A current frontier model runs on a part of a rack of GPU accelerators. That's roughly $200-900K capital cost for exclusive access 24x365. With distilled and optimized models costs are lower. The highest that the market is likely to bear is something like $100K per year for exclusive access (like the access you want). Above this point I think most employers would currently prefer hiring another human. I wonder at what point people will start seizing GPUs and interconnects from data centers, diverting them during shipment, or using alternative hardware, to run less capable open weights models.
One argument for the ending list is around point 1: labs could be so scarce on compute that when thinking about what products to work on they go for the ones with the highest marginal profit which may well not be $20k/month SWE agents simply because there might be fewer buyers for that Vs cheaper agents in different verticals.
For the same reason, right now it's not worth it for labs to work on building RL environments for things with small TAM
'There was a time when everyone used Github Copilot. It used to cost $10 per month, or free for students...This world is already partly gone; the cheapest usable tier of Claude Code is $100/mo.'
This seems like an odd framing to me. You can currently spend $10 a month and get coding assistance much better than Github Copilot was two years ago. Or for that matter you can get that for free, using small open models run locally. You can *also* spend more money to get something even more useful than that, but that doesn't mean that the situation has gotten worse! This is the usual trajectory of new technologies in the marketplace, so it doesn't seem surprising.
Two cases in which that framing would seem more justified to me:
1. Competition for compute increases so much faster than it can be built that cost of compute actually goes up in absolute terms. This could become the case in future, but so far it looks to me like cost of compute has continued to drop (see eg https://epoch.ai/blog/training-compute-of-frontier-ai-models-grows-by-4-5x-per-year, https://www.accio.com/business/h100-gpu-trends). Maybe the scaling labs are mostly focused on using the latest and fastest chips, so there isn't that much competition for compute on older chips?
2. Coding assistants become a positional good. I can imagine this perhaps becoming true at some point in the future also, if/when software development becomes so fiercely competitive (and coding assistants so capable) that only those companies or individuals with access to the top-of-the-line assistants can succeed in the marketplace. My guess is that this is a temporary situation, though, during the interval when coding agents are vitally important but haven't yet automated the whole process. After that transition there may be other serious economic and distributional issues, but I don't expect them to look like 'the best coding assistants are too expensive'.
To be clear, I haven't done much thinking or research about either of those cases; these are just my thoughts at the moment.
In the meantime, it seems to me like there's still an abundance of affordable and increasingly effective coding tools for the easily-foreseeable future (FWIW I personally have been plenty happy with Claude Code on the $20/month plan).
My claim is that, in 2027, there will be AI coding tools which are much more expensive and somewhat more effective (or maybe much more effective) than the $20-$200 agents most people will use. The usefulness of the lower-priced tools will continue to improve; although I do not expect anything being priced lower than $8/month due to SaaS market dynamics.
The tone is like this because I felt like "you cannot possibly do research as fast as the researchers in the labs because you don't have the autonomous RE coding agents, or the money to pay for those" is a world worth preparing for, for people interested in technical work on AI safety (be it researchers, grantmakers, even open source tinkerers). Maybe I can outline possible consequences of that state of the world in another post.
A current frontier model runs on a part of a rack of GPU accelerators. That's roughly $200-900K capital cost for exclusive access 24x365. With distilled and optimized models costs are lower. The highest that the market is likely to bear is something like $100K per year for exclusive access (like the access you want). Above this point I think most employers would currently prefer hiring another human. I wonder at what point people will start seizing GPUs and interconnects from data centers, diverting them during shipment, or using alternative hardware, to run less capable open weights models.
One argument for the ending list is around point 1: labs could be so scarce on compute that when thinking about what products to work on they go for the ones with the highest marginal profit which may well not be $20k/month SWE agents simply because there might be fewer buyers for that Vs cheaper agents in different verticals.
For the same reason, right now it's not worth it for labs to work on building RL environments for things with small TAM
Bro I’m just out here trying to generate as many deepwikis as I can while it’s free 🫡
'There was a time when everyone used Github Copilot. It used to cost $10 per month, or free for students...This world is already partly gone; the cheapest usable tier of Claude Code is $100/mo.'
This seems like an odd framing to me. You can currently spend $10 a month and get coding assistance much better than Github Copilot was two years ago. Or for that matter you can get that for free, using small open models run locally. You can *also* spend more money to get something even more useful than that, but that doesn't mean that the situation has gotten worse! This is the usual trajectory of new technologies in the marketplace, so it doesn't seem surprising.
Two cases in which that framing would seem more justified to me:
1. Competition for compute increases so much faster than it can be built that cost of compute actually goes up in absolute terms. This could become the case in future, but so far it looks to me like cost of compute has continued to drop (see eg https://epoch.ai/blog/training-compute-of-frontier-ai-models-grows-by-4-5x-per-year, https://www.accio.com/business/h100-gpu-trends). Maybe the scaling labs are mostly focused on using the latest and fastest chips, so there isn't that much competition for compute on older chips?
2. Coding assistants become a positional good. I can imagine this perhaps becoming true at some point in the future also, if/when software development becomes so fiercely competitive (and coding assistants so capable) that only those companies or individuals with access to the top-of-the-line assistants can succeed in the marketplace. My guess is that this is a temporary situation, though, during the interval when coding agents are vitally important but haven't yet automated the whole process. After that transition there may be other serious economic and distributional issues, but I don't expect them to look like 'the best coding assistants are too expensive'.
To be clear, I haven't done much thinking or research about either of those cases; these are just my thoughts at the moment.
In the meantime, it seems to me like there's still an abundance of affordable and increasingly effective coding tools for the easily-foreseeable future (FWIW I personally have been plenty happy with Claude Code on the $20/month plan).
My claim is that, in 2027, there will be AI coding tools which are much more expensive and somewhat more effective (or maybe much more effective) than the $20-$200 agents most people will use. The usefulness of the lower-priced tools will continue to improve; although I do not expect anything being priced lower than $8/month due to SaaS market dynamics.
I see. I agree that that seems likely as a descriptive claim. The introduction suggested a somewhat different tone / framing to me.
The tone is like this because I felt like "you cannot possibly do research as fast as the researchers in the labs because you don't have the autonomous RE coding agents, or the money to pay for those" is a world worth preparing for, for people interested in technical work on AI safety (be it researchers, grantmakers, even open source tinkerers). Maybe I can outline possible consequences of that state of the world in another post.