Azhar’s view of redesigning organizations around AI is a lot of what the Pope argues against recently (that AI should be designed around humans—or in his words, Christ).
Do you think that what AI is really showing us is that the coding/ writing/ reading that we think slows us down, is not actually what limits productivity? Or in operations speak, are LLMs just piling up WIP in front of the bottleneck?
@Karl Gantner - strange to say this - but I 100% agree with the Pope (at the level of “designing around humans”).
I think we look too much at the industrial revolution as an analogy for where our society is. I think you have to reflect on what humanity was emerging from at the time Watt spun up his first steam engine.
Life was often nasty, brutal, and short. Whatever you want to say about Watt - I think what motivated him and the other industrialists was to drive us out of mere survival. In that domain of experience, it makes sense that we organized around technology as the central design principal. The suffering was so abundant. Suffering in a factory is better than starving to death if the rains didn’t come.
But, like, that’s not where we are now. Our means are so great, we need new ideas.
I personally think life could be a lot more fun. And I think fun is a much purer and stronger motivator than survival. We should design the future around fun.
That has a fun downstream consequence. The best fun happens when everyone feels safe. So we need both things - safety and fun. The same thing in my mind.
“Do you think that what AI is really showing us is that the coding/ writing/ reading that we think slows us down, is not actually what limits productivity?”
I think LLMs are actually showing us the limits of myopic optimization. I think people over the last century have come to believe that the world is closed and determined. If that were true - then LLMs would be all you would need. But I don’t think it’s true. And I think this is a technical argument to that effect.
> Their generative value is in their unreliability. If you turn temperature down to zero, you get a deterministic machine
This is a not entirely correct take. A system can be reliable even when sampling is taking place. At the same time, sampling can be deterministic even with non zero temperature. And finally, you can run a model with zero temp in production.
Hi @empiko - you are almost certainly correct technically. My experience with LLMs is largely as a product manager and I built GPT-2 out of that book that walks you through it. I am not current on the way the models function in actual production.
By “break” I don’t mean “won’t function”. I mean “won’t deliver on their value proposition”. A functioning product is a necessary, but not a sufficient condition for technology to have utility. I would defend vigorously that the generative value in LLMs is derived from their unreliability. Which is what the argument ultimately rests on.
I’ll adjust the text in the post to make this clearer.
Great article.
Azhar’s view of redesigning organizations around AI is a lot of what the Pope argues against recently (that AI should be designed around humans—or in his words, Christ).
Do you think that what AI is really showing us is that the coding/ writing/ reading that we think slows us down, is not actually what limits productivity? Or in operations speak, are LLMs just piling up WIP in front of the bottleneck?
@Karl Gantner - strange to say this - but I 100% agree with the Pope (at the level of “designing around humans”).
I think we look too much at the industrial revolution as an analogy for where our society is. I think you have to reflect on what humanity was emerging from at the time Watt spun up his first steam engine.
Life was often nasty, brutal, and short. Whatever you want to say about Watt - I think what motivated him and the other industrialists was to drive us out of mere survival. In that domain of experience, it makes sense that we organized around technology as the central design principal. The suffering was so abundant. Suffering in a factory is better than starving to death if the rains didn’t come.
But, like, that’s not where we are now. Our means are so great, we need new ideas.
I personally think life could be a lot more fun. And I think fun is a much purer and stronger motivator than survival. We should design the future around fun.
That has a fun downstream consequence. The best fun happens when everyone feels safe. So we need both things - safety and fun. The same thing in my mind.
“Do you think that what AI is really showing us is that the coding/ writing/ reading that we think slows us down, is not actually what limits productivity?”
I think LLMs are actually showing us the limits of myopic optimization. I think people over the last century have come to believe that the world is closed and determined. If that were true - then LLMs would be all you would need. But I don’t think it’s true. And I think this is a technical argument to that effect.
> Their generative value is in their unreliability. If you turn temperature down to zero, you get a deterministic machine
This is a not entirely correct take. A system can be reliable even when sampling is taking place. At the same time, sampling can be deterministic even with non zero temperature. And finally, you can run a model with zero temp in production.
Hi @empiko - you are almost certainly correct technically. My experience with LLMs is largely as a product manager and I built GPT-2 out of that book that walks you through it. I am not current on the way the models function in actual production.
By “break” I don’t mean “won’t function”. I mean “won’t deliver on their value proposition”. A functioning product is a necessary, but not a sufficient condition for technology to have utility. I would defend vigorously that the generative value in LLMs is derived from their unreliability. Which is what the argument ultimately rests on.
I’ll adjust the text in the post to make this clearer.
Thank you for the feedback.