Can Large Language Models (LLMs) predict founder success?

Generate my results

Can LLMs predict the future for startups? Sounds impossible, but is it?

We published our first research report outlining the founder traits most associated with company success a few years ago. In the last few months, we have experimented with using LLMs to try to predict company outcomes, based on founder psychology alone (yes, we know there are 1000 other variables).

This could be a big deal, if it works. It could mean that for any given company, we would have better odds in identifying winners, and founders would know how to increase their odds of success. Even if the predictions are only 20% better, the difference could amount to significantly stronger financial outcomes for all stakeholders and an overhaul of the way startups are funded.

Here’s a little more on what we did and what we learned. At the end, you will be able to enter your own data and immediately see what your odds are based on our model.

Generate my results


By using founder psychology attributes as the only inputs, the outcomes predicted by the model had a moderately positive correlation to our actual portfolio financial performance (IRR & MOIC).

What makes a strong founder with the highest chance of success? Fast and continuous learning, strong founder-market fit and true conviction are the most essential ingredients.

Not all successful founders rate highly on all variables. For example, storytelling can yield advantages in hiring, fundraising and selling, but while more money gives founders more shots on goal, they still need to score.

Of course market externalities (co-founder splits, regulatory changes etc) can kill even the strongest companies. Conversely, weak companies can still find success, propelled by exceptional demand.  


Our original report laid out a framework with six founder archetypes associated with successful startup outcomes. We received lots of emails from founders excited about this approach as it focuses on who they are as people (vs. who or what they are associated with). Many self-identified with our named superpowers. Since then, Large Language Models (LLMs) have undergone major advancements and we wondered: can we make this even better?

We found ourselves asking:

Experiment Setup

We built a dataset from our portfolio of 37 companies from our first fund (it is a small sample as we are a new firm; we would love to train with your data if you are willing to share with us). We recorded how each company performed longitudinally based on financial metrics.

We trained our LLM based on a training set of our founder data. We tasked our model with predicting the likelihood of success for the remaining companies, based solely on our dataset of founder attributes.

We repeated this for each company ten times to simulate randomness in startups. We then ranked their simulated outcomes and compared it to our actual portfolio rankings.

Attributes LLMs Found To Matter

We were pleasantly surprised to find that our LLM was able to successfully find patterns in the data that were predictive of company outcomes. Among the founders ranked highest by the LLM, here is the relative importance of each founder attribute:

This replicates our prior research! For any given company, we can simulate company outcomes with a reasonable degree of accuracy, given reliable founder data.

Universal Superpowers

Similarly to last time, we found that quick-learning, or the ability to continually learn and rapidly iterate, is vital, especially in a market where there is little margin for error.

The most successful founders have strong founder-market fit in their respective fields, with unique insights and experiences that give them an advantage solving a given problem.

True conviction — a new variable this time — is both rare and differentiating. It goes beyond believing, it means willing an often non-obvious vision into existence.

Nuanced Superpowers

The simulation also captured the nuanced superpowers accurately. Many types of founders can be successful.

For example, while humility is largely associated with success, the lack of humility can be counter-balanced by high customer empathy.

Similarly, storytelling can give companies an advantage in hiring, fundraising and selling, but ultimately it only goes so far and its usefulness can start to fade in the long-run. It gives founders more shots on goal, but they still need to score.

Company Deep Dives

To help contextualize our findings, we dug into the stories behind a few of the outliers in our portfolio. To protect their identities, we have anonymized the companies.

At company X, this founder has impressed us from day one on their exceptional execution ability, speed of learning and steadfastness in sticking true to what they believe in. They are consistently over-delivering. Their execution is top 0.1% of what we have seen, day in and day out for years.

They stand out in their focus and prioritization of business drivers. The product-market fit (PMF) is so strong that the company gave their customers something that they did not know they wanted.

Company Y had one of the biggest deltas between predicted and actual rankings. They are predicted to be doing well, but they performed poorly in reality.

This founder has an impressive and pedigreed background, checking all the traditional VC boxes. Over time, we realized that there is a gap between their perception of themselves vs reality. The expectations for them are so high, both externally and internally, that it became hard for them to learn and grow. This company has never once hit a target.

This departure from reality manifested in hubris, an inability to learn, and ineffectiveness in managing the company day-to-day, leading to failure.

At Company Z, the founder’s exceptional conviction — even though unpopular — impacted their trajectory towards success.

The company did not hit product-market fit right out of the gate and, in fact, pivoted multiple times. The founder had conviction and grit to pursue an unusual idea in a daunting market they had little experience in. This conviction propelled them through challenging periods, even when the market seemed to defy logic.

Where LLMs Differed From Actual Data

We did find some surprises in the data. Some predicted high-performers are actually struggling, often caused by tough market conditions, co-founder splits, and other external factors.

Conversely, a few companies our model predicted to struggle are actually thriving. Many of these have been pulled by strong market demand, despite their clear weakness in execution.

What About Me?

At this point, you must be wondering, what about me? Or will so-and-so be a successful founder? Play with our model! Given our model is trained on ratings given to founders by close third parties, your results will be most accurate if you ask someone relatively unbiased, who you’ve worked with closely.

This is part of our continued efforts to understand founders, beyond the most obvious attributes, to help us be better partners. Our goal is to pursue truth and greatness, and to keep learning, together with our founders. Please give us feedback.


We just closed a new fund with $185M dry powder to focus on finding and growing with founders who are going to change the world. Since the founding of the firm, we have believed that technology that generates outsized value will transform our lives and improve productivity from offices to factories and farms. Some call the driver AI, but we are just as at home with those who say mathematical models and statistical simulations.

We are a new firm built from the ground up, enabled by the latest technology. We believe in a computational approach to understanding our world, rooted in an understanding of human psychology — empowering the best founders to build is just step one.

This experiment was inspired by Generative Agents: Interactive Simulacra of Human Behavior and Using Large Language Models to Simulate Multiple Humans and Replicate Human Subject Studies.