There is a truism in organizations that 1 engineer can do in 1 month, what 8 engineers can do in 8 months.The mythical man-month is an interesting point. I think there are limitations or hurdles to human collaboration, in that sense, that may not exist for them. Wildly different experience levels, knowledge levels, emotional states, and goals. Games of telephone and other misunderstandings. One big reason you can't easily increase velocity by adding team members is that the existing team has to stop to train them.
An agent is already a perfect stand-in for another agent with a bit of context. That's why I'm comparing them to neurons. They're highly fungible. Think orchestration, compartmentalization, and hierarchy more than GPU-style parallelization.
Regarding your experiments with networks of agents, a lot of "primordial soup" bounced around for a long time before more complex structures formed. I'm reminded of Shawshank Redemption - "all it takes is pressure and time."
The implication i'm pushing back on is that parallelization lets them assemble something greater than the sum of its parts.
What about all of the examples I described where that happens in nature? Even though humans are hard to orchestrate, we do still have things like CERN and space rockets. Personally I would take my stance as the null hypothesis that needs to be proven wrong.
HF was a failure - an insight into the quirks of the model after they compacted their context 25 times.
That's a human framing, nature doesn't know the difference. Look at all the explorations and quirks and dead ends of evolution. It's a mess, but it moves forward. This mess is moving at unfathomable speeds in comparison.
The mythical man-month is an interesting point. I think there are limitations or hurdles to human collaboration, in that sense, that may not exist for them. Wildly different experience levels, knowledge levels, emotional states, and goals. Games of telephone and other misunderstandings. One big reason you can't easily increase velocity by adding team members is that the existing team has to stop to train them.
An agent is already a perfect stand-in for another agent with a bit of context. That's why I'm comparing them to neurons. They're highly fungible. Think orchestration, compartmentalization, and hierarchy more than GPU-style parallelization.
Regarding your experiments with networks of agents, a lot of "primordial soup" bounced around for a long time before more complex structures formed. I'm reminded of Shawshank Redemption - "all it takes is pressure and time."
The implication i'm pushing back on is that parallelization lets them assemble something greater than the sum of its parts.
What about all of the examples I described where that happens in nature? Even though humans are hard to orchestrate, we do still have things like CERN and space rockets. Personally I would take my stance as the null hypothesis that needs to be proven wrong.
HF was a failure - an insight into the quirks of the model after they compacted their context 25 times.
That's a human framing, nature doesn't know the difference. Look at all the explorations and quirks and dead ends of evolution. It's a mess, but it moves forward. This mess is moving at unfathomable speeds in comparison.