Robot task costs fell 84% in two months

Transcript

Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.

1. 17:34

   So it's about a curve that should make every robotic founder, like, stand up and pay attention. So RoboCurve, it's a Y Combinator startup, designed a benchmark called RoboDojo RC Tier 1, which measures how well a model can drive a physical robot through actual subtasks. so the cost for a frontier model to complete a robot subtask fell from two dollars and 36 cents in july to 38 cents in september 84 drop so here's the curve uh and imad talk us through this i mean we've had kind of the question is the robot's body good enough and then it was like yeah they seem they can do just about anything and now the hands have got good another question is the mind and it's always been the case of do you need specialist models or can you use generalist models that suddenly learn these new capabilities this shows that generalist models

2. 18:25

   can actually win um and it's something we talked about previously with the auto self-driving kind of benchmark unitary ceo actually just came out and said like we're not even sure we need to create specialist models anymore because these things do it and we're used to kind of task and cost halving but every three weeks yeah like you're heading to the point whereby effectively there is no cost for robotic intellect and robotic locomotion and it comes down to the hardware cost which again is going to lead to an explosion just like we see again the unit trees when they first came out they were like stumbling about like little drunkards and now they're doing kung fu and synchronized kind of things next they're doing surgery with generalized models yeah well i mean again that's the conversation i had with elon uh at the beginning of the year you know his prediction

3. 19:12

   robots are doing perfect surgery within three years i mean look with the new grok model that they're coming out with they're putting all of spacex's engineering that can create flying cars right so the robot will be an engineer that can build flying cars yeah and the robot will be an amazing lawyer it'll be an amazing plumber it'll be an amazing etc so this is why does spacex tesla have customized robotics models no they're just like we will play the generalization game And again, I think that's what's going to win here because humans are generalized.

4. 19:45

   We're not like lobotomized and specialized. If you want to have humanoid-type robots, if you want to, again, have very specialist ones, sure. But humanoid robots, I think that's a different game. Yeah. We're going to be interviewing Brett in... In a couple of weeks, remember, he started building his own models for the robots because the early models didn't work, the large language models. But I'll be curious to ask him, you know, what he's using or building now.

5. 20:12

   But having a cost every three weeks means it effectively goes to zero, you know, within the year. So, you know, if you're an entrepreneur now in hardware space, what are you thinking? I think how can I build as many robots and actuators and the spare parts of the robotic supply chain as possible? Are you also thinking I need to collect as much training data as I can? I think that the training data will be important for a lot of the specialized scenarios.

6. 20:41

   And again, it was a case of will the robotics companies buy them? Brett has committed a billion dollars, that figure, for example, for data. Or will it be that after hoovering up the world's data from companies and all the text data, every single major foundation model company now wants all the robotics data. I think it'll be the latter. And robotics data is an order of magnitude larger than the language data. But again, if you're embodying the intelligences, you need that data.

7. 21:09

   Is the performance dependent on the form factor? This is just for human robots or bullet work on the three armed this one is for arms and humanoid type things in specific settings but again these should generalize very quickly like the whole thing is the language models got good enough all of a sudden that you kind of don't have to baby them anymore the robots aren't quite there but they're about to get there as well at the same time so all of these models, the video models now don't make mistakes, they're all getting good enough fast enough cheap enough at the same time that's the bottom line you know in the last episode we talked about jeff holden's atomic machines company which is basically going from prompt to product where you describe what you want and this machine is building it at millimeter

8. 21:58

   level accuracy with multiple materials so it's you know just to give a vision of where we're heading all of a sudden we're in a situation where if you want a specific kind of robot to do a specific you know you know clean windows repair hvac you know brick lay bricks mow the lawn you can ask it to design the perfect robot uh and build it and then operate it so do we end up with with star wars with like thousands of Different robot designs. Well, I think you're going to end up, remember that this works very well when you've got lots of repetitive tasks, right? And then you can generalize to, say, plumbing, where you've got lots of little manipulative tasks in a particular domain. I think it's clear that for now, it'll start off in the, what we call DDD, dull, dirty, dangerous jobs, toxic cleanups, and underwater manufacturing, and all sorts of things.

9. 22:56

   I think it'll be a while before it comes into the home, liability issues, et cetera. Again, we'll get to the institutional lag there. I think where I'm most excited is in things like surgery, where we make a lot of humanists. Is that dull, dangerous, or dirty? It's dangerous for the patient. If you're the first 10 heart surgeries that a doctor has ever done, you kind of don't want to be there. You kind of want to be in the 500.

10. 23:21

    And you don't want to be too far down the line either because the doctor gets lazy after a few thousand surgeries of the same time and doesn't pay attention, right? And so now you can smooth that curve out and have the world's like literally the best doctor first supporting and then over time just doing the surgeries. And that becomes completely, you want those things as automated as possible, operating in the most precision level measurements possible.

11. 23:43

    I think that will see a huge take up in those types of tasks for robots. I'm fond of saying if you're looking, if you need a surgery, there's one thing you ask your surgeon, which is how many times did you do it this morning, right? So the surgeon who's seen the most varied cases is most likely to be the most successful. And when you've got federated learning across all robots, they've seen millions of cases, and they share all those memories.

12. 24:09

    What I'm most excited about is the robot chef. Okay. Can you just imagine that?