Elon Musk’s Approach to AI Could Make His Robotaxi a Long Shot

[00:00:00] How do airplanes fly? What's in this box? What does this thing do?

[00:00:05] Kids are curious about everything, including guns. Learn how to store your gun securely and make your home safer at nfamilyfire.org. Brought to you by N Family Fire, Brady and the Ad Council.

[00:00:19] Welcome to Tech News Briefing. It's Monday, November 4th. I'm Julie Chang for The Wall Street Journal.

[00:00:24] If you're tired of searching through endless recipe sites looking for creative dinner ideas, maybe an artificial intelligence chatbot can help.

[00:00:33] That's what one WSJ editor did when planning meals for his toddler.

[00:00:38] And Tesla says it plans to launch its robo-taxis in the coming years. But how ready is it really for primetime?

[00:00:46] Art Tech columnist Christopher Mim says CEO Elon Musk is going about it all wrong.

[00:00:52] Those stories, coming up.

[00:00:57] AI chatbots can do lots of things, from writing emails to creating code.

[00:01:02] For parents, they can also be used to outsource certain tasks, like meal planning for kids.

[00:01:08] That's exactly what WSJ AI editor Ben Fritz has been doing for the past few months.

[00:01:14] He's here to tell us about the upsides and downsides to tapping Anthropik's Claude chatbot as a helping hand in the kitchen.

[00:01:21] Ben, take us through the steps here. You opened up Claude, and then what?

[00:01:26] I'm going to Claude, and you start off, I really found with AI, the best thing to do is talk to it almost like you're talking to a person.

[00:01:31] So I didn't have to give it a very specific instruction, like, you know, this is exactly what I need.

[00:01:37] And here's step by step.

[00:01:38] I started off by saying, like, hey, it's a struggle for me to come up with meals to cook for my daughter every week.

[00:01:43] Is that someone you can help with? And it immediately said, yes, I'd be happy to help design a meal plan for an 18-month-old.

[00:01:51] And then it started giving me ideas right away. And I was like, okay, well, that's great.

[00:01:55] And then I said, let me give you the specific kind of guidelines for our family as far as what kind of meals you want to give her and what kind of ingredients we like to use.

[00:02:03] And I said, we're trying to give her a nutritious variety of meals, and I want stuff that I can cook on the weekends and freeze that I can heat up during the week.

[00:02:10] And it just throws ideas at me.

[00:02:13] It gave me a baked salmon and risotto cup.

[00:02:16] It's like I sort of made mixed salmon and risotto in like a muffin tin and baked them.

[00:02:21] And that was obviously certainly not something I would have ever come up with my own.

[00:02:25] And my daughter really likes it.

[00:02:26] It gave me a bunch of different kinds of like healthy pancakes I can make for her in the morning, like almond flour, blueberry, zucchini pancakes.

[00:02:36] So those are two things that I've like made recently that have gone really well and were easy to freeze and reheat.

[00:02:40] You wrote that you got more specific with your chatbot queries. How so?

[00:02:45] At times I'd be like, hey, I've given her a lot of meals with fish recently.

[00:02:48] Could you give me meals that don't have that in it?

[00:02:50] Or sometimes I would be like, I can't go shopping.

[00:02:53] I need to make something now.

[00:02:55] Here are some ingredients I have in my refrigerator or my pantry.

[00:02:58] Can you come up with anything to make with it?

[00:03:00] And I would just start listing ingredients.

[00:03:01] And then it would give me ideas based just on that.

[00:03:04] Or the other day it was the morning and I had to make lunch for my daughter in a hurry.

[00:03:08] I only had like 10 minutes.

[00:03:09] I was like, give me something fast and easy.

[00:03:12] And here's like a bunch of ingredients.

[00:03:14] And it came up with stuff.

[00:03:16] So this all sounds very helpful.

[00:03:17] But are there any downsides to using AI in this way?

[00:03:21] Sure.

[00:03:22] I mean, look, AI makes mistakes.

[00:03:23] So if you have allergies or really important dietary concerns, you have to double check.

[00:03:28] It will sometimes give you a recipe that might have nuts in it.

[00:03:32] If your kid can't eat nuts, for example.

[00:03:34] Or if you keep kosher, you have to be cautious about that stuff, I would say.

[00:03:38] Sometimes it comes up with, for lack of a better term, weird stuff.

[00:03:42] I'm like, I can't even believe it thinks this is a recipe.

[00:03:44] It was like sardine and oatmeal cake or something.

[00:03:48] You mix them together.

[00:03:49] I was pretty confident that would be gross.

[00:03:51] I didn't even make it.

[00:03:52] So Ben, you've been using AI for a few months now to plan meals.

[00:03:55] Did you have any takeaways?

[00:03:57] My takeaway is that meal planning is something that takes a lot of mental bandwidth.

[00:04:02] I'm a fine cook.

[00:04:03] It's not that I can't cook.

[00:04:04] It's that I don't like to spend the time doing the research and being like, what am I going to make?

[00:04:07] What have I not made recently?

[00:04:09] What is my shopping list going to be?

[00:04:10] And there are a lot of tasks like this in our lives that just take mental energy, but they're not like that creative.

[00:04:18] They don't feel like they're that high level and important.

[00:04:20] And these are things that, in my opinion, it's a really great thing to outsource to AI.

[00:04:23] I sort of thought of it as like, if I was wealthy and I had an assistant, I would have my assistant do this for me.

[00:04:29] That was WSJ AI editor Ben Fritz.

[00:04:32] After the break, how Elon Musk's robo-taxi AI approach differs from other major players in the self-driving world.

[00:04:39] Stay tuned.

[00:04:50] Tesla says it plans to roll out its fully autonomous robo-taxis in the next few years.

[00:04:56] On a recent earnings call, CEO Elon Musk said that large-scale production of the cyber cab would happen in 2026.

[00:05:04] But WSJ tech columnist Christopher Mim says it's a long shot.

[00:05:08] He says there are many reasons to think Musk's approach to self-driving and AI won't work.

[00:05:14] Christopher is with me now to explain further.

[00:05:17] So by now, listeners may know how Tesla's robo-taxis differ from others, such as Waymo and Cruise.

[00:05:25] Instead of using LiDAR technology, that's a kind of laser-based radar that a lot of self-driving car technology is based on,

[00:05:32] Tesla relies heavily on cameras.

[00:05:35] But Christopher, you wrote that Musk's plans center on what he has called end-to-end artificial intelligence.

[00:05:41] How does AI fit into the picture here?

[00:05:44] So Elon Musk's big idea here is that Tesla has millions of hours of recordings of driving from the cameras on every Tesla.

[00:05:56] And that if they can just shove all of that data into a big enough AI supercomputer using modern techniques,

[00:06:05] it will sort out how to drive more or less on its own.

[00:06:09] So the whole idea here is scale.

[00:06:11] Just show it enough data, enough hours of driving, and it's going to learn how to drive.

[00:06:16] Got it.

[00:06:17] And competitors like Waymo and Cruise are also using AI.

[00:06:20] How are they using it?

[00:06:22] They use it in a way that is much more modular in the sense that they have created self-driving systems,

[00:06:30] which initially most of their components were coded by humans.

[00:06:34] It was a bunch of rules.

[00:06:35] If you see this kind of scenario, do this.

[00:06:38] Don't hit the bicyclist, that sort of thing.

[00:06:40] And then gradually they have incorporated AI into every part of that process that starts with perception

[00:06:46] and ends with the vehicle actually taking an action.

[00:06:49] And because they have broken up the problem into basically chunks, there is just a lot more transparency for them.

[00:06:58] If their vehicle does something unexpected, the safety driver behind the wheel can log an exception

[00:07:05] and then the engineers can go and try to correct that.

[00:07:08] Whereas with Tesla, it is much more with this end-to-end AI, a black box.

[00:07:14] It's just a huge spaghetti of neural connections, digital, artificial neural connections.

[00:07:20] And so when it does something strange or undesirable, it's hard to know why it did that.

[00:07:26] So the companies are taking very different approaches here.

[00:07:29] Musk envisions an AI system that learns by watching people drive.

[00:07:34] Waymo and other companies are teaching their vehicles by correcting them as they do the driving themselves.

[00:07:40] How do those approaches impact results?

[00:07:43] Tesla's approach, which computer scientists call imitation learning, seems to yield a system that it's almost like chat GPT.

[00:07:53] There are scenarios in which you can do amazing things, but then sometimes it hallucinates.

[00:07:58] And you really don't want a self-driving system to hallucinate in a literal sense.

[00:08:04] So the result has been, you know, it's all over social media.

[00:08:08] Every time there's a new rollout of the latest version of what Tesla calls full self-driving supervised,

[00:08:16] social media fills with videos of Tesla drivers finding their vehicle trying to turn into oncoming traffic

[00:08:25] or not stopping for a train because there's fog or getting blinded by the sun because it only uses cameras to see what's outside of it.

[00:08:35] And this doesn't really seem to be getting better.

[00:08:38] Tesla claims, according to their internal data, that already human driving with its full self-driving system is safer than a human driving on their own.

[00:08:47] But independent evaluations have found that's really not the case.

[00:08:52] What does the research say about how effective each of the company's approaches are?

[00:08:56] The so-called imitation learning that Tesla seems to be banking on in simulations have found that those systems can get into trouble

[00:09:05] when they get too far outside of the data that they've been trained on.

[00:09:09] And it's not hard for them to do that.

[00:09:11] So that's a big issue with their approach.

[00:09:15] Whereas the reinforcement learning that Waymo and others are relying on, that kind of data is harder to gather.

[00:09:23] It's more labor intensive.

[00:09:24] It involves driving thousands and thousands of miles with safety drivers behind the wheel and feeding that data back to engineers who then tweak the system.

[00:09:33] But research suggests that that yields a system that is more transparent and safer.

[00:09:40] And most important of all, a system that can gradually improve over time.

[00:09:46] So the number of accidents that it gets into or disengagements that occur, you can continually shrink that amount.

[00:09:55] Whereas with Tesla's approach, it's not clear.

[00:09:57] It seems like it's kind of a random walk.

[00:09:58] It's the system from one version to the next.

[00:10:03] New, strange features of its behavior crop up kind of inexplicably.

[00:10:08] And we should note Tesla didn't respond to several requests for comment on its self-driving systems.

[00:10:13] Okay, so Christopher, you've been following the story.

[00:10:16] What will it take for Musk's robo-taxi vision to come true?

[00:10:20] Engineers who take Musk's approach seriously say that if we're going to be able to create a self-driving system that can learn just by observing humans driving, we might actually need a breakthrough in AI technology.

[00:10:37] We might need something that is closer to human level intelligence for it to be able to learn entirely through observation and then transfer that learning to the real world.

[00:10:49] And it's not clear that we are close to any such breakthrough.

[00:10:52] That was WSJ tech columnist Christopher Mims.

[00:10:56] And that's it for Tech News Briefing.

[00:10:58] Today's show was produced by me, Julie Chang, with supervising producer, Catherine Millsop.

[00:11:03] We'll be back this afternoon with TNB Tech Minute.

[00:11:06] Thanks for listening.