On why AI is coming for my job next

I’ve written dozens of posts over the years on Y Combinator startup trends and how we use data and algorithms at Rebel Fund to invest in the top 10% of new YC companies. We’ve invested millions of dollars into our data automation infrastructure, proprietary machine learning algorithms, and internal software to quickly identify and invest in the best new YC startups each year. Not only do we invest in hundreds of the most cutting-edge technology startups coming out of Silicon Valley, but we also invest in developing and using the most cutting-edge technology possible to do so.

It should thus be no surprise that we’ve been keeping a close eye on the latest developments in AI, and figuring out how to merge its fast growing capabilities into our existing technology infrastructure. In this post, I’ll explain some of the mind-blowing capabilities that AI has developed over just the past few months for investors, and why I think in the near future, early-stage investors who don’t leverage AI in their due diligence and selection process will be left in the dust by those who do.

First off, if you’re not already familiar with our proprietary Rebel Theorem 3.0 machine learning algorithm, I’d suggest reading that post first, since it gives some background on how we’ve used millions of data points to train an ML model to accurately predict seed-stage YC startup success.

When we developed the previous version Rebel Theorem 2.0 in 2022, we were limited to using data features that were either easily quantified (like years of founder education or work experience) or that could be easily coded numerically (like where a founder studied or worked in the past). By the time we released Rebel Theorem 3.0 in 2024, we could use ‘older’ AI models like GPT-4 to incorporate new AI-generated data features, like inferred founder personality profiles or demographic characteristics based on other data we had collected on them. We found that each time we re-trained the algorithm with deeper data insights on companies and founders, the more accurately it could predict startup outcomes.

The most recent advance in AI is a new series of reasoning models, like OpenAI’s o1 and o3 models, which can think more deeply, logically, and flexibly than their predecessors. These have unleashed yet another set of new features that we can incorporate into our Rebel Theorem algorithm, bringing its predictive performance (and presumable the quality of investments we can make with it) to even greater heights.

These models have created an entirely new set of capabilities for early-stage startup investors who know how to use them properly. Because they take the time to think through inputs logically and consistently, given the right training, prompts and context, they can accurately answer more nuanced and complex questions like “Does the CAC to LTV ratio for this startup make sense given its GTM strategy?” or “Is this founding team’s experience aligned with their product vision?”

In our early testing of new AI reasoning models’ output, we’ve been blown away by not only their performance vs previous generation models, but also our own human experts (including me). Below is a humbling chart that slows the correlation between OpenAI’s o1 reasoning model’s predictions of Rebel portfolio startups valuation growth vs our internal partner ratings, based solely on information available on each startup at the time we invested, presented to the model in a certain structured format that we developed internally:

At Rebel, we’re a top-tier venture fund with some of the smartest partners in Silicon Valley, co-founding companies now valued over $100B in total and investing of hundreds of startups together, yet a properly trained AI reasoning model fed the proper context on a company and its founders, with the proper format and prompting, can make better predictions of startup outcomes than even us. In fact, I don’t think any human investor can make such highly accurate valuation growth predictions at such an early stage of a startup’s lifecycle.

Even more impressive than the model consistently outperforming ratings from me and my partners is that its predictions show an increasingly strong correlation with startup valuation growth for earlier vintages, which have had more time to mature (or, in data science terms, more time for valuation outcomes to disperse). This aligns perfectly with what we’d expect from a model genuinely effective at forecasting startup success.

In another backtest, we trained the o1 reasoning model to assess and score historical YC startups across various dimensions of interest to early-stage investors like us. This is very similar to how human investors think through whether to invest in a given company.

As you’ll see in the chart below, not only were the reasoning model’s overall scores more predictive of startup success than our internal partner ratings, but some individual scoring dimensions like team quality, rate of progress, and product strength were much more predictive. As an experienced VC with 300+ startup investments under my belt, I noticed the dimensions where the model ratings were most predictive align perfectly with those that I’ve learned over the years to weigh more heavily when making investment decisions.

Now that I’ve hopefully got your attention on just how powerful these new AI reasoning models are, let’s talk about the implications for early-stage venture capital investors in the near future.

1. VCs can no longer afford not to use AI in their screening decisions

I’ve long felt that using data, technology, and algorithms has given us an edge at Rebel, but I now believe that using them are table stakes for VCs. We already live in a world where the latest AI models, if used properly, can make better startup success predictions than not only the average VC, but the very best VCs, and that gap will only widen as the underlying models evolve, and we get better at model data preparation, fine-tuning, prompting, etc. Remember, this is the worst these models will ever be.

This is not to suggest that VCs should rely entirely on AI models in making their investment decisions — at least not yet. At Rebel, we use ML/AI heavily in our startup screening process, but every investment decision is still ultimately made by a partner, as there are still nuances that AI can’t assess. We’ve been slowly incorporating more data signals into our algorithms as we backtest and validate them, and slowly handing over more influence to our ML/AI models as they prove themselves better at certain aspects of diligence than humans, and will continue doing so indefinitely.

2. VCs must start viewing ourselves as technologists, not just investors

Most venture funds view internal use of technology as an afterthought, despite our universal conviction that technology is the future. There’s an irony in manually drafting an investment memo for a cutting-edge AI startup investment in Microsoft Word (at Rebel, we use AI for that too)

Venture investing is often viewed today as an artisanal craft, requiring years of apprenticeship to master. That was true before the advent of advanced AI, but it’s not true today. VCs should stop thinking of ourselves as trained masters of our craft with impeccable tastes, but rather as orchestrators of a technology that’s far smarter than we are.

At Rebel, we’re uniquely situated to do this for YC startup investing with the most comprehensive dataset that exists of YC companies, founders, and outcomes for model training, plus a top-notch data science and software engineering team to make use of it — but every fund should hire at least one engineer and start training, backtesting, and using their own models based on their own data and investment strategy.

3. LPs need to start asking fund managers different questions

I’ve raised hundreds of millions of dollars from LPs over the years, so have a pretty good feel for the types of questions they ask venture fund managers. These questions tend to cover some combination of fund thesis, partnership team, portfolio strategy, track record, deal selection process, access advantages, GP experience/network, and portfolio support.

These all are relevant topics to cover, but noticeably absent are questions around data & technology. In a world where advanced AI can outperform the best human investors in many regards, LPs should start getting comfortable asking questions like:

• How do you use ML/AI in your screening and due diligence process?
• What is the background of your data & software engineering team?
• How much do you spend on AI inference and which models do you use?
• How would you describe your data automation infrastructure?
• How much have you invested into R&D since fund inception?

The top-performing venture funds in the world of AI won’t just invest in cutting-edge technology companies, they will be cutting-edge technology companies, and LPs will need to learn to diligence them as such. In this strange new world, a manager’s investing track record from 5+ years ago won’t be nearly as predictive of their future success as how well they’re harnessing AI technology today, at least in terms of startup due diligence and deal selection (though overall fund strategy, GP networks, deal access, and portfolio support certainly still matter)

I’ll conclude with one final thought for my fellow early-stage venture investors who remain skeptical: If you believe AI will revolutionize every industry, why wouldn’t it revolutionize ours?