The AI Moment Part III: Application Layer
Does Microsoft just win everything?
Since we published Part II, OpenAI finally released GPT-4 after a long-awaited and often-frenzied hype cycle. The early results are impressive, including improved performance on exams, less hallucinations, etc. although I’ve yet to see anything that makes me think it is a step-function in intelligence to GPT-3. From my perspective, scaling in large language models has begun to show more predictable improvementthat is beginning to taper off and show diminishing returns.
Some mounting evidence for this includes Sam Altman’s recent pronouncement that Open AI is not currently training GPT-5 and is instead focused on smaller models that do a better job (ie, solve use cases more cheaply). While some took this to mean OpenAI is so ahead of the field that they can afford to rest on their laurels, my view is that they are already seeing limits to scaling and that GPT-x is not going to bootstrap itself into an AGI.And as I was writing this piece, Sam just announced that he feels like the era of giant AI models is already over and new ideas will be needed to drive AI forward.
With that being said, GPT-4 is likely a better hammer for the productivity nail that OpenAI is currently focusing on for its enterprise use cases. Microsoft is starting to roll out some integrations while other incumbents like Notion use OpenAI competitors like Anthropic to power their own workflows.
After looking at the infrastructure layer in Part II, we’ll now go to the two-part questions around the application layer for Part III.
Does value accrue to startups or incumbents applying AI?
What startups might make sense given our views so far?
An amazing piece trying to answer the first question was written by Elad Gil here. You should go read it, but here’s the TL;DR:
Every technology wave has a split between incumbent winners and startup winners
Incumbents were the big winners in the previous AI wave.
This wave might be different due to 10x better technology which creates new infrastructure and workflow opportunities that are not currently well-served by incumbents.
We’ll try not to rehash too much of what was said in Elad’s article here and keep it focused to the how we see the incumbent-startup landscape shaking out and where interesting opportunities may lie.
Beating incumbents as a startup is hard. Marcus Ryu from Guidewire memorably puts it this way, “When you’re a startup, you have no assets like a brand, product, reputation, references, or capital…the one asset you have is you can nail the headpin of demand more precisely than incumbents. If you get that right, that may be worth more than the rest combined.”
While this approach worked really well for Guidewire in a legacy industry, the issue for new startups looking to use AI to enable workflows is that the “incumbents” in the productivity technology space are fast-moving enterprise (Microsoft) or late-stage startups that are eagerly adopting and using the technology (Notion, Glean, etc.) It’s been repeated ad nauseam, but it’s worth nothing that every technology investor fears a repeat of the Slack-Teams story.
With their close ties with OpenAI, the question of “What if Microsoft does it?” feels less like a dumb VC question and more of a real threat that comes up in every single investment committee conversation.
My current views on enterprise applications can be summarized as follows:
Microsoft will win the majority of large enterprise and F500 use cases for the AI-enabled productivity suite.
There will be a non-Microsoft ecosystem serving the tech world in particular, with potential winners like Notion, Replit, Glean, etc. filling various holes of the stack. Mid-2010s late-stage startups are moving fast to adopt this technology and have a giant head start in distribution and data over new startups. It’s hard to imagine new startups competing well versus them without some sort of truly unique insight.
Systems-of-record like Salesforce have the opportunity in front of them to use their entrenchment to start eating workflows built on top of them and re-bundle customers into their ecosystem. That being said, this may be more obvious in theory than in execution. We’ve seen more than one sclerotic incumbent fumble obvious opportunities. When you look at a company like Salesforce specifically compared to Microsoft, there’s less to suggest there’s been a cultural reinvigoration to ship and move quickly once again.
Here’s where new startups could potentially thrive instead:
Highly regulated industries with difficult-to-get customers like healthcare, finance, and government
These industries often have strict restrictions around privacy, data security, etc. which make them less likely to integrate with a vendor like OpenAI. Even if OpenAI begins to change its data collection and model improvement strategy, it’s reasonable to expect that serving these types of difficult customers won’t be a burning priority for some time. There’s also an interesting moat in getting these type of difficult customers: once you’re in, it’s very, very hard to rip you out.
Net new opportunities not well-served by fast-moving and AI-adept incumbents
Education: Maybe the one I’m most excited about currently. Homework and one-size-fits-all education destroys the love of learning. Personalized AI tutors that help students discover their interests and learn at their own pace allow you to envision reimagining education from the ground up. I’d be interested in a Montessori-type school that guides kids down various rabbit holes and follows their interests rather than following a particularly strict curriculum. Regardless of your theory of education, AI plus software enables startups to potentially scale to millions more kids and potentially help solve Bloom’s Two Sigma problem. Adding an in-person layer on top of the software may decrease margins but probably creates more sustaining value for students and increases defensibility.
Biotech (drug discovery + creation, better clinical trials, better management and utilization of new and existing data)
Defense (copilots in the operating field, improved threat detection, etc.)
Robotics (when do we got Rosey the robot in our homes? Self-driving might finally be here, just not evenly distributed, and manufacturing improvements may finally unlock growth that actually shows up in the GDP numbers)
Consumer, which has been a dead zone for VC for the last few years
New social networks need to enfranchise creators who were previously disenfranchised and don’t currently have social capital. While companies like Midjourney and Leonardo.ai are generally viewed as infrastructure tools enabling art or game asset creation, they’re also “franchising” new people who weren’t famous artists or game designers. Someone like me who can’t draw to save his life could become a popular Midjourney creator. A different perspective may be that these companies or any future competitors are actually budding social networks where new classes of users compete for social capital accrual. So instead of competing on who has the best “workflow” tool for enterprise customers, the real question is who gets to network effects and how quickly. It’s very possible that the next real competitor to Tik Tok or Instagram has AI tooling as a core part of its creation workflow.
AI-powered therapy, nutritionists, coaches, or friends, etc (TBD if these end up capturing any value, but I definitely want to explore ideas in this area)
The most obvious consumer use case is one I would never invest in for moral reasons, but it doesn’t take a genius or a perusal of most AI art “playgrounds” to understand what the majority of content creation is.
AI is a powerful technology form factor, but one that is largely enabling for companies with large amounts of data, distribution, and already entrenched in workflows. My general thesis is that the most interesting startups will not be competing versus Microsoft or well-funded late-stage startups particularly in the productivity suite. They’ll instead find whitespace in industries and opportunities not well-served by incumbents.
Some of those are listed above and I’d love to meet any founder building in these areas or even something totally new where only you see the future vision.Email me at pratyush [at] susaventures [dot] com.
From the OpenAI announcement about GPT-4: “GPT-4 training run was (for us at least!) unprecedentedly stable, becoming our first large model whose training performance we were able to accurately predict ahead of time.”
There continues to be much confusion in the debate around intelligence and what exactly that means. Computer programs are already far more competent than a human in a variety of tasks and many have more knowledge than any one human could hold in their head. So what do we exactly mean when we say computers are not *yet* intelligent? I’ll try to briefly explore this more in Part IV, but the TL;DR is I think we need to decompose intelligence into elements like knowledge/aptitude, learning capacity, self-improvement ability, agentic behavior, etc. As Wittgenstein said, the limits of our language is the limits of our mind and we are encountering this head-on in the current AGI debate.
One might protest that Slack was sold for $27B, but that was a blend of perfect timing, zero interest rates and a booming bull market, and a less-than-savvy acquirer. It’s hard to imagine many companies that face a similar existential risk getting that type of outcome these days.
As you can see, I’m heavily influenced by Arda Capital whose views on these topics I’ve found quite helpful and insightful.
Clubhouse, BeReal, Locket, Poparazzi, Dispo…
If you think I’m totally wrong about the future of enterprise application software and that Microsoft and other companies like Notion, Figma/Adobe aren’t in a strong of a position as they seem, I’d always curious to hear a well-articulated argument for why.