The State of AI

March 27th, 2024


I am excited to share this interview with Nathan Benaich, the Founder & General Partner of Air Street Capital, a venture firm focused on AI-first tech and life science companies.  Nathan was an early visionary in the AI space, having launched his Guide to AI newsletter in 2015 and then his annual State of AI report in 2018.  Today, he has more than 25,000 newsletter subscribers, 50,000 followers on Twitter, and a new $121 million fund to invest.

Nathan shares his views on the development of AI, including thoughts on the strategies that companies like Apple, Meta, and Nvidia are taking.  He also discusses a number of Air Street’s portfolio investments and outlines the opportunity for AI in life sciences.

I hope you find value in this interview.  Enjoy!

Best Regards,

William C. Martin

P.S. Please tentatively save the date, Thursday, September 19, 2024, for our 3rd Annual Fall Conference in NYC.  More details to come soon!


Topics in this Issue of An Entrepreneur’s Perspective:


Interview with Nathan Benaich, Air Street Capital: The State of AI

Welcome!  To begin, can you tell us a little bit about your background.  What got you interested in artificial intelligence (AI)? 

My original plan in undergrad had been to go to medical school and become a physician-scientist working on translational research in cancer and stem cell biology. During summers working at the Whitehead Institute at MIT, I saw the incredible engine of the Boston biotech ecosystem firsthand. I was captivated by its ability to translate inventions in the lab into spinouts to create products that make a difference in the world.

As I pursued my PhD in experimental and computational cancer research at Cambridge in the UK, I became more and more fascinated with technology and startups beyond biotech. Products like Dropbox, iPhones, and Twitter were launching during my undergraduate studies and had properly scaled by the time I got to grad school. I started immersing myself into the tech and venture capital scene.

As I started to get involved in VC, a few different trends were starting to create the right conditions for an AI boom. These included the global proliferation of smartphones that had created unprecedented volumes of data, rapidly declining compute and storage costs, and a number of fundamental advances in AI research. Back in 2015, I wrote a piece making the case for investing in AI, which I think still holds up well almost a decade later.

You launched your Guide to AI newsletter in 2015 and the State of AI Report in 2018 (sign-up here:  Today, these have enormous followings.  Tell us a bit about that success.

In all honesty, these both started as a leap of faith. In fact, the first installment of my newsletter was an unsolicited email to six friends in 2015.

Both the newsletter and the report were driven by the belief that if you want to be a part of the AI community, then you have to be an active contributor. AI is a vast field, so dedicated researchers and practitioners often don’t have the time to follow every twist and turn outside their own immediate field of study. Years of regularly reading and analyzing AI research on a regular basis also helps when you’re judging potential ideas.

At the same time, people don’t just want a list of research publications or start-ups, so it’s important to bring some opinionated analysis to the table. That’s why we make predictions in the State of AI Report every year and then measure ourselves against us. We won’t get everything right, but it’s much more compelling than playing it safe.

AI has caught the world’s imagination.  What recent developments are most exciting to you?  What innovations are you paying most attention to moving forward?

Multimodality is emerging as the new frontier in AI research. Language alone doesn’t obviously capture the full scope of human reasoning or action, so we’re increasingly seeing powerful models that combine different kinds of data. For example, Google’s Med-PaLM 2 exceeded expert physician performance on the US Medical Licensing Examination after being trained on both medical questions and their accompanying images. The team then took this base model and trained it on real-world diagnostic conversations and medical reasoning to build a conversational system.

We’ve also seen UK self-driving start-up Wayve build a model called LINGO-1, which combines videos of journeys with expert commentary on driving behavior and the scene. You can also ask the model questions via natural language. As well as improving reasoning and planning, it potentially marks a big step forward in the explainability of end-to-end driving models.

Air Street Capital, your venture firm, has also raised two funds, most recently $121,212,121 in September 2023.  Can you tell us about your investing focus?

I’m unashamedly optimistic about the potential of AI to unlock a new era of economic progress and scientific discovery. At Air Street, we look for founders building AI-first companies that tackle real-world challenges. By AI-first, we mean that AI sits at the heart of what they are building, and without it, the product wouldn’t function.

I look for two primary traits in founders. Firstly, deep insight into their customers’ operating context, pressures, pain-points, and how new technology would fit into their way of working. Secondly, technical brilliance matched with pragmatism when it comes to selecting the right tools to use or build. Even in 2024, not every problem is solved with a large language model.

What are some examples of companies that you’ve invested in and are excited about?  Can you tell us briefly about them?

There are a few different themes that we’re looking at closely right now, particularly in defense, techbio, robotics, and enterprise automation. Some recent investments include:

  • Lambda Automata, a Greece/UK start-up producing AI-first hardware and software product that automates the intelligence, surveillance, target acquisition, and reconnaissance process.
  • Sereact, who are building computer vision software for robotic arms and other platforms to enable warehouse automation starting with “pick and pack” jobs.
  • Profluent, a techbio company building an integrated wet lab and computational lab to discover, design and develop protein-based therapeutics with a focus on precision design.
  • Interloom, a new business built by the team behind Boxplot, which was acquired by Hyperscience. Their software combines connectors across enterprise silos, a contextual knowledge graph, and a task orchestrator (AI + human in the loop) with a focus on unstructured tasks.

You have a long-standing interest in biology.  What are the opportunities there?  How do you manage the long, expensive, and risky new product approval timelines?

Historically, significant amounts of work in areas like drug discovery has essentially been grounded in trial and error. Back in 2019, I argued that biology was having its ‘AI moment’. I see AI-first biology as a way of harnessing technology’s capabilities across automation, high-throughput and high-resolution experimentation to help us move from discovery to design.

There’s obviously risk associated with any scientific enterprise, but if a start-up gets the business model right as well as the technology, then the potential upside in sectors like pharma is considerable. Ultimately, part of the appeal of applying AI in biology is precisely the potential to cut the timelines and the costs. 

What type of AI business models do you think will be most successful in the future?  How will startups play with the big tech giants?

Back in 2019, I wrote an essay making the case for building a full-stack machine learning company. I argued that too many new companies working on AI were building part of the stack and then licensing it out to incumbents, rather than creating a fully-integrated product that solves the problem end-to-end.

For example, a company that licenses out a model to a big pharma company will capture significantly less economic upside than one that builds an end-to-end drug discovery platform that owns drug assets.

In terms of who will be successful, there will end up being a mix of big tech companies, start-ups, and open source projects. The AI ecosystem is competitive but also highly interdependent. For example, Mistral uses Microsoft’s cloud infrastructure to support its go-to-market, while ChatGPT massively scaled up the transformers architecture originally published by Google. We’ve seen this dynamic since the early days of the current AI ecosystem.

What do you think of Meta’s (NASDAQ: META) approach to open source its LLM models? 

We recently wrote an essay arguing that open source AI has been one of the biggest drivers of progress in AI and that the modern ecosystem wouldn’t exist without it. It increasingly feels like there are two versions of the AI community: a growing, vibrant open source ecosystem versus a group of big labs that are publishing less and less technical information.

While there may be legitimate safety concerns, there are definitely commercial motivations at work. A highly regulated, closed-source ecosystem will naturally benefit a small handful of well-funded incumbents. Competition, innovation, and safety research would all suffer as a result, while new opportunities for rent-seeking would present themselves. We’re wholly supportive of Meta as they emerge as a leader of the fightback.

Can Google (NASDAQ: GOOGL) protect its search monopoly in this new world?

Google is definitely facing a challenge from tools that produce semantic responses directly to user queries and can filter out large numbers of irrelevant results. These kinds of tools don’t rely on the kinds of user-driven network effects that have created the winner-takes-all dynamic we’ve often seen in digital markets.

That said, Google has two points going in its favor. Firstly, despite its now notoriously poor red-teaming, Gemini shows us that Google is producing some of the most impressive work in GenAI. It’s inevitable that this will be integrated more deeply with its product offering. Secondly, newer entrants to the market need to demonstrate their stickiness. At the moment, many of them still lag incumbents on user retention and ability to convert free trials into paid sign-ups.

How do you think Apple (NASDAQ: AAPL) will attack this space?

Apple has so far approached this space cautiously, but we should expect their offer to follow their standard playbook. This usually involves rolling technology out across their entire product line-up, centering on privacy, and opting for hardware integration over a reliance on the cloud. For example, we know they’ve been exploring the potential of running (and potentially updating) LLMs on-device.

Interestingly, the company is also looking to swerve the copyright battles with publishers that have hit OpenAI, Meta, and others, by preemptively striking content partnerships. It’s been rumored that they’re prepared to pay newspapers in excess of $50M to be able to use their archives for training purposes.

Finally, everyone has been focused on NVIDIA (NASDAQ: NVDA) and the data center space.  Is NVIDIA’s moat defensible?  Are you investing in semiconductor and networking companies to try to capitalize on the opportunities to re-architect data centers for AI?

While we could have an endless philosophical argument about whether NVIDIA’s present valuation is sustainable, it’ll prove challenging for competitors to dislodge the company. Not only did NVIDIA spot the future before its competitors, it understood the importance of adaptability.

Instead of trying to predict the future direction of AI research or customer demand and build bespoke solutions, they produced programmable GPUs that users could customize as they wished, while building an ecosystem of readily-available tools, libraries, and frameworks.

By contrast, a lot of AI chip start-ups focused on building more rigid, bespoke hardware, while underinvesting in the accompanying software. There’s a reason that NVIDIA chips are used in 19x more research than all of their competitors combined. Overturning this kind of lead, while theoretically possible, is a tall order.

Considering the dynamics of the market and how capital intensive a business it is, I’m not actively exploring semiconductors at the moment.

You have actively invested in European-based AI companies. I was curious if you had any observations on how the European startup scene is developing?

I’ve had my fair share of criticisms of the European tech scene, especially around caution and lack of ambition, but the tide is beginning to change. We’re hearing from more and more founders who want to swing big, but are frustrated by the regulatory burden they face and conservative local investors.

We’re also starting to see a broadening of the ecosystem. For years, the UK had the lion’s share of capital, talent, and exciting start-ups, but that lead is looking increasingly fragile. After years of people talking about the potential of the French ecosystem to topple the UK, it’s now looking like that could actually happen.

Whatever macro complaints we have about Europe, there are great founders everywhere if you know where to look. That’s ultimately why specialist investing is so important for early-stage AI-first companies, especially when you’re outside the Bay Area. You’re not going to spot a promising robotics research group in a European university from the 52nd floor of a Wall Street skyscraper.

Thank you for being so generous with your time.  Good luck!


Favorite Books & Media

Liberty in a Cold Climate: Niall Ferguson Lectures at Princeton University

One of my favorite history writers, Ferguson delivers a compelling pair of lectures on the history of liberalism while delivering a stark assessment of the current “cold” climate for open debate on college campuses today.  Ferguson is an excellent thinker and his perspectives and the historical context he brings to the table are valuable.


Conversations with Tyler: Marc Rowan on Financial Market Evolution & University Governance

One of my favorite podcasters, Tyler Cowen, interviews Marc Rowan, the CEO of Apollo (NYSE: APO).  Rowan explains Apollo’s unique, insurance-driven business model, digs into the pros and cons of the fast-growing private credit markets, and tackles university wokeness at places like the University of Pennsylvania.


Interview with Barry Sternlicht

This is a really interesting interview with Barry Sternlicht, the founder of Starwood.  Barry shares a ton of great stories and nuggets from his years investing in real estate, as well as thoughts on the current environment.  He is very transparent and this is a fun listen.


BG2 Podcast: Tesla FSA 12, Imitation AI Models, Open vs. Closed AI Models, and more

This is a super fascinating discussion between two power investors, Bill Gurley & Brad Gerstner, on Tesla’s new self-driving AI model. Tesla basically rebuilt their model from the ground up over the past few years, leveraging new neural learning models & actual driving data from their cars.  Given the amount of training data that Tesla has access to, due to its large fleet of cars on the street, the flywheel implications of this approach are very interesting.  Also, this podcast series is worth listening to in general.


Bill Martin on (More) Bank Shorts

Real Estate Pain for U.S. Regional Banks Is Piling Up, Say Investors


Bill Martin, Who Called SVB’s Collapse, Adds to His NYCB Short and Says It Will End Up a ‘Zombie Bank’


A Selection of Recent Tweets from @RagingVentures:




“Don’t waste your time on small game when there are big beasts in the woods.”

– Theodore Roosevelt