Depending on how you measure "improvement" they already have or they never will :-/
Measuring capability of the model as a ratio of context length, you reach the limits at around 300k-400k tokens of context; after that you have diminishing returns. We passed this point.
Measuring capability purely by output, smarter harnesses in the future may unlock even more improvements in outputs; basically a twist on the "Sufficiently Smart Compiler" (https://wiki.c2.com/?SufficientlySmartCompiler=)
That's the two extremes but there's more on the spectrum in between.
you can also execute larger tasks than this using subagents to divide the work so each segment doesn’t exceed the usable context window. i regular execute tasks that require hundreds of subagents, for example.
in practice the context window is effectively unlimited or at least exceptionally high — 100m+ tokens. it just requires you to structure the work so it can be done effectively — not so dissimilar to what you would do for a person
How to organize code like you said, and how agents interact with it, to keep the actual context window small is the fundamental challenge.
I looked at that response by GP (rgbrenner) and refrained from replying because if someone is both running hundreds of agents at a time AND oblivious to what "context window" means, there is no possible sane discourse that would result from any engagement.
Doesn't change my point: the amount of code the agent can operate on is very large, if not unlimited, as long as you put even a little bit of thought into structuring things so it can be divided along a boundary.
If you let the codebase degrade into spaghetti, then the LLM is going to have the same problem any engineer would have with that. The rules for good code didn't disappear.
It's like like if your context window with one agent is n, your context window with 10 agents is n/10. It is some skill, but that is also where a lot of the advances are coming in.
Assistant: “I propose A”
User: “Actually B is better”
Assistant: “you’re absolutely right”
User: “actually let’s go with C”
Assistant: “Good choice, reasons”
User: “wait A is better”
Assistant: “Great decision!”
They certainly have. The only impactful improvement in the last year is "just run it in a loop until it gets it right" lmao
Which, of course, only works as long as the costs are subsidized by the companies vying for market share.
If LLMs stop improving at the pace of the last few years (I believe they already are slowing down) then they will still manage to crank out billions lines of code which they themselves won’t be able to grep and reason through, leading to drop in quality and lost revenue for the companies that choose to go all-in with LLMs.
But let’s be realistic - modern LLMs are still a great and useful tool when used properly so they will stay. Our goal will be to keep them on track and reduce the negative impact of hallucinations.
As a result software industry will move away from large complex interconnected systems that have millions of features but only a few of them actively used, to small high quality targeted tools. Because their work will be easier to verify and to control the side effects.