We invest in AI companies with a clear differentiated advantage, not just a thin wrapper on a company's foundational model’s API.

The market is chasing too many thin wrappers on LLMs that have been trained predominantly on the public internet, which is fundamentally the wrong architecture for many of these applications. LLMs feel like pure magic, unless you are technical enough to understand what they can or can’t do, it’s too easy to conclude “they’ll be able to do everything eventually.”

In particular, many people are using LLMs for health care applications, diagnosis, etc. That said, training an LLM on the public internet and publicly available data is not the way to do this. You actually need specific diagnostic data that is not public in order to train a virtual healthcare provider.

If that were not the case, one could go to medical school, read textbooks, and be a physician. But you can’t—you need a residency to see how patient order flows, imaging, labs, and EMR data all interact in real life. This is why we invest in companies like Knit Health or Dandelion Health, which have unique, differentiated data that more closely mirrors what doctors learn in training, allowing them to build fundamentally different, world models, compared to the limitations of generic LLMs.

I think that all VCs claim to be the most helpful, so why should LPs believe us, or any other firm?

In our experience, firms that are led by people with deep operational experience, not just a couple of years as a junior PM, are the ones that add true value. Amongst our team, we’ve run companies, led thousand-person engineering and product orgs, hired (and fired) hundreds, led PR/marketing/brands for major consumer brands, built end-to-end AI products, and used the latest products to write code on the weekends.

At the risk of tooting our own horn, we are often told “we’re the most helpful firm on the cap table” or that “we punch way above our weight class.” We’re proud that founders consistently reference us highly on this point—but it’s hard to just believe us when we say it.

We recently looked at a company with high-pedigree founders working on data services for large AI training companies. At first glance, strong founders, playing in an area with potentially large budgets in AI, seemed compelling. They even had an early seven-figure contract. Very impressive early momentum.

That said, we’re always on the lookout for what early signs signify the start of an impressive and durable growth curve, vs. what early signs will top out early, too soon to grow into a big company. One thing that matters a lot to us as well is founder speed, and the depth at which they can go into very detailed responses to our questions. So when founders are slow, or aren’t able to go very in-depth on details of their GTM, it ends up reinforcing the thesis to pass.

This was a hard decision for us, but after talking to similar buyers at other (non-early customer) companies, we walked away, after learning that there probably wasn’t enough strong buying demand to build a massive business here.

Using all these new coding AI assistants last summer, I was convinced that the “we don’t need junior software engineers anymore” argument was overblown. I thought it was just a narrative pushed by VCs who were trying to hype their companies.

I build side projects and internal tools mostly for fun, but also to stay current. My experience was that the tools felt like having an idiot-savant for a colleague: occasionally amazing, and then, 30 seconds later, completely incompetent (“How can we leave this tool undersupervised or expect it to do anything non-trivial!?!?!”). Anyone who has used these tools can relate.

That said, all of this changed last fall, primarily with the release of Opus 4.5 from Anthropic and similar models from competitors. Almost overnight, the ability to give it much more substantial tasks, and not get stuck in infinite debug loops on the wrong track, changed dramatically for the better.

I’m now convinced that the market for junior engineers is collapsing. At the same time, the amount of software being built, and the barrier to do so, has drastically changed. The need for higher-level software engineering is greater than ever. It’s very unclear how we will train people to be senior engineers without the historical apprenticeship of being junior engineers.

There is too much money chasing and too few potentially durable and long-lasting AI companies. Unfortunately, there are too few VCs that understand the technology well enough to discern what is really novel and defensible, versus who is just any departing engineer from OpenAI that wants to command a $1B valuation pre-revenue.

Technologically speaking, we are far beyond the AI winter of the last decades. That said, we might be entering an “AI funding winter.” If some of these companies blow up spectacularly after commanding massive valuations, investors who don’t know better will wrongly conclude “we shouldn’t invest in this space”, making it harder for great companies to raise follow-on capital. We advise all our companies to always have some knobs they can turn to get to at least break-even, at the expense of growth, should that come to pass.

Some of these new companies are founded by truly world-leading AI experts and researchers, but unfortunately many others (and even some of those) will flop. A related contributor to many startups blowing up will be how many use cases are captured by the foundation model companies themselves.