Moe

i don't really see an argument against AI acceleration: back in debate, if you could tie your argument to the most lives affected in a plausible link (w/o K's or non-trad args), you'd usually win, after all you had the biggest impact. thus, i make my impact calculus: for model acceleration: 1. super-intelligence (given a 5-10 year timeline), is the only argument you need: *as a counter to a lot of the against arguments (3), if we end up in an ultron-type scenario, vision is the best counter. thus, having a more powerful, aligned model is a naive but realistic argument. 1a. solve all forms of disease, we're seeing early progress in cancer treatments, wet lab experiments, and scientific development. this feels like an inevitable. we will solve medicinal research undoubtedly with LLMs. in fact, LLMs will allow for specialized medicine to become possible (individual cancer treatments to target different types, etc). 1b. given western dominance, we solve all geopolitical issues. this one is more complex, but from the way i view LLM progress, RSI means that any given lead MUST be maintained and will continue to grow. if SOTA makes the next model better, faster, iteration cycles will continue to accelerate, and being 3 months behind now, will be years in the future. it is IMPERATIVE that we maintain that lead geopolitically. after all, if any belligerent foreign actor gets access to a model that can outperform ours, we run the risk of bio-weapons we can't understand, shifts in technological domination, robotics & military technology at levels we can't combat, etc. imo, almost everything is downstream of intelligence, and most countries used to brain drain into the west. now, people aren't the bottleneck to intelligence/research, models are, so giving up the lead functionally is the same. 1c. the most downstream of all of these is we just solve all problems. math, biology, etc, all forms of open research will be solved. we can do immensely powerful things via this technology. in the next 10 years, i want to be able to rip a prompt to the SOTA model that says something like "build me a rocket ship and take me to my friends on the moon", and i would want that mode to exist. the only way we get there is via more intelligence and investment into physical AI. every luxury we could ever possibly want has come via some sort of intelligence, whether it be taste-based or science-based. 1d. a lot of the above arguments have a very clear bottleneck, physical materials. i personally am under the belief that intelligence will help us get around these constraints to an extent. for example, if a certain material is too difficult to get and/or too expensive, i believe having a smarter model may help synthesize a new version or a replacement. there are ways to lever intelligence in different angles to accelerate all development. against acceleration (similar timeline and probabilities imo): 1. job loss / identity loss. there seems to be two-parties here and both are valid. job loss is your average person losing their knowledge work job (accountants, lawyers, etc) to a model. identity loss (more frequently in SF), is people who fear not being able to contribute to the world at all anymore. identity loss seems to be a bigger fear in my mind, given that if people were to lose all forms of identity it leads to heavy civil unrest / revolution. 2. environmental issues. data centers & disruption to the community and environments there. electricity costs from models. 3. models turning evil and killing all of us or something along these lines. given all of the hacks + fear mongering from labs and politicians alike, i think this is a really big concern, not sure if it should be though. alignment & safety are problems that undoubtedly need to be understood and worked on though. 3a. i think the bigger argument here is that by us not understanding interp and potentially not being able to understand and analyze traces perfectly, it may be difficult to catch or understand latent patterns in models that could be dangerous. as models get smarter we understand them better, and hopefully more aligned, i think this will become less of an issue. moreover, bad actors using good models (cyber attacks, bioweapons, etc) for bad things, is my biggest concern and the most probable of negative outcomes, and the most tangible. the naive argument against all of 3 is that a smarter, more aligned model just counters though. 4. hemming of people's intellect & an over reliance on LLMs to do all thinking, thereby stunting the growth of students/people in general. valid concern and i think this needs some level of thought on how to structure new learning systems around LLMs. for the first time ever, students can ever-growing, customized teachers, we should empower that. i wish i had this technology growing up. however, used incorrectly, i can see this happening. 5. power concentration in a few companies. tbd on how aligned those companies are. this is a bit of a red herring, but i think the alternative to western concentration is belligerent concentration, which is worse. though i think this is a valid argument (not sure if there is any single government or company that should have this kind of power though, which is why the race b/w model companies is actually good). 6. fake information/deepfakes/etc. this is a byproduct of technology getting better, not really sure there's a way around this. regulation, watermarks, filtering, and education are gonna be important here, similar to internet safety lessons imo. although i can see both sides, it feels like for acceleration clearly outweighs.