Maciej Cegłowski's talk "Web Design: The First 100 Years" (which my friend Alec recently sent to me) does not entirely stand the test of time, but there's a nice parable about airplanes that is worth contemplating. As a short summary, Boeing's development of the 747 (circa 1965) was preceded by many years of continual improvements in airplane speed and capacity. There was therefore some sentiment that airplanes would get faster and faster, to the point of speculating that commercial space travel was around the corner. Cegłowski discusses how the 747 was actually sort of a second-fiddle project to Boeing's main focus of developing a commercial supersonic aircraft. However, ~60 years layer, airplanes largely resemble the 747 and supersonic travel never really took off. It turns out it is just too expensive to build and fly supersonic airplanes, not to mention the loud booming noises they make. Halving the duration of a transatlantic flight is just not worth the dramatically increased cost and inconvenience.

I wonder if language model development will follow a similar trajectory. Many claim that language models will continue to improve, leading to predictions that language models will attain superhuman intelligence in most or all computer-based tasks (not to mention predictions of recursive self-improvement and so on). Recent progress suggests this could be possible, and scaling laws provide some scientific basis for these claims. However, scaling requires exponential investment. Epoch AI's analysis suggests that model training costs are increasing by roughly 5x each year. The costs of training frontier models is already staggering; for example, OpenAI's leaked financial documents suggest training costs of roughly $10B and, more broadly, a worrying picture for the financial sustainability of continued scaling. At what point does the improvement provided by scaling not justify the exponentially increasing cost - not only economic costs, but also in terms of externalities such as energy usage?

Meanwhile, although their capabilities are certainly “jagged”, current language models are reasonably helpful for many things (citation needed, but hopefully this does not require much justification). In areas where they are already quite helpful, it is a bit hard for me to imagine any groundbreaking benefits from improved capabilities. Simon Willison tells an interesting story about his experience asking Fable to fix a simple UI bug. Fable ultimately spent an inordinate amount of time and compute building a custom framework for running the app and taking screenshots before it ultimately produced a two-line CSS change to fix the bug. I would posit that this task would have been better accomplished with a “weaker” model that was provided with appropriate tooling (so that it could avoid building out a custom debugging framework). In this case - and perhaps increasingly many - what do increased capabilities buy us other than the ability to be lazy? Additionally, while the improvements claimed by new models are indeed impressive, I wonder how relevant they are to the majority of use cases. Do we really need our models to be able to (dis)prove Erdős conjectures when we ask them to give us a taco recipe?

Even “small” models far from the “frontier” are, apparently, quite useful. For example, there are many reports that recent open-weights models, when harnessed correctly, are good enough to be useful coding agents, approaching the utility of late-2025 “frontier” closed models while being runnable locally on modest consumer hardware. If, for some tasks, the movement of the frontier in the past year mainly allows for laziness, then perhaps smaller, open models have already reached “good enough” for those tasks. Perhaps instead of being lazy we can put additional effort into better "managing a genius" and figuring out how to build systems around good-enough models to allow them to work as effectively as possible. As on example, the recently developed AI worm is, apparently, a somewhat mediocre (albeit undisclosed for security reasons) local model, surrounded by what sounds like a sophisticated harness. Additionally, in tandem with increased scaling, Epoch AI estimate that the amount of compute required to attain a certain level of performance has roughly halved every 8 months. An interesting scientific question related to “when will models be good enough?” is “how big is big enough?”; I suspect the answer is quite a lot smaller than current frontier models. Perhaps the most prudent path for language model development is not scaling, but rather redoubling our efforts to shrink down good-enough models.

Thanks to John Kirchenbauer, Gül Sena Altıntaş, Derek Tam, Alec Jacobson, and Yoav Artzi for feedback on this post.