Ad Interim: Provisional Thoughts on AI, Issue #5

Welcome to the fifth edition of Ad Interim, the occasional newsletter on AI from Inner Loop Capital.

Inner Loop Capital is a pre-seed and seed venture capital fund investing at the Frontier of Infrastructure Software and AI. Inner Loop’s thesis is that the $600B+ Infrastructure Software sector is being simultaneously accelerated and disrupted by AI. We are looking to partner with early-stage start-ups in both AI for Infra and Infra for AI. Inner Loop is based in Baltimore and the current portfolio spans from San Francisco to Israel. Inner Loop Capital is a solo-VC firm managed by Justin Label and is currently investing out of a $30m-target Fund II.

What is it that you do, again?

When I talk to others about Inner Loop’s investments thesis – AI for Infra and Infra for AI – especially to others who are not active daily in tech and AI, I sometimes feel that the thesis does not fully land. I want to take a moment to clarify what I mean by this pithy turn of phrase.

AI for Infra

The Infrastructure Software industry – Cybersecurity, Cloud Tech, and Data Science – has about $660 billion in global revenue, split roughly equally across those three categories. It supports $1 trillion in public market cap across just the 30 largest publicly traded companies.

None of these companies have built their entire tech stack from scratch. They are all building atop the commercial and open source projects that have come before them. When Hashicorp went from founding to $6.4 billion acquisition by IBM in 12 years, it didn’t reinvent the basics of deploying and managing software. It recognized where today’s pain points were for its target customers and alleviated them. It built upon existing tech and techniques, somewhat commoditizing lower layers in the process.

So it is with every important company in the sector. They find the “coal face” of their customers’ current work and improve it. The core work that infrastructure software does hasn’t changed in 40 years – managing software, networks, and data efficiently and securely. The “how” keeps changing, and AI is now on the cutting edge of that change.

AI is creating accelerated opportunities for infrastructure software start-ups to rapidly scale and to disrupt incumbents. The incumbents, of course, are not standing still either, often adopting and deploying AI even faster than more nimble start-ups. This is a bit of an arms race, but not a linear one. I believe no one is exactly sure where or how AI will prove most effective in improving infrastructure software. This creates a very fluid situation, which the best founders will find a way to turn to their advantage.

So I have described an adequately large ($660 billion) and rapidly growing (double digit percent) market with the likelihood of coming significant shifts in market share. This is the perfect fertile ground for founders, start-ups, and early-stage venture investing. This is the core of our AI for Infra thesis. Bringing new AI-based innovations to the Infrastructure Software categories that Inner Loop has been investing in since our launch, and where I have investing experience reaching back to 2001.

Infra for AI

The infrastructure software sector where Inner Loop focuses is a small component of the total tech industry. The tech industry globally is about 16 times larger than our infrastructure sector – around $10 trillion, 9% of global GDP. As noted in the prior issue of this letter, Inner Loop does not play for the biggest outcomes in this industry. We are not positioned to fund the next Alphabet, Meta, or Microsoft. We will have to make do with maximum outcomes that are about 10x smaller than those, still plenty large enough to generate outlier returns for our small funds.

However, our fortunes are tightly tied to the broader tech market, including because tech companies are often the first adopters of our innovations. And that $10 trillion tech industry is currently undergoing a significant platform shift, now toward AI. For over 20 years, no analyst has seriously questioned Google’s grip on the search and online ad industry, until the launch of ChatGPT in November 2022 and the arrival of new competitors like Perplexity. That is the kind of tectonic change that comes from platform shifts.

And this platform shift is massive. It’s the first since the dual platform shift 18 years ago to Cloud and Mobile, marked by the launch of AWS in 2006 and the introduction of the iPhone in 2007. It is the largest shift since the commercial Internet (Netscape was founded in 1994), and possibly bigger than that. Crucially, this is the first platform shift when the technology industry is as large as it is today and as influential culturally. I would argue the world may not have a model to understand what happens when a $10 trillion industry disrupts itself. I believe we will be learning new things along the way of this journey.

That platform shift was ushered in by ChatGPT in November 2022, but it was underpinned by a lot of new inventions and improvements at the data management and compute layers: transformer models, vector databases, next-generation GPUs, CUDA, modern data pipelines, and more. This is all infrastructure software! (Except for the chips, and I will happily leave hardware startup investing to people smarter and braver than me.)

The core Infra for AI thesis at Inner Loop is that these were not one-time innovations that are now over, leaving all future innovations to occur at the application layer. Quite the opposite. I believe the infra available for AI in 2022 was comparable to the TCP/IP stack available to Netscape in 1994.

That 1994 infrastructure stack enabled Microsoft to launch this website:

Not bad for what you could do with Window NT Server 3.5 and EMWAC’s HTTPS, I imagine, but hardly what we would consider sophisticated today. (“If your browser doesn’t support images” 🤣) Tonight you can livestream the 2024 Paris Olympics for free, from multiple angles, with real-time statistics and projections. The infrastructure software developed over the last 30 years has played a key role in enabling that.

Our Infra for AI roadmap projects that advances coming over the next decades in capturing and managing data, building and managing models, coordinating local and cloud computing resources, and securing and making sense of it all in a transparent and ethical way will enable AI-based services that will make ChatGPT look like the Microsoft page above. I believe that in a few decades, there will be no significant tech service that isn’t an AI service, that the infrastructure layer to support that is being developed now, and that it will be re-developed over several decades yet.

Long-term, the Infra for AI roadmap will substantially outgrow our AI for Infra roadmap. But in the short term, it is a fraught space. There are a lot of breathless claims and a lot of money sloshing around. It’s Inner Loop’s job in this environment to get rational exposure (you have to be in it to win it!) to founders and innovations that we believe can make meaningful progress on reasonable funding sizes and will be valuable to customers at any point of the hype cycle. It has been invigorating to be focused on that opportunity over the last year and a half.

Recent Investments in Infra for AI

For all that great theory, the rubber meets the road when we make an investment. I hope to further illuminate our Infra for AI investment thesis by diving into the two most recent pure-play Infra for AI companies we have added to our portfolio in 2024.

MLtwist. Before AI models can be produced, and even more so before they can be moved into production, the data that the models consume must be prepared for the AI use case.

Data is almost always coming from multiple sources and in multiple formats. It needs to be normalized, formatted, labeled, and prepped for AI.

Historically, enterprises have had few of the necessary tools or skills to complete this process in-house.

The market leader for this function today is Scale AI. Founded in 2016, Scale has raised $1.6 billion in venture capital, most recently at a reported $13.8 billion valuation. When Scale was founded, few tools to manage an AI data pipeline existed or were mature, and the skills needed to manage this process were quite rare. Scale was founded as a tech-enabled service: Give us your enterprise data in its current form, we will label it and prepare it through a combination of internal tools and mechanical-turk-style human processes, and we will return it to you ready for AI. It has clearly been a rousing success and a key piece of infrastructure in the recent AI boom.

But the founding premise of Scale AI is no longer fully true. The AI tools ecosystem has matured dramatically since 2016. See, for example, this Cap Gemini market map which includes Data and Developer Tools at the bottom – including MLtwist. And while AI skills are still highly sought after, they aren’t impossible for an enterprise to acquire. Increasingly, enterprises have options. The tool sets include commercial third party, open source, and even internally-developed tools. And the enterprise may have some key internal skills and resources that should be brought to bear in the process. Scale is not set up for this, as it resembles a one-size-fits-all black box. Hand us your data, we’ll do what we do, and we’ll get it back to you ready for AI.

MLtwist provides customers the same value proposition, but in a way that leverages the growing AI tool ecosystem, and gives customers more input and control over how the process runs. MLtwist’s core software is workflow management and data pipeline integrations that allow customers to move their data through the various steps of AI preparedness, using whichever tools and processes best suit them. MLtwist also provides software for certain of those functions where the ecosystem is still maturing. To my knowledge, MLtwist is providing the only platform that allows customers to automate and manage their AI data preparation with a choice of which tools are used at which step, including possibly internally-developed tools.

While MLtwist has features of “disrupting” Scale’s business model, this is far from a zero-sum game. AI data preparation has grown so fast – and, core to Inner Loop’s investment thesis, will continue to do so – that there is plenty of room for alternative methods to serve customers in this space. I would, however, like the market to see MLtwist as an opportunity to get some diversification in how their data is handled, and perhaps bring some of their own insights to the process, where appropriate.

We were introduced to MLtwist in March by a Seed/Series A firm that had decided the business was too early for their investment. After a few months of conversation, and meeting the founders in San Francisco, I decided it was a good fit for Inner Loop’s Infra for AI thesis.

MLtwist has several live customers and is generating revenue today. Their commercial customers are generally unannounced, but they have announced collaborations with Palantir and with Carahsoft to serve joint Federal customers. I am excited to be invested in MLtwist as they continue to mature and play an important role in their customers’ AI journeys.

Cerbrec. If MLtwist aids in preparing data for AI applications, Cerbrec assists in designing and tuning the models themselves. Cerbrec brings a drag-and-drop graphical UI to model customization and tuning.

Much like Wix and other modern web development platforms democratized web site design – while still retaining the full code underneath – Cerbrec makes it possible for subject matter experts without the requisite ML coding skills to directly interact with model inputs and weights.

Cerbrec’s TechStars Demo Day video (4 minutes) is available here, to give some first-hand exposure to the company.

Early customers have told Cerbrec that the product accelerates model troubleshooting from weeks to hours, and models are developed with 80% fewer bugs. I believe that increasing the speed and richness of collaboration between subject matter experts and ML experts on the models can have a substantial impact on the quality and value of the entire AI process.

Moreover, I believe Cerbrec might be able to play an important role in AI governance and safety. Given Inner Loop’s background in Cybersecurity, founders working in “Security for AI” have found their way to us since the early days of the current AI boom. For almost two years, LPs and other partners have been asking me whether we are making investments in the space.

To date, Inner Loop has not made a pure-play “Security for AI” investment. This is for several reasons. Some of the hottest companies in the space (e.g., HiddenLayer, Protect AI) were founded before ChatGPT fired the starting gun, and have now raised large rounds at high prices – rounds that are not appropriate for Inner Loop’s fund size and strategy. And so far, we have not found a newer entrant we want to commit to. While there are some promising and important approaches in this start-up ecosystem, most fall into one of two categories – fairly straightforward extensions of existing security fields to AI (data loss prevention, identity management), or black box systems with big but unverifiable claims.

Of the two, Inner Loop is warmer to the first category. Every underlying platform shift unlocks new opportunities for tried and true security approaches. Just see Wiz applying configuration and compliance principles to the Cloud. DLP, IAM, Zero Trust, etc. for AI are important principles and it’s entirely possible good companies will come out of this space. We continue to look here.

The more systemic startups in AI ethics/safety/security have been a little underwhelming to my eye. The way LLMs and other modern AI models work are so complex, the developers themselves often have little insight into what drives their behavior. I have been skeptical that an independent system, sitting outside the model, is likely to provide much insight or any useful guardrails. (Recall, this newsletter is Ad Interim – for the present. I reserve the right to change my mind on any of this at any time!)

I have come to see Cerbrec has a compelling approach to a more straightforward form of AI Safety – trial and error! If models are behaving strangely – hallucinations, offensive responses, sharing secret information, or allowing unacceptable business processes – I am so far unconvinced that an AI-based review of the model is likely to fix it. However, if models can be more quickly tuned, with more input from subject matter experts, and perhaps even with some structural automation and human oversight, problems can be remediated more quickly without taking the entire system down and losing the value it was providing.

While this AI Safety value proposition is yet unproven for Cerbrec, and while I believe the investment stands alone on the model building and iteration that it provides its customers today, I am also excited to have some exposure to possible fixes for the ongoing surprises that continued AI rollout throughout society is certain to bring.

If Not Now, When? If Not Us, Who?

Goldman Sachs and Sequoia Capital each threw some cold water on AI hype in recent reports. The skepticism is well warranted and mirrors some of the fun we had at ChatGPT’s expense in the very first issue of this newsletter in April 2023.

As in any platform shift (or what Mike Maples might call an inflection), caution is appropriate. And even with caution, there will be some tremendous public failures and misallocation of resources. However, none of that dims my confidence that this platform shift will indeed prove to be as valuable to global business as the Internet. It will surely take decades to play out – which is the reason we need not feel rushed to have a perfect investment thesis today. Regardless, I strongly align myself with Anshu Sharma, co-founder of SkyFlow, who made a compelling reply to these two reports. Anshu concludes:

“We are living in a unique time in history where one of the largest tech revolutions is “obvious”, and we get to all assume it’s going to happen — and go build tools and apps on this new platform.

I am building the data security & privacy layer for AI, and we are betting our company on it. What are you building or investing in assuming the AI revolution is low risk and inevitable?“

To be clear, Anshu isn’t saying that any particular implementation is inevitable, and certainly no start-up investment in this (or any) environment is low-risk. But his point is that the writing is on the wall. Those of us who have lived, built, and/or invested through the Internet, Mobile, and Cloud platform shifts have a very strong sense of inevitability about the coming disruptions, dislocations, value add, and value shifts of AI. In a turnabout from traditional tech culture, I often feel that the 50-somethings are more excited about AI than the 20-somethings. Perhaps if you came up in a world where the Internet, Mobile, and Cloud were already built, doing AI just feels like doing more tech. If you saw how the world, business, and culture changed from the Apple II to TikTok, your imagination is unleashed by the thought of another platform shift.

I do, however, think that Anshu has a certain set of blinders on. In the tech, start-up, VC, and Bay Area worlds, the AI revolution may seem obvious. I distinctly had this sense in San Francisco in December 2023. It really felt alive again, like the early days of Cloud or “Web 2.0”. The “what” is obvious, the “how” is far from obvious, and everyone is scrambling to figure out if they can be part of the “who” who writes this chapter of history. It felt real, I believe in it, and I think in hindsight it will be obvious.

But it’s a big world out here, and normies go entire weeks on end without discussing AI, if you can believe it. There is an opportunity, and perhaps a need, to make what is obvious to Anshu more tangible to everyone else. Which is why I like Anshu’s final query: “What are you building or investing in assuming the AI revolution is inevitable?” I cannot promise that your work will pay off exactly as you set out for it to. But I suggest, with very high confidence, that in 2030 you will feel that whatever tinkering, building, collaborating, or investing you did in AI in 2024 was extremely valuable. I do not think you will wish you had waited longer to get started!

Whatever excites you in this space today, especially at the infrastructure and tooling layer, I hope we can collaborate on it together.

AI Industry Updates and What We’re Reading

Bessemer Venture Partners devotes a meaningful portion of their annual Cloud roundup to the AI Cloud. They offer a deep dive into five key trends, with five associated predictions for the AI Cloud. Trend #3, Transforming Humans’ Relationship to Software, felt the most impactful to me. Extrapolating the impact software has had on business and culture, just in a fairly traditional UI over the last 20 years, to an Agentic AI future can spur one’s imagination.

Glasswing’s Rudina Seseri shares her substantial expertise in this video titled “AI Is So Much More Than ChatGPT”. (Registration required.)

Charles Hudson at Precursor Ventures echos many of my feelings in his post, “Today’s Era of AI Investing Reminds Me of 2000 Internet Investing (but not because it’s a bubble)”. Excerpted, and emphasis added:

“When the commercial internet started to take hold in the late 1990s, it felt as meaningful of a shift to me as AI does today. Many people (including me) staked their careers on the belief that the internet would be a transformative force that would remake many parts of society.

That being said, AI in 2024 feels like where the commercial internet was in 1999 or 2000. You could feel things were changing. You could imagine a future where the internet was deeply woven into everything from commerce to community to communications. The only problem is that we were collectively wildly off in how quickly those changes would occur.

There were so many good ideas in 2000 (buying books online, buying groceries online, digital payments, social networking, online publishing, etc.), but they didn’t all work on the first try. It was just too early. Not enough people were ready for these new behaviors. We didn’t have the plumbing we needed. New tech needed to be created. People needed time to catch up with this rapid era of change.”

I associate myself with Charles’ comments, although I think for various reasons (including incorporating lessons learned from 1999), AI’s promise will be realized more quickly than the Internet’s. One thing that is similar, and which I am staking my career on, is that we need continued innovation in the plumbing and core technologies of AI in order to fully realize its business potential. That is the “Infra for AI” part of the Inner Loop investment thesis that I believe will yield fruitful startup investment opportunities for the next 15-20 years. We aren’t too late, but we wouldn’t want to wait any longer to get going.

Justin Label
Managing Director
Inner Loop Capital