AI as an exo-skeleton
So Alpöge and Claude just proved the GRH (Generalized Riemann Hypothesis), which is extremely impressive. I think this really says a lot about the era we are currently in. Brilliant people with the help of AI tools can achieve much greater heights than previously imagined. If you have sufficient experience, context, and critical thinking skills you can guide the LLM in order to speed up research, development and search by many, many times.
The doomers push the narrative that the more effective AI is the more likely it will lead to doom, but I prefer the cautiously optimistic view that improvements in AI isn’t necessarily a negative. Many things will change, but many things won’t. We have had the technological ability to eradicate mankind for almost a century but so far mutually assured destruction has kept the status quo more or less stable.
What I think is that AI will come to live as a layer on top of the computing capabilities we already have. Local models will eventually find their way to all capable consumer hardware, while frontier cloud models will be rented to companies with decent RnD budgets. There has always been a discrepancy in available computational resources since the beginning of the era of the computer. Large companies and universities had access before everyone else and were able to run experiments, research, and tasks to achieve their goals. Later computing became cheaper and now everyone has the supercomputers of the past age in their pockets. Who is to say that the same won’t hold true with LLMs? In a couple of years I think every smartphone will ship with reasonably capable LLMs to fit various use-cases. Sure, frontier models will never fit on consumer hardware but does it have to? Does everyone need a PhD-level frontier model at their immediate disposal? Just like the average person had no immediate need of Quantum Supercomputers in the pre-LLM era, I’m not sure that the ecquivalent will be needed going forward.
I’m also not convinced the AI will vaporize all white-collar jobs either, there are enough problems to be solved and opportunities in any era to accomodate most people. Also an experienced enginer + a frontier LLM will always beat out a single LLM running on some vague prompt like “You are a senior engineer at $CORP, do good work and make money for the corp”. Someone will always need to be watching and grading the work of the clankers, how else do we know that what they are doing is good?
If I tell an LLM to make me a millionaire and it starts scamming people and eventuall put me in jail, was it worth it? I think not, no one can trust a probabilistic algorithm 100%.