Weekend Read in AI
Exploring a new format: things I've seen this week in AI that caught my attention.
I’m experimenting with a new format for the Substack. My posting frequency here has been relatively low - partly out of consideration for you, dear reader! However, with the rapid pace of developments in AI, it feels worthwhile to provide a weekly summary of the most interesting insights I’ve come across. While I won’t send a post for every detail, this new format aims to strike a balance, keeping you informed without overwhelming your inbox.
I will generally try to send these posts on Fridays, just in time for some relaxed weekend reading. I hope you find them valuable!
AI in education
Earlier this week, I came across a post from François Chollet on Bluesky that caught my attention. In it, he considers the possibility that “AI would not only keep students ignorant, but in fact make them fundamentally incapable of learning anything”.
This idea strikes me as awkward for several reasons. How could AI keep students ignorant? It’s arguably the ultimate knowledge tool. I assume this is tied to the second part of his statement, that AI might render us fundamentally incapable of learning. The fear, I assume, is that if people rely on AI for answers - which it increasingly delivers with precision -they might never engage in the messy, critical process of understanding.
But this concern isn’t new. It has echoed through every technological revolution for generations. When a groundbreaking tool emerges, some members of the intellectual elite - those who grew up without it - worry that the next generation will somehow be worse off. I’m not suggesting that François Chollet necessarily subscribes to this view, but the broader trend is hard to ignore.
What surprises me about this argument is how consistently history has proven it wrong. Humans have only become more intelligent over time, and technology has played a critical role in advancing our capabilities. Calculators eliminated the need for mental arithmetic, yet we found new ways to apply our cognitive energy. Similarly, Google made memorizing countless facts unnecessary, and allowed us to focus on using information instead of just storing it.
AI will likely follow the same trajectory. By answering almost any question instantly, and adapting to our individual knowledge levels and communication preferences, it will free us to use our brains for more complex, creative, and uniquely human tasks. Just as we’ve done before, we’ll adapt, evolve, and find new ways to thrive.
Indeed, I’ve in the meanwhile come across two highly relevant studies in the meanwhile. An empirical 2024 study from University of Cologne and Rotterdam School of Management on LLMs in coding education. The main finding: How you use AI matters.
Using AI as a tutor (asking for explanations) → improved learning
Using AI as a solution generator → decreased learning
Overall, this is in line with my experience, but not only for AI. If you use a tool to help you in your learning process, it's useful. If you use it to just copy-paste and save time, obviously you will not learn.
A similarly interesting came from a new randomized controlled trial by the World Bank of students using GPT-4 as a tutor in Nigeria. Teacher-guided AI tutoring, conducted after school for six weeks, boosted test scores by 0.3 standard deviations - matching what students typically learn in two years! The program proved more effective than 80% of other educational interventions, with particularly strong benefits for girls who started at lower performance levels.
AI for drug design
AI for drug design is an an incredibly active area. There were two papers that caught my attention out of many.
🐍 The first is a great story that was published in Nature: Scientists use AI to design new proteins that can neutralize deadly snake venom toxins. Snakebites claim over 100,000 lives annually! Using deep learning methods, a team created synthetic proteins that effectively neutralize three major types of snake toxins. These proteins showed complete protection in animal studies, even when administered after venom exposure.
Combine this with rapid snake identification via smartphone image recognition, and we're looking at a major public health breakthrough thanks to AI. Really amazing stuff with substantial real world impact.
The second was a paper by EPFL AI Center colleagues, also in Nature, on AI-designed protein interactions that can be controlled by drugs. They developed a new computational tool that can design proteins that only bind together when a specific drug or molecule is present. Why is this important? In nature, proteins often work together by binding to each other. Sometimes these interactions need to be carefully controlled - for instance, you might want a therapeutic protein to only become active when you give a patient a specific drug.
The really cool part is that the tool can learn general principles about how proteins and small molecules interact, rather than needing to be specifically trained on each new case. Amazing new opportunities that are opening up in this space! 🤩
Y Combinator, AMLD, and a new podcast
Lots of cool things happening at EPFL. First, the Applied Machine Learning Days, AMLD 2025, is around the corner. This is our 9th edition; tickets are available at https://2025.appliedmldays.org/ We have a stellar program with over 50 tracks and workshops, and amazing keynotes.
Speaking of YC keynotes: One of our guests last year was Jean-Philippe Fricker, EPFL alumnus and cofounder of AI chip unicorn Cerebras Systems - the company behind the world's largest AI chip. I interviewed him for our new EPFL AI Center podcast “Inside AI”, and he had so many great quotes in our conversation that I had a hard time picking the best five. But here are some of my favorites:
1️⃣ "We wanted to be 1,000 times faster, not just 20%. Because if you don’t disrupt, the competition will catch up." - on how to build something truly groundbreaking.
2️⃣ "For three years, I had to tell our board: ‘Sorry, the packaging failed again.’ Yet we kept going, because that’s how breakthroughs happen." - on perseverance in the face of failure.
3️⃣ "We pitched to investors with no slides, just a story. Ten minutes in, we were already talking term sheets." - on fund-raising in Silicon Valley.
4️⃣ "In Europe, even starting is sometimes tough. And it's made tough because they start too small. If someone else could go faster than you, then ask for the money you need to go faster. [...] They will outperform you because they will be able to invest slightly more and go faster than you." - on being first and being fast.
5️⃣ "Where’s the ‘CERN of AI’ in Europe? If you don’t build your own infrastructure, you’ll end up just consuming someone else’s tech." - on technological sovereignty.
I can highly recommend the episode, where he takes us behind the scenes of founding and scaling a hardware startup that built the world’s largest AI chip. JP recounts how a group of former colleagues, some free time, and a leap of faith led to a disruptive solution that secured multi-million-dollar VC funding - at a time when nobody wanted to invest in hardware. He highlights the thrill and terror of repeatedly failing at wafer-scale engineering before hitting a breakthrough, the critical role of team chemistry, and the rapid evolution of AI that’s reshaping everything from infrastructure investments to global competitiveness. If you’re fascinated by moonshots in AI hardware, startup life, and embracing ambitious risks, this conversation gives you a fascinating and unique insider’s view.
Listen to the conversation on "Inside AI" on Apple Podcasts or on Spotify.
That’s not all re AMLD: One of the keynote speakers this year is YC partner Nicolas Dessaigne. In a special event in the evening, he and EPFL alumnus Daniel Yanisse will also be talking about building Checkr and Algolia, how to start a company, and how AI has opened up a whole world of opportunities.
1️⃣ Do you need an AMLD ticket for the evening event? No (but you'll want one anyways 😅)
2️⃣ If you have an AMLD ticket, does that mean you have a guaranteed entry to the YC evening event? No - the evening event is organized by YC, and you will need to apply for it at YC directly: https://events.ycombinator.com/ycatepfl
Varia
There were lots of other things, but one thing that I would like to specifically highlight is an extremely interesting and open-source reasoning model, Sky-T1-32B-Preview.
From the announcement: "We introduce Sky-T1-32B-Preview, our reasoning model that performs on par with o1-preview on popular reasoning and coding benchmarks. Remarkably, Sky-T1-32B-Preview was trained for less than $450, demonstrating that it is possible to replicate high-level reasoning capabilities affordably and efficiently. All code is open-source."
Incredibly impressive results by team at UC Berkeley - bravo.
CODA
This is a newsletter with two subscription types. I highly recommend to switch to the paid version. While all content will remain free, I will donate all financial support to the EPFL AI Center.
To stay in touch, here are other ways to find me:
Social: I’m mainly on LinkedIn but also have presences on Mastodon, Bluesky, and X.
Podcasting: I’m hosting an AI podcast at the EPFL AI Center called “Inside AI” (Apple Podcasts, Spotify), where I have the privilege to talk to people who are much smarter than me.
Conferences: I’m an organizer of AMLD, the Applied Machine Learning Days - our next large event, AMLD 2025, takes place on Feb 11-14, 2025, in Lausanne, Switzerland.