Future of Discretionary/Fundamental Hedge Funds

So AI can do everything except the smart things you guys do… Do you realize how biased that sounds? You guys are so smart that of course nothing you do can be automated but artists are all idiots. You stop believing nonsense like this when you do literally any job outside of the industry. The reality is any paid profession has nuance and skill and sure you guys get paid a lot but I don’t believe it’s because the job actually is that difference in terms of knowledge or reasoning ability required if you think about it on a bit more grand of a scale and put intelligence in perspective. 
 

chess players thought what you guys think now. So did artists and writers. Then programmers. So did videographers and physicists. I really think a lot of you are in for a surprise because a lot of what you do is trivial. A lot of what you do is protected right now by limitations in multi-modal input and agentic behavior for these models. You won’t feel the way you do when these constraints go away. 
 

there’s no reason why humans are the end of history or why we are the last species on the planet. Think on a scale of intelligence in the universe with maybe like atoms on the left hand side then like rocks then plants then like bacteria then apes then us on the right. On this scale, I don’t think we are all that much smarter than apes and the distinction between the intelligence of humans looks small. But think of how much a relatively small amount of intelligence produces for a species - it means we essentially became the dominant species on the planet. And I don’t think there’s a philosophical reason why the scale of intelligence needs to stop with humans. 
 

AI speaks the language of autism and doesn’t differentiate between what is “easy” and “hard” for humans because of inherent abstraction or nuance. Like an autistic kid I think abstract philosophy and booking a flight are equally easy or hard. It’s just as easy or hard for it to understand how to book a flight versus reconfigure a Kubernetes server or do an LBO model. 
 

a lot of the assumptions people on this site hold are not really true for modern deep learning. For traditional statistical learning, increasing model parameterization decreases out of sample performance due to overfitting, but for deep learning there’s “double descent” phenomenon (increased parameterization increases, then decreases, then increases then decreases performance, which is bizarre). People used to be like “bro idk if we can have 3 variables in this regression what about overfitting” or like “bro idk if we can have a polynomial of high degree here bc overfitting.” Have you stepped back to think about what these parameter counts mean when people say “X model from OpenAI has 1 billion parameters”? How did we go from thinking a few variables was overfitting to a billion parameters being ok? This is honestly such a paradigm shift in how we think about how machine intelligence works that it’s mind blowing because it implies that the traditional problem of overfitting is no longer a hard constraint provided you work around it in an intelligent way. 

also most people’s intuitions about modern machine intelligence expressed on this site are based on the approaches from 10 years ago. So I guess they are intelligent opinions but they are just totally outdated. It’s actually not at all true that you need large samples to learn from in order to generate domain specific intelligence. That’s just not how modern models work. The modern approach is to take a foundation model with massive data then adjust it to a domain specific use case with parameter efficient fine tuning approaches and there’s actually little evidence that suggests you need more than a few hundred samples. 

even complex phenomenon can be captured by these models and they don’t need a million examples anymore. Have you ever really thought about how these deepfaked onlyfans models work? The implications of that are actually insane because historically you could not get a consistent image of one person from a diffusion model but now you can, which is what allowed these AI influencers to now exist. It can with one example of a mode produce enough content to fool all the simps that pay for onlyfans. The implication here is these llm models can understand specific context provided (eg a base image) and work from within that context, though now it mostly can only do this with images. 

Sure, right now you need a human to babysit a model and chat with it. It can’t easily take continuous input from the user’s screen or from external sources because of the costs of running the model GPUs at inference time. (A continuous feed of external input would be too costly basically since each update would essentially be another query.) step back and think about this - a lot of the value of humans in many domains will exist primarily because GPUs may be too expensive so we need a human in the loop to help allocate GPU usage. But this constraint will not exist forever. Eventually multi-modal and agentic systems will exist that are open source and can be run on rentable GPU units. When this happens a lot of things will change. 
 

tldr I basically think most highly paid knowledge workers should just buy NVDA, MSFT and similar stocks as a career hedge. perhaps kind of for similar reason as someone might buy gold as an inflation hedge. The scenario where a lot of workers are redundant imo is a real possibility and maybe it doesn’t materialize but if it does I would at least buy a hedge.