A long-form overview of our thoughts on our funds, venture markets, our core categories of focus, and building our investment firm.
Below is a slightly redacted version of the annual letter we recently sent to Compound investors.
Dear Compound Investors,
We recently sent through a more traditional update on the portfolio alongside a collection of our latest research and writing. In addition to that, we wanted to relay our views on the current market and our areas of focus.
Compound Fund Summary
Compound I investors received a distribution last week. You’ll also notice a few markdowns in statements as we close out the year. We felt it prudent to be conservative while also remaining consistent with our valuation methodology for audit purposes. We have now called $45.4M and invested $49.1M from the $48M fund.
Despite this environment, we remain very excited about the prospects of this fund as we have strong early liquidity and DPI with multiple companies that we feel could return the fund or more.
Compound II continues to be deployed only into our highest conviction theses at a slow but steady pace. In the past 2 months we have closed 1 new investment and have another term sheet signed that we expect to be closed in the coming weeks. Including these investments, we have now called $12M and invested $8.6M since September 2021.
Broader Venture Market
At this point it feels as if we’ve been a broken record in our updates but…the broader market continues to soften, with mid-stage and growth-stage financings continuing to be nearly frozen except for the top-tier, most-consensus investments, as founder pricing expectations and VC skittishness continues to result in a (slowly closing) gap.
In addition, we have yet to see material startup mortality. We expect this will come in the second half of 2023 as companies run out of runway from their 2020/2021 financings.
While seed has been less impacted by the slowdown, we are happily starting to see diligence processes take more normal lengths of time (1–2 weeks versus 1–2 days). This allows us to build stronger conviction, and as importantly, more meaningful relationships with founders.
Our thesis-driven approach is built around the idea that we can move faster and go deeper than generalists when conditions are moderately normal and everyone is doing some amount of prerequisite diligence. This advantage has strongly re-emerged in this current environment relative to the past few years where our approach allowed us to “keep pace” with VCs yeeting money at unprecedented pace and scale.
While we are starting to increase pace of deployment after a very slow 2+ years, we are still being diligent about pricing and runway of our investments.
This entails us making sure the milestones for our new investments are properly conservative, as we believe that Series A+ rounds are still “shifting goal posts” in terms of what investors want to see to fund a company in this very uncertain market.
Gradually, and then Suddenly — On AI and Everything
“How did you go bankrupt?” Two ways. Gradually, and then suddenly.” - The Sun Always Rises, Ernest Hemingway
Gradually and then suddenly…or sometimes stated as slowly, then all at once is an often-used phrase of recent, and a dynamic we feel more and more in our world of technology and investing. Some might refer to this at a high level as “the rate of innovation” and will show charts like the one below, showing the rate of adoption of various technologies. To add more nuance, we would say this can be defined as the rate at which we collectively build excitement around niches, momentum takes hold, and then consensus continues to accelerate and a given concept takes the world by storm.
We’ve discussed this dynamic in parts under the concept of Inflection Points, where gradual changes suddenly escalate, catching most off guard time and time again. Yet, obvious breadcrumbs are often sitting out in the open for long periods of time for those looking closely and willing to do the work of reasoning, first principles thinking, and skillful mapping of early breakthroughs’ cascading impacts.
“While the whole world was having a big old party, a few outsiders & weirdos saw the giant lie…& they saw it by doing something the rest of the suckers never thought to do…They looked.” - The Big Short
These breadcrumbs are both positive and negative.
They can be found by merely sitting and doing the math to understand when numbers don’t add up. They can be found by reading the fine print. And sometimes they are like GPT-3, published to the world in the open since 2020 just waiting for everyone to reason through what was about to happen.
At Compound, we speak often about caring only for a subset of areas and paying close attention to the adjacencies of these areas, all in an effort to reason our way through, in a detailed manner, what the various futures we believe in look like. We care about what are the technological inflection points, societal shifts, catalysts of adoption, and potential fork in the road moments across these areas. We stay in these areas through ups and downs and study them when they are coming apart. We then utilize the luxuries of power laws and long time horizons to take many risks over and over again as we wait for these futures to be possible and then brought to reality by the people we (hopefully) partner with.
Many of you have reached out to chat about Artificial Intelligence over the past few months as it has entered the zeitgeist due to things like OpenAI’s ChatGPT and StableDiffusion (which was co-created by our portfolio company Runway). While this moment is exciting and one we think will be seen as a potentially world-changing inflection point, we continue to be fairly steadfast in our beliefs we laid out years ago. Beliefs we formed by looking closely, following the breadcrumbs, and navigating the cascading impacts of things like larger models, scaled compute, more mature infrastructure, pockets & proliferation of talent, commoditization curves, and much more.
Following these breadcrumbs has led us to partner with a variety of AI/ML companies. We watched closely as the “first” wave of deep learning (enabled by GPUs) in image classification began to inflect in performance, and thought that audio classification (or speech to text more specifically) could see similar dynamics. We thus partnered with Deepgram in 2014.
We watched the birth of GANs and Variational Auto-encoders launch an entire “cottage” research industry around creative AI, leading us to incubate Shadows in 2017 and lead Runway’s seed round in 2018. We watched AV companies struggle to scale with a modular approach and felt more novel end-to-end models could solve this problem, leading us to co-lead Wayve’s seed round in 2017. We watched academics begin to experiment with simulation and synthetic data in an increasingly data hungry world and partnered with AI Reverie (acq. by Meta) in 2018, and on and on we could go if our egos were slightly larger.
Through all of these investments (and many more) over the years people often told us we were too early and sometimes, if you can believe it, told us we were too late.
Now, back to our beliefs on AI companies.
In our 2018 Annual Letter we said:
On a macro level, 2018 was another banner year for machine learning as we continued to see a breakneck pace of research being published. With that said, we think that the competitive environment for “AI-first” companies has become significantly more complex. We believe the market misjudged both the pace at which AI would improve, as well as the willingness by large companies to own horizontal types of machine learning at low-costs to the consumer. In addition, we believe many of the low-hanging fruit use-cases for deep learning were over-invested in in 2017/2018 and thus have picked specific categories to stay away from in the sector.
All of this has pushed us towards a few key beliefs with respect to investing in AI-first companies moving forward:
- Companies building on the basis of being “state of the art” today will need more than technology to win markets.
- There is an opportunity in being an expert at adopting the bleeding edge of AI research into a commercialization and productization plan.
- Companies must either be bleeding edge with an applied focus, very narrow with a long-term data advantage, or a pickaxe, as anything in-between these feels commoditized.
As AI has matured, we believe our intuitions have been increasingly validated over time. We think it’s very possible a large number of AI startups are about to be steamrolled by “general purpose” Foundation Models (and their APIs) over the next 6–18 months.
At the same time, we’ve seen existing teams with great distribution be well-poised to take advantage of the payoffs of these APIs or fine-tune models to improve their products, increase monetization, and box out competitors with less distribution.
We also remain bullish on the potential for teams who can gather truly unique datasets and serve complex industries/market segments. We’ve invested against this thesis in companies such as Orbital Materials who is building Foundation Models for Material Design, while also seeing early signs of this in larger organizations like Bloomberg with their latest paper BloombergGPT.
We have a variety of industries and use-cases we’re tracking related to this view.
As the bubble continues to inflate around AI startups, we’ve seen our peers deploy aggressively into a variety of companies with the framing of “AI” at increasingly high prices. We believe a massive number of companies are going to pivot around as they look for a nail to hit with the hammer of the GPT-4 (etc.) API. We worry those used to running a vertical SaaS playbook from the most recent bull cycle will face a far different dynamic with AI-enabled software, perhaps suffering from the death by the thousands cuts dynamics we’ve written about, or realizing that with multiples compressing, these niche applications will have difficulty materially scaling revenue and generating venture scale outcomes across hundreds of companies.
We wouldn’t be surprised if the words repeated about AI being over-invested in our 2018 and 2019 letters is a similar rhetoric that is echoed in ‘24/’25 as the commoditization curve of AI continues to destroy thin layers of applications with high churn and low defensibility.
As we think about what additional heuristics we would apply to these companies, we continue to come back to the view that the best companies will be those that can build into the future of where the broader AI industry innovates.
This could mean startups building AI-enabled products that skate to where the puck is going in AI (taking advantage of industry-wide innovation) while building unique product primitives relative to prior incumbent software.
It also could mean building custom models with formidable moats, likely in the form of unique data in a horizontal model, or in the form of a range of highly specialized models (the latter point requires a very strong point of view on performance and commoditization curves of broader foundation models).
We also are looking closely at novel techniques that could emerge. In creative AI we went from GANs to diffusion models for image generation. Could there be something better like OpenAI’s recently published Consistency models? Will Transformer models be the final model to rule them all?
While it is currently unclear to us what will change, we have a hard time believing that nothing will. We’ll have to keep looking closely.
These moments are some of the most fun for us, as we feel well positioned to help serve founders as they navigate an unprecedented pace of innovation that is creating a level of existential uncertainty we haven’t seen in many categories for years. Put more simply in a text the other night, it’s no longer a matter of if these things will work, but moreso about how you capture value and at what scale.
As certain technologies mature they move from being considered a vertical to a more horizontal enabling technology or platform (web/mobile/AI/Crypto). Much like other emerging technologies, robotics has been in a multi-year trough of disillusionment largely due to the inability for companies to scale utility and ROI in a meaningful way.
Years ago we had the belief that AI was going to make its way into robotics quickly and meaningfully change this. This led to our investment in (redacted commentary on investment that has struggled due to ML/Robotics timing).
Our learnings from 2016–2018 led us to believe that while the hardware to enable lower-cost and scalable robotics companies was reaching the market, the software and intelligence was not. This led to the best robotics companies being effectively strong mechanical engineering and system integration organizations instead of being built at the intersection of robotics and robust artificial intelligence. This guided our investments in Hyphen and SparkAI (acq. by John Deere), among others.
Since forming this view we have been actively tracking the space in hopes of seeing the early signs of a shifting tide, which has resulted in fewer robotics investments over the years.
We now believe we have seen those signs that give us confidence something important is about to happen.
In 2017, a now-famous paper was published called Attention is All You Need that debuted Transformers. The authors of this paper went on to found some of the most prominent AI labs today and the Transformer architecture is what underpins most of the innovations that have captured attention in AI over the past year. This paper, along with a few other less popular ones (most notably in our eyes One Model to Learn Them All) signaled a clear inflection point in the AI industry to us. We noticed this in real-time.
Over the past year there have been multiple other papers that we feel could signal similar shifts in robotics, most notably papers like RT-1 and Code as Policies stand out to us. While narrow in use-case both of these papers show novel approaches to communicate with and teach robots complex tasks with natural language. Since then, there have been a slew of others that we are tracking internally and some we publicly track here.
If we were to summarize our views on robotics moving forward it would be four key points:
- We believe robotics today is at a similar place that AI was at in ~2017. History suggests that these next 24–36 months is when we should be seeding the next wave of robotics companies.
- While many believe that we are in need of more data for robotics, each day we continue to see teams utilize existing LLMs to lower the burden of training data needed for embodied intelligence. We are seeing this internally at Wayve across a variety of use-cases and believe this will become a common narrative in the next year.
- We are less prescriptive than normal at this moment about where these venture-scale companies will be made. Ultimately we believe this is partially a product creativity problem as much as it is a technology one thus are intrigued by founders that come from more product-centric backgrounds versus only hardcore technologists.
- Related to this, there is currently a bit of a disconnect between the rate of progress in intelligence versus hardware. Put more simply, a large number of teams are aiming to build the humanoid modality as the “hardware” due to our world being built for this modality (built around humans). Others are stuck using less-innovative robotic arms or other end effectors that may not be advanced enough for this next step-change in intelligence. We think this will change but it is an area of concern.
Investing in Healthcare and Biology companies has been a perpetual learning process for us at Compound, as areas spanning the spectrum from clinical health all the way to traditional biotech have evolved materially in the past few years. Interestingly this evolution has pushed companies on each side of that spectrum to move closer to one another.
Today, we are most drawn to healthcare companies that have unique views on how science and novel care models can impact outcomes. This can take the form of what Tia has done by pioneering a cycle-connected care model, to a recent term sheet we gave to an EEG hardware company, to our perpetual research on biomarkers for mental healthcare, and many companies in-between and adjacent.
In the middle of the spectrum we are seeing an explosion of software solutions for biologists. Many tech VCs are investing in this part of the stack recently due to the comfort with software, but our belief is that the nuances of this field are very difficult to grasp without a larger focus on bio broadly. Many products are point solutions, working in an already small market with long sales cycles that can lead to fragmentation of product use.
We believe the big winners will be horizontal solutions, with compounding data value. We’ve made an investment under this thesis in a company called *redacted* which we discussed in our last update.
On the “bio” end of the spectrum we have the (now obvious) trend of biotech companies inching closer to software and technology businesses as we move from a search-and-discover methodology to a more targeted and design-centric approach.
The promise of this future is far too enticing for investors to pass up, and because of this we see investment drying up for bio companies without this tech enablement, with tech VC funds only investing in what’s now known as TechBio.
There are of course debates as to how effective this shift has been in actually getting new products to market, however we expect the synthesis and deployment of COVID-19 vaccines to be a proverbial GPT moment, ahead of a gradually then all at once moment for computational biology. These moments could be enabled by things like taking single cell data from an organism over its entire development, leading to foundation models for whole organisms, to breakthroughs like Med-PaLM which are foundation models for clinical health data that can outperform doctors.
Before this happens however, lab automation, decreased experimental costs, and data aggregation and mining infrastructure are all needed to create/capture data and make it searchable and interoperable.
Although on the bio side we are largely partnering with companies which have the above phenotypes, we’ve noticed our long-standing commitment to AI, Robotics, and Biology allows us to bring a unique lens to these companies. We’ve also seen the founder profile continue to evolve, with many founders coming from the first wave of now-public TechBio companies and big tech companies. We believe this trend will lead to the biggest companies of the next 5–10 years being built by engineers and/or teams that have a blend of both engineering and science competency. We see this as an opportunity for us to sit at this intersection and are building a community specifically to bridge the gap between AI engineers and scientists interested in building in bio.
While the dynamics at play today for investing in early-stage health and bio are multi-faceted, on roads paved with regulatory, computational, and efficacy uncertainty, we are confident that large change and value creation (both societal and financial) is inevitable.
The promise of this future is both better and faster outcomes, and likely this is led by ambitious, multi-disciplinary teams, bought into the premise that most bio and health research will be conducted on computers, drastically decreasing the bar for scientific discovery and n-of-1 trials.
Candidly, we feel the need to discuss crypto with you all with each update during this bear market as the first question asked upon us 1:1 is “Do you all still believe in crypto”. We expect this question to continue until the next bull market, and we expect this answer to continue to be “yes”.
Each time we meet, we also become more convinced by world events of our vision for a more crypto-centric future: A distrust of scaled centralized entities, a desire for self-sovereignty, an increased economic dominance of smart contract-driven models in certain industries, and more.
Since we’ve last spoke we’ve had a bank run on one of the largest banks in the United States to the point where the government had to intervene to stop the broader US banking system from collapsing, concerns about centralized entities accumulating power as they build AGI, governments threatening to move off of the US Dollar, threatening broader American global dominance (we wrote about this dynamic here), and continued hyperinflation concerns in a variety of countries.
Thus, while the venture market continues to be slow, we remain very convicted that these moments we speak of above in our other industries are repeatedly happening in smaller forms within Crypto, while perhaps being held back by a lack of regulatory clarity. With this in mind, 0xSmac has joined our team as an investor focused on crypto as we continue to seek early teams building novel protocols in areas we care about.
Investment Approach Amidst Consensus
Time and time again we describe our strategy as getting to emerging categories early, going deep in them, and exploiting our advantages over multiple fund cycles. This allows us to make non-consensus investments with a hopefully better view on time-to-readiness for complex technologies. As time goes on we continue to gain conviction in this approach, and also notice where we can improve in the long-term.
While we’ve shown strong competency in building and investing in futures we believe in ahead of consensus, the next difficulty is then maintaining and defending that moat as these areas become consensus. In some areas it helps to have a very visible company that is leading the category, but this only goes so far as prices increase alongside industry consensus.
Instead we must hone our ability to pick why certain areas will have last-mover advantages (perhaps making non-consensus investments after a consensus winner has been chosen, as we did with Wayve) or have deeply convicted, non-obvious views on unique value accrual or competitive positioning mechanics for the startups we partner with.
We mentioned in our September 2022 update that we had spent a large portion of the year turned inward, doing research, refining theses, and working with our portfolio. In this time of relative stillness (and market chaos) we also had many moments to reflect on the firm we were building and the work we do at Compound as early-stage investors.
Over the past few years much ink has been spilled about the impending disruption of venture capital as new entrants were “breaking” the traditional model. As it turns out, the thing we’ve learned is that venture wasn’t so much “broken” as it was slightly evolving.
At the early-stage we still take massive risk on nearly impossible futures to imagine and do unscalable things to find and help our partners and communities bring these to life. These futures sometimes take far too long to come to fruition, and other times happen seemingly overnight and some firms capture alpha while others act as beta depending on the market environment.
We also talked about a feeling of being in this calm before The Most Important Century storm. Our core categories of focus continue to feel on the precipice of a magnitude of change and impact that we (perhaps naively) feel is very unique for our lifetimes. These perturbations cycle, with AI now cycling into the zeitgeist, Crypto cycling out of it, and Bio and Robotics remaining difficult to haphazardly parse and deploy capital into.
Over the next few years it’s likely much ink will be spilled about the impending disruption of our lives as a series of technological waves continue to crash over humanity. The thing that we know is that the future will continue to get weirder, filled with more opportunity, and more excitement. We hope we can notice those moments, follow the breadcrumbs, and participate in the creation of this future.
Regardless of if it takes too long, or if it comes faster than any of us are ready for.
Michael, David, & The Compound Team