Herbie Bradley says memory bottlenecks limit actual AI speeds.
People sometimes confidently claim that humans would keep making major decisions even if AIs are >100x faster.
Imagine you could only chat with your boss on one day per year! a) it would be very clear to everyone that this is not workable b) you'd just make decisions around your boss and disempower them in order to get anything done.
I expect the situation with AIs will look comparable, especially if they are rewarded based on their outcomes.
Positive users endorse the claim that faster AIs will disempower human decision makers by praising illustrative blogs, while negative users reject the argument as obviously flawed in its framing.
yep
People sometimes confidently claim that humans would keep making major decisions even if AIs are >100x faster.
Imagine you could only chat with your boss on one day per year! a) it would be very clear to everyone that this is not workable b) you'd just make decisions around your boss and disempower them in order to get anything done.
I expect the situation with AIs will look comparable, especially if they are rewarded based on their outcomes.
@MariusHobbhahn Precisely! I really like this blog for illustrating this point - https://control-inversion.ai/3-tale-of-the-slow-mo-ceo/
A few anecdotes: - OpenMined's board meets 4x per year. They're my bosses and (imo) they do a great job. - Measuring the speed of intelligence is tough. If we just say "raw token output" as a reasonable/generic proxy. I type 100 words/minute. Claude code types around 6000 words/minute (60x faster than me). And I still spend 100:1 time waiting for claude code to finish doing what I asked it to do.
Tree based delegation of tasks is a highly powerful/scalable abstraction. It's probably ready to absorb AIs that are 100x faster.
People sometimes confidently claim that humans would keep making major decisions even if AIs are >100x faster.
Imagine you could only chat with your boss on one day per year! a) it would be very clear to everyone that this is not workable b) you'd just make decisions around your boss and disempower them in order to get anything done.
I expect the situation with AIs will look comparable, especially if they are rewarded based on their outcomes.
I'm sorry, but this is just such an obviously flawed argument I'm surprised it's still here
"one day per year" framing is one that by construction ignores compute bottlenecks or feedback loop speed! it's like looking at how fast the GPU kernel can calculate matmuls then extrapolating that to expected FLOPs with 0 consideration of the speed of data movement between VRAM, L1, and L2 cache, which obviously becomes the main bottleneck
People sometimes confidently claim that humans would keep making major decisions even if AIs are >100x faster.
Imagine you could only chat with your boss on one day per year! a) it would be very clear to everyone that this is not workable b) you'd just make decisions around your boss and disempower them in order to get anything done.
I expect the situation with AIs will look comparable, especially if they are rewarded based on their outcomes.
@MariusHobbhahn Consider a legislative body is responsible for regulating the activities of a 300,000,000 person economy. They certainly can’t micromanage, but they’re nonetheless responsible for major decisions.
@MariusHobbhahn boards meet ~4 times a year and orgs stay roughly aligned
@MariusHobbhahn But I do think there's a foundation of truth to what you're saying. There's some sort of... control vs speed tradeoff.
A few anecdotes: - OpenMined's board meets 4x per year. They're my bosses and (imo) they do a great job. - Measuring the speed of intelligence is tough. If we just say "raw token output" as a reasonable/generic proxy. I type 100 words/minute. Claude code types around 6000 words/minute (60x faster than me). And I still spend 100:1 time waiting for claude code to finish doing what I asked it to do.
Tree based delegation of tasks is a highly powerful/scalable abstraction. It's probably ready to absorb AIs that are 100x faster.
@MariusHobbhahn nah, if my boss only checked on me once a year, i’d just chill
@MariusHobbhahn Your argument seems to assume oversight means approving things in real time, but what about reviewing logs after the fact? Do you think speed asymmetry breaks that kind of oversight too?
@MariusHobbhahn And I think you're right that people will gamble here... and experiment with taking their hands off the wheel of some lower level decisions (in a way which could be dangerous) in exchange for a lot more output. It'll be a messy time.
@MariusHobbhahn But I do think there's a foundation of truth to what you're saying. There's some sort of... control vs speed tradeoff.
@pressed_tin @MariusHobbhahn There is a lot of weekly back and forth between board committee members and senior staff in functional orgs.
true but an architectural preference layer can built on top of the predictor in which humans provide feedback signals (at our slow glacial pace) to guide actual sys outputs and rewards from outputs (not prediction rewards); like solving general robotics with local and global human feedback on the output, path and future selection layers
@_mattneary @MariusHobbhahn they only *make* the rules, executive branch *enforces* it