A visit to Cerebral Valley

I spent yesterday mingling with a couple hundred leaders from the AI industry in the San Francisco neighborhood of Hayes Valley, which has recently been dubbed “Cerebral Valley” due to its concentration of AI startups and hacker houses.

The appropriately named Cerebral Valley AI Summit was thrown by my friend Eric Newcomer, who has an excellent Substack about the VC industry you should subscribe to if you haven’t already. Speakers included Stability AI CEO Emad Mostaque, Quora CEO and OpenAI board member Adam D’Angelo, Databricks CEO Ali Ghodsi, and Hugging Face CEO Clement Delangue.

Below, I have notes from the day’s onstage talks and hallway chatter, including fears about AI killing us all, how people really feel about Elon Musk’s open letter, why open-source AI is having a moment, and more.

Early in the day, a startup founder mentioned that the term “generative AI” wasn’t a thing until the second week of her company existing. Her company is nine months old.

While the industry is still incredibly new, it’s quickly sorting itself out into two general categories of companies: the “picks and shovels” providing fundamental platforms, which includes LLM providers like OpenAI and Anthropic, and the companies “digging for gold” on top of them by building consumer applications.

While AI work was largely cordoned off to the world of academia for years, it is quickly becoming productized in what is now an era of “proliferation.” As the CEO of Adept (and a former OpenAI VP), David Luan, put it, “We have entered the industrialization age of AI. It’s now time to build factories.”

The companies building factories already are seeing insane growth. Hugging Face, which is essentially GitHub for open-source AI models, already has 3,000 customers paying for premium features out of 15,000 total companies using its platform, according to CEO Clement Delangue. Meanwhile, Databricks CEO Ali Ghodsi said that over 1,000 customers are using the company’s open-source LLM called Dolly. It was announced a week ago.

While a lot of the attention has been on using AI for text, an area that is moving particularly fast is generative AI for video. Hugging Face distributed the model that made a recent viral video of Will Smith eating spaghetti. While that particular video isn’t fooling anyone (it’s pretty disturbing actually), Delangue made the prediction onstage that text-to-video models will start producing content at the same level of quality as what ChatGPT does with text in just a “few months.” The implications of that prediction coming to pass are hard to comprehend.

While it’s clear that OpenAI is the industry leader right now, no one seems to think that will necessarily be the case forever. There’s a camp that’s saying these models will be commoditized and another saying it will be a winner-take-all market.

“It’s pretty clear right now GPT-4 is in the lead,” Quora CEO and OpenAI board member Adam D’Angelo said onstage. “But the entire field is moving so fast it’s very hard to know where things will go in the future.”

Quora recently put out Poe, a web and iOS app that lets you use GPT-4, Anthropic’s Claude, and other models in a simple chat interface. Though D’Angelo was quick to acknowledge his conflict of interest with OpenAI, he was clear that he isn’t in the winner-take-all camp. “If it was just that they are going to be totally dominating everything, I wouldn’t be building Poe.”

In another panel, Kleiner Perkins partner Bucky Moore compared the current state of AI model competition to the beginning of the cloud industry, which resulted in Amazon Web Services not being able to keep up with all the form factors companies wanted. “The surface area of how all this tech is applied is just so impossibly broad for any one company to cover,” he said.

The wild card here is how successfully OpenAI may or may not be at becoming more than just a model provider to other companies. With the recent launch of its platform for outside developers to hook into ChatGPT via plug-ins, its aspirations are clearly to also be (to borrow Ben Thompson’s parlance) an aggregator of attention. “It may be in a world where OpenAI is providing this powerful model and developers are building on top of it,” said D’Angelo. “A diversity of AI apps is not in conflict with OpenAI having the best model.”

I found one of the most interesting parts of the day to be D’Angelo talking about his vision for Poe to become akin to a web browser for discovering AI models. He made the comparison to the early days of the web, when people thought Yahoo was the whole internet. Eventually, people learned to understand websites as destinations, and browsers became critical for discovering them.

“With Poe, we want to be similar to a web browser in that we want people to access many different forms of AI,” said D’Angelo. “We want to get into recommending the best bot for people to talk to for their use case.” He said Quora is building an API to let other LLM providers easily integrate their bots into Poe. “I think there’s going to be this massive ecosystem similar to what the web is today… I could imagine a world in which most companies have a bot that they provide to the public.”

In the here and now, there is a race underway for AI companies to find distribution through companies that already reach millions of people. No one is under the illusion that another startup interface like ChatGPT will attract that many eyeballs as quickly as OpenAI did. According to Moore from Kleiner Perkins, there are lessons to be learned from how the software enterprise market has played out. Bundling is powerful. “Slack and Notion have bumped up against the benefits of distribution… If I was competing with Microsoft that would scare me a little bit.”

Of course, Microsoft has already picked its horse in the race with OpenAI, and the message throughout the day was that it won’t be long before all the key AI providers have linked up with the key Big Tech players. As Stability AI CEO Emad Mostaque put it, “You partner with people who can give you scale because you won’t be able to build scale in time.”

It’s clear that the people working on generative AI are uneasy about the worst-case scenario of it destroying us all. These fears are much more pronounced in private than they are in public, though there was definitely an underlying sense of unease throughout most of the panels. Early in the day, nervous laughter filled the room after a founder quipped that “we’re going to be dead” by the time OpenAI is on GPT-10.

There doesn’t seem to be a shared definition for artificial general intelligence, which many, including OpenAI’s Sam Altman, believe to be the end state of this work. Ghodsi from Databricks argued that we’ve already achieved it, while others demurred on how long it would take to build an AI that is truly as advanced as humans. At one point, a VC nervously laughed at the prospect of a model being trained on his writings to create the perfect investment pitch he would say yes to. “I am concerned about that both personally and globally,” he responded.

It was striking to talk to founders building AI products who sincerely believe in the doom and gloom narrative, like one who said over lunch with a straight face that he is savoring the time before he is killed by AI. “Does anybody here want to build something that could kill us all?” Stability AI’s Mostaque asked the room at one point. Thankfully, I didn’t see any hands go up.

Perhaps the most talked about topic of the day was the open letter published earlier this week asking for a temporary halt on generative AI research. The fact that Elon Musk signed it and his foundation is a key funder to the group behind the letter has led many to believe that there is an ulterior motive at play, given the reporting that Musk is gearing up to compete directly with OpenAI.

“I think there’s a variety of different motivations people have for suggesting [the temporary halt in development],” OpenAI board member Adam D’Angelo said. “I hear he [Musk] wants to build his own model,” Databricks CEO Ali Ghodsi said later in the day when asked about the letter.

“The pause for six months seems destructive,” Bucky Moore from Kleiner Perkins said. “At the same time, there are serious problems in front of us we need to solve.” The most prominent speaker for the day who actually signed the letter was Stability AI’s Emad Mostaque, who said he did it to send a message.

“I have no problem with private models,” he said. “What I have a problem with is that they [OpenAI] have a technology that they say could wipe us out… and there is zero transparency or governance there… If there’s ever a time to discuss it, it’s now.”

To be sure, there is reason for alarm. “The main concern is the concentration of power,” Hugging Face CEO Clement Delangue said later in the day. “We’re in a fast moving field that needs some regulation. We need more transparency into training sets.”

An early-stage AI startup founder mentioned at one point that her co-founder wasn’t allowed to use Microsoft’s Copilot for coding when he worked at Meta. Beyond the obvious need for a counterweight to concentration of power, that pretty much sums up why open-source models are going to only grow in importance.

OpenAI’s entanglement with Microsoft has already made room for competing model providers like Anthropic to attract customers who are wary of Microsoft as a competitor. Many of the panelists expect the biggest companies to eventually create and implement their own models, though there was a general agreement that open-sourced AI, as a category, will win out in the long run.

“You need open source models in your cloud you can use for proprietary data,” Emad Mostaque said during his interview onstage. “You can’t send your private, regulated data to black boxes. it just doesn’t work.”

His firm, Stability AI, is perhaps the most influential industry player in open-source AI right now, having kicked off the creation of Midjourney and some of the other biggest generative tools in use. An iOS update in December added Stable Diffusion support for Apple’s Neural Engine, which many think will accelerate the development of on-device models as well.

With rumors of a huge funding round in the works, it was clear Mostaque came to the conference to beat his chest. (After his interview was over, I noticed he beelined right to Ron Conway, who was sitting in the audience.) “I’m going to IPO,” he proclaimed at one point. “I’m going to build one of the biggest and best companies in the world.” What about showing a business model first? “You’re going to see my business model properly in a year or two when I give away all my secrets.”

I was frankly amazed that no one onstage had much to say about the growing copyright concern around generative AI. “I don’t think anyone has the answer right now,” Cristóbal Valenzuela, the CEO of an AI startup called Runway that generates videos and images. Stability AI is being sued by Getty for allegedly using photos it didn’t license to train its models. Emad Mostaque declined to comment on the case itself beyond pointing out that he hired “the person who wrote the book on fair use leading our defense.”

There was a lot of talk throughout the day about the importance of letting people opt out of being included in model training, though it’s clear copyright has been an afterthought for the industry to date. I have a feeling that will come back to bite everyone who isn’t already taking this seriously.

The buzz around AI has really breathed new life into San Francisco in a visible way. I talked to an AI startup founder who lives in Toronto and has started coming down to the city multiple times a month to meet with others working in the space. It has given SF that “you have to be here” energy the city has sorely been lacking.

There are multiple AI-related events and mixers seemingly every night, and many of the people at the conference I attended said they saw each other at events earlier in the week. The general excitement around AI is just too infectious to ignore. “It feels like the rest of tech is black and white,” Deepa Seetharaman, a veteran tech reporter for The Wall Street Journal told me during a coffee break. “And this is like seeing in color.”

“This has been a symbiotic kind of relationship” – Apple CEO Tim Cook speaking at the China Development Forum in Beijing last weekend.

The timing of his visit is incredible: just a few days earlier, TikTok’s Shou Zi Chew was skewered in front of Congress for four hours. Cook, meanwhile, was greeted to applause and a friendly meeting with China’s minister of commerce. It was his first visit since 2020, and this year marks the 30th anniversary of Apple having operations in China.

Cook deserves credit for building and maintaining such a delicate and lucrative relationship with China. It will go down as one of the greatest business relationships of all time. At the same time, I’m still amazed that Apple’s brand in the US remains untainted by the censorship it does on behalf of the Chinese government. Cook’s balancing act is something to behold.

That’s it for this edition, which was brought to you without the help of AI.

I’ll be back next Thursday. In the meantime, if you have any feedback on this edition or stories I should be looking into, let me know. Thanks for subscribing.