Is it really likely that a "recursive self-improvement" capability would lead to a great acceleration of AIs capabilities?

Isn't the preponderant bottleneck in improving the models the need to train them at scale to verify the hypotheses, and the time and cost that it takes?

Or does someone think that they could get magically able to predict big improvements without training?

I don't think anyone really knows.

I consider these scenarios:

1) We stumble onto an algorithmic improvement in intelligence. This isn't just "what humans do but faster", its "better than what humans do". I've got no idea what that might mean (it could be fundamentally different heuristics, it could be that we've got some intellectual blind spot that they cast off). It doesn't matter, the instant this happens AI is smarter than us and we won't be able to keep up. We're intelligencing at O(n^2) and they're doing O(n log(n)).

2) AI gets good enough at physics and engineering that they can really quickly use up all "the room at the bottom" as Feyman put it. They design and build a factory that produces a mystery metal amalgam that computes at some small percentage of the minimum predicted by the Landauer principle, within a few percent of Bremermann's limit. It's not "smarter" its just suddenly tens-of-orders of magnitude faster. But those orders of magnitude matter: there's only 8 billion of us, and there's plenty more than a factor of 10 billion "at the bottom".

3) It turns out that this is a "sum is greater than the parts" situation. No human can be an expert in all subjects, but we eventually build a big enough AI that it is. Turns out, you don't need extreme speed or different algorithms, just knowing everything all at once is enough to catapult AI dramatically beyond our grasp. Always knowing the best statistical test to apply, the best mathematical techniques, and relevant physics means that AI never makes a mistake, and can learn with maximum efficiency.

I built a self-learning recursive agent that finds academic research about using options data to trade, re-creates the research, and then probes and tests for gaps and potential strategies testing against over one year of out-of-sample trading data with one of several strategies that beat SPY by 10x. [0]

One rule is that if a position is opened using the historical data, it can't close the position until the next morning so it isn't a day trading strategy.

I'm curious how this self-learning recursive agent would have preformed in the past 4 months? I don't feel like shelling out $200 to access the data. Do you think that trading strategy will collapse? Whatever the case, if this agent really can perform like that and there isn't a look ahead bias leak in the backtesting (which is definitely a possibility or more likely what happened even though I spent days trying to harden against that), it is game over!

[0] https://github.com/adam-s/alphadidactic

The preponderant bottleneck is inventing new architectures to make AI actually good at human and superhuman tasks. For example, AI agent harnesses add tool calls and long term state management, allowing AI to autonomously complete complex tasks. Once these are in place, finetuning models with examples of good tool calls helps, but somebody first needed to invent the fundamental capability. Now try to implement humanlike long term memory for AI to be your coworker or life assistant working on tasks that last month. Even if necessary low level technologies are already there, structuring them to be practically useful is non trivial.

Humans gain knowledge through experiments. Without a physical body it has no chance of performing the same. That it can update it's training weights does not seem particularly significant.

Weird, there was a good comment about this but it vanished