do OK (e.g., Google's using something analogous to GPT-4 to improve their Search service except that unlike GPT-4, Google's model can cite its sources) but OpenAI is not doing that: their focus is to increase capabilities trusting that it will become clear in the future how to use those increased capabilities to produce economic value, but that plan doesn't work very well if the government imposes a limit on the increase of capabilities. Don't say I didn't try to warn you. ADDED. The leader of OpenAI says that it cost OpenAI roughly 100 million dollars to provide the computing resources necessary to create GPT-4. Once that "foundational model" model was created, it could be run (deployed) on much humbler hardware; some open-source competitors to GPT-4 for example can run on a beefy Macbook. Naturally there is no practical way to stop people from running a model once it has been released to the public, so regulation will target the "training runs" such as the aforementioned training run that produced GPT-4. The point is that there are many ways to "fine tune" a basic model to increase its economic value. GPT-4 by itself for example is useless as a search engine, but Microsoft fine-tuned it to produce something (Bing Chat, a.k.a., Sydney) that has some hope of becoming a popular search engine. You probably want to restrict your AI investments to companies good at this fine-tuning process.