4 ms·

But I can actually use an LLM to do useful work for me and no one gets hurt. I think there’s too much hype and money, but LLMs are useful right now.

Fricken 1y ago
SAE Level 2 automated vehicles are useful now, just not so useful that it's safe to take one's eyes off the road and hands off the wheel. LLMs are in a similar place, and still a long way from doing the whole job reliably by themselves. The current state of AI is one or several unknown unknowns away from real AGI. We're missing something fundamental. Likewise with autonomous vehicles, the mythical SAE level 5 vehicle that goes anywhere and everywhere there is a road is still very much science fiction.
- fellowniusmonk 1y ago
  I see you were very precise with what you wrote about SAE 2. I ride in waymo all the time, they are SAE 4. I know many people who take a waymo every day. What percentage of driving represents the delta between 4 and 5? SAE 4 level LLM/Ai (if we can really even make that comparison) would have far less difficulty in deployment and would be far more disruptive in a far shorter period of time than SAE 4 self driving cars.
  - Fricken 1y ago
    An L4 LLM in my mind, would be one that can perform fully autonomously at some domain specific job, a job being a collection of tasks that have to come together coherently to achieve a desired goal. Waymo's robotaxis, likewise, are able to do the driving task in geo-fenced areas. Waymo does a lot of hand-built code and testing to deal with particular problems, such as the 5-points in the Cairo district of SF, where it's a completely unique intersection, there's no other intersection quite like it. A ton of effort went into dealing with just that one intersection, and the bespoke effort doesn't generalize to any other intersection. So if you want to be able to say, have a platform for producing working video games out of prompts, well, I believe that can be done with our current AI, but it will depend on a lot hard work making tools and hand-built code that do not generalize to other domain specific jobs. Now if you want to make a movie worth watching out of prompts, that could be done too, but it depends on solving a whole different set of bespoke problems that once again need to be solved the hard way using more conventional software.

Fricken 1y ago

SAE Level 2 automated vehicles are useful now, just not so useful that it's safe to take one's eyes off the road and hands off the wheel. LLMs are in a similar place, and still a long way from doing the whole job reliably by themselves. The current state of AI is one or several unknown unknowns away from real AGI. We're missing something fundamental. Likewise with autonomous vehicles, the mythical SAE level 5 vehicle that goes anywhere and everywhere there is a road is still very much science fiction.

fellowniusmonk 1y ago
I see you were very precise with what you wrote about SAE 2. I ride in waymo all the time, they are SAE 4. I know many people who take a waymo every day. What percentage of driving represents the delta between 4 and 5? SAE 4 level LLM/Ai (if we can really even make that comparison) would have far less difficulty in deployment and would be far more disruptive in a far shorter period of time than SAE 4 self driving cars.
- Fricken 1y ago
  An L4 LLM in my mind, would be one that can perform fully autonomously at some domain specific job, a job being a collection of tasks that have to come together coherently to achieve a desired goal. Waymo's robotaxis, likewise, are able to do the driving task in geo-fenced areas. Waymo does a lot of hand-built code and testing to deal with particular problems, such as the 5-points in the Cairo district of SF, where it's a completely unique intersection, there's no other intersection quite like it. A ton of effort went into dealing with just that one intersection, and the bespoke effort doesn't generalize to any other intersection. So if you want to be able to say, have a platform for producing working video games out of prompts, well, I believe that can be done with our current AI, but it will depend on a lot hard work making tools and hand-built code that do not generalize to other domain specific jobs. Now if you want to make a movie worth watching out of prompts, that could be done too, but it depends on solving a whole different set of bespoke problems that once again need to be solved the hard way using more conventional software.