What's Next in Phase Two of Generative AI?

0:00Today on the AI Breakdown, we're talking about what's next in Phase 2 of Generative AI. 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 YouTube, and our newsletter. Welcome back to the AI Breakdown. Today, we are doing another trend-type story and exploration that I think aligns with a lot of the things that we've been talking about recently, and that in many ways is trying to sum up some of where we actually are. Now, if you listened to yesterday's show, you heard me talk about Professor Ethan Mollick's essay, The Shape of the Shadow of the Thing.

0:37The way that he started was that he saw us reaching, quote, the culmination of the first phase of the AI era that started with the launch of ChadGPT. We talked a lot about what that meant, about my belief that we had been sort of converging around unleashing and exploring the full range of capacities of GPT-4, and that a lot of the big questions were what comes next. Now, the other big thing, though, that I've been talking about a lot over the past couple weeks based on the news of autumn AI product announcements represented not the race to bigger, better foundation models, but instead the integration of AI into our workflows, into our spaces.

1:12So just to quickly go through some of the examples of that, there was, of course, Meta's announcement, which they framed literally as introducing new AI experiences across our family of apps and devices. This included AI stickers in messaging. It included an advanced conversational assistant that would live inside WhatsApp, Messenger, and Instagram. And of course, it included their 28 characters that were played by a variety of cultural icons that were supposed to bring fun AI personalities to the experience of different users. Then, of course, there was YouTube. At its Made on YouTube event now about three weeks ago, the company once again released a set of AI-powered tools that would live inside its software and would make it trivial for video creators to use AI both to be more productive but also to do new things.

1:54DreamScreen is a feature which can include AI-generated images and video as the background of YouTube Shorts. The company is integrating an AI-powered music recommendation engine. Really powerfully, there will be a new AI dubbing feature that will give creators the ability to put their content into other languages, which if it works well, could totally break down how we think about linguistic barriers. And there's even a new assistant style feature, which will help people generate ideas and even outlines for new videos based on what's working with their audience, as well as the other types of videos outside of their content that their audience is watching.

2:25Again, the consistent theme here is this AI integration era. These are really powerful features, but what makes them extra useful is that they are so fully enmeshed in this creator suite that people are already using. LinkedIn has also announced a set of new AI integrations. There's a personal assistant for learning, new AI-powered recommendations around recruitment, including suggesting candidates that might not necessarily fit with what a recruiter thinks that they're looking for, as well as AI-integrated writing tools for marketing and sales roles. Just this past week, Google Assistant got a barred upgrade, bringing the power of generative AI into that popular mobile assistant experience.

2:59Canva had an absolute slew of new AI tools integrated into their software. Again, in a video earlier this week, I showed off a number of what they call their magic editing tools, their magic media tools. One really exciting thing is their partnership with Runway, which brings text to video generation directly into this suite. And again, just to really beat this dead horse, integration era, integration era, integration era. Canva has 150 million users, and now at the very center of their creative process is this suite of AI tools, which while now kind of presented as shiny new things, almost certainly are just going to start to blend into the background as the way that you interface with Canva.

3:34Finally, there is ChatGPT Vision. The new upgrade to ChatGPT makes it multimodal in a meaningful way. Images have become inputs, which opens up a variety of new use cases, and users can also speak to ChatGPT and have it respond via voice. This isn't an upgrade to the underlying GPT-4 model, but the modalities that it opens up are so significantly different that it feels like an entirely different experience. And that brings us back to Ethan's piece, this idea of reaching the culmination of a first phase of the post-ChatGPT AI era. And it gets me to this recent post by Investor Sequoia called Generative AI's Act 2.

4:09This is just a couple weeks old, and the framing of the piece is this. One year ago, we published a hypothesis that generative AI would become a profound platform shift in technology. Then came the firestorm. So in some ways, this is a reflection on that first post-ChatGPT phase that Professor Malik was also talking about. Sequoia writes,

4:34And yet they said, quickly AI excitement turned to borderline hysteria. Suddenly, every company was an AI co-pilot. Our inboxes got filled up with undifferentiated pitches for AI Salesforce, AI Adobe, and AI Instagram. The$100 million pre-product seed round returned. We found ourselves in an unsustainable feeding frenzy of fundraising, talent wars, and GPU procurement. And sure enough, the cracks started to show. Artists and writers and singers challenged the legitimacy of machine-generated IP. Debates over ethics, regulation, and looming superintelligence consumed Washington, and perhaps most worryingly, a whisper began to spread within Silicon Valley that generative AI was not actually useful.

5:10Now, Sequoia goes on to compare the crooning from critics that followed to the very early days of the internet, referencing Paul Krugman's famous declaration in 1998 that by 2005, quote, it will become clear that the internet's impact on the economy has been no greater than the fax machines. So then how does Sequoia see where we are now? Well, they say, generative AI's first year out of the gate, Act I, came from the technology out. We discovered a new hammer, foundation models, and unleashed a wave of novelty apps that were lightweight demonstrations of cool new technology. They go on. We now believe that the market is entering Act 2, which will be from the customer back.

5:43Act 2 will solve human problems end to end. These applications are different in nature than the first apps out of the gate. They tend to use foundation models as a piece of a more comprehensive solution rather than the entire solution. They introduce new editing interfaces, making the workflow stickier and the outputs better. They are often multimodal. Now I will pause here to quickly go through again the announcements that we have recently seen. YouTube Dream Screen, AI-suggested content ideas, and AI-powered dubbing. Canva's Magic Studio with editing and media creation. LinkedIn's AI tools for writing and marketing.

6:12Meta's in-app conversational assistant that lives inside WhatsApp, Messenger, and Instagram. Pretty good evidence of this thesis of customer-back, right? Now, this piece is really comprehensive from Sequoia and is well worth a read on your own, but I'm going to skip over some parts, including the generative AI market map, the generative AI infrastructure stack, and get to some of their reflections on what they got wrong and what they got right. Under what they got wrong, they list that things happened much faster than they thought. Last year, they write, we anticipated it would be nearly a decade before we had intern level code generation, human quality videos or human quality speech that didn't sound mechanical.

6:43Obviously, we have all of these things. And by the way, although maybe a little bit out of scope of this video, this I think is one of the most compelling arguments from the AI safety folks, that the fact that no one seems to have a sense of how fast these things are going to move and that we have to keep revising our goalposts because things keep moving faster than anyone thinks is possible, is a compelling reason to try to take a moment, to take a breath even, and ask if they're going too fast. But back to Sequoia's reflections, they were surprised that access to GPU was a bigger problem than end-user demand.

7:12They note that it reintroduced consumers paying as an actual business model, which could be a net positive in the long run. There's a couple more, but the other one that I wanted to mention is number five, the moats are in the customers, not the data. They wrote, we predicted that the best generative AI models, companies could generate a sustainable competitive advantage through a data flywheel. More usage, more data, better model, more usage. While this is still somewhat true, especially in domains that vary specialized and hard to get data, the data moats are on shaky ground. The data that application companies generate does not create an insurmountable moat, and the next generation of foundation models may very well obliterate any data moats that startups generate.

7:44Rather, workflows and user networks seem to be creating more durable sources of competitive advantage. This again is something we've talked a lot about on this show, that if the choice is Canva's integrated version of these tools versus a totally new design suite that requires you to go change all of your workflows, even if it's 10 or 15 % better in terms of outputs, might not be sufficient to get users to change their behavior. That's even more true on the enterprise level, where we see a lot of companies wanting to use providers that they already trust with their data rather than trusting a whole new set of startups.

8:12But what about where things are now? One thing that Sequoia notes is that although these tools have grown faster than basically any other type of tools in internet history, a lot of it seems to be driven by novelty. They have a chart which you might have seen floating around Twitter or X around the retention of social media giants versus the new AI apps. Retention of companies in the old school, like YouTube, Instagram, TikTok, Snapchat, WhatsApp, etc., has a 63 % 30-day median. In other words, one month after they first started using those tools, 63 % of them stick around. That goes all the way up to 85 % for YouTube, by the way.

8:42Now for AI-first companies, that average retention is only at 42%. ChatGPT is the highest of that cohort at 56%, but at 56%, that's lower than Roblox, WhatsApp, Snapchat, TikTok, Instagram, and YouTube. Now, I think it's fair to ask if that has to do something with not just the novelty of these new tools, but also the fact that they are tools. They're largely for productivity versus these social experiences like YouTube and Instagram, which are much more about content consumption, but it's still an interesting thing to see written out this way. Sequoia also notes that user engagement is lower.

9:09Consumer companies tend to have 60 % to 65 % daily active and monthly active users, whereas the median for AI-first companies is only 14%. I could probably spend an entire show speculating about why this is and how much of a problem I think it is, which TLDR I'm less convinced than Sequoia seems to be, but I still think they're really interesting statistics. And finally, where the piece leaves off is with an estimation of how entrepreneurs and developers are converging around a shared sense of how to actually build these new applications for maximum utility. They reference, for example, a number of what they call emerging product blueprints.

9:40They discuss new generative interfaces and point out that new form factors are entering the mainstream. They talk about an entirely new set of editing experiences that a variety of different AI companies are getting people used to. And one that will be of interest to many people on this channel, they discuss increasingly sophisticated agentic systems. Generative AI applications, they write, are increasingly not just autocomplete or first drafts for human review. They now have the autonomy to problem solve, access external tools, and solve problems end to end on our behalf. We are steadily progressing from level 0 to level 5 autonomy.

10:09Based solely on download figures and viewer numbers, I think there's a strong argument the thing that people are most interested in, whatever this Act 2 is, is the move from just tools to actual AI agents. Anyway, trying to sum up a bit, what's interesting to me about this is this growing sense among lots of different parties that some first phase of this new world that shifted the moment that ChatGPT was released is coming to its culmination. And the next thing that comes after that is in the early stages of being born. it is inherently a liminal period. And part of the chaos right now is that so much does feel up in the air.

10:44I think that that's reflected positively in entrepreneurial energy and negatively in public anxiety. But in any case, generative AI at its first birthday is screaming forward. And what comes next, whatever it is, seems likely to be even more interwoven into our lives than what we've seen over the past 12 months. Anyways, guys, interesting food for thought for your weekend. I hope wherever you are, you are having a great one. Until next time, peace. Thank you.