Policy scholar Dean W. Ball argues 2023 tech industry forecasts underestimated the steep trajectory of frontier AI capabilities

IMO diminishing returns to capabilities is a big deal and mental models here hinge more on the effects of that rather than whether model capabilities peak at some level or continue to increase.

I agree on heavy-tailed excellence! But in a lot of areas, there are enough tasks with high human premium that I expect heavy-tailed excellence in competitive markets to come from differentiation in the *less* automatable tasks, rather than going from 90% to 98% automated in the rest (e.g., Gans & Goldfarb "focus effect").

The fact that any frontier lab is spending any time lobbying or PRing against open weights strongly suggests that your position here is false.

We are in the early stages of the AI industry, but it's just computing. We're in the early stages of AI computing. You think you're Intel. But you're IBM. (Which is pretty good!)

@herbiebradley Agree with your second paragraph, but I think a lot of that differentiation in less automatable tasks will look like humans with machine augmentation rather than humans alone. I suspect frontier intelligence will be quite useful in many of those contexts.

@deanwball Having got my puff-puff-pass of Fable (sans the pass), I can confidently assert that anyone who thinks this stuff is commodified at present is Unintentionally Confessing they can’t properly evaluate intelligence.

@deanwball Also explains why they’ve taken this insane, radically anti-regulatory stance towards frontier AI risk

@daniel_271828 yep. increasingly their discourse feels like it requires disbelieving your eyes, imagining eg a political economy where open-source models profoundly more capable than Mythos/fable are encouraged by governments just as governments begin hardcore regulation

@daniel_mac8 I guess I shouldn't really share my thoughts on how this will resolve since any insights about this particular strategic question will probably soon constitute trade secrets of openai group pbc

@deanwball I believe the commodification question will come down to if the value mainly accrues in the intellectual/informational layer ie. the learning algorithm vs. the hardware layer.

The market is going crazy because it's having trouble with whatever folksy narrative catches hold. The AI narrative has tanked SaaS valuations, but have LLMs reduced companies' dependencies on SaaS? Do companies have the expertise and talent to divorce themselves from SaaS vendors? Isn't it already well established that enterprises struggle to leverage AI? People are believing in narratives that have yet to be established to be true.

@deanwball been watching this play out in real time and it's wild how many people keep underestimating what's coming

the "good enough" mindset is such a trap - like saying the internet was good enough in 1998

Perfectly said. I've long been mystified at why consensus is so confident on model commoditization. Heterogeneity is near impossible imo - intelligence is a very subjective metric and tastes differ. Room for differentiation is infinite. People tend to see land-grab driven intense competition and think it leads to commoditization - but that's wrong . Personally think the labs will succeed far beyond expectations , and legacy giants will struggle to match their pace .

Your argument is that because I will work at OpenAI in a few days and wrote bearish analysis about open-weight models, it must be wrong because it was bearish? Also for the record I have written a ton in support of open weight, not to mention helped write a section of the country's current AI strategy devoted to embracing open weight. I don't think you know who I am at all.

@deanwball we’re going to look back at ai today like we look back at 8-bit atari games. it doesn’t take that much imagination to imagine much richer, smarter, and more ergonomic ai.

@deanwball Most people are REALLY bad at imagining what a life/world that isn't their own must be like.

The only thing that could possibly stop the acceleration is a world war.

@deanwball we are not a car!

@deanwball Once the models can write a novel equivalent to LoTR, then they will be "good enough" lol

@deanwball We can barely serve 5T models and human parity is at least 100T. There is a lot of juice left.

I was definitely a hesitant member of that community - refusing to fully learn the bitter lesson - until o3 opened my eyes to the direction of travel.

The key fissure was seeing "tool use" in Chain of Thought - so I was ready for the earthquake of reliable instruction following and 90% on TerminalBench.

@deanwball @AndrewCurran_ Cc @lessin @MoreorLessPod maybe invite Dean on to discuss frontier business model and adjacent points in evaluating ai startup business models and opportunities?