Has Open Source Screwed a Whole Generation of AI Startups?

0:00Today on the AI Breakdown, we're asking whether open source software has had some negative impacts for AI startups. Before that on the brief, a leaker reports that the DALI 3 model is being tested. 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. The first story on our list today is potential leaks from OpenAI's forthcoming DALI 3 model. A user first appeared on Discord in May claiming to be part of an alpha test of the latest text-to-image model from OpenAI.

0:41Now, the leaker showed off a few photos then, and just recently they have popped back up on Discord again with a new set of images that show off what they say DALI 3 can do. According to the leaker, the model is currently only accessible to about 400 people. There are a few notable things about this model. First of all is that, as you can see from the image on your screen, which if you're listening to this not watching, I highly encourage you to check out the video as well, involves some images that might not make it to the final sanguinized version that is released to the public. Among other things, it includes copyrights.

1:11In this photo that we're looking at, there is a Subway sandwich with the Subway logo clearly available, but it also is purportedly able to handle text better. For example, there's an image of an angel that says, be not afraid above. This is obviously one of the big things that people have been excited about with the latest Stable Diffusion XL release as well. Another thing that this new model reportedly does really well is handle more complex prompts with more small details. For example, a wombat sits in a yellow beach chair while sipping a martini that is on his laptop keyboard. The wombat is wearing a white Panama hat and a floral Hawaiian shirt.

1:40Out of focus palm trees in the background, DSLR photograph, wide angle view. That would be a lot of instructions to get right for current models than it seems to here, again assuming that the leaker is actually being truthful. Another example comes from the prompt, a group of farm animals, cow, sheeps, and pigs, made out of cheese and ham on a wooden board. There's a dog in the background eyeing the board hungrily. This prompt has a lot of what's called concept spillover. In other words, the image model mixes different content concepts. However, this output of DALI 3 is able to distinguish between the dog that is actually a dog in the background and the animals made out of cheese that are in the foreground.

2:14When Decoder did the same prompt, there's a bunch of cheese on a table and a couple dogs in the background, one of which has cow horns. This wasn't the only interesting AI image generation news. NVIDIA has also released research about its new perfusion image generator model, which is notable for its tiny size and the fact that it reportedly only takes four minutes to train. The perfusion model was presented in a recent research paper that was created by NVIDIA and Tel Aviv University. Its main technique or new idea is called key locking. As Decrypt describes it, this works by connecting new concepts that a user wants to add, like a specific cat or chair, to a more general category during image generation.

2:51This helps avoid overfitting, which is when the model gets too narrowly tuned to the exact training examples. Overfitting makes it hard for the AI to generate new creative versions of the concept. By tying the new cat to the general notion of a feline, the model can portray the cat in many different poses, appearances, and surroundings, but it still retains the essential catness that makes it look like the intended cat, not just any random feline. So in simple terms, key locking lets the AI flexibly portray personalized concepts while keeping their core identity. It's like giving an artist the following directions.

3:20Draw my cat Tom while sleeping, playing with yarn, and sniffing flowers. Trying to super simplify it, I think what's conceptually exciting about this is that by training the model on a specific object that you want to be represented in the output photos, it opens up the possibility not just of a generalized image of a concept, but of a very specific example. Again, it's not just any cat in these images, it's your cat. So overall, lots of exciting things happening in the AI image generation space. Next, we have a story at the intersection of AI and health. In a study conducted over the course of about a year and a half, radiologists supported by AI were 20 % more able to detect breast cancer than were their colleagues who weren't using AI.

4:02The study was published in Lancet Oncology and looked at scans of more than 80 ,000 women in Sweden who had a mammogram between April 2021 and July 2022. 40 ,000 of them were assigned to a group where AI had read the mammogram before a radiologist looked at it, and the other half had their scans read by two radiologists but without the use of AI. All the radiologists in both samples were considered highly experienced. Overall, the screening that was a single radiologist supported by AI detected six per 1 ,000 screened women, compared with five per 1 ,000 for the team of radiologists that didn't have AI.

4:33Now, importantly, this wasn't AI being oversensitive. The AI-supported radiologists did not have a higher false positive rate than the two radiologists who weren't using AI. What's more, the group that was using AI had a reduced reading workload of 44%. Summing it up, AI led to better results with less time, which is sort of the dream. Next up, we move to the world of social media and content, where YouTube is testing using AI to summarize videos. Basically, YouTube is trying to solve a problem which has plagued any platform that handles long-form content, be it videos or podcasts. And that is discoverability for new videos and trying to get people to try out things that they haven't watched before.

5:13The experiment has AI auto-generating video summaries so that, quote, it's easier for you to read a quick summary about a video and decide whether it's the right fit for you. Right now, the test is limited to a small handful of creators as well as a small handful of users. Speaking of video, one project that I'm watching closely is still just at the demo and testing stage. It comes from a company called Sync Labs, and is basically software that translates video into other languages, and then overlays lip syncing on the video of the speaker. The most recent example that Prady, the founder of Sync Labs, posted was from David Sachs from the All In podcast Speaking Hindi, a language which he doesn't speak, and in which this novel translation and lip sync was done in less than five minutes.

5:53I think one of the most interesting and positive outcomes for AI when it comes to content is the breaking down of linguistic barriers, so I am keeping a close eye on this project. Finally, moving over to markets, earnings season continues, and yesterday it was AMD's turn. Perhaps not surprisingly, given the white-hot AI chip space that it's in, sales and profit both exceeded analyst projections, and the company also reported that their accelerators, which is a type of processor that speeds up the development of AI software, is drawing even more interest than anticipated from customers. The company is looking to ramp up production of its MI300 chips over the course of this year, and one of the things that they're looking into is developing a less powerful chip that can be exported to China under current export controls.

6:34This is something that NVIDIA does, but is also coming under increasing scrutiny in Washington. Just a couple days ago, Reuters ran with the headline, U.S. lawmakers urge Biden administration to tighten AI chip export rules. The story was about a bipartisan open letter written to Commerce Secretary Gina Raimondo asking the U.S. to further strengthen chip export rules, tightening especially restrictions on AI chips even further. This is a great example of where AI is meeting geopolitics in a big, challenging mess. Overall, Wall Street investors are mixed. On the one hand, people were impressed by AMD's results, but overall, there is definitely a growing sense that the AI-driven rally may be coming up against its limits.

7:13As of around noon today, AMD shares were down 6%, NVIDIA was down 4%, and the PHLX Semiconductor Sector Index was down more than 3%. So friends, that is going to do it for today's AI Breakdown Brief. If you enjoyed it, please hit the like button below, or if you're listening, come check out the YouTube channel, and I'll be back soon with the main AI breakdown. Hey guys, before we get to the main show, I want to tell you about today's sponsor, Giscard. Giscard is tackling one of the most important challenges in our new AI world, which is detecting hidden vulnerabilities in machine learning models.

7:46So what do I mean by hidden vulnerabilities? Well, I mean things like performance bias, data leakages, spurious correlations over confidence issues, basically all the things that could negatively impact the performance of a machine learning model. Giscard is compatible with all Python frameworks including PyTorch, HuggingFace, Langchain, and more, and works for both tabular models and LLMs. It's an open source framework that also has enterprise and hosted options, and it's quick and easy to install with just four lines of code. I really believe that this type of testing framework that Giscard offers is so important, as more and more ML models impact the applications we interact with every day.

8:26So to learn more and try it out, go to giscard.ai. That's G-I-S-K-A-R-D dot A-I. Thanks again to Giscard for sponsoring the show. And with that, let's get back to the episode. Today on the AI Breakdown, we are exploring something that I've been thinking about a lot lately, which is the relationship between open source models, enterprise customers, AI startups, and the wider economy. Now, what prompted this is the beginnings of a narrative shift that we started to see in July. And there were two big things that got this narrative shift happening. The first was the reports that for the first time, ChatGPT traffic had actually gone down.

9:08In June, fewer people went to the ChatGPT website and mobile app than they had in May. And the second big piece of the narrative shift came when it was reported that Jasper, as well as a smaller company, Mutiny, but definitely Jasper, had cut a number of jobs in the previously assumed white-hot AI space. These things together brought up a big glaring question of whether the hype in AI was dying down. However, for careful market observers, it was less about overall hype per se and more learning about where value is actually going to accrue in the AI space. Right around this time, AI entrepreneur Sam Hogan wrote a really compelling explanation for what was going on when it came to the AI startup space.

9:46He said six months ago, it looked like AI and LLMs were going to bring a much-needed revival to the venture startup ecosystem after a few tough years. With companies like Jasper starting to slow down, it's looking like this may not be the case. Right now, there are two clear winners, a handful of losers, and a small group of moonshots that seem promising. So the losers he characterized as the companies that had raised a huge amount of money, even though basically they were just a user experience layer on top of some other company's API, as he called it essentially a generic thin wrapper around open AI.

10:17The other category of losers, he argued, were application layer AI companies that raised a bunch of VC money between December and March, based on all the hype surrounding ChatGPT, assuming that if nothing else, a plausible exit would be selling themselves to later stage or enterprise companies. Sam characterized these startups as typically having products that are more focused than something very generic like Jasper, but still not having a real technology moat. In other words, the products are easy to copy. I think that what Sam is arguing here actually has a lot to teach us about how the overall AI space is developing.

10:49And effectively, what he's arguing is again less about whether the patina of the artificial intelligence space as a whole is wearing off, and more about, on the one hand, whether old VC models are savable, and on the other hand, how enterprises are interacting with the AI space. Let's discuss that first dimension first, whether AI startups can save the VC model. For this, let's turn to a piece from investor Sam Lesson, also from last month. It was published in The Information, and the way he teed it up on Twitter was, Seed investing isn't coming back, at least not as it existed in the last decade.

11:23Sam writes, Seed investing can't turn back on unless the public market changes how it values run-of-the-mill tech companies, and that ain't happening. About 15 months ago, I wrote a post on how seed investing was pretty clearly going to be in an 18-month timeout, that the capital factory line would be shut down until the inventory of dramatically overmarked late-stage private deals got worked through, washed out, or expired on the line. This is basically how the world has looked for the last almost year and a half, with the noted exception of an AI death spasm, where a bunch of funds decided to pour untold amounts of capital into AI companies on the factory model they were used to, with even higher valuations and more hype.

11:57The thing I think seed investors need to come to terms with at this point is that this isn't an 18-month timeout, it is likely much, much longer. And perhaps even the death of systematic and thematic seed investing as we knew it between 2010-ish and 2022-ish. Why? Because run-of-the-mill public tech companies just aren't worth that much, it turns out. And if the bulk of so-called unicorns can't get public and or do and are disappointing, the whole model of seed investing starts to look way, way less attractive as an asset class. Now from there, Sam gets into some of the details, but I will divert our attention to the macro for a moment.

12:29One of the things that was remarkable to me, living for a decade in Silicon Valley, was how little the average VC took the time to understand just how much public market dynamics and macroeconomic dynamics impacted their industry. In the middle of the teens, everyone was noticing that valuations of startups were going up and up and up. The blame was put on things like Y Combinator, driving a premium for the latest startups. What was almost never discussed was the fact that after six or seven years of living in a zero interest rate world, capital had to move farther and farther out on the risk spectrum in order to find yield.

13:03For the entire teens decade, the entire period that Sam Lesson is here acknowledging as this time of seed investing, billions and billions in capital that hadn't been exposed to venture capital or even private equity was coming into those asset classes because they simply had to. It doesn't take an economics major to understand what happens when what is a relatively constricted supply of top startups meets a massively growing demand in the form of excess capital. Valuations go up, round sizes go up, companies stay private longer, and some weird dislocations start to happen. In the case of venture capital in the teens, that was funds not really having to return in practice, but just being able to raise ever growing funds on the supposed IRR based on increased valuations in later rounds.

13:46This is the factory line that Sam was referring to. That factory line stopped when inflation started to rip upwards and the Federal Reserve reversed course entirely, and we went from an era of quantitative easing and zero interest rate policies to an era of quantitative tightening, i.e. the Fed removing liquidity from the system and the fastest rate hiking cycle in 40 years. As that has happened, there has been a massive contraction in risk capital across all asset classes that deal with risk, venture capital included. VC is simply not immune to macroeconomic changes. And so to some extent, I think it's reasonable to look at the flood of VC money into AI over the last six months as a last gasp of trying to hold on to that system.

14:28There's obviously also been a comparable frenzy in public markets where AI enthusiasm has really been one of the only things holding the markets up and has provided a strong countervailing narrative to all sorts of other insecurities throughout the year, from banking crises to debt ceiling negotiations to government debt being downgraded this week. So one part of the challenge for AI startups is that even as enthusiastic as venture capitalists have been, they're still not immune to the broader changes that are going on across the startup and venture capital landscape. But there is another really interesting dimension to this, which reveals a lot about how the AI field specifically is evolving.

15:02Remember, one of the categories of losers that Sam identified were companies who expected that they might be able to be acquired by enterprises. In his post, Sam effectively argues that these companies would prefer to build their own tools rather than become the customers or the acquirers of unproven AI startups. As he put it, an engineering leader would rather spin up their own Langchain or Chroma infrastructure for free and build tech themselves than buy something from a new unproven startup. Now, I think there are two reasons why enterprises might be heading in this direction. One has to do with risk.

15:34Yesterday, we covered McKinsey's State of AI report. And one of the interesting pieces of that report was what threats organizations consider relevant. Across the survey participants, inaccuracy was the most relevant risk with 56 % of organizations considering it a risk. Cybersecurity had 53 % of organizations concerned. IP infringement had 46%. Regulatory compliance had 45%. And those were the top categories. Now keep those concerns in mind as we imagine a world in which there are two broadly speaking, ways that an enterprise company could get into generative AI. Option one is through some sort of vendor, think ChatGPT's forthcoming business version, or on the other hand, a second pathway is spinning up one's own proprietary AI tools.

16:17For example, a large language model that's trained on proprietary data, but that's hosted perhaps on-premise and is specifically customized to an enterprise, their needs, their data, their security. It's not hard to see if one's concerns are inaccuracy, cybersecurity, IP infringement, and regulatory compliance, why that latter model of something custom spun up with data controlled by and already available to the company might appear to be a better choice than some new startup. And of course, you have to think that that has only increased after we've seen companies like Samsung ban their employees from using tools like ChatGPT after discovering them leaking sensitive data, which then becomes part of the data set that ChatGPT trains on.

16:55So on the one side, we have the risk and concern reason that enterprises might be looking away from startups and towards their own solution. But then we also have the availability side, and this gets to the title of this episode. The fact that we now live in a world with high-performant, commercially available, open-source or open-source-ish models like Llama 2 means that those enterprises aren't starting from scratch. In fact, not even close to it. There is an incredible amount of infrastructure that they have to actually go build solutions in ways that with previous technology movements, they just haven't been able to.

17:28When that leaked Google note, we have no moat and neither does open AI came out, the leaker wasn't really talking about enterprise business models and what big corporations might do. But interestingly, their argument that the open source community was going to eat the lunch of Google and open AI and companies like them seems more likely for the fact that those big companies are in fact adopting open source models and customizing them for their own use rather than working with startups. Now, of course, for these enterprise companies, the landscape of choices is not actually as binary as startups on the one hand or totally custom spun-up solutions on the other.

18:02The companies in the middle are the cloud providers and tech partners that already have deep relationships with those enterprises and who already host or interact with lots of their proprietary data. The Wall Street Journal published yesterday a piece called Companies Weigh Growing Power of Cloud Providers Amid AI Boom. A wave of partnerships between AI model makers and cloud providers is leading tech chiefs to assess the benefits of convenience versus becoming too reliant on any one vendor. The piece writes, For many businesses, the primary choice isn't which AI model to use, but whether they stay within the AI ecosystem offered by their cloud providers.

18:36If a company chooses a single AI ecosystem, it could risk vendor lock-in within that provider's platform and set of services. Companies say the problem with vendor lock-in, especially among cloud providers, is that they have difficulty moving their data to other platforms, lose negotiation power with other vendors, and must rely on one provider to keep its services online and secure. Now, perhaps in response to that, you're seeing companies like Amazon and Google offer an approach where they create a sandbox or managed service environment where enterprises that trust either Amazon or Google can interact in a single place with all sorts of various AI models.

19:09Amazon's Bedrock platform, in other words, isn't locking people into some Amazon LLM, but is instead creating a safe enterprise space for companies to interact with lots and lots of different models from open AI to stable diffusion and beyond. This may seem in some ways like deep insider baseball, but the way capital flows from investors to startups to enterprises to big tech companies is going to have, I believe, a dramatic impact on how the artificial intelligence space develops. And so understanding how those flows are evolving, I think, is a super valuable thing. Anyways, guys, let me know what you think in the comments, or come join us on the AI Breakdown Discord.

19:47You can find a link to that on breakdown.network, and I can't wait to see you guys there to discuss this further. Thanks for listening or watching as always, and until next time, peace.