Startup Research: Oxymoron or Key to Success?

Should early-stage AI startups do research?

Image generated with Runway.

I used to have a mantra: “startups shouldn’t do research”. My argument was that research burns someone else’s money and your equity while you try to figure out your problem.

Times have changed and so has my thinking.

My old thinking went like this: academic spin-offs have a big advantage that many people don’t realize. Academic funding is a “fixed-dilution” investment. Your technology licensing office is going to take a percentage of your company when you spin out. But it doesn’t matter whether you spent 1 year, or 5, or 10, developing the technology. They’ll still want the same amount. So stay “spun in” as long as necessary to de-risk the technology. Spin-out only when you know it works and are ready to commercialize it.

While the above is still true for many fields, the crazy pace of AI has changed my thinking. Whatever technology you’re working on in the lab today is likely to be outdated in 6–12 months. So what an AI startup needs is a team that can quickly absorb and take advantage of the latest breakthroughs. Successful AI startups are running on the constantly evolving bleeding edge of the field.

To win you have to push that evolving edge further and faster than your competition, and you’ll have lots of competition. By definition, this means that a successful AI startup is doing research. There’s no alternative.

It would be a mistake, however, to think about this like academic research. Here are some “principles” about how an AI startup should think about research. (Note: my focus here is on early stage, deep-tech, AI software startups with fewer than 60 employees.)

• You need a core science team that is immersed in the academic literature. They need to be following the latest papers and constantly evaluating where new ideas may be disruptive.
• You need a culture that is not scared of such disruption but is excited by it. You need to be sharing papers that challenge your assumptions. This is very much like an academic culture.
• To do this successfully, you need the best people. The best scientists are not threatened by new ideas. They’re excited by them. They see opportunity.
• Adaptability is the key skill. This differs from “traditional” academic research, which rewards people who stick to an idea over the long run and eventually emerge victorious. AI is changing so quickly that one has to change with it.
• But adaptability needs to be controlled. I’ve seen startups panic at the rise of GenAI and flip their entire business model (unsuccessfully). You need to know what’s constant and what’s evolving. Your business problem and customer needs are your rock. When technology changes, your customers don’t go away. They still have the problem they had in the beginning — the problem that motivated you.
• Work backwards from your customer. This Amazon principle is as true in the AI world today as it was before. The key is to see how evolving technology can be used to better solve your customer’s problem. So don’t panic when the technology changes. If you picked the right problem, your customer is still there and still needs you. Figure out how to use the latest technology to serve them better.
• To hire the core ML team that can play on the bleeding edge, you need to have a culture of publishing. Every top scientist wants to share their knowledge — they live and breath arXiv. Publication has to be supported and easy. This is actually a competitive advantage for startups. BigTech companies have arcane and time-consuming processes for clearing publications. Startups can be very nimble. Decision making is easy and they can act quickly to protect IP and get work out.

To put a pin in it: you need the best scientists to win. The best scientists want to be with other top scientists. They want to publish their work at top conferences. They want to work at places where they can do top work and publish it. If you don’t support this, you lose.

How do you avoid becoming an equity-burning academic lab?

Three things: customer, customer, customer. Everyone in the organization, including the scientists, must understand the customer problem and work backwards from that. In contrast, academic research often takes the following forms

• Come up with a clever problem that nobody has thought of before. This is great for your academic career if others follow you. You can create a new ‘area’ or field.
• Jump on the latest fad and try to push the idea further. There will be lots of interest but also lots of competition.

Startups have to avoid both of these like the plague. Your problem is your customer’s problem. They’ve got mice and you’re building a better mousetrap. Can you use the latest, hottest, GenAI method that you saw on X to catch mice? No? Can you adapt it to catch mice? No? Then move on.

Why startups have an advantage over Big Tech.

AI researchers today are focused on “impact”. They’re more likely to measure success based on “stars” on GitHub than citations on Google Scholar. While it’s true that, in a big company, there’s the possibility your work may impact millions of customers, it often doesn’t. The dynamics inside a big company are such that you are a tiny speck in a large system. Any one person’s technology contribution is unlikely to move the needle. Most work in a big company never sees the light of day. In a startup, everything you do has impact. Big companies are ocean tankers and startups are like kayaks. Every movement in a kayak has an impact. If you make something work, it will get out.

The science of engineering.

This focus on impact means that it is easy to align science with products. One needs a culture of relentless delivery. Build it, publish it, ship it. Engineering and science have to be aligned and this is one of the trickiest parts of the puzzle.

It’s natural to break people into teams based on skills — science vs engineering. This is death. Whatever you do, don’t build an ivory-tower academic team inside your startup. BigTech can afford a “research lab” if they have an effective monopoly, but a startup can’t.

As soon as you build a divide between science and engineering, you drastically slow progress. Scientists will want to “keep their hands clean” and throw code over a wall to engineering. Don’t build a wall. I mean this literally. Don’t separate people physically. Keep Slack channels open to everyone. Get the whole company excited about the latest tech appearing on arXiv — product, sales, engineering, everyone.

Relentless innovation has to be matched with relentless delivery.

Where startups have an advantage over academia.

Scaling laws are called laws for a reason. Break them at your own risk. With our existing AI technology, scale matters — data, model size, compute. Scale costs money. Data, people, and GPUs are all expensive. Too expensive for many researchers in academia to compete.

The most prestigious grant a young academic can get in Europe is the ERC Starting Grant. This gives 1.5M EUR over 5 years. This is peanuts in AI today. If you’ve got a good idea that will have customer impact, you could raise much more than this in a startup. Then you have a real chance of success.

Summary.

Startups are the sweet spot today for doing impactful AI research. You can get the resources and have direct connection to customers. Managing research in a startup, however, is very different from managing it in academia or a big company. Get this right and you can attract and motivate the top talent that will keep you ahead in a fast-changing tech landscape.

A brief bio as it relates to managing industry research.

My first experience managing industrial research was during the last big AI boom in the late 1980’s at a Bay Area AI startup called Advanced Decision Systems (ADS). That taught me that a business based on government contracts was no fun. After my PhD, I managed reseach at Xerox PARC, where I learned the good, the bad, and the ugly parts of managing research in a big company. John Seely Brown was an inspiration. Managing a basic research lab in a big corporation takes guts. You have to make big bets on a bunch of crazy people, protect them, and then trust that something good will come out. Good did come out but Xerox was famously bad at making it pay off. My first academic spin-off was Body Labs in 2013 and this was my first experience with venture funding. The lessons I learned there will fill a whole other blog post or two. When Body Labs was acquired by Amazon, my main goal was to integrate the people and technology into Amazon and to ship great products. We worked on Halo and Made for You, among other products and got our technology into the hands of customers. What I learned about successfully integrating an acquisition into a big company would also fill another blog post. Being in project reviews with Bezos was like a crash business course. He is the most customer focused person I’ve ever met and Amazon’s success was built on his deep insights into customers. Amazon’s leadership principles also changed how I approach research in academia, which I didn’t expect. Now I’m working with an amazing team at Meshcapade and it’s the most exciting time in my career because things really work. We have the tools now to solve big problems that were only a dream in my previous roles.

For additional tools and resource ➡️ Visit StartupStash
Zendesk is giving $75,000 in credits and perks for startups! ➡️ Apply Now!