> Maybe in the future but certainly no evidence of this anytime soon
Here's some anecdotal evidence from me - I cleaned up multiple GPT 4.x era vibecoded projects recently with the latest claude model and integrated one of those into a fairly large open source codebase.
This is something AI completely failed at last year.
Maybe you should try something like this or listen to success stories before claiming 'certainly no evidence' in future?
There are untold billions of dollars to be had if you can make this future come to pass. You don't need AGI to make it happen either. You just need to keep making the context windows bigger and keep coming up with updated training data. It's not the outcome I want, but it really does feel within reach. The only limiting factor is going to be token count and cost to process/generate those tokens. But if you don't particularly care about quality, costs are going to have to go up by several orders of magnitude before you start to regret firing your software engineers.
I don't know what happens in a decade when there are no junior engineers, skilled senior engineers are becoming rare, and the only data left the train LLMs on is 200th-generation slop. But AI slop being qualitatively slop is not enough of a obstacle to prevent that future from coming to pass. And billions of dollars will be "saved" along the way.
Companies are already putting billions out there just to secure and produce training data. And that's the isseue; spending X billions to make X-Y) billions isn't a profit, it's a gamble hoping Y becomes negative (or at least close to zero with a commodity that is profitable) . Real profits have not been made directly from the work on AI as of now. It's made from marketing a narrative of AI working.
That's what makes this whole house of cards dangerous. The prescription to psychosis is profitable. Aka, selling a grift.
I have personally had success telling Claude that some AI-written system is too complicated and ask it to rewrite it in a more logical way. This sometimes results in thousands of lines of code being deleted. I give an instruction like that if I see certain red flags, eg:
1) same business logic implemented in two different places, with extra code to sync between them
2) fixing apparently simple bugs results in lots of new code being written
It’s a sign I need to at least temporarily dedicate more effort to overseeing work in that area.
I somewhat agree with the AI psychosis framing of the OP. It takes some taste and discipline to avoid letting things dissolve into complete slop.
I'm no expert, but the skeptic's opinion I've heard would be to ask:
What evidence is there that we're not at or close to a plateau of what LLMs are capable of? How do you know the growth rate from 2023 to present will continue into 2029? eg. Is it more training data? More GPUs? What if we're kind of reaching the limits of those things already?
Ultimately, you are describing a fundamental problem with induction -- Hume's problem of induction to be specific. How can we know that anything that has been shown empirically in the past will continue to be true - we can't. Best to investigate mechanistically:
I don't see why we would assume that we are at a plateau for RL. In many other settings, Go for instance, RL continues to scale until you reach compute limits. Some things are more easily RL'd than others, but ultimately this largely unlocks data. We are not yet compute/energy/physical world constrained. I think you would start observing clear changes in the world around you before that becomes a true bottleneck. Regardless, currently the vast majority of compute is used for inference not training so the compute overhang is large.
Assuming that we plateau at {insert current moment} seems wishful and I've already had this conversation any number of times on this exact forum at every level of capability [3.5, 4, o1, o3, 4.6/5.5, mythos] from Nov 2022 onwards.
I think we're close to the plateau of what LLMs can do, but they will keep improving. IMHO the results are already showing diminishing returns.
The (leading) LLMs work by consensus, like Wikipedia, Openstreetmap, web search engine or opensource movement.
What I mean is if I ask LLM "create a linked list", its understanding (of what I want) is already close to the expected ideal. Just like Wikipedia article on linked list, for example.
But the LLMs will continue to improve in breath and depth of understanding the world, although technically (what they CAN do) they probably already peaked. Similarly, OSS movement technically peaked in the 90s with the creation of compiler, operating system and a database; doesn't mean that new opensource isn't being created.
There is so much money at stake, and so much money pouring into AI development, that I think we are going to continue to see gains for a while. People keep coming up with new agent harness techniques like chain of thought, tool calling, and memories. And then the big LLM companies figure out how to actually train their models to optimize the use of those techniques. To claim that we are reaching the top of the plateau is to claim that we are out of effective ideas for improvement. I think that's a ridiculous claim, the technology is too new. And because of the strong incentives to keep making these things better, it's pretty much a given that people will continue to explore ideas until we really are out of effective ideas. I don't think anyone apart from professional AI researchers have any idea where this is all going to settle.
Well depends what you mean by peak. I was answering parent's question of what LLM's CAN do. It's not about peak of technology or humanity itself.
LLMs (or specifically GPT algorithm) are 8 years old. It has matured as a technology. I am not sure how you imagine it being significantly improved, from a user point of view, without some kind of paradigm shift (i.e. something significantly different from GPT or LLM).
Although I can imagine one important social innovation yet to come - a generally available big public LLM, that "anybody can train". We had a technology of "encyclopedia" for years (famously Brittanica); yet the concept of Wikipedia has been a truly new take on encyclopedia.
Also, new kinds of AI might emerge - for example we might formalize all types of human reasoning and build a reasoning AI, as well a model of human language, from scratch rather by training via GPT (and thus, more understandable and potentially smaller). But that won't be an LLM.
One major axis on which LLMs could get better is energy (and in general material) efficiency, doing the same stuff they do now but with fewer inputs. I actually fear that we are very early on this curve. The period of time between electric arc lamps and the current state of affairs where electric light is almost free was more than a hundred years. Lots and lots of investment is taking place under the assumption that LLMs will ride that curve lots lots faster. If it doesn't -- or if there is some physical law which means we're already close against some asymptotic limit -- then we're talking about a generational misinvestment, and that's only one of the underlying but potentially false assumptions of the AI investment boom.
> I am not sure how you imagine it being significantly improved, from a user point of view, without some kind of paradigm shift
I proposed how. New harness techniques and new training data/techniques, so the harness gets better and the LLM can be trained to work better with the harness. There's no reason to believe we're out of momentum for improvement in that direction.
Since we're not experts, we treat it as a black box. What are the results? Is the quality of the results improving? Is the improvement accelerating or decelerating?
And the answer appears to be that the improvement is accelerating. So how could it be stopping?
I don’t think improvement is accelerating. We went from “computers can’t do these things at all” to “now they can” in a few years with the discovery of transformers, and now we get “it can do the same things, except incrementally better, at a drastically higher cost” every few months.
I don’t think that the current AI paradigm has infinite headroom for improvement, similar to how every other AI approach before it eventually hit a limit.
Incrementally, higher cost? A model I'm running on a 10 year old entry level computer is better at programming than GPT4. Those are multiple orders of magnitude of improvement in a few years.
And the link I posted shows the amount of work a query can do increasing non linearly. You can explore the site for more detail and a graph that shows error rates getting halved every couple of months.
No one said anything about infinite. It doesn't mean we don't have headroom to spare.
Software itself took 80-120 years to get where it is today depending on how you count. Time is on AIs side here.
AI is currently, actively getting better. That might stop in a year, and you can argue whether it’s linear or exponential, but it’s very difficult to argue that it won’t get better in the near term. On the other hand, arguing that we hit a plateau at this time is just ignoring reality.
Here's some anecdotal evidence from me - I cleaned up multiple GPT 4.x era vibecoded projects recently with the latest claude model and integrated one of those into a fairly large open source codebase.
This is something AI completely failed at last year.
Maybe you should try something like this or listen to success stories before claiming 'certainly no evidence' in future?