The Exchanges

Q then the question there is, you know, all these companies, all the startups are built on top of these model companies, right? And these models are also learning what are the workflows that these startups are, are doing. And we have seen in the past, right? For example, with what happened with Windsurf, right? And there are more examples that these model companies are now launching their own vertical Absolutely.

A So look, that's, you know, it's a full circle to where we started off, 30 minutes ago, which is, Every startup in the AI application space, every, you know, service as software startup has to first and foremost internalize that their vendor is also their competitor. The model provider is both vendor and competitor. And that's a hard place to be. And they are, they are actually educating their vendor on how to compete with them. You know, that's, that's the inherent challenge. And so, if you recognize that challenge up front, then you have to start thinking about what is the data that you're not going to send to them, send to the model provider? What is the data for which you're going to use an open source model versus a closed source model? What is the role of reinforcement learning in sort of post training of a model for a customer specific situation? So, for example, I have a company called Player Zero, And while they do use GPT-V and other, ah, proprietary models, the core code base of the company for their customers, what they, what, what Player Zero does is it helps you debug problems in your code, and it actually simulates a pull request, so you can proactively identify likely problems. For both of those, uh, they're using reinforcement learning extensively, so they build a graph of your code. That graph of your code, they don't share, because no one wants their code to g…

AI assessment note: “Every startup in the AI application space... has to first and foremost internalize that their vendor is also their competitor.”

Q But coming back to the infra, like LLMs have short-term memory. Context graphs require long-term persistent memory. So do we need to solve memory first before we talk about context graphs?

A Look, memory is inherent in context graphs, but, and memory is inherent in technology. So when you say, do we have to solve memory? I think we have solutions. A database is memory. You have relational databases which capture a point in time. You have evented stream databases. There are databases which capture, you know, the time dimension of memory. You have structured and unstructured databases. So I think many of the components exist. There is also a ton of innovation that is being done and has been done around model memory, around memory between, you know, how we have short-term state versus long-term state. And yes, there are many, many fundamental questions about memory that are unsolved. So what makes this opportunity exciting and challenging, both at the same time, is the context graph of today may not be the context graph of tomorrow, because the stack is changing. And innovations in memory, whether it's model memory, or agent memory, or system memory, whether it's short-term or long-term memory, innovations in memory will change the nature of the context graph.

AI assessment note: “So when you say, do we have to solve memory? I think we have solutions.”

Q Are you using any agents which are autonomous right now for yourself?

A You know, I experiment with a lot of things. I have agents that, you know, look at my email. I have agents that read my email. I have agents that help me, you know, prepare for meetings. So I'm trying a lot of things. Autonomy is a very, you know, broad term. Am I comfortable today letting an agent respond to an email on my behalf? The answer is no. But that's also a function of the fact that, you know, my work emails have value. It's like, you know, if I told a founder by, you know, the agent by mistake said, yeah, we're ready to sign a ground sheet. That would be problematic. Uh, but there are, you know, the value you place on emails is different for different people. So I think people will start using agents to respond to emails sooner than I would.

AI assessment note: “Am I comfortable today letting an agent respond to an email on my behalf? The answer is no.”

Q the companies winning today or seem to be winning like hundred mill era today are more attacking the green field right now, which is building new use cases for prosumers or consumers like lovable, open AI, carved out new budgets for people as well as companies, right? So should founders attack the existing industries and try to replace ServiceNow, Salesforce, or should they focus more on the green field side?

A You know, I think the, the beauty of technology is that both are opportunities. If you, if you look at, if you, if you look at, ah, code generation tools, where there has been the most zero to a hundred, you know, the fastest zero to a hundred ramp have been code generation tools. You know, I think you could argue it both ways. You could argue that code generation is a new category, because no one could generate code before it. There was no lovable before lovable. Or you could argue that Lovable is, you know, eating into categories like Wix and, you know, other site creation tools. Because we've had traditional configuration type of tools that allowed you to do self-service website creation or self-service mobile app creation. So what is Greenfield versus Brownfield itself is unclear in today's world with AI. But ultimately, I would say the fastest growing companies find a wedge in the market that is high value. They find a wedge where, in addition to being high value, the pain is acute enough that people are willing to take some risk and, you know, sign a check quickly. Now, sometimes those checks, in the case of Lovable, are relatively small, and you need many, many, many tens of thousands, even hundreds of thousands of customers to get to a hundred million. And in other cases, you know, the checks are large, and you have more of a traditional enterprise sale, and you can get…

AI assessment note: “I think the, the beauty of technology is that both are opportunities.”

Q And when you wrote this article, why did it hit such a, such a nerve? Why did it go viral?

A So Jay and I, and, and, you know, Jay has been my sort of co-conspirator on this for, for the last year. For the last year, we've really been asking ourselves this question of, uh, why Why have agents struggled to realize the potential? The promise of agents, you know, which we were all making 1218 months ago, uh, has struggled. And while we are seeing improvements, we're seeing improvements in single player mode. We're seeing improvements in single chat mode. Uh, what you really want is a multiplayer mode with state and memory across time and more automation. Uh, we have seen our portfolio companies Work through these issues. And as Jay and I worked through this with companies like Maximoor in finance automation, Player Zero in production engineering automation, uh, and many others, what we saw was that there was a common thread to the answer. And that common thread is the context graph. And so our belief is that the most successful AI startups of the future will be those that are able to build a moat by building a unique context graph.

AI assessment note: “we've really been asking ourselves this question of, uh, why Why have agents struggled”

Q So in, in many cases, uh, this context is a moat of an organization, right? So why would an organization, you know, typically try to share it with the startup that they are working on? Because then, you know, the moat is, is out there.

A Look, the most valuable asset, Of a company's context graph. And the reason context graphs have not been stolen was it's very hard to build, as you said, and companies with organizations are rightly nervous about, hey, if I let everyone build a context graph of me as a customer, what is my organizational value? So first and foremost, you know, context evolves, context will evolve, and so my first response to that is, relax, human beings are not that easily replaceable. That said, on a more serious note, I do think data security, data governance, Uh, access controls. Which agents can access what parts of a graph? See, the downside of stitching everything together is now it's all available.

AI assessment note: “relax, human beings are not that easily replaceable. That said, on a more serious note”