Tech Brew Ride Home - Thu. 02/22 – Nvidia’s Earnings Are Epic

Starting point is 00:00:00 On April 4th, 2023, around 2 in the morning, a man was found stabbed multiple times on a sidewalk in downtown San Francisco. Hey, who did this to you? What happened next turned the story into a political firestorm. Reports have identified the victim as Bob Lee, the founder of Cash App. From Bloomberg Podcasts, this is Foundering, the Killing of Bob Lee, beginning April 16. Welcome to the Tech meme right home for Thursday, February 22nd, 2024. I'm Brian McCullough today. The most important tech company in the world right now reported earnings yesterday and they were historic. Stability AI is previewing stable diffusion 3.0. Google has to fix some of its AI image generation details. Amazon is getting aggressive about bringing sports to streaming and has a startup we've never mentioned made a big AI breakthrough. Here's what you missed today in the world of tech. As Intelligencer put it last night, the most important stock on planet Earth just got even bigger. After the bell yesterday, NVIDIA reported Q4 revenue up 265% year-on-year to $22.1 billion.

Starting point is 00:01:19 Data Center revenue, which is where all of that AI stuff lives, was up 409% to $18.4 billion. And NVIDIA's going forward, Q1 revenue forecast came in above estimates. So let me repeat that. Overall revenue grew 265%, and data center revenue grew a whopping 409%, and that's on a year-over-year basis. And it's not like those were small numbers to begin with, like going from millions of dollars to tens of millions. This was tens of billions of dollars going to bigger tens of billions. To quote Gunjan Banerjee on threads, for context, none of its mega-cap peers, Apple, Microsoft, Amazon meta, have ever expanded revenue. that fast from a similar starting point, end quote. Invidia's stock jumped more than 10% overnight

Starting point is 00:02:07 on the news, and to put that in context, going up 10% when your market cap is $1.7 trillion, suggests that Nvidia is adding $170 billion in value overnight. The average market cap of a company in the entire S&P 500 is $50 billion. Quoting CNBC, NVIDIA has been the primary beneficiary of the recent technology industry obsession with large artificial intelligence models, which are developed on the company's pricey graphics processors for servers. NVIDIA CEO Jensen Huang addressed investor fears that the company may not be able to keep up with this growth or level of sales for the whole year on a call with analysts, quote, fundamentally the conditions are excellent for continued growth.

Starting point is 00:02:49 In 2025 and beyond, Huang told analysts, he says demand for the company's GPUs will remain high due to generative AI and an industry-wide shift away from central processors to the accelerators that Nvidia makes. Nvidia chief financial officer Colette Kress said that while the company had improved supply of its AI GPUs, it still expected them to be in short supply, especially the next-generation ship called B-100, expected to ship later this year. We are delighted that supply of Hopper Architecture products is improving, Kress said, on a call with analysts.

Starting point is 00:03:20 Demand for Hopper remains very strong. We can expect our next generation products to be supply constrained, as demand far exceeds supply. Whenever we have new products, as you know, it ramps from zero to a very large number, and you can't do that overnight, Wong said. The company's gaming business, which includes graphics cards for laptops and PCs, was merely up 56% year-over-year to $2.87 billion. Graphics cards for gaming used to be Nvidia's primary business before its AI chips started taking off, and some of Nvidia's graphics cards can be used for AI, end quote.

Starting point is 00:03:52 Another day, another one of these. Stability AI has released a preview of Stable Diffusion 3.0. Its next flagship text-to-image model aiming to offer improved image quality and better performance. Quoting Venture Beat, the new Stable Diffusion 3.0 model aims to provide improved image quality and better performance in generating images from multi-subject prompts. It will also provide significantly better typography than prior stable diffusion models, enabling more accurate and consistent spelling inside of generated images. Typography has also been an area of weakness for stable diffusion in the past, and one that rivals including Dolly 3, ideogram, and mid-jurney have also been working on with recent releases.

Starting point is 00:04:35 Stability AI is building out Stable Diffusion 3.0 and multiple model sizes ranging from 800 million to 8 billion parameters. Stable Diffusion 3 isn't just a new version of a model that Stability AI has already released. It's actually based on a new architecture. Stable Diffusion 3 is a diffusion transformer, a new type of architecture similar to the one used in the recent Open AI SORA model. Imad Mastock, CEO of Stability AI, told Venturebeat, it is the real successor to the original Stable Diffusion. Stable Diffusion did not have a transformer

Starting point is 00:05:05 before, Mastock said. Transformers are at the foundation of much of the Gen AI Revolution and are widely used as the basis of text generation models. Image generation has largely been in the realm of diffusion models. The research paper that details diffusion transformers explains that it is a new architecture for diffusion models that replaces the common used UNET backbone with a transformer operating on latent image patches. The approach can use compute more efficiently and can outperform other forms of diffusion image generation. The other big innovation that stable diffusion benefits from is flow matching. The research paper on flow matching explains that it is a new method for training continuous normalizing flows,

Starting point is 00:05:46 CNFSs, to model complex data distributions. According to the researchers using continual flow matching with optimal transport pass leads to faster training, more efficient sampling, and better performance compared to diffusion pass, end quote. So on that improvement in terms of spelling, you know how when you ask an AI image generator to show a Tesla outside of a Tesla factory, it can occasionally get the Tesla logo kind of right and maybe even the font, but it can't actually spell the word Tesla on the signage. Well, apparently with the new transformer architecture and some new text encoder, Stable Fusion 3.0, is capable of writing even for full sentences. In their blog post announcing this, they even have images spelling out Stable Diffusion

Starting point is 00:06:27 3. Meanwhile, Google says it is working to fix Gemini's, quote, inaccuracies in some historical image generation depictions after some users complained about them, which is a euphemistic way to say this happened, quoting PCMAG. Do AI generated images have to be historically accurate down to the racial identity of the characters created? Some users of Google's generative AI tool Gemini think so, have taken to social media platforms like X and Reddit to complain. Google's senior director of product Jack Krochek, who's overseeing Gemini's development, wrote Wednesday that the Gemini team is working to tweak the AI model so that it generates more historically accurate results. We are aware that Gemini is offering inaccuracies and some historical image generation depictions,

Starting point is 00:07:16 and we are working to fix this immediately, Krochek said. Some Gemini users posted screenshots claiming that Gemini considered a Native American man and Indian woman to be representative of an 1820s-era German couple, an African-American founding father, Asian and indigenous soldiers to be members of the 1929 German military, and diverse representations of a medieval king of England, among other examples, end quote. This was followed by news that Google is going to, quote, pause the image generation of people via Gemini, full stop, and quote, re-release an improved version soon while working to fix its race inaccuracy, quote, issues, quoting the verge. Google's decision to pause image generation of people in Gemini comes less than 24 hours after the company apologized for the inaccuracies and some historical images its AI model generated.

Starting point is 00:08:04 Some Gemini users have been requesting images of historical groups or figures like the Founding Fathers and found non-white AI-generated people in the results. That's led to conspiracy theories online that Google is intentionally avoiding depicting white people. The Verge tested several Gemini queries yesterday, which included a request for a U.S. senator from the 1800s that returned results that included what appeared to be black and Native American women. The first female senator was a white woman in 1922, so Gemini's AI images were essentially erasing the history of race and gender discrimination. Now that Google has disabled Gemini's ability to generate pictures of people, here's how the AI model responds if you request an image of a person. We are working to improve Gemini's ability to generate images of people. We expect this feature to return soon and will notify you in release updates when it does. Google first started offering

Starting point is 00:08:49 image generation through Gemini, formerly barred earlier this month, in a bid to compete with OpenAI and Microsoft's co-pilot, end quote. Sources say Amazon is paying $120 million for exclusive rights to show a 2025 NFL playoff game on Prime Video. NBCU paid only $110 million for the January 2020-4 NFL playoff game on Peacock from mere weeks ago. Looks like that was a success. Quoting front office sports. Out of the big four tech giants, Amazon is the most ambitious about invading live sports. The giant streamer is opening its nearly bottomless wallet for the most valuable programming in entertainment postseason NFL football. Prime Video is paying an estimated

Starting point is 00:09:35 $120 million for exclusive rights to a single NFL playoff game after the 2024 season. Sources familiar with the deal tell me. That's more than the $110 million that NBC Universal's Peacock paid for an AFC wildcard playoff game between the Chiefs and Dolphins on January 14th, and more than the $100 million prime itself paid to stream the league's first ever Black Friday game between the Dolphins and Jets on November 24th. The Wall Street Journal's Joe Flint first reported Prime had snagged the rights to next season's playoff game after passing on the game that ended up on Peacock. The giant streamer won't make the same mistake again. Despite fierce backlash to the NFL putting the game behind a paywall,

Starting point is 00:10:12 the first ever live-streamed NFL playoff game on Peacock delivered the goods. With an average audience of 23 million viewers, it ranked as the most streamed live event in U.S. history. That figure was up 6% from the previous years, comparable wildcard telecasts, which drew 21.8 million viewers on the NBC broadcast network and other platforms. Meanwhile, Peacock posted its biggest single day ever with a record 16.3 million concurrent devices. The game generated 30% of the day's internet traffic. Prime didn't have the same success on Black Friday. The Tech Giants' exclusive stream drew only 9.61 million viewers the day after the NFL's three Thanksgiving Day games averaged a monster 34.1 million viewers. The NFL has not confirmed whether there will be a

Starting point is 00:10:53 second Black Friday game on Prime in 2024, but given how pleased the two, two partners were about the marriage of football and e-commerce, I bet that game becomes another tent-pole event on the NFL calendar, end quote. Following up on this, Reuters says that Reddit has inked a deal with Google to make its content available for training the search giant's AI models. A source says the contract is worth around $60 million a year, so it was Google, not Apple or OpenAI, who signed the deal with Reddit. Quote, the deal underscores how Reddit, which is preparing for a high-profile stock market launch, is seeking to generate new revenue amid fierce competition for advertising dollars from the likes of TikTok

Starting point is 00:11:34 and Meta Platform's Facebook. Last year, Reddit said it would charge companies for access to its application programming interface or API, the means by which it distributes its content. The agreement with Google is the first reported deal with a big AI company. San Francisco-based Reddit, which has been looking for a stock float for more than three years, is preparing to make its initial public filing this week, which would detail its financials for the first time to potential IPO investors. The filing could be available as early as Thursday, two of the sources said. The company, which was valued at about $10 billion in a funding round in 2021, is seeking to sell about 10% of its shares in the offering. Reuters has previously reported.

Starting point is 00:12:11 Makers of AI models have been busy clenching deals with content owners in recent months, aiming to diversify their training data beyond large scrapes of the internet. That practice is rife with potential copyright issues, as many content creators have alleged that their content was used without permission, end quote. Finally today, one more potential AI breakthrough here, and I'm stressing potential. Magic recently raised $117 million from the likes of Daniel Gross and Nat Friedman, and they say that their AI coding assistant has a larger context window than rivals, but also that they may have made an AI reasoning breakthrough. A reasoning breakthrough would be a completely new paradigm, quoting the information. Former GitHub CEO Nat Friedman and his investment

Starting point is 00:12:59 partner Daniel Gross raised eyebrows last week by writing a $100 million check to Magic, the developer of an artificial intelligence coding assistant. There are loads of coding assistance already, and the top dog among them is Microsoft's GitHub co-pilot. So what did Friedman and Gross see in Magic? The answer goes beyond. The company's claim that it will soon be able to furnish its customers with fully automated coding coworkers, not just semi-automated assistants that finish segments of codewriting as GitHub co-pilot already does. The startup has created a new type of large language model that can process a huge amount of data known as an input or context window. Magic claims to be able to process 3.5 million words worth of text input, five times as much

Starting point is 00:13:39 information as Google's latest Gemini-LLM, which in turn was a big advance on OpenAI's GPT4. In other words, Magic's model essentially has an unlimited context window, perhaps bringing it closer to the way humans process information. That could be especially helpful in a field like software development, where such a model would be able to ingest and remember a company's entire code base to generate new code in the company's style. Just as important, Magic also privately claimed to have made a technical breakthrough that could enable, quote, active reasoning capabilities similar to the Q model developed by OpenAI last year, according to a person familiar with its technology. That could help solve some of the major gripes with LLMs,

Starting point is 00:14:17 which is that they mimic what they've seen in their training data rather than using logic to solve new problems. As for how Magic developed its LLM, this person said it took some elements of transformers, a type of AI model that powers consumer products like ChatGBT, GBT, and coding assistance like co-pilot, and fuse them with other kinds of deep learning models. Magic's co-founder and CEO Eric Steinberger has grappled with the problem of getting AI models to reason before. He previously worked at meta-platforms conducting research on how reinforcement learning, the machine learning technique that many speculate is behind the impressive performance of open AIs' LLMs, can help AI models find the optimal solutions to problems even with imperfect

Starting point is 00:14:55 information. His ambition with magic is bigger than a coding co-worker. Steinberger wants to develop AI superintelligence the same way OpenAI and Google do. Sadly, he didn't respond when we reached out to ask him about this. While GitHub has access to Microsoft's hordes of Nvidia AI chips, something which is hugely valuable nowadays, the tech giant has had tough choices to make when deciding how to allocate chips to internal teams like GitHub and OpenAI, its most important partner. Last October, for instance, Microsoft staffers had to ask higher-ups for access to more AI-specialized servers after a surge in traffic knocked the model that powers Bing and GitHub co-pilot offline. For what it's worth,

Starting point is 00:15:33 Magic's website says it has thousands of AI chips to train its models. It's not clear where they're from, but a likely possibility is the Nvidia-made AI chip cluster, Friedman and Gross have made available to their portfolio companies. We'll be curious to see how Magic takes on the tech giant. AI coding has emerged as one of the most effective early uses of LLMs and GitHub co-pilot has become a critical part of Microsoft's grand plan to make gains against Amazon Web Services in renting out cloud servers. It's hard to imagine Microsoft will take Magic's challenge lying down, end quote. So for the first time, I can announce to you that the Ride Home Fund has had an exit on an investment we made. Cadre, which we invested in almost exactly a year ago,

Starting point is 00:16:23 has been acquired by Yield Street. Investors in the Ride Home Fund for the quarter in question were notified of the transaction yesterday. This was almost the textbook ride home investment. A listener to this podcast turned me on to cadre raising around a year ago, and so they will participate in the carry when that distribution happens. Periodic reminder that any accredited investors can invest in the ride home fund at any time because it's a rolling fund. You can invest this week, next week, next month,

Starting point is 00:16:52 as long as the fund remains operational. And you participate in any deals we do while you're, an investor in the fund. The minimum terms are you have to invest for at least four consecutive quarters and invest at least $5,000 per quarter. So a $20,000 minimum investment, which is less than the ride home AI funds, $100,000 minimum investment, though that fund is also still accepting new investors as well. And the ride home fund has to participate in every deal the ride home AI fund does, so you can piggyback on all of our AI investing. More info, as always, at ridehomefund.com. Talk to you tomorrow.

Tech Brew Ride Home - Thu. 02/22 – Nvidia’s Earnings Are Epic

There aren't comments yet for this episode. Click on any sentence in the transcript to leave a comment.