The Road to Superintelligence
From DeepMind's founding engineer to raising $130M for groundbreaking AI
This is Startup Pirate #117, a newsletter about technology, entrepreneurship, and startups every two weeks. Made in Greece. Here’s what we recently explored:
If you’re not a subscriber, hit the button below and come aboard:
The Road to Superintelligence
“When you are young, take as many risks in your career as possible. Join the craziest, most ambitious startups because this is where most breakthroughs happen. This is the best way to work on things that matter and have an impact.”
Ioannis Antonoglou is one of the very few people who have had such an impact on the AI breakthroughs of the last decade. Ioannis was a founding engineer at DeepMind, playing a key role in industry-defining moments from Deep Q Network and AlphaGo to Gemini. He recently left DeepMind with a banger announcement: raising $130M to build superintelligent autonomous systems with Reflection. We discussed his journey and the past and future of frontier AI. Nuff said. Let’s get to it.
Alex: You had worked as a founding engineer at DeepMind for over a decade before starting a company to build superhuman AI agents. If I'm not mistaken, DeepMind was the first company to embrace the term AGI, which has become somewhat bloated nowadays. What does AGI mean to you?
Ioannis: You're right. Starting in 2010, DeepMind was the first company whose goal was to build AGI. Artificial General Intelligence describes a system that can do everything a human can, given the same inputs and action space—a system with human-level intelligence that can perform meaningful knowledge work the same way a human can. You can delegate tasks to it and receive the results without human involvement.
However, as we got closer to AGI with technologies like AlphaGo or ChatGPT in the past years, we started to move the goalposts. The tendency now is to talk about AGI as a system that's an expert in every single field—the best that a human could be in all domains. Yet, this trait is highly unlikely to be found among humans. On the contrary, this behaviour is better represented by what we define as a "superintelligent" system or ASI, Artificial Superintelligence. Several companies have adopted this term recently. Unlike AGI, where the system achieves human-level intelligence, ASI is a system that is better than the best human expert.
Alex: I want to take you back to 2012. You have just graduated from university and joined a small startup developing computer programs to master video games. Was it like a handful of people back then? What led you to DeepMind?
Ioannis: I graduated from Aristotle University of Thessaloniki in 2011 and then pursued postgraduate studies in machine learning and AI at the University of Edinburgh. In 2012, I moved to London and joined DeepMind. We were 25 people then and joined as employee #6 in the Research team.
I wasn't sure anything groundbreaking would come out of it, but it was a very ambitious team. As a young engineer, it seemed like the coolest project, and I decided to embark on this journey instead of the consensus path of joining an established tech company where the scope of my role and potential impact would be limited. It was the right time to make a bold, risky move in my career. I remember thinking, "I'll just spend a couple of years doing really cool stuff, and then I'll get a real job". Well, things played out quite differently!
The company's mission was to "solve intelligence and use it to solve anything else". We started with video games like Breakout, Pong, and Space Invaders. Video games are self-contained worlds with clear rules and measurable outcomes, providing an ideal environment for training algorithms to learn, adapt, and strategise.
We went from systems that outscored professional video game players to winning the #1 ranked player in the world at the Chinese board game of Go—a game considered far too complex for machines, to predicting protein structures better than the most high-esteemed biologists and much more. It was a fascinating journey, to say the least!
Alex: You played a key role in all these industry-defining moments for AI. Your name is under the Deep Q Network, AlphaGo, AlphaZero, MuZero, and Gemini research papers. I don't want to sidetrack, so I'm leaving a link to the recent video I put together about DeepMind's story here for anyone who’d like to learn more. At what point did you feel that AI was reaching levels of superhuman intelligence?
Ioannis: I'd say there were three significant milestones. Deep Q Network (DQN) was the first aha moment. In 2013, we developed an algorithm that could master everything from shooters to car-racing games. It might be evident in hindsight, but we were still scratching our heads then. We weren't sure it was the right path. Seeing something working was an incredible moment and gave us more confidence that we would get intelligence out of these systems.
Then in 2016, AlphaGo won against Lee Sedol. Lee Sedol was one of the best human professional players in the game of Go. Three months before AlphaGo came out, people thought that beating a human professional in the game of Go was 10 years away. Three months later, we had this big game in Seoul that showed how these systems were already superintelligent. At that point, I could see we had achieved narrow superintelligence, in the sense that for something which requires a lot of intelligence, we had a system that does much better than humans. Of course, it had a narrow scope. It could only play Go, and then you had to take the algorithm and train it on chess or shogi. It wasn't the general system which could solve many different tasks.
Then ChatGPT came out, and also through my involvement with Gemini, I saw the generality of systems being able to complete a wide range of tasks. An AI program that knows where to start from, how to handle a task, and has some minimum level of intelligence to comprehend it and start trying to solve it. At that point, I felt we had all the ingredients to build superintelligence. This is what Reflection is all about.
Alex: We’ll get to Reflection in a minute, but before we do… What were the breakthrough technologies that enabled all this?
Ioannis: Several breakthroughs happened in the previous 10 years. Reinforcement learning and the ability to take these systems and optimise them using a reward model is undoubtedly one of them. However, several other technologies came together at the right time for general systems to emerge. The idea of next token prediction had been around for a long time in AI, but better deep learning architectures like the transformer, which came out in 2017, enabled the advent of LLMs. But ChatGPT wouldn't be possible unless we could amass internet-scale data and have the compute power to train these systems. All these couldn't have been possible 20 years ago.
Richard Sutton, in "The Bitter Lesson", argues that there are two methods we know will always lead to better performance in AI systems: learning and search. Today, they come together at a great scale, which can create general superintelligent systems. Big models, lots of data, better reinforcement learning methods on one side, and inference-time compute (leveraging more compute at inference time) with reasoning and thinking models like OpenAI o1 on the other.
Alex: You left Google DeepMind with a banger career announcement: Raising $130m to build superintelligent autonomous systems, starting with coding agents. I'd love to understand why you chose coding as a testbed and your vision with Reflection.
Ioannis: At Reflection, we want to build superintelligent autonomous systems. The way we think about it is that it's an AI model coupled to a computer, and you can delegate tasks to it. It's like giving AI access to your computer, and while you go on with your life, AI performs the tasks and gets you the results.
Then the question becomes: What's the best way for this intelligence to interact with the computer? Is it via the mouse movements? Via your keyboard? Is it the same way humans interact with computers? Humans have been shaped over thousands or hundreds of thousands of years of evolution. However, these systems were trained differently; they have different priors and biases. To them, coding comes more naturally. The natural way for AI models to interact with the computer is programmatically via code.
So, we concluded that the root node problem, which needs to be solved first to achieve superintelligence, is the ability of a system to reason and have a way to interact with the world that's well defined. And that's autonomous coding agency. The ability to try things, receive feedback, correct your assumptions based on this feedback, self-correction, in-context learning… All these capabilities are present in existing models, but we need to make them stronger to achieve superintelligence.
This is what we aim to do with Reflection, so join us if that vision resonates with you.
Alex: Playing devil's advocate here, several plausible bottlenecks could slow things down, e.g. limited compute, algorithmic progress becoming more challenging, not enough data. Going forward, what are the main challenges in building AI systems that are vastly smarter than humans and capable of novel, creative, complicated behaviour we couldn't even begin to understand?
Ioannis: We need to derisk several areas across data, compute and algorithms. There's a clear roadmap that shows us how to get there, which is something you can also sense in the industry. No one questions whether we'll get there. It's more of a matter of time. Some people say we'll get there in 2-3 years; others say it's more like 5-10 years. But there's an agreement among most experts that we have the ingredients, and it's a matter of putting them together in the smartest possible way to overcome current bottlenecks.
Yes, we should find ways to generate synthetic data, scale to even more compute, better algorithms, scale reinforcement learning further, train for agency, etc. But, I cannot foresee any significant blockers. It's work and the necessary steps that need to happen to get us to the next level of intelligent systems.
Alex: What are some AI research areas you are most excited about today and paying attention to?
Ioannis: I have been excited about reinforcement learning for a long time. Reinforcement learning is a paradigm of machine learning inspired by how humans and animals learn from rewards and punishments. The idea is to create systems that don't need to be explicitly programmed for every task. Instead, these systems can teach themselves by interacting with an environment.
After ChatGPT came out, up until last year, reinforcement learning had taken a backseat. The industry was all about pre-training and scaling data and compute. But I think now people have started to realise reinforcement learning is the way forward that will take us from something that just works to the superintelligence required to solve real-world problems. It's something we also saw with AlphaGo. Training big models was a big part of it, but being able to think more or less before taking an action was equally important.
Alex: You worked on Gemini before, and Reflection focuses heavily on language models. What’s the best way to approach benchmarks and evaluate the progress we make with LLMs?
Ioannis: We are getting to the point where the benchmarks we use to evaluate the performance of these models are saturated, and we have to come up with better benchmarks. But you know what? Academic benchmarks don't matter. What matters is solving problems for your customers and having an impact in the real world. This is why we have structured Reflection to be 50% product and 50% research. We are not aiming to build a big model that is good in benchmarks but does not provide value.
The other day, Satya Nadella, Microsoft's CEO, said something along the lines of "It doesn't make much sense to talk about AGI and superintelligence in an abstract way. We should measure its impact on the world economy. When the world economy starts growing by 10%, we know we are getting closer to AGI because we have unlocked something that wasn't there before." And I agree 100% with that. We can use benchmarks to keep track of our research progress, but what matters is the impact these systems have in the real world.
Alex: Before we wrap up… How would you advise a recent graduate who wants to get into frontier AI? Where would the 23-year-old Giannis finishing off his MSc degree start today?
Ioannis: There are many good opportunities for young people nowadays. I would strongly encourage people to join startups. When I graduated with my master's degree, I interviewed with many established tech companies like Microsoft. But I just decided to join a startup. When you are young, it makes sense to take more of these high-risk, high-reward bets and see where life takes you.
Especially among Greeks, there is this tendency to go for the safer bet. This is perhaps a cultural thing. Most families believe there is too much uncertainty in the country, so they teach their kids to go with the safer, low-risk path for everything they can control.
I recommend the opposite: When you are young, take as many risks in your career as possible. Join the craziest, most ambitious startups because this is where most breakthroughs happen. This is the best way to work on things that matter and have an impact. If you join an established tech company, you end up another cog in a big machinery. But if you enter a startup, you can immediately hit the ground running, as they say, and have an impact.
Alex: That’s a fantastic way to end this chat. Thank you!
Ioannis: Thanks, Alex.
Jobs
A list of over 400 startup jobs in Greece from 100 companies hiring! Check it out here.
News
Runway (video AI) raised $300m Series D led by General Atlantic.
Galatea Bio (genomic research & clinical genetic testing) announced $25m.
Brainomix (AI imaging technology for healthcare) secured €16.7m Series C.
Wheelsys (SaaS for car rental & leasing) announced funding from Halcyon Equity Partners.
Wealthyhood (investing platform) raised €3.6m Seed.
Plugsecure (cybersecurity for electric vehicles) secured €720k led by Apeiron Ventures and Genesis VC.
Schema Labs (civic engagement) received a €100k grant from Helidoni Foundation.
Squaredev (AI custom solutions) was acquired (majority stake 51%) by QNR.
Resources
Sanne Goslinga and Marathon launching the 3rd edition of the Greek Tech Compensation Report. If you're a Greek tech company, you can participate here.
The distorted reality of startup success stories by Dimitris Glezos, founder of Transifex.
How Skroutz became Greece’s biggest online marketplace with George Hadjigeorgiou, co-founder & CEO of Skroutz.
The new role of PM in the age of AI by Babis Makrynikolas, SVP Product & Pricing at Blueground.
Nikos Drandakis, founder of Sync, on launching a car rentals marketplace.
Usage is the moat by Konstantine Buhler, Partner at Sequoia.
How a $1B SaaS turned user content into a growth engine by George Chasiotis, Managing Director at Minuttia.
Events
Join us today for Open Coffee Athens #122!
Talent Days 2025 by linq on Apr 5-6
Unleashing the Power of Gemma by GDG Cloud Thessaloniki on Apr 7
Bootstrapping a SaaS by foobar io on Apr 8
Harnessing GenAI on AWS Bedrock by AWS User Group Thessaloniki on Apr 8
ProductTank Athens meetup #7 on Apr 10
Silicon Valley Greek Biohub Kickoff Meetup on Apr 10
Practical AI Development with Python, Ollama, LangChain & RAG on Apr 10
Learn how to earn from gaming by ath3ns | community & web3 on Apr 10
AI Hackathon for Devs by Microsoft on May 3-4
That’s all for this week. Tap the heart ❤️ below if you liked this piece—it helps me understand which themes you like best and what I should do more.
Thanks for reading,
Alex