Was Elon's Tesla FSD Demo the ChatGPT Moment For Self-Driving Cars?

0:00Today on the AI Breakdown, we're asking whether Elon Musk's Tesla full self-driving demo from Friday represented the chat GPT or GPT-4 moment for self-driving cars. Before that on the brief, a look at why Wall Street says AI hype might not be as dead as mainstream media might want to suggest. The AI Breakdown is a daily podcast and video about the most important news and discussions in AI. Go to breakdown.network for more information about our YouTube, our Discord, and our newsletter. Welcome back to the AI Breakdown Brief, all the AI headline news you need in around five minutes. If you are a regular AI breakdown listener, you will know that over the course of the summer, there was a bit of a cooling, an ebbing, if you will, of the hype around AI.

0:49Now, as we discussed, there was, I think, an open question around the extent to which this AI hype cooling was actually AI hype cooling versus just a convenient counter-narrative as the next thing for media to talk about as related to AI. And at least according to Wall Street, the hype is certainly not gone. Not by a long stretch. We are just coming off second quarter earnings season, and the discussions around AI were a dominant theme across a huge array of companies, not just big tech companies. According to the Washington Post, more than 1 ,000 companies mentioned AI in their quarterly reports.

1:29Now, that number is fairly similar to last year's number, but about 50 % higher than the period between 2020 and 2021, and up from just over 30 a decade ago. Now, of course, the highlight company in this is NVIDIA. NVIDIA's second quarter earnings report, even with incredibly inflated expectations, still blew those expectations out of the water. It was called the earnings guidance heard around the world, and was even called a 1995 internet moment. What the Washington Post points out, however, is that it's far from just tech companies or infrastructure companies that are talking about artificial intelligence.

2:03Some examples they give, Fidelity talking about the technology as a way to help detect fraud, Alaska Air using it to find more fuel-efficient flight paths, medical companies like Hologic using it to identify certain conditions, in this case, precancerous lesions, the owner of KFC and Pizza Hut using it to better connect online orders with brick-and-mortar stores, and Ulta Beauty using it to power its, quote, virtual try-on and skin analysis tools. Overall, they say one in seven public companies talked about AI in their most recent filings. Now, that said, some companies are mentioning it not just as something to get investors excited, but also in the part of the report that deals with coming risks.

2:41Companies like Adobe and Zoom noted that regulation of AI could disrupt their business models. William Sonoma cited intellectual property risk. Now, we will have some tests upcoming of just how far hype gets a company on Wall Street. On Friday, The Verge published a piece called Arm's IPO will tell us how much AI hype matters. The subheader reads, SoftBank is hyping the AI potential, but its public filing shows a slowing mobile market. Now, Arm, which I always used to refer to as ARM, but apparently people just refer to like the body part, is a company that designs chips and licenses the designs to other people who are actually building the chips.

3:18So as The Verge writes, Arms IP is licensed by companies such as Apple, Qualcomm, and NVIDIA, which use Arms blueprints to design and fabricate their chips. Now, the reason to be concerned about this IPO, according to The Verge, is that demand for mobile devices, which has long been Arms bread and butter, has been slowing somewhat. For example, according to the filing, the company's revenue fell 1 % in the fiscal year that ended on March 31, 2023. What's more, its net income for the quarter that ended in June was less than half of last year's. At the same time, there is growing interest among many of the companies that already work with Arm, such as Apple, Amazon, Google, to custom fabricate their own AI-specialized chips.

3:58So in many ways, you have what is set up to be a referendum from investors on whether the slowing growth in the mobile market is more important to the company's destiny than the growing potential for its AI business. Another company with growing Wall Street notice is South Korea's SK Hynix. The Wall Street Journal this weekend wrote a feature piece called This Company is NVIDIA's AI Chip Partner, and its stock is soaring. The WSJ writes, The hardware powering the current artificial intelligence craze is most closely linked with NVIDIA, but packaged alongside NVIDIA's Brainy H100 processors are specialized memory chips that enable the mind-boggling number of near-instantaneous computations behind AI applications.

4:35SK Hynix is the main provider of the latest high-bandwidth memory chip for NVIDIA's top-line AI processor chip. The WSJ goes on, Despite a severe downturn in the broader memory chip world, caused in part by slumping sales of smartphones and computers, SK Hynix's stock price has risen by almost 60 % since the start of the year. You'll notice one of the things that I frequently say is that what's most interesting to me about an article is not so much what's contained within it, but the fact that it exists at all, and this is a great example of that. On any given day, if you search AI on Google, for example, a lot of the top results are going to be around AI-related stocks, stock picks, stock predictions, And I think in some ways, this article is an example of the sophisticated highbrow Wall Street Journal version of this, which is identifying a company that is perhaps little known or at least less known than some of its larger peers and giving it big exposition as opposed to, for example, Motley Fool stock picks or something like that.

5:28But really, I think it's part of the same consumptive impulse for investors who are coming to grips with the AI trend, understanding which companies might help them actually profit from it. Now for the back half of this brief today, I wanted to look at two sets of studies that I thought were interesting. One comes from the global small business platform Xero, who surveyed over 3 ,000 small business owners from countries including the US, Australia, Canada, New Zealand, Singapore, and the UK. Now on the concern side, small businesses seem to be most concerned around sensitive information disclosure and data privacy violations, each having 41 % of those surveyed say they're concerned with those things.

6:02That's followed in third place by worker displacement, with 38 % of small businesses calling it the biggest ethical challenge of AI. At the same time as a crowdfund insider, Summary Peace says, Data privacy concerns don't reflect actions. Only 32 % of the small businesses surveyed aren't taking any proactive steps, while the majority are either experimenting or investing or working with third-party vendors. Interestingly, right now, 51 % of small businesses surveyed said they trust AI with identifiable customer information, while 45 % say they trust AI with their sensitive commercial information.

6:33In terms of the big question about its impact on jobs, 14 % of the businesses who are currently using generative AI have already seen a reduced headcount. In terms of overall sentiment, there is a pretty even split, with 30 % reporting being excited, 32 % reported being intrigued, and 31 % feeling anxious. Now, the second survey that I want to discuss comes from Salesforce. It wasn't only about AI, but AI was one part of it. The report was its sixth annual State of the Connected Customers report that surveyed 11 ,000 consumers and 3 ,300 business buyers to get a feel for some big questions such as how changes in inflation in the economy are impacting buying decisions and in general what people think about business performance among the companies that they do business with regularly.

7:17The headline AI stats, 68 % of respondents said that advances in AI make trust even more important. However, only 51 % of consumers said they trust companies in general, and only 45 % of consumers trust companies to use AI ethically. Now, a last metanarrative note before we head out. The title of this piece in Fast Company was Salesforce, a surprising number of consumers trust companies to use AI ethically. But then that seems to be contradicted by the body paragraph where they frame it as when it comes to AI, only 45 % of consumers trust companies to use AI ethically. Perhaps this is some sort of weird reflection of media not really being able to make up its mind about how it wants to interpret data, but then again, it could just be not the best editing.

7:59In either case, that is going to do it for today's AI Breakdown Brief. I'll be back soon with the main AI breakdown. Welcome back to the AI Breakdown. The video that you are watching was live streamed by Elon Musk on Friday evening. And what it is, is Elon and a Tesla showing off their new full self-driving mode. Now what makes this version of full self-driving or FSD different than some previous iterations of this software powering the Tesla is that unlike previous models that use specific software with specific logic embedded in code to tell the car what to do in specific situations, this version is powered by artificial intelligence.

8:40It has been fed, in other words, a huge, huge amount of video data about good driving so that it learns what good driving is and can make better decisions in the moment. People who are paying attention closely say that this is an absolutely huge moment and represents a sea change. And so that's what we're going to explore today. Now, to get a sense of that big, bolsterous opinion, Robert Scoble wrote, Our world changed tonight. In 10 years, we will look back at the first public demo of a robot that learned to move around the world by watching only videos. This is a paradigm shift in how software is built.

9:15At one point, Elon Musk took over because the AI made a mistake. He said the fix is to feed it more videos. Multimodal AIs are here at full scale. This speeds up the humanoid robot for me. Imagine you showing your robot how to make grandma's recipe, and from then on it can make it every night if you want. Cameras just had a paradigm shift. Now, there are a couple things going on here. One is an assessment of this demo's place in history, but let's leave that one aside for the moment. That is inherently an unknowable thing, and I think if one wanted to quibble with that, it might distract from the broader point, which I think is pretty unarguable.

9:49And that's Scoble's point that this is a paradigm shift in how software is built. Effectively, what Scoble is saying is that instead of a world where software is pre-programmed with logic from humans, We're increasingly seeing software that's designed to learn from inputs and make decisions for itself based on that learning. Farzad Masbahi actually had a really good post on this in more detail. He writes, This is how Tesla's FSD V12 learns. As humans with driver's licenses, we know what to do at an intersection. We've been taught that a red sign with the word stop means that once we arrive at the sign, we need to stop and check for cross traffic.

10:24We've also been taught that the white lines on either side of the car are barriers that tell us where the car should be as we approach the perpendicular line. We've also learned about crosswalks, right-of-way rules, speed bumps, rain, snow, cyclists, not to smash into oncoming traffic, navigating through crap or missing lane markings, etc. Some of this we've learned by reading a book and then practicing said things on the road, and other things we've learned by driving around and experiencing the environments on our own. With Tesla's latest V12 FSD update, Tesla's vehicles learn in a similar way, but it's actually much broader than you think.

10:52Tesla isn't telling the AI that trains the FSD system what a stop sign is. It isn't telling it what white lines are, what pedestrian sidewalks are, what other cars look like, what red brake lights mean, etc. Instead, as for this intersection example, Tesla is feeding the AI a ton of video depicting what proper driving looks like at a stop sign, with drivers coming to a stop while slowing down at a reasonable speed while centered between the lines. With this footage, the AI says to itself, okay, one thing I'm noticing is that every time the car comes to a stop, the surrounding areas have these stop sign things on either side every single time, and the car is always centered between white lines on the road when it approaches these signs.

11:28The AI then writes code for the car to behave correctly at every stop sign it encounters. This is how Tesla's system learns. This means that for Tesla to reach self-driving under any condition, it needs to collect all driving conditions that a human encounters with many examples of each. It needs to see stop signs that are just on one side, stop signs that are partially covered by a tree, stop signs that have been vandalized, etc., etc., etc. Luckily, Tesla is able to do this because it has a fleet of 4 million cars driving around the world today, and this fleet is growing exponentially. This means that every condition a Tesla finds itself in, the footage can be used by the AI system to learn the proper behavior based on how Tesla drivers navigate that scenario.

12:05This means that the Tesla Elon Musk was driving on FSDV-12 learned from all other Tesla drivers in the world driving their own cars. With this data, the AI system was able to generate commands for the steering wheel, accelerator, and pedal to navigate around its own environment as good as a human could, and possibly significantly better than a human as Tesla collects and processes more data. Imagine having a car with this AI system that never gets tired, never makes a mistake, is always paying attention, is constantly monitoring every angle around the car. This is what Tesla has achieved with V12.

12:32As we finish out the decade, Tesla has plans to reach an annual goal of 20 million cars sold per year by 2030. All of these cars will be outfitted with the camera systems used to collect the data that the AI system used to train itself. Tesla is also investing billions of dollars in training compute to dramatically increase how much data the AI can process at once, which will allow the company to make improvements quicker and be able to process every conceivable scenario that a driver could face on the road. If Tesla AI successfully learns how to drive under any condition, this will mark one of the greatest technological achievements of our time.

13:01The age of the self-driving car is finally here. Now, interestingly, if you go in and see what Elon has been replying to after his demo, he responded to Scoble and talked about the need for inference compute power. Inference in this case being the ability to make decisions quickly. He writes, what is also mind-blowing is that the inference compute power needed for eight cameras running at 36 frames per second is only about 100 watts on the Tesla-designed AI computer. This puny amount of power is enough to achieve superhuman driving. It makes a big difference that we run inference at INT8, which is far more power efficient than FP16.

13:34This requires us to do very difficult quantization-aware training at FP16 in order to infer at the lower resolution of INT8. But think about that for a minute. INT8 only gives you a numerical range from 0 to 255, and yet the car can still understand the immense complexity of reality well enough to drive. Same caveats here. Reaching superhuman driving with AI requires billions of dollars per year of training, compute, and data storage, as well as a vast number of miles driven. Tesla also has over 4 million cars on the road, capable of training the AI. In a few years, we will have roughly 10 million.

14:04AI Authority responds, Tesla is really just an AI training software company, huh? Now, this makes comments from Cathie Wood back in May make a little more sense to perhaps the skeptic who heard them at the time. Back on May 12th, the Observer wrote an article, Cathie Woods dubs Tesla's controversial FSD tech as, quote, most impactful AI project. Now, Cathie Wood is, of course, probably the longest duration Tesla bull outside of Elon Musk himself. And in the beginning of May, Wood and her ARK Invest predicted that by 2027, Tesla's stock price could be about 11 times higher than its current level.

14:37The piece writes, Wood believes Tesla is close to achieving true autonomous driving because of its unique underlying technology. Most driver assistance systems on the market guide a vehicle's movement by pre-mapping an area using a LIDAR, a radar using light instead of radio waves, but Tesla's system relies on cameras that capture a real-time view around a vehicle. An algorithm then processes these video streams in order to guide motion. Said Wood of Elon Musk, he is almost there. I think it's the most impactful AI project out there. Now, there are a few things that make this such an interesting story to me.

15:08The first is this idea expressed by many that this is a seminal or inflection point moment. Boris here, for example, called it the GPT-4 moment for self-driving and real-world robotics. I think obviously there is an incentive for people, and particularly media, to hype up moments as moments where everything changes. and so those designations are always worthy of skepticism, but it's still interesting to see when people think they happen. Now, obviously, the implications for self-driving itself are pretty immense. Self-driving cars are one of those questions and debates as a society that we've only had partially or in fragments, because frankly, we just haven't had to yet.

15:42The technology hasn't quite been at the place where we need to really deal with the full complexity of good and bad, or not even good and bad, but just change that they represent. It appears to me now that that conversation is a lot closer than maybe some people thought. And in that way, it shows that artificial intelligence applied in almost any domain is going to raise big questions. Questions of trust, questions of humanity, questions of loss of control. Obviously, Tesla's self-driving AI is very different than the LLMs that power ChatGPT. But in either case, whether it's writers in Hollywood worried about being replaced or reduced by script-writing AIs like ChatGPT, or truck drivers worrying about being replaced and reduced by self-driving trucks, there is a shared tapestry of societal-level questions that are being generated by basically every instance of this technology.

16:30I also think Scoble's point about the paradigm shift in how software is built is a really salient one. One of the things that people have noted in Silicon Valley is that one of the outcomes of the rise of ChatGPT is a huge shift in where founders are applying their attention. Paul Graham from Y Combinator recently posted that the founders they're seeing are more technical in general and radically more likely to be focused on AI than they were just a year ago. The mental model of designing software to learn from inputs, to be trained and then make decisions for itself, could have really fascinating implications for the next generation of things that get built, whether or not we view them as quote-unquote AI startups at all.

17:11In fact, in some ways, this full self-driving demo might be seen as evidence of the AI-ification of everything that this is, as some have suggested, a fundamental shift in the computing paradigm. So when all is said and done, I don't know how history will look back at that demo, but I don't think that people are crazy to identify it as a fairly significant moment. Anyways, guys, let me know what you think. Is this going to change how software is built? Is it already changing how software is built? How significant is Tesla's full self-driving in the history and lineage of this technology coming to consumer society.

17:45If you want to get deeper into the discussion, come join us on the AI Breakdown Discord. You can go to bit.ly slash AI Breakdown, or you can always just leave a comment here on YouTube. Or if you're listening on the podcast, I think you can go leave a comment on Spotify as well. In any case, I appreciate you guys listening or watching as always. Until next time, peace.

18:11Thank you.