Encyclopedia Galactica
Venture capital in the age of LLMs
Marc Andreessen thinks venture capital could be one of the last remaining jobs performed by humans. Meanwhile, David Beisel is worried that AI is already an existential threat to the profession.
The truth, as usual, is somewhere in the middle; Marc may be right about David, and David may be right about Marc.
In essence, an LLM can be viewed as a near-complete library of human achievement. What we have theorised, discovered, calculated and ultimately recorded. What we know about life on Earth, and the little we know about what lies beyond.
Venture capital exists to fund the expansion of that frontier, and profit from it. Never looking backwards, always looking forwards for the next opportunity.
Like the many explorers in history who discovered new lands, all of whom used existing maps and navigation science to aid them in that process.
Through this lens, AI and venture capital exist in harmony.
As investors take pitches about anything from asteroid mining to plutonium batteries, AI is the expert partner that costs $20 a month and doesn’t disappear for ski trips or try to fire founders. Agents can break cold inbound down into the most practical format, handle reporting for LPs, synthesize investment memos and package market research. It is the perfect teammate.
“Individuals working with AI showed a substantial 0.37 standard deviation performance increase over the baseline of working alone without AI. This finding suggests that AI can effectively substitute for certain collaborative functions, acting as a genuine teammate by granting individuals access to the varied expertise and perspectives traditionally provided by team members.”
The Cybernetic Teammate: A Field Experiment on Generative AI Reshaping Teamwork and Expertise
Thus, the era of AI is also the era of the high-agency solo capitalist. It will empower an army of rugged individuals who leave no stone unturned in their search for potential.
It’s also a turning point for an industry that has been stuck in a decades-long rut.
Maxims for Revolutionists
“The reasonable man adapts himself to the world: the unreasonable one persists in trying to adapt the world to himself. Therefore all progress depends on the unreasonable man.”
Consider that venture capital is fundamentally a hunt for the purest source of alpha; outlier founders and companies, the weirder the better.
However, it’s also a deeply insecure strategy. Success is revealed on distant horizons, incremental metrics are mostly useless. Investors never really know how well they are doing at any given moment.
The consequence of this insecurity is herd behaviour. Because performance across the industry is understood on a relative basis, there is an incentive to do the same thing as everyone else, even when participants recognise it’s likely to be wrong. So, venture firms become enmeshed with each other, literally and ideologically, and alpha decay pulls down returns.
The exceptions are “unreasonable” individuals, by definition. Angel investors or solo GPs with conviction in a perspective that precludes herd behaviour. Their idiosyncrasy is a shield, and their eccentricity prevents them from ever fitting into the group too comfortably. The features which make them stand out will quickly sort them into failure or excellence, avoiding the sweaty anxiety of the crowd.
Essentially, this archetype is the inversion of an LLM. They spend their time thinking about what could be, rather than what is. What people should be doing, rather than what they are doing. They may or may not study history, but they certainly have strong opinions on the future.
Even if an LLM was able to tell them what to invest in, they would likely ignore it.
“The future isn’t written in training data. Early-stage investing lives in what hasn’t happened yet. Humans are still the best at this and I can’t imagine AI being better on any alarmist timeline.”
Vested Interests
To the extent that an AI can predict the future, it will do so by extrapolating the past; assuming a continuation of the status quo. Ask an AI how to deploy capital in the stock market and it will recommend buying the index because markets are mostly efficient and everything knowable, everything in the model, is already priced in.
Definitionally, this is a beta-generating strategy that supports today’s incumbents. There is no variance, positive or negative. And for almost everyone, almost all of the time, it’s the right thing to do.
Apply the same logic to venture capital, and the result is something like a scaled venture platform. Their strategy is to index market themes, to reduce the brand-related risk of missing “obvious winners” and juice fund metrics with momentum. Implicitly, they are incentivised to support a stable status quo that produces clear winners and predictable outcomes, rather than messy disruption.
This strategy sacrifices the outliers; the startups building novel solutions in obscure categories. It’s an acceptable compromise, as the bigger firms can always pay up for access later as the opportunity becomes more legible.
As a result, it’s easier to see how AI could identify promising investments within this strategy, based on defined parameters. Thus, unlike boutique venture capital, automation is clearly more viable as AI continues to improve.
Autonomous Allocation
Early experiments with applying LLMs to venture capital deal selection already appear to improve the overall rate of success. This effectively reproduces the scaled venture platform “smart beta” strategy, and would offer considerable efficiency to those firms. It comes with the same sacrifice of outliers, as the model’s inherent biases in the training data, but this could be reduced over time.
A similar result comes from studies looking at how AI can be used to categorise and filter opportunities. The LLM-based approach was 537 times faster than the human analyst without any loss in quality. Again, outliers will be lost in the margins of this process, but that is already accounted for in the scaled venture strategy. The uplift in efficiency and reliability is likely to be more attractive on a practical basis.
If the goal of scaled venture capital is to pull more money into private markets, an automated approach would dramatically reduce market friction. It would also allow for a significant reduction in the associated management fees, resolving some of the incentive-related issues at the larger end of the market.
“The fee drag on gross-to-net returns is estimated at 0.1x to 0.7x, equivalent to an annualized impact of 5% to 8% on returns. This impact is notably higher for buyout funds (7.9%) and venture capital funds (8.5%) than previously documented. […] This drag raises concerns that high fees may significantly erode investors’ returns, potentially outweighing the benefits of investing in private capital markets.”
In practice, a reduction of management fees to 1–1.5% would significantly lift the performance of this part of the market, making it even easier to draw in LP capital, thus pulling down the cost of capital that gets passed on to the founders and their companies. Today this cost of capital often becomes the killer over time, as it jacks up valuations beyond any reasonable appetite for exit markets.
It is rational that scaled firms should operate with reduced management fees compared to smaller funds. This makes sense both in terms of expected returns and in terms of fee income relative to workload.
Cybernetic Teammates
“Every great business is built around a secret that’s hidden from the outside. A great company is a conspiracy to change the world; when you share your secret, the recipient becomes a fellow conspirator.”
On the other end of the spectrum, life is very different.
Solo GPs and small partnerships are the fundamental interface between capital and talent. It is inherently a human job, as it often requires going where knowledge is inaccessible to larger firms or digital platforms. It is the business of secrets; of building connections and winning trust. As a result it is best achieved by a larger field of independent actors.
Here, AI is the cybernetic teammate. The partner specialised in everything. The always-online associate. The agentic IR department that makes larger LPs more accessible. It amplifies the impact that a solo capitalist can have on the world, and expands their pool of accessible capital.
In their nature as libraries of what is known, and what exists, LLMs cannot tell these investors what to look for, or where to look for it. Origination and selection remain uniquely human skills which are amplified by the idiosyncrasies of the individual. AI would only drag them to the crowded centre where competition destroys returns.
For the best investors in this category, there is room to increase fees without a meaningful drag on performance. Indeed, by making life with a smaller fund much more manageable, it incentivises GPs to raise right-sized vehicles that can deliver attractive multiples. Again, this produces a better product for founders by reducing market friction and embracing idiosyncrasy.
A Wedge for Innovation
“Ironically, innovations in venture capital haven’t kept pace with the companies we serve. […] As chips shrank and software flew to the cloud, venture capital kept operating on the business equivalent of floppy disks.”
For a long time, and for reasons related to everything from fund structure to economics, it has been clear that venture capital is bizarrely behind the times. As an industry that supposedly obsesses over innovation, it clearly doesn’t spend nearly enough time on introspection. As a result, it has become top-heavy and stagnant.
Seth Levine first wrote about the bifurcation of the venture capital market in 2010, reflecting on the divergence into small and large firms. That idea has continued to evolve over the last 16 years, and today it is irrefutable. However, very few people have asked what this should mean for the “2 and 20” structure, which dates back to Alfred Winslow Jones in 1949.
That lack of critical thinking about compensation is a red flag for market capture; conditions that incumbents have clearly grown comfortable with. Indeed, new GPs remain reluctant to innovate on fund strategy or economics for fear of scaring off risk-averse LPs, having been warned it would induce adverse selection.
“Consistent with the literature on institutional behaviour, we find that the continuous use of boilerplate provisions is a form of ‘collective conservatism’. We argue that these results help explain why venture capitalists typically maintain using suboptimal boilerplate provisions. Venture capitalists may be hesitant to abandon these provisions out of concern that their peers and major rivals prefer to use them in their partnership agreements. This suggests that the often ineffective ‘2 and 20’ rule may persist for this reason.”
There is a clear argument that AI should finally disrupt this status quo.
For scaled, systematic allocation there are obvious gains to be made in efficiency and productivity, which would allow for reduced fees and a lower cost of capital for entrepreneurs. This would help keep the late-stage market more rational, and preserve the viability of paths to exit and liquidity for LPs.
For boutique investors, there is reason to feel more bullish than ever. Their potential to accelerate the development of new technologies, by identifying them and onboarding them to private markets, is immense.
As history illustrates, with the returns of the greatest angel investors, solo GPs or small partnerships, the ceiling on that potential is already hard to perceive.
If the core constraint is venture capital’s ability to recognise greatness at the earliest stages, an era of empowered solo capitalists will only raise that ceiling further.
Watch the latest episode of Going Solo, with Adam Besvinick of Looking Glass Capital
Run your investment firm from your phone, with Odin
