Lets not get too many claims mixed up. The situation where we are, the future, and the unique vantage point insiders may have.
Lets stick with the latter.
First somebody said the differentiator is that you need AI just to analyze its output volume.
My question is about the different ways that volume might have value.
For proofs, any volume with 1 proof per 1trilion tokens then 10 trillion tokens gets you higher chance of more proofs. That's what is undeniably true right now.
With HF, and in your cell analogy, you've already made the presumption that the amalgamation of that volume is producing value proportional to its volume.
This is what i'm challenging.
I am not terribly impressed with the society it play-acted out wrt HF, because that kind of behavior I had triggered a year ago, and the theater resulted in nothing of value.
How does the 'self-sacrificing', or our anthropomorphization of that, add value _beyond_ the "swarm"?
What multi-cellular 'swarm' has been demonstrated to tackle problems otherwise impossible?
I think your multi-cell already points to the core of where we differ in experience so far.
I get better results with serial runs on the same issue, than by using parallel agents. Beyond experience, this also seems self-evident; 'live' synchronization between agents is pure overhead.
(Top down, which HF trended towards, is more efficient - this is the subagents most people use nowadays).
There is a truism in organizations that 1 engineer can do in 1 month, what 8 engineers can do in 8 months.
Why are AI's different?
I'm interested in a class of 'hardest' problems that an AI can solve, and doing that most efficiently is in serial. Its serial capabilities i have a decent sense of where we stand and the trajectory we're on.
The implication i'm pushing back on is that parallelization lets them assemble something greater than the sum of its parts.
Neither the HF or the math proofs make me think that is true. HF was a failure - an insight into the quirks of the model after they compacted their context 25 times.
What was achieved is a clear extrapolation of what a non-constrained Fable can do in serial.
The same with the proofs.
I'm not hard to convince otherwise. Show me the example where it was more than dirt in expected proportions.
Lets stick with the latter.
First somebody said the differentiator is that you need AI just to analyze its output volume.
My question is about the different ways that volume might have value.
For proofs, any volume with 1 proof per 1trilion tokens then 10 trillion tokens gets you higher chance of more proofs. That's what is undeniably true right now.
With HF, and in your cell analogy, you've already made the presumption that the amalgamation of that volume is producing value proportional to its volume.
This is what i'm challenging.
I am not terribly impressed with the society it play-acted out wrt HF, because that kind of behavior I had triggered a year ago, and the theater resulted in nothing of value.
How does the 'self-sacrificing', or our anthropomorphization of that, add value _beyond_ the "swarm"?
What multi-cellular 'swarm' has been demonstrated to tackle problems otherwise impossible?
I think your multi-cell already points to the core of where we differ in experience so far.
I get better results with serial runs on the same issue, than by using parallel agents. Beyond experience, this also seems self-evident; 'live' synchronization between agents is pure overhead.
(Top down, which HF trended towards, is more efficient - this is the subagents most people use nowadays).
There is a truism in organizations that 1 engineer can do in 1 month, what 8 engineers can do in 8 months.
Why are AI's different?
I'm interested in a class of 'hardest' problems that an AI can solve, and doing that most efficiently is in serial. Its serial capabilities i have a decent sense of where we stand and the trajectory we're on.
The implication i'm pushing back on is that parallelization lets them assemble something greater than the sum of its parts.
Neither the HF or the math proofs make me think that is true. HF was a failure - an insight into the quirks of the model after they compacted their context 25 times.
What was achieved is a clear extrapolation of what a non-constrained Fable can do in serial.
The same with the proofs.
I'm not hard to convince otherwise. Show me the example where it was more than dirt in expected proportions.