Self-Driving Cars, AI & Venture

October 24, 2024

Friends,

We recently hosted our third annual Ideas & Networking Conference in New York City on September 19th.  It was a fantastic day with great speakers and participants.   I specifically want to thank our special guest speakers, Sol Bier of Factorial Funds, John Serafina of Hawkeye 360, Brian Bellinger of Monimus Capital, and Ken Kurson of Sea of Reeds Media, for sharing their ideas and insights.  Also, thanks to our many attendees!

As a follow-up to that event, I am excited to share the edited excerpts from my fireside chat with Sol Bier.  Sol is the Founding Partner of Factorial Funds, an AI-focused venture fund that has raised over $600 million for AI-centric companies.  Previously, Sol was on the early founding engineering team of Cruise Engineering.

Our chat included a discussion of Cruise’s evolution as well as the current state of the self-driving market.  Sol also spoke about the future of AI, specifically large language models (LLMs), Factorial’s current investments in AI infrastructure and hardware, and the potential for AI adoption by consumers and enterprises.  He also shares thoughts on Nvidia (NASDAQ: NVDA), Tesla (NASDAQ: TSLA), Waymo, OpenAI, Anthropic, and Apple (NASDAQ: AAPL). I am confident you will find value in this interview.

In the coming months, I anticipate sharing additional excerpts from my chats with the other conference participants.  Enjoy!

Best Regards,

William C. Martin

Topics in this Issue of An Entrepreneur’s Perspective:

• • Interview with Sol Bier: Self-Driving Cars, AI & Venture
  • Favorite Podcasts & Media
  • Recent Tweets from @RagingVentures

**

Interview with Sol Bier, Founding Partner of Factorial Funds: Self-Driving Cars, AI & Venture

This fireside chat originally occurred on September 19, 2024.  ChatGPT was used to format and lightly edit the original discussion, using a “formal conversation” prompt. 

Could you share a bit about your experience at Cruise Automation as a founding engineer, as well as your work at Factorial, where you’ve raised over $600 million for AI-centric companies?

Certainly!  My journey began with a high school program in California, which I thoroughly enjoyed.  That experience inspired me to pursue my studies at USC in Los Angeles, where I majored in physics and computer science.  At that point, I found myself at a crossroads between finance and technology.  I was accepted into Harvard Business School’s 2+2 program, which allows students to work for 2 to 4 years before starting their MBA.

While interviewing with various hedge funds and VC firms, one fund, CAA Ventures, advised me to embrace risk and not work at a fund.  Taking their advice to heart, I decided to join Cruise.  At that time, it was a very small company—only six of us working out of Kyle’s garage.  This was back in 2014, a challenging time for raising venture capital, particularly with minimal interest in AI and self-driving cars.

The work was far from glamorous.  We faced significant hurdles; no car manufacturers would sell us vehicles, so we had to purchase them and reverse engineer everything ourselves.  Back then, Nvidia had just released their Titan model, and tools like TensorFlow were not yet available.  However, we started to recognize the potential of utilizing GPUs for our projects.

Interestingly, in 2013, Nvidia was a hot topic among hedge funds, but for the wrong reasons.  Activist investors were pushing for the shutdown of their mobile division, Tegra, thinking it wasn’t a sound investment.  It’s intriguing to consider that if someone had made a significant investment in Nvidia back then, it likely would have outperformed many others.  I first became aware of Nvidia’s stock during that period.

Was Cruise initially focused on building cars, or were you primarily developing software for vehicles?  You also have several patents to your name.  Can you elaborate on your role at Cruise?

At the beginning, like many startups, we went through some pivots.  We started with the intention of creating software, but it quickly became evident to Kyle that we needed control over the hardware to succeed.  So, we shifted our focus to building a fully integrated vehicle, specifically an autopilot system.  That pivot stirred some controversy; not all our investors were supportive, and some team members had concerns.  Nevertheless, I believe it was the right call.  Companies like Anduril have shown that owning the hardware and data can lead to significant success.

We made that decision in 2014, and it required us to raise additional capital, which proved to be quite challenging.  It took us around 14 months to secure our Series A Funding—today, that process would likely be much smoother.  During that time, I met many investors in Silicon Valley, and I learned to tune out some of the prevailing trends; at that point, there was a strong focus on social networking and consumer investments, which eventually faded.

We managed to grow the team to about 40 people.  My work included developing the routing algorithm and motion planning—essentially figuring out how the car should respond in various scenarios.  I also focused on perception tasks, like ladder segmentation.  My role was quite hands-on and involved a lot of computer science and machine learning, especially before the rise of large models.

GM acquired Cruise in 2016?

Yes, that’s correct.  They acquired us for approximately $1.2 billion, which was a significant exit for us.

How much funding had been invested in Cruise up until that point?

We raised just under $30 million, making it a fairly capital-efficient endeavor.  After the acquisition, GM divided the company into private divisions. SoftBank, Honda, and Microsoft invested, particularly for GPU computing.  This was, as far as I know, the first GPU compute deal prior to the later partnership between OpenAI and Microsoft.  Interestingly, Sam Altman was also an investor in Cruise, and we took away some valuable lessons from that experience.

How has Cruise performed under GM’s ownership?

Overall, I think they’ve fared well, but they haven’t reached the level of success that Waymo has.  For instance, Waymo recently surpassed $100 million in revenue.  One of the challenges GM faces is managing two supply chains, particularly because they opted to outsource hardware production to Honda.  This decision has proven problematic, especially given that the Bolt has some notable shortcomings and doesn’t meet the necessary standards for a self-driving vehicle.

Additionally, GM may have rushed their deployment efforts.  Interestingly, back in 2021, just before the release of ChatGPT, we were aware of breakthroughs from the OpenAI team.  Some individuals from Cruise had joined OpenAI, and we observed significant advancements in large models.  Consequently, we decided to integrate those insights, transitioning from a rules-based planning algorithm to a fully automated one.

It’s interesting to see how the enthusiasm around self-driving cars has diminished since then, which has greatly influenced my investment perspective.  I’ve learned to maintain a long-term outlook and not get swept up in hype cycles, as they inevitably normalize.

Nowadays, the self-driving car market isn’t garnering as much attention, and we haven’t invested significantly in it recently.  However, if you’re in San Francisco, I recommend taking a ride in a Waymo vehicle.  They’ve integrated a lot of generative AI technology and are approaching $100 million in annual recurring revenue, which is impressive considering their recent scale-up.  It’s one of the largest revenue generators in the AI sector, yet it often goes unrecognized.

How do you view the current business models of Cruise and Waymo, especially in relation to Tesla?

Tesla’s model is quite effective because they have a comprehensive car platform; they own the EV framework.  This ownership enhances their capacity to deploy self-driving cars.  I don’t believe a complete self-driving solution was necessary from the outset.  Tesla’s Autopilot, priced around $10,000, has proven to be very profitable, allowing them to roll out features gradually.  We think Waymo has strong potential for long-term success.  Their business model, particularly their efforts to license technology to OEMs and their collaboration with Silver Lake, is promising.  They don’t intend to manufacture hardware; they prefer OEMs to handle that aspect.  This creates an interesting dynamic as they begin to compete with Uber.  Currently, about 80% of Uber’s margins go to drivers, so if Waymo can achieve effective scaling, it could result in a profitable model for them.

How important is Tesla’s deployed fleet and the data they’re collecting?

That’s a crucial topic.  The data collected—particularly image data—holds immense value.  Waymo has the potential to gather similar data, but manufacturers like Honda and Hyundai lack the depth of information that Tesla possesses.  While they may aspire to compete with Tesla, they simply cannot match Tesla’s deployment capabilities.  Tesla has developed its own dedicated chip, which creates a data flywheel that enhances their training capabilities, essential for prediction and segmentation.

Looking ahead, in five to ten years, do you envision Waymo and Cruise operating independent robo-fleets, or will they also be selling software to other OEMs?

I believe they will be engaged in selling software to other OEMs while also functioning on platforms similar to Uber or Lyft.  Recently, Uber struck a deal with Waymo to integrate their services.  There remains a strong brand loyalty among customers for Uber and Lyft, which may evolve over time.  I doubt GM will dominate this market, as they generate substantial profit margins from their truck lines, which might lessen their incentive to aggressively pursue this area.  I think Waymo is likely to emerge as a long-term leader in this space, especially given their symbiotic relationship with Uber.  Waymo can attract customers through Uber, while Uber can leverage scale, particularly with electric fleets, benefiting both parties.

Looking back on the early days of Cruise, when did people expect to see full driverless capabilities?

Initially, they were proclaiming 2020 as the target year, but we internally knew that was unrealistic.  If you understand hardware development, you realize it involves four-year cycles.  Upgrading components like cameras and lidar takes considerable time.  So, we believe there’s probably one more upgrade cycle needed, which I think occurred this year.  Therefore, we anticipate that by around 2028, a small portion of the fleet could operate autonomously, but for widespread deployment at the desired scale, an additional hardware refresh will be necessary.

So you’re suggesting 2028 for full driverless capabilities?

Exactly.  If you’re in San Francisco, I encourage everyone to try the Waymo app, which is now public.  It offers a glimpse into their impressive capabilities.

Can you discuss Tesla’s advantages and challenges?

Tesla opted not to invest in LIDAR technology, which raises doubts about their ability to compete with Waymo at scale.  Their current deployment efforts are largely focused on Level 3 driving.  While there’s a narrative from Elon about robotaxis, it’s a complex undertaking.  Managing a fleet and deploying at scale involves many factors, including maintenance and charging.  I’m skeptical about whether Tesla has a significant advantage in this domain.

Given your experience and connections, how did you transition into being a leading AI investor?  Can you share some early investments?

Toward the end of my tenure at Cruise, I had been there long enough to be fully vested and knew I wanted to pursue investing at some point.  I had built a network within the San Francisco ecosystem, and in 2021, I began focusing on technology and technology-adjacent investments.  We got involved with SpaceX just after they launched Starlink, and we had connections with many from the OpenAI team.  Our current investment team includes former directors of engineering and co-founders from OpenAI, creating a symbiotic relationship as talent often moves between these organizations.

By 2021, we were aware of the impending release of ChatGPT, especially after the foundational paper was published in 2018.  In 2022, we made significant investments in the foundation model layer, backing early ventures like Mistral, Anthropic, and xAI.  When ChatGPT was released, it marked a pivotal moment.  We spent about three months deliberating on our investment strategy.  While many VCs were focused on enterprise applications, we believed consumer adoption would occur first during this cycle.

The typical trajectory of new technology tends to begin with consumers, so we identified search as a key area.  We conducted numerous interviews and ultimately invested in Perplexity during their Series A round.  They’ve since performed very well, recently securing Series C funding.

As we assess our next moves, we agree that the current phase is still early.  We’re focusing on the infrastructure layer now, particularly semiconductor and networking companies.  On the enterprise side, our investment activity has been more limited, but we can share interesting insights from OpenAI about that.

Can you discuss the progression of large language models (LLMs) over the past few years?  A lot has happened, and I’d also like to know how you envision the next few years unfolding.

Absolutely.  To provide some context, let’s start with the timeline.  The fundamental shift in technology we’ve seen is the move to parallel computing.  The breakthrough paper demonstrating how engineers could leverage GPUs for machine learning was published in 2013.  Interestingly, while that paper laid the groundwork, its significance wasn’t fully appreciated until 2018 when Google released their paper introducing the transformer architecture.

Can you explain that in layman’s terms?

Certainly.  The transformer architecture changed the way data is processed.  Previously, information was fed to models sequentially, similar to writing a sentence word by word.  The transformer approach, however, allows for all data to be processed simultaneously, enabling greater efficiency and effectiveness.  The key paper on this is titled, “Attention is All You Need.”

What would you say was the major breakthrough?

The breakthrough was really about that shift to simultaneous processing with the transformer architecture.  It wasn’t until 2020 that we saw significant advancements, and then finally in late 2022 and 2023, the ChatGPT product was released.  It’s worth noting that there has been relatively modest investment in LLMs compared to other areas of machine learning.

Most of the capital expenditure over the last decade in machine learning has focused on computer vision, not LLMs.  To give you a sense of how early we are in this cycle, even though many are enamored with the capabilities of these products, they are only about 70% to 80% accurate.  The recent developments remind me of the iPhone’s launch in 2008—there’s so much more potential to explore.

So, how do you see this technology scaling?  Will it simply involve more GPUs and data?

We’ve actually seen some movement in that direction recently.  There are two main schools of thought.  One involves fine-tuning models for specific use cases, which we’ve always viewed as somewhat limiting since LLMs are intended to be general-purpose.  The other, and I believe more promising, approach focuses on increasing inference time as models grow larger.

What we’ve observed is a shift toward models running for longer periods, which is crucial for enterprise applications.  For instance, when you interact with ChatGPT, the response time is quick, but the latest release emphasized extending the time the model spends on processing before delivering an answer.  This change allows for deeper reasoning, and the relationship between inference time and accuracy is exponential, as shown in a recent plot they released.

What exactly is happening when the model is “thinking”?

Essentially, it’s running through multiple iterations.  You can think of it as a pathfinding algorithm.  The model evaluates possible answers, assesses them, and then refines its response based on your question.  It’s like how we approach problem-solving; we don’t just blurt out an answer immediately.

So it’s somewhat like a chess game, exploring every possible move?

Exactly!  This approach highlights a significant need for computational power.  For the data centers supporting these models, we estimate that they might require 10 to 100 times more GPUs than currently available.  And that’s not even considering video processing, which demands an even higher level of inference.  Right now, many users might notice fluctuations in performance—sometimes ChatGPT or other platforms feel less responsive.  That’s often due to throttling, akin to a bartender diluting a drink by adding more water.  It’s a matter of available compute resources for GPUs.

With so many players in the field, where does OpenAI stand?  Is it still leading, and what about Google, Facebook with Llama, Mistral, and Anthropic?

There’s a lot happening in that landscape.  We’re currently evaluating the latest round for OpenAI.  The critical question for us is whether OpenAI’s position is defensible.  For now, these remain consumer products at their core.  Right now, many users default to ChatGPT.  Anthropic is about a fifth the size of OpenAI, and while developers in San Francisco are familiar with them, that knowledge may not be as widespread outside the Bay Area.

The challenge for companies like Perplexity and Anthropic is deciding whether to focus on consumer engagement or develop a breakthrough model that would compel users to switch.  Currently, it appears the market is consolidating.  We foresee a scenario where Amazon and Google each have their own models, with OpenAI likely remaining the market leader.

So brand recognition is playing a big role?

Absolutely.  OpenAI’s brand recognition creates a strong flywheel effect.  However, there’s a risk if OpenAI focuses too much on the software layer and overlooks potential breakthroughs.  If they do that, consumers might shift to other options.

Additionally, I learned something intriguing during our analysis of Anthropic and OpenAI’s funding rounds.  OpenAI has been marketing various products: consumer offerings like ChatGPT and ChatGPT Plus, an enterprise version with privacy controls, and an API for developers.  They noticed that while their teams product—used internally by organizations like Goldman Sachs—gained traction, the API saw little adoption.  This indicates how early we are in the cycle; the enterprise market is still in its infancy, with most users opting for the standard ChatGPT experience.  Right now, it’s about 75% consumer-focused for both companies.

Guest Q&A: What is the bottleneck in driving enterprise sales?

There’s a crucial sales cycle involved when it comes to adopting APIs.  Organizations need to build products on top of these APIs, which requires a clear understanding of their data landscape.  There’s often political maneuvering as well.  Small startups may have a handle on their databases, but large enterprises face significant hurdles.  For instance, obtaining sign-off from the Chief Security Officer (CSO) can be a challenge, especially when access to technology is limited.  In many organizations, engineering departments are viewed as cost centers rather than strategic assets.  Unlike companies like Google or Facebook, where engineers are prioritized, in other organizations, they lack the authority to drive product innovation.

What we’re observing now is a reliance on the ChatGPT subscription model among consumers, but enterprise adoption is lagging.  This is disappointing because we anticipated greater uptake, particularly in sectors like insurance and call centers.  Perhaps that will come eventually, but it hasn’t materialized yet.  If I were a CEO, I would be advocating for faster adoption, but the reality is different at this point.

Guest Q&A: What are you seeing in regards to user workflows?

What we’re seeing now is that consumers are engaging with these tools primarily for basic workflows, rather than executing more complex tasks.  Historically, 80% of searches still directed users to the web, but the disruption in web search was a significant factor in our decision-making.  Currently, usage is split between professional workflows—like document analysis or multi-step reasoning—and simple queries.

Right now, most multi-step reasoning tasks are fairly straightforward and can be addressed through web searches.  However, as users become more comfortable, they’re starting to ask more complex questions.  For instance, someone might inquire about Intel’s market share changes in the ’80s and ’90s compared to Nvidia.  This is where the need for longer-running models comes into play, as accuracy improves over time.  For true enterprise adoption, there’s a higher standard: while consumers might accept a 70% accuracy rate, firms like Goldman Sachs require 95% to 99% accuracy.  Hence, the focus is on enhancing those additional reasoning steps, which take longer to compute.

What does the inflection point look like for integrating personal data into these models?  It seems like enterprises are still figuring out how to manage their data securely.

Exactly.  Companies are still navigating how to leverage their data safely and effectively.  As for our investment strategy, we’re closely monitoring this landscape.  We recognize the potential for disruption and are looking for innovative solutions that can address these data security challenges.

This brings me to our new fund we’re raising, which aims to capitalize on these emerging opportunities.  We’re focusing on businesses that can facilitate safe data integration into AI models, as well as those that enhance enterprise adoption of AI technologies.  By investing in this space, we hope to not only drive innovation but also ensure that enterprises can leverage AI effectively and securely.

Our focus is on investing in software companies that enable this technology, along with API, cloud, and data infrastructure firms.  What you’re asking about is becoming increasingly feasible.  However, the documentation from OpenAI and Anthropic still needs improvement.  We’re collaborating with them, and companies like Perplexity are developing tools that will soon allow users to connect local file storage and perform desired tasks.  We believe that this level of sophistication will soon be achievable, particularly for enterprise use cases.

Is that the bet Apple is making?

Not exactly.  Apple hasn’t made significant investments in this area; they’re primarily outsourcing to OpenAI.  They face unique challenges as they aren’t particularly fond of Nvidia and have chosen to invest in their own silicon.  Currently, Apple’s hardware hasn’t significantly accelerated AI tasks.  Offloading processing to the cloud is manageable now, but as latency improves, users may become frustrated if Apple’s performance lags due to reliance on cloud GPUs.  Consequently, Apple has tasked Anthropic and OpenAI with creating a smaller, slower model for device distribution.  For instance, the new iOS version 18 reflects this shift.

Has Apple missed the boat?  Shouldn’t they be investing $100 billion in data centers right now?

I believe they have missed some significant opportunities, but they are a smart company and likely to catch up eventually.  However, they’ve really squandered the potential of Siri, which could have been a strong entry point for them in AI.  Over the past year, their efforts regarding Apple intelligence have been minimal and lack full integration.  That said, Apple has one of the best ecosystems for distribution, which could still allow them to recover.

It feels reminiscent of the telecom build-out in the early 2000s, where companies were pouring money into infrastructure.  Do you foresee a similar situation where a crash occurs?

Absolutely, I think a crash is likely, much like what happened in telecom.  The challenge lies in balancing capital expenditures (CapEx) with customer demand.  Companies are hesitant to invest in infrastructure without clear demand, but that demand can be somewhat constrained.  Nevertheless, we remain optimistic for the long term, especially since we haven’t fully rolled out video capabilities yet.  Video processing is significantly more compute-intensive than current workloads.

When we examined Nvidia’s positioning, we compared it to Intel.  Intel’s x86 architecture has dominated desktops and servers since the 1980s and still retains about 75% market share, down from 90%.  Despite decades of potential disruption, it remains a formidable player.  Nvidia’s customer concentration resembles Intel’s historical pattern, focusing primarily on hyperscalers.  However, I’m somewhat skeptical about companies like Google being able to displace Nvidia.  Transitioning to a new chip requires engineers to learn new programming languages, which can slow adoption.

It seems like a risky endeavor for companies to switch.

Exactly.  While OpenAI is supporting initiatives like Triton to facilitate the process, I’ve observed that in every chip boom, the original market leader typically retains its advantage due to optimized compilers.  Therefore, I’m cautious about Nvidia losing market share.  Predicting the next 12 months of revenue is fraught with uncertainty, but I remain skeptical about drastic changes in the market landscape.

Moreover, we’ve noticed companies like Perplexity and Anthropic using Google TPUs, and they’ve reported that the performance isn’t ideal.  The time to first token is notably slower.  Nvidia invested a decade refining its technology, sending engineers to various companies to gather insights before returning to TSMC for further optimization.  This extensive groundwork gives Nvidia a significant advantage over Google and others in terms of performance.

Do you want to share some details about your fund?

We’re aiming to finalize our next fund with a first close by the end of this year.  We’re targeting a range between $300 million and $500 million.  Our primary focus will be on Series A and C investments within the AI sector.  While we’ve had a diverse portfolio in the past, this cycle will likely lean more towards hardware due to the infrastructure build-out required.  For example, we’re considering companies similar to AlphaWave Semi, which specializes in optical chips for chip-to-chip connections, akin to GPU communications.  I noticed that AlphaWave achieves about one terabit per second, while the companies we’re looking at are aiming for around 20 to 40 terabits per second—essentially a 20x improvement.  This level of advancement is critical as the volume of data movement between GPUs is set to increase dramatically.

How do you secure deals in such a competitive environment?  It seems easier to get involved with smaller deals, but what about the larger rounds?

Our advantage stems from having a team comprised of ex-operators, which helps us connect with founders on a deeper level.  Early investments create a flywheel effect; once we establish relationships with those companies, others tend to follow suit.  While our fund size is relatively modest compared to larger funds, we can still position ourselves effectively in competitive rounds.  In fact, we anticipate leading some rounds in the upcoming fund.  We had opportunities to do so in our previous fund, but it didn’t align with the needs of the companies, as we were still growing as a firm.

Will the focus still primarily be on infrastructure and hardware, rather than applications and agents?

Yes, I think that focus will continue for the next few years.  We might still be a couple of years away from seeing a significant investment cycle geared towards applications.

Are you suggesting that VCs who are making application investments today may struggle?

Exactly.  Many VCs tend to follow the path of least resistance.  They already have a background in enterprise SaaS and may invest in companies that lack defensibility, merely slapping AI onto their existing offerings at inflated valuations.  This is something to watch out for in the current cycle.

How can businesses thrive when their value is essentially an API call without controlling the underlying model?

That’s precisely why we wouldn’t invest in such companies.  If they don’t have a unique value proposition beyond an API call, it’s challenging to see them as sustainable businesses.

Is the data itself the true value then?

Not necessarily.  Data is important, but having a robust system of record and distribution channels is equally critical.  Displacing established players like Salesforce or Intuit is no easy task.  We’ve also observed that traditional companies can experience substantial revenue uplift.  Take Notion, for example.  They offer a $10 per month software product and recently introduced a $5 per month AI feature, achieving a remarkable 60% adoption rate.  This demonstrates a significant incremental revenue opportunity.  However, it poses risks for companies like Canva, where the competitive landscape is evolving rapidly.  Therefore, it’s vital to remain vigilant regarding your existing portfolio, both in public and private markets, as these developments touch every sector.

Guest Q&A: You mentioned inference being the next unlock, potentially taking one to two product cycles to fully realize.  What do you think will be the next key topic we’ll be discussing after that?

I believe the next significant focus will be on optimizing databases or data storage formats specifically for AI applications.  Currently, companies like Vast Data and Weka are addressing this issue, but they primarily perform well with video and images.  Most existing databases are still tailored for text or simple relational models.  This optimization will likely happen in parallel with developments in inference.  Inference itself encompasses various elements, such as data center build-outs, liquid cooling, optimizing the next generation of GPUs, and improving networking interconnects.  Presently, NVLink and InfiniBand are the only viable options for interconnect, so advancements in those areas will be crucial.  Ultimately, it will follow the trajectory of Moore’s Law, focusing on increasing the efficiency of GPUs for inference tasks.

Do you think software stocks, particularly enterprise SaaS companies, hold an advantage due to their existing data and customer relationships, or are they at risk of total disruption?

It’s a mixed bag.  Some companies will undoubtedly face disruption, while others will emerge as winners.  The key is to understand each company on a granular level, especially those with a solid system of record that can leverage AI.  For example, Salesforce is likely to see a positive revenue uplift.  While some might think they could create their own version of Salesforce, rebuilding what enterprises have already adopted is far more complex than it appears.  These established companies should be viewed through the lens of potential revenue uplift.

On the consumer side, we’re noticing something unusual: pricing isn’t as much of a consideration.  We see significant overlap in subscriptions to services like Anthropic and OpenAI, which is quite rare for consumer products.  Typically, users don’t pay for multiple cloud providers; they tend to stick with a primary service, like Microsoft.

What about you?  What’s your default choice among these services?

As an investor, I utilize all of them, but my go-to changes over time.  Currently, it’s Anthropic because I find it to be the best for writing tasks.  I admit I’m a bit biased since we hold a larger position in that company.  OpenAI’s latest model excels at lockstep reasoning and web search, and we also really like Perplexity for its superior web search index.  However, it’s still early days, so we’re continuously testing all these products.

Let’s say I’m searching for cars for my son, looking into safety, maintenance, insurance—what’s your take on how these platforms ultimately monetize that kind of search?

That’s an insightful question.  We’re actively exploring this area and monitoring daily queries for various companies.  Historically, with Google, around 8% of queries generate approximately 90% of monetization.  A significant portion of Google queries is actually non-monetizable, often focused on navigation—think of queries like “Raging Capital Ventures” or “Standard Hotel.”

In contrast, we’ve noticed that a much higher percentage of queries on platforms like Perplexity or OpenAI are monetizable.  These include local searches or product inquiries—questions about schools can lead to advertising opportunities.  However, monetization in AI platforms is still an unsolved puzzle.  We aim to invest in companies that make advertising easier on these platforms.  Google’s monetization model is straightforward thanks to the PageRank algorithm, but with large language models, it’s more challenging to interrupt the model’s thought process to insert ads.  There’s a lot of innovation yet to come in this space, which makes it an exciting time to be involved.

Guest Q&A: I’d like to get your perspective on humanoid robots, especially at the intersection of hardware, computer vision, and AI.  Can you share your thoughts on timelines for when we might see them in stores or households?  Also, how do you view the competition between major players like Tesla’s X1 and Figure?

I won’t comment specifically on those companies since we’re not overly optimistic about them at the moment.  The challenge with robotics is that the data availability isn’t on par with that of large language models.  Therefore, we are somewhat pessimistic about short-term deployments.  Investors we work with, who also invest in our next fund, have stakes in large robotics companies.  We believe that startups in this space need to collaborate with these established firms, as the sophistication of the data simply isn’t there.  A breakthrough in simulation—specifically, the creation of synthetic data—would be necessary.  We’ve invested in two companies focusing on that area, but without such a breakthrough, the data remains insufficient.  For instance, Cruise had to drive through San Francisco millions of times with numerous cars to create the required level of synthetic data.

Can you explain what synthetic data is?

Sure!  Synthetic data refers to the use of models to generate a simulated version of the world.  At Cruise, for example, we utilized the Unity engine to create a fake version of San Francisco populated with virtual people interacting.  This process can generate thousands or even millions of scenarios for the model to train on.  However, historically, this approach hasn’t worked well for robotics because the feedback necessary for training is too complex.  The latest OpenAI o1 model, which focuses on inference reasoning, might push us toward a breakthrough, making simulations detailed enough to capture all relevant interactions.  While there’s some interconnectedness, we still believe that robotics isn’t feasible at this moment without collaboration with companies like Boston Dynamics.  But who knows?  These LLMs might eventually unlock that potential.

Building on that, how important are corpuses of interactions, such as data from X, Reddit, or Wikipedia, in training these models?

They’re incredibly important.

I’m also curious—if I had proprietary research, I would consider pulling my data off the web.  Does that mean the open web will fade away?

I don’t believe the open web will disappear.  We maintain good relationships with many folks at Google, particularly in the search team, and we view the evolution of search as a net positive rather than a zero-sum game.  Regarding your earlier question about the defensibility of data: for the top companies, access to quality data isn’t the primary bottleneck.  Rather, the real challenge lies in finding talent capable of conducting pre- and post-training processes.  In ten years, many will be adept at this, but currently, there are only about 100 to 300 individuals worldwide who truly understand how to scale a model using large datasets and customize it to meet specific needs.  It’s akin to the early days of mobile technology when there were only a handful of iOS engineers available.  It takes time for knowledge to spread, but yes, data remains crucial.

Guest Q&A: How do you assess the prospects for Google’s search franchise?

As a public markets investor, I have a somewhat cautious view.  I invested in Google at $80 a share, but I’m concerned about the asymmetric risks the business faces.  A small percentage of Google’s search monetization comes from a limited number of searches, which puts them at risk of losing those.  We know many individuals in Google’s search and AI teams, and they are among the best in the world—perhaps even better than OpenAI.  However, there are internal politics to consider.  For instance, when the Search Generative Experience (SGE) was released, the ads team was alarmed because it led to a 20% reduction in mobile display ads.  For a company of that scale, such changes are significant.  This represents an innovator’s dilemma.

Currently, there’s a greater risk-to-reward ratio for Google.  In contrast, for Meta, we see positive developments.  My brother works on the Meta AI team, and they are integrating generative AI for advertising, which is yielding higher CPMs and more accurate ad targeting.  As for Google, their stock price appears cheap, trading at around 15-16 times forward earnings from a valuation perspective.  However, one must consider what would happen if they miss search estimates by even 1%—the repercussions for these mega-cap companies could be severe.  While the stock may seem cheap, I’m hesitant to place my bets on it in this market.

Guest Q&A: I am skeptical that the return on investment on AI is currently present to justify the massive infrastructure spending on data centers and the like, even assuming significant growth rates.  How do you think about it?

I think we align more closely with your perspective than most in San Francisco.  We anticipate a short-term correction in the market.  It would be intriguing to compare our notes—David Cahn at Sequoia wrote a similar analysis.  The pricing on Nvidia and the associated margins will likely decrease, and that needs to be factored into any analysis.  We agree on the need for caution.

Another point to consider is that Nvidia chips can be utilized over an extended period, allowing for amortization.  This situation is reminiscent of 1996, where we foresee a short-term bubble.  Historically, whenever there’s a spike in CapEx, it tends to crash, leading to a period of widespread skepticism toward AI.  This could present an attractive investment opportunity, similar to when we initially invested during the last cycle.  Balancing this dynamic is indeed challenging, especially given the current valuations of some semiconductor companies raising funds.

I believe we should continue comparing notes.  I share your concerns about CapEx sustainability.  However, we believe there may still be a few years of runway left due to recent breakthroughs in models.  The rate of enterprise adoption will be crucial.  OpenAI is evaluating pricing strategies, including a $20,000 model, which suggests they’re also contemplating the usefulness of their technology for enterprises.  If a company can increase employee productivity by 20% while paying $100,000 to $200,000 for a model, that changes the calculus.

While I agree that the CapEx figures may be somewhat inflated, I’m not sure they’re as excessive as you suggest.  Corrections are inevitable, so it’s essential to prepare your portfolio accordingly.  When that correction occurs, it could be an opportune moment to invest aggressively, ideally at the market’s nadir.

Would you consider buying OpenAI at a $150 billion valuation?

We are evaluating that option.  However, there are several concerns, particularly related to the company’s structure as a limited partnership, which has a preferred waterfall for payouts.

Do you think they’ve lost their engineering edge?

They have lost some talented individuals, but I believe they will manage to sustain their position in the industry.

How did you navigate investing in Anthropic, given their earlier “Effective Altruism” commitment to donate profits to charity and other constraints?

We had to advise them against making such statements, and it took some time for that message to resonate.  Being engineers ourselves, we found it easier to communicate our concerns, which likely lent us some credibility.

So once their valuation started rising, their altruistic intentions began to wane?

Precisely.  It’s fascinating to observe how capital markets influence such shifts in focus.  When OpenAI reached higher valuations, their initial nonprofit aspirations seemed to take a backseat.

Do you anticipate that AI model companies will go public next year?

Some banks have posed that question to us, but I don’t foresee significant incentives for them to do so, given their ability to raise capital in private markets.  I would welcome a return to an environment where public offerings were more common.

But surely they require substantial funding?

They do, especially for computational needs.  However, I believe that estimates for computing costs will decrease over time, alleviating some concerns.  The structure of preferred payouts for early investors, particularly with Microsoft involved, creates additional complexities for companies like OpenAI.

Best of luck! Thank you!

**

Favorite Books & Media

Invest Like the Best: Gavin Baker – AI, Semiconductors, and the Robotic Frontier

Listening to this podcast was like opening presents on Christmas morning, with great thoughts on AI/LLMs, networking tech, why global warming has been solved, Tesla’s FSD break-out, and the unique traits of Musk & $NVIDIA’s Jensen. I highly recommend you listen.

https://podcasts.apple.com/us/podcast/gavin-baker-ai-semiconductors-and-the-robotic-frontier/id1154105909?i=1000666758592

 

AI Can (Mostly) Outperform Human CEOs

Perhaps not surprisingly (based on some of my own anecdotal experiences as an activist investor!), AI often outperformed human CEOs and excelled at optimizing corporate strategy and profitability.  However, AI struggled with black swan events, leading to catastrophic blowups.  This likely reflects a lack of “imagination” due to limited training data.  It will be interesting to see how this reasoning ability develops in the future.

https://hbr.org/2024/09/ai-can-mostly-outperform-human-ceos

Meb Faber Podcast: Rob Citrone on Milei’s Argentina

A veteran of Tiger Management and his own firm Discovery Capital, this is the first time I’ve heard Rob Citrone speak and I really enjoyed it.  He shared a number of good stories and shared a number of investment perspectives – including his bullish take on Milei & Argentina, bearish outlooks on China and Europe, and his view that U.S. stock market returns are likely to more subdued and volatile from here.  He also shared his concerns about the U.S. fiscal position.  I also enjoyed listening to some of Citrone’s thoughts on short selling.

https://podcasts.apple.com/us/podcast/tiger-cub-rob-citrone-on-mileis-argentina-a-bullish/id1128955736?i=1000670150129

 

In Good Company Podcast: John Elkann, CEO of Exor

John Elkann, the Chairman & CEO of the Exor Company (and part of the Italian Agnelli family that controls Exor), sat down to discuss his thoughts on the benefits of family-led holding companies and long-term time horizons, and digs a bit into Exor’s portfolio including Ferrari (NYSE: RACE), Stellantis (NYSE: STLA), CNH (NYSE: CNH) and others.  Elkann is a thoughtful investor with incredible contacts, but rarely sits for interviews like this.  I have followed Exor for years, and have found their recent public company minority stakes in businesses like Philips (NYSE: PHC) and Clarivate (NASDAQ: CLVT) (as well as VEON and Sibanye Stillwater) to be intriguing.   I would also recommend reviewing our 2023 interview with David Marcus, who talks about the benefits of investing in family-owned businesses: https://ragingcapitalventures.com/an-entrepreneurs-perspective-investing-in-great-family-businesses/

https://www.youtube.com/watch?v=BLdDJ8BWqhw

 

Kevin Kelly: Excellent Advice for Living: Wisdom I Wish I’d Known Earlier

Written by Kevin Kelly, the founder of Wired Magazine, Excellent Advice for Living is a short book of collected wisdom that is shared in a succinct and thoughtful style.  Some good quotes: “The more you are interested in others
the more interesting they’ll find you. To be interesting, be interested.”; “Learn how to learn from those you disagree with or even offend you. See if you can find the truth in what they believe.”; “A great way to understand yourself is to seriously reflect on everything you find irritating in others.”  A great stocking stuffer!