The AI Daily Brief: Artificial Intelligence News and Analysis - AI Chip Wars: Tesla Building "Bunkerlike" Home for Dojo Supercomputer

Episode Date: October 11, 2023

Tesla is building a massive facility for their Dojo supercomputer. NLW explores how it could impact what they do in AI. Before that on the Brief: Adobe releases their Image 2 model. TAKE OUR SURVEY ON EDUCATIONAL AND LEARNING RESOURCE CONTENT: https://bit.ly/aibreakdownsurvey ABOUT THE AI BREAKDOWN The AI Breakdown helps you understand the most important news and discussions in AI.  Subscribe to The AI Breakdown newsletter: https://theaibreakdown.beehiiv.com/subscribe Subscribe to The AI Breakdown on YouTube: https://www.youtube.com/@TheAIBreakdown Join the community: bit.ly/aibreakdown Learn more: http://breakdown.network/

Today on the AI breakdown, Tesla is setting up a new bunker-like facility for its dojo supercomputer down in Austin, Texas. Before that on the brief, Adobe releases a new image generation model. 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 Discord, our newsletter, and our YouTube channel. Welcome back to the AI breakdown brief. All the AI headline news you need in around five minutes. And today we kick off with a fun new tool. something that I think a lot of you listeners will end up using,

and that is Firefly's new image model Firefly Image 2. Adobe Senior Creative Evangelist Chris Kastanova tweeted yesterday, We released a new model. As a scientist, I am in disbelief that such a model is possible with the smaller dataset as we had. We compensated artists for training it. We worked very hard on it. Please enjoy.

So this was announced at the Adobe Max event. Max is Adobe's big creative event, and it's actually still ongoing today. As part of the announcement, Adobe said that it's Image 2 model, creates significantly improved images, especially compared to its predecessor. Colors are more vivid, images come in higher resolution, and it's much better with texture and other details. Now, interestingly, in addition to just the new model, image 2 also comes with a new set of AI-powered editing capabilities. So, for example, photo settings like adjusting depth of

field can be added to images after they've been created. Adobe also introduced a feature that allows people to match the outputs of their prompts to either a pre-selected list of images that Adobe provides, or by uploading a reference image. Adobe also introduced a variation on the model that was specifically for vector images. This is being built now directly into Illustrator, and from directly inside that application, you can describe the vector graphic that you're looking for, for example, an astronaut dog in a spacesuit, and then have access to all of Illustrators' editing controls right from the same software experience.

Ever since this announcement yesterday, the legion of AI content creators have been testing to see how Firefly's Image 2 model compares to Mid Journey and Dali. My favorite of these threads came from Chase Lean at Chase Lean, T.J on Twitter, who organized it into a set of different categories. On realistic photos, he found that Firefly 2 was very good at generating photorealistic images, particularly of humans. When it came to product photos, he found a lot of similarity between the three models. Same with interior designs where there seemed to be parity. Now, when it came to text generation, Dolly 3 continued to be light years ahead of both Mid Journey and Adobe. But while Mid Journey produced nonsense non-letters, at least Adobe sort of got the R from Rachel in their output.

based on Chase's prompt. When it came to landscape photos, Chase found that Mid Journey was ahead followed by Firefly, but that both were ahead of Dolly 3, and where Adobe's image 2 really shined, was in close-ups and photo realism around humans. Now, the other interesting bit of this announcement, more from a broad societal standpoint, was the new content credentials icon as part of the Content Credential's new icon of transparency is an invisible digital watermark that explains how an image was created, with what software, and when. Now, of course, with these types of standards, just a technology question, but a social adoption question, which is why it was interesting that Adobe announced the number of partners who are adopting this credential, including

Microsoft, publicists, Lyca Camera, and Nikon. All and all, some really interesting stuff coming out of Adobe. And I think for me, one of the big questions continues to be, does the fact that Adobe's model was trained on images that they theoretically have rights to change the equation for particularly enterprise or corporate users? Are companies going to start getting notes from their legal department that says, you have to use Firefly, not mid-Jurney, because that's more legally defensible from an IP standpoint. And more broadly than that, if that does happen, how much space between state-of-the-art and safe is too much for people to comply and go with the safe option. Lots of interesting questions for the near future. Now, moving on to a couple other

stories today, following announcements from both Spotify and YouTube over the past few weeks, that they were going to start inserting automated dubbing to videos and other content, 11 Labs has become the latest startup to launch their own dubbing service. The new tool allows users to upload files or point to something from YouTube, TikTok, Twitter, Vimeo, or another URL, to then be localized across 29 possible languages. Now, part of the reason that people are excited about 11 Labs version of this service is that they're widely considered to have the most advanced voice generation technology so that when it comes to the promise of preserving the original speaker style, even though they're translating to another language, there's a lot of optimism that

11 Labs can deliver that. Now, of course, this is a white hot space right now. Wondercraft AI is another company coming out of Y Combinator that does something similar. And so to me, What it all adds up to is a very likely scenario where a few years down the line, it will be simply the standard that when you create content in one language, it's automatically translated and dubbed into other languages as simply a matter of course. In our next story, Business Insider has followed up time releasing their own AI 100, what they're calling the top people in artificial intelligence. Insider says that their considerations included people who were successfully reinventing a business model with AI, tackling some human condition problem to use AI to make people happier, healthier, etc. people who were focused on providing guardrails and checks and balances for the industry. Now, my observation going through this list was that there were a lot more folks from the corporate and enterprise world than, for example, we got on that Time 100 list.

And that frankly, I think a lot of the developers and entrepreneurs out there would like to see. I think that's fairly defensible if for no other reason than what we found over and over again is that a lot of the dynamics of the artificial intelligence space give existing players and incumbents more of a leg up than they've had in previous startup spheres, be that capital and access to compute, or existing relationships with customers that create a trust bridge to a new area that uses a lot of customer data, maybe insider's corporate focus list actually functionally makes sense. Lastly, today, a little narrative watch that I have seen as absolutely inevitable. The New York Times published a piece called AI could soon need as much electricity as an entire

country. This begat three or four other mainstream articles talking about the same issue. And of course, for anyone who's lived in the Bitcoin world at all, these research pieces that point to the consumption of energy by countries are an extremely potent headline generator. Now, that's not to say that we shouldn't be asking questions about the electricity and compute costs of artificial intelligence. If anything, I think it's another reason to actually ask what we want out of this technology, and remember that we as a society get to determine what technology is actually good for us and what we want to leave on the table. But overall, I just think it's interesting that we are now seeing the same sort of energy and climate comparisons that we've seen levied on other technology spaces in the

past. Anyways, friends, that is going to do it for today's AI breakdown. Next up, the main AI breakdown. Welcome back to the AI breakdown. Today we're talking about a story that touches a lot of different aspects of the artificial intelligence space, from the AI chip wars to foundation model supremacy to real world AI. And that story is, of course, about Tesla and a report from the information that they're building a new, quote, bunker-like structure to house their dojo supercomputer. So let's look at the story first and then let's talk about what some of the interesting

implications are. As has been so often the case recently, the story comes from the information first, and the way that they describe this bunker-like structure is as something that could, quote, one day help move the company beyond electric vehicle manufacturing. Continues the information, the Austin Dojo Project, details of which haven't been previously reported, reflect an audacious plan by Musk to take greater control over the technology it needs to run the AI software at the heart of his products. So one dimension of this is, of course, just the battle, around access to compute and the challenge of AI chips. Like so many other companies, Tesla is dependent on Nvidia.

Nvidia's chips currently power Tesla's full self-driving software that sits inside Tesla vehicles. And while that works for now, it seems very likely that Tesla's need for compute is going to do nothing but increase as it adds new types of vehicles and even other types of products, one of which we'll get into in a moment, to the set of things that it builds. Even before we get into the Optimus Robot, Tesla is making plans to expand its fleet of cars. that includes a robotaxy as well as a new entry-level Tesla. Now, like other big tech companies, Tesla is also designing its own AI chips. The Dojo supercomputer is going to be powered by Tesla's D-1 chips.

Now, last month, news outlets from Taiwan reported that Tesla had recently increased its order for the manufacturer of these D-1 chips by double from TSM, who's actually doing the fabrication for Tesla. Now, at this point, not a ton is known about the D-1 chip. Tesla does have some information on their website about it, and they even have a downloadable technical white paper, but there are a number of benefits that analysts and observers expect Tesla to get from its own chips. One comes from a note from Morgan Stanley, which suggests that the D1 chip will give Tesla a better ability to control how much energy it uses to run AI software, and in that

same report, the analysts from Morgan Stanley expect that the D1 may be optimized to process video data faster than with the Nvidia chips. That would make sense given that the first and most important use case for Tesla is as a part of their full self-driving suite, which has to take in a ton of video data very quickly. Tesla has said that it expects that Dojo will be able to reduce training time for full self-driving workloads from a month down to a week, and overall Morgan Stanley estimates that the D-1 could save Tesla $6.5 billion over the next few years. Now, outside of just cost savings for Tesla itself, Tesla and Musk haven't exactly been cagey about the fact that they want other companies to be able to use their full self-driving technology as well. As Musk said, we're not

trying to keep this to ourselves. To the extent that other companies and vehicle manufacturers are licensing Tesla self-driving system, that would obviously increase the need for compute even farther. Even beyond just full self-driving, Musk has suggested that Dojo could be used to expand access to saleable AI compute services. In April, he told investors, quote, Dojo has the potential to become a sellable service that we would offer to other companies in the same way that Amazon Web Services offers web services, even though it started out as a bookstore. So I really think that, yes, the dojo potential is very significant. Now, of course, there is a ton going on in this AI chip space. Yesterday, we discussed reports that Microsoft

will be debuting their AI chip that has been built under the code name Athena as soon as next month at their annual developers conference. Back in August, Google unveiled the latest version of its tensor processing unit and then had to deal with a number of different news reports that suggested that they were also trying to get away from Nvidia. And more recently at the end of September, Amazon announced an investment in Anthropic, which while starting at 1.25 billion could go all the way up to a $4 billion investment. A big part of that seemed to be collaboration around Amazon's custom AI infrastructure, including their chips. From the Anthropic blog post,

AWS will become Anthropics' primary cloud provider for mission-critical workloads, providing our team with access to leading compute infrastructure in the form of AWS Traneum and Infersia chips. Together, we'll combine our respective expertise to collaborate on the development of future Traneum and Infersia technology. And then, of course, just last week, we heard that Open AI was also exploring making its own chips, and that while the company hadn't made that decision officially yet,

They had gone so far as to evaluate a potential acquisition opportunity in the space, suggesting that even if they don't make that decision ultimately, it is a very, very serious consideration. Now, one other bit of actual news from today around the AI chip space is the latest move by AMD to catch up to the distant current leader, NVIDIA. Now, a big part of why NVIDIA has become the default choice over the last decade is not just the quality of their hardware, but also the software that surrounds it. It makes sense, then, that AMD is acquiring an AI software startup to better invest, in the software ecosystem around the company's chips. The company that they're acquiring node.

A.I sells tech to large data center operators and other types of customers to help them deploy AI models that are tuned for AMD's chips. The company will be absorbed into the 1500 strong engineering group inside AMD that's focusing on software around AMD's chips. Now, of course, the other reason that people are interested in Dojo is the way in which it might be used to power up Tesla Optimus. Optimus is Tesla's humanoid robot that has quietly been making pretty significant advances over the last few months. In an update shared at the end of September, Tesla announced a few advances on the optimists, including the fact that it's able to calibrate its limbs in the real world, understanding where its arms and legs are in space, and that it can now sort objects autonomously.

It has a fully onboard neural network that means that just with the video input in, of objects with different colors, it's able to sort them correctly as the desired output. Now, even in this field of robotics, there is intense competition between different companies in the space. In May, various news outlets reported that an open AI backerner robot startup had quote-a-quote beaten Tesla to deploy humanoid robots in the real world. That company was called 1X and their robot is called Eve and in April of this year had been deployed as security guards in two industrial locations. The company said that the robots were going to be deployed in hospices and assisted living facilities next. Now, the big theme, of course, for Tesla and Elon when it

comes to AI is this idea of artificial intelligence in the real world. Fully self-driving vehicles, autonomous humanoid robots, and all of it powered by a supercomputer with chips of their own design. When you take a step back and look at it in context, it starts to make Kathy Woods assessment of the situation make a little bit more sense. Over the next, between now and 2030, we think artificial intelligence is going to add more value than any of our other technology platforms. In fact, it's going to catalyze them. We talk about Tesla all the time. It actually is the biggest artificial intelligence play we believe right now in our portfolios. It is the largest position in our flagship portfolio, ARKK. Why is this? Because autonomous taxi platforms, we believe globally,

will deliver by 2030, $8 to $10 trillion in revenue from almost zero right now. Think about that. Eight to $10 trillion in revenue is almost half of the size of of the US economy. We think that's a global autonomous taxi platform opportunity. And we think it's going to submit to natural geographic monopolies. Tesla certainly United States and perhaps elsewhere. So you're going to be surprised at seeing who's going to, many people think it's just hardware and software stocks widely advertised.

But Tesla, many people think is an auto stock. We don't. We think it's much more than that. But we think it's one of the biggest AI opportunities out there. And so all of the pieces of the Tesla and Elon AI story continue to come into view just a little bit more. Next up, maybe we'll get more information about what XAI is up to. But that is where we will leave the story for today.

Thanks as always for listening to your watching. And until next time, peace.