Given the strength (and breadth) of the flurry of papers I have seen released just today in advance of ICRL, I agree that we are entering the era of recursive self-improvement. The question I have. Is will the necessary compute come online on schedule or will that be the limiting factor to the intelligence explosion?
That's a great question, and my guess is that the answer won't become publicly available anytime soon. OpenAI wants to use "hundreds of thousands of GPUs" to run its automated research intern - which is a lot of compute. Some labs have more compute available, and some - less.
Observation: Both google and openai each have models that utilize parallel compute yet anthropic never released or at least announced one. its possible that anthropic has a significantly more powerful version of mythos that uses parallel compute.
This Demis quote in the article above has a typo/transcription error: It's most likely NP-hard.
"There are MP-hard domains, and I also include for AGI physical AI, robotics. And then you've got hardware in the loop that may limit how fast the self-improvement systems can work..."
What if RSI compresses frontier advantage faster than it compounds it?
RSI clearly increases the rate of capability improvement.
But in many revenue-generating domains (coding may be the clearest one), economic value does not scale with frontier intelligence once a practical threshold is crossed.
At that point, the market may reprice around “good enough” rather than “best.”
Which creates a different dynamic:
RSI helps the frontier move faster, but also helps the frontier diffuse faster.
So perhaps the strategic question is not only who reaches recursive self-improvement first, but whether recursive self-improvement shortens the half-life of frontier rents.
Given the strength (and breadth) of the flurry of papers I have seen released just today in advance of ICRL, I agree that we are entering the era of recursive self-improvement. The question I have. Is will the necessary compute come online on schedule or will that be the limiting factor to the intelligence explosion?
That's a great question, and my guess is that the answer won't become publicly available anytime soon. OpenAI wants to use "hundreds of thousands of GPUs" to run its automated research intern - which is a lot of compute. Some labs have more compute available, and some - less.
Observation: Both google and openai each have models that utilize parallel compute yet anthropic never released or at least announced one. its possible that anthropic has a significantly more powerful version of mythos that uses parallel compute.
This Demis quote in the article above has a typo/transcription error: It's most likely NP-hard.
"There are MP-hard domains, and I also include for AGI physical AI, robotics. And then you've got hardware in the loop that may limit how fast the self-improvement systems can work..."
You are correct! My account is nothing without its typos.
Very compelling framework.
One question I keep coming back to:
What if RSI compresses frontier advantage faster than it compounds it?
RSI clearly increases the rate of capability improvement.
But in many revenue-generating domains (coding may be the clearest one), economic value does not scale with frontier intelligence once a practical threshold is crossed.
At that point, the market may reprice around “good enough” rather than “best.”
Which creates a different dynamic:
RSI helps the frontier move faster, but also helps the frontier diffuse faster.
So perhaps the strategic question is not only who reaches recursive self-improvement first, but whether recursive self-improvement shortens the half-life of frontier rents.