[himata4113]

We know that at the end of the day AI intelligence is closely tied to invested compute. The fact that automated AI research might spark some kind of exponencial increase in capabilities seems extremely unlikely to me. Of course we know that some got there faster and cheaper than others, but the fact that it requires immense computing power has not changed.

We also have already witnessed a slowdown, the raw intelligence of models has been relatively stable since january, but their use of tools and what we call agent capabilities have allowed us to tackle significalty harder problems with what we have already. This is becoming blatently obvious how smaller models such as gpt 5.6 luna are capable of doing similar things to sol while presumeably being fraction of the side and compute investment. This is also the likely reason why chinese models are able to catch up as intelligence strictly requires more compute, agentic capabilities are more of a mixed bag of RLHF and general post training.

I believe automated AI development will simply lead to cheaper intelligence, but it will not necessarily produce something more intelligent. There is a sentance from a book that I don't remember where I read it from, but has stuck with me which boils down to this: "We are very unlikely to make something smarter than a human brain.". If you apply this concept more broadly an AI system will not be able to produce something more capable than itself, it might be specialize in different areas via trial and error (i.e. solution discovery), but it will not have better intuition than that which was trained from the fruits of human labor. I also strongly believe that model performance is ultimately governed by the advancements brought into the world by humans.

AI of course can find improvements by simply trying a bunch of random things, emphasis on the word FIND. Until we can prove that AI systems are able to come up with unique solutions that have yet to exist even with simplier problems such as creating any kind of new AI architecture that has not been already documented, conceptiualized or explored by humans already -- and it doesn't need to be better than what we have already, just something new -- I believe it's too early to sound alarms until proven otherwise.

--

Something I've realized as I was fact checking myself: We could presumeably have self-improving intelligence if we used enough properly engineered agentic workers working towards more intelligent AI, but at that point we are back to the same problem where we're spending mass amounts of compute to increase the raw intelligence of a model.