“From PhDs to AI Start Ups” Panel Discussion— Key Takeaways
On Tuesday, 14th August, with the support of Machine Intelligence Garage, we organised “From PhDs to AI Start Ups” Panel Discussion at Digital Catapult. Panelists were Ross Harper from Limbic, Jameel Marafie from Headlight AI and Will Jones from Heterogeneous. They kindly shared stories of their journey, transitioning from PhD studies to launching companies.
Below are some unique insights they offered to the audience:
Talk about your transition from PhD to start ups — which skills proved to be valuable and which didn’t? [11:06]
Will Jones: It’s a good question actually — it’s kinda hard because PhD makes you very, very specialised and if you are too broad, they don’t tend to like that — at least that was my experience. Universities tend to put you in a box, where you are an expert of the world, but it can often be the case that no-one really cares about that particular area. If you learn applicable skills there can be a mismatch of value in academia and the start up world, at least in my experience. For me working with software was the best takeaway I had. My particular area was also useful, as it was in AI, so that was helpful. But AI for Heterogeneous is a long term bet, as it hasn’t yet worked in genetics. Overall, I think that there can be synergies but often there can also be a mismatch.
Ross Harper: I think that a large amount of expertise that you build in PhD by its very nature of the specialist structure of that process means that it is unlikely to be applicable to what you end up doing, especially in a start up. Unless you end up doing a PostDoc in that area, it is unlikely to be that applicable. However, I wouldn’t undersell it too much. On one hand, a lot of PhD students underestimate what they do know. They compare themselves to other PhD students, but their baseline, foundational knowledge in some areas is actually quite expertise I have found. Depending also on the type of PhD that you choose, personally from my perspective, project management and work ethic, switching between different projects to make sure an overall goal comes together are very transferable. Also public speaking and communication skills, ability to explain complicated concepts is really really useful, especially when you try to explain what it is you are trying to do with your business.
Jameel Marafie: I would say the most transferable skill for me was project planning, prioritisation. These became super important in a start up, because there are just so many things going on and you need to know what to do next on your list. It will always go on and never end. This is similar to PhD, where there is always another experiment you can do and you can keep going on and on and on… Again some elements of public speaking and your own time management and dedication.
How did you go about validating the pain problem you are trying to solve for your customers? How did you secure your first clients? [23:47]
Will Jones: We could see a lot of research organisations spending six figure sums on projects that we frankly badly organised, badly managed and really ineffective in getting data into researchers hands. So that was the first clue that something wasn't right. Then we started talking to pharmaceutical companies and we realised that they were spending a lot of money recruiting people for their studies. They would also resequence a lot of people because they couldn’t find the data or it was a new project. We realised that we could do the same thing better, faster, cheaper on the cloud infrastructure, adhere best security practices and pay patients directly. At the moment this is handle by many different middlemen organisations that do not add much value.
Ross Harper: We believe there is some tech that should exist that doesn’t today. Now emotional recognition is a novel concept when you try integrate it with AI. It has been around for a while and has been gathering momentum in recent years. But typically companies who do this use computer vision and natural language processing as a way of solving this problem. Our belief at Limbic is that it could be so much bigger and the way to make this bigger is to use the third pillar of the motion recognition, which is to use physiology or biometric information. We are now at an exciting time where more and more people are wearing devices on their body, measuring your heartbeat. So we now have a dataset that we didn’t have access to before.
To create models we need, we require training data. To have training data, we need customers to be using our platform. So where does this cold start problem begin? The way we went out and got our first clients is by saying we want to build this system for you, but right now we can’t do it — right now we can offer you an MVP and for us the easiest one to offer is stress detection. We used existing solutions from academic literature, we packaged this as a platform for developers to use and they started to use it. This validated our hypothesis that the makers of digital products do want a tool that can help them understand phycological state better. So they start using this and begin pushing data to our servers. This means that we have the data we need to begin developing the core tech.
Jameel Marafie: We knew what we were capable of doing — we knew a lot about materials and AI models. I know about many problems and markets, but wanted to find out more from the people working in these industries. We started by finding companies and contacting people that were facing these problems. I started talking to water contractors, who were going and inspecting tunnels and sewers. One of the key things was to just ask them about problems they experience. It soon came to light that they were using a lot of old technology and there was a lot of space to innovate, especially with new forms of sensing and autonomy. People were very keen to tell us about their problems and soon there was a pull, where they were asking us to help them and were willing to pay us money to look into this. That’s how you know you are onto a problem that you can begin to solve.
If you are short on time, you can listen to all their answers on SoundCloud.
Alternatively, you can watch the full video below.
Feel free to drop us an email, if you want to be interviewed or collaborate with us.