My personal strategy is specialization, because I've noticed that I cannot work with AI in hard domains that I haven't got a background in. For example, I've tried to use claude opus to understand a phd thesis in quantum physics from a friend. My background is in engineering and later compilers/static analysis. Despite opus helping a lot to give me the high level idea behind it, I could see a few problems with my lack of background: - I couldn't verify if opus was right or wrong in its explanations - Even when asking the AI to simplify explanations, there was so much prerequisite knowledge I needed to absorb that there was little point to bother with its output - It was very hard to collaborate with the AI to understand what predictions can be made from the PhD. The model could produce a lot of output but I didn't really understand its answers. Asking follow-up questions did not really solve the issue because it felt like I was missing a mental model.
My thinking is that, no matter how much LLM intelligence grows, for sectors with inherent complexity, people will still need specialized expertise to understand, evaluate and use their outputs. I also doubt that the most efficient future is one where humans don't understand the outputs and delegate everything, because it will be both be hard to understand if the machines are aligned to the benefit of their users.
So specialization sounds like a good strategy for the future.
I'd like to hear some opinions around this topic. Have you observed the same? Do you disagree based on other experiences/data?
thewebguyd•1h ago
But I think that's highly dependent on if we've hit diminishing returns on LLMs yet, or if the US Gov (if you are US based) is going to continue to restrict models over a certain parameter count, or models trained on a certain level of compute, etc. and pull additional models after they've pulled Fable/Mythos.
If Opus 4.8/GPT 5.5 ends up being the strongest model the government allows to be publicly available, then yeah, specialization will be helpful.
If we haven't yet hit diminishing returns, and models continue to improve, and they don't get pulled off the market, then I think the opposite may become true in ~5+ years and that being a generalist will be more beneficial. Knowing enough to be dangerous across a huge variety of verticals, with AI assistance for more obscure stuff.