'No' to ban on killer robots
Lethal Autonomous Weapon Systems (LAWS) should be regulated rather than banned says Wadham Fellow Tom Simpson in a new Policy Memo for the Blavatnik School of Government.
Date Published: 19.11.2026
New Tutorial Fellow in Computer Science talks about joining Wadham, helping the College face AI, and the revolutions disrupting his field.
Mahmoud Abo Khamis joined Wadham this Michaelmas, becoming the College's first dedicated Computer Science Tutorial Fellow in roughly twenty years. Alongside Mahmoud, Wadham has appointed a new Junior Research Fellow in Computer Science, Joris Nieuwveld, and plans to appoint an additional Tutorial Fellow in the field later in the year.
2026 is a historic year for the College to expand its computer science expertise, having seen a slew of headlines about AI, from hacks, to insider warnings, to mathematical breakthroughs, shift public consciousness on the seriousness of this new technology. We asked Mahmoud to share his thoughts on joining Wadham, the nature of his research, and the impact of recent events on his field.
"Since the day I arrived, everyone at Wadham has gone out of their way to make me feel at home. Wadham seemed to be a very diverse and welcoming community, bringing together people from all sorts of backgrounds, cultures, life philosophies, and schools of thought. I wasn’t surprised to learn that, despite its humble nature, Wadham was actually the birthplace of the Royal Society. It is this kind of diverse and futuristic thinking place where revolutionary ideas are born. I moved here from a similar place, California, that has always been ahead of its time, and now I’ve never felt closer to home!"
"Tutoring! I really enjoy this style of teaching where I get to interact with a very small group of students, viewing a technical problem as a puzzle, inspiring them to explore different ways to visualize this puzzle, untangle it, and invent new puzzles. My first tutoring experience was with my nephews. Since they were little kids, I used to play various logic puzzles with them, and then I started introducing them to the concept of “programming”, where now our aim is to teach the computer how to “think” on its own to solve such a puzzle. It was my nephews who convinced me that I could succeed as a mentor. My oldest nephew is now starting as a freshman at UC Berkeley. He is the same age as my first cohort of students.
I am also looking forward to the college community: The college brings together experts from various disciplines, e.g. philosophy, literature, medicine, engineering, etc. Those disciplines might seem unrelated to one another, but now more than ever, it is quite the opposite. For example, my own field is Computer Science: If you had asked me many years ago “What is the point of Computer Science?”, I would have answered “to build faster or more intelligent computers”. But if you had asked “But why? What is the human value in that?”, I wouldn’t know how to answer. Today, I still don’t know the final answer, but at least, I know who can help me figure it out: Our humanities experts."
"I am an advocate of maintaining a well-balanced attitude towards AI that combines excitement with concern."
"One important challenge that I am hoping to tackle, together with our Fellows from all fields, is Artificial Intelligence (AI). We need to understand the impact of AI on both college education as well as academic research and adapt our educational and research infrastructures accordingly. I am an advocate of maintaining a well-balanced attitude towards AI that combines excitement with concern: We want to remain optimistic about the potential of this rapidly evolving technology, while being genuinely concerned about the risks and planning for them. In the context of tutoring, we need to make sure our students don’t lose their own critical thinking skills to AI. Simultaneously, we want to ensure that they are fully prepared for a future where humans and AI collaborate on research and development. I don’t know the answers yet, but, as a college community, we need to get this conversation going.
I was very pleased to see our Fellow in Mathematics, Prof. Mark Mezei, start an initiative aimed at protecting the learning process in the age of AI, and I am hoping to see more soon."
"It is a bit unfortunate that, in recent years, the term “algorithm” started becoming associated in the media with social networks. The term “algorithm” is actually much wider and much older than that. An algorithm is originally a mechanical step-by-step procedure to compute something. For example, the method we use to multiply two numbers on a piece of paper is one of the earliest examples of an algorithm. In the computer era, the word “algorithm” became almost synonymous with a computer program or software. You can think of every app we use, e.g. navigation, as an algorithm (or a collection of algorithms)."
"I work on databases and “automated algorithm design”.
In the past, database systems used to be simple tools to store tables of data, like Excel spreadsheets. But since then, they have evolved to automatically perform far more intelligent tasks that used to be exclusive to human experts. For example, in the past, a database in the college context used to be a place where you could store student information and answer simple questions like: how many students were taking a particular course? Nowadays, a “smart database system” is trivially expected to be able to take in all time constraints and preferences for students and tutors and somehow produce an optimal tutoring schedule that makes everyone happy.
In the past, solving such a scheduling problem would have required the college to hire a software company to manually design a piece of software for this specific problem. If the requirements of the college were to change, e.g. some tutors now have extra departmental teaching, then software engineers would need to tweak the software by hand to handle the new requirements. But nowadays, a smart database system should be able to automatically design and run an entire scheduling solution on the fly, without any human intervention, and should be able to update it whenever the requirements change. This is what “automated algorithm design” is for databases.
In the above example, a scheduling solution might be able to run smoothly at the college scale, but might struggle to run at the university scale, where the database system is now trying to come up with a global schedule for all students and staff across colleges and departments. A decent database system should be able to automatically produce the most efficient solution that works on a large scale. “Efficiency” is one of the main challenges facing automated algorithm design (but not the only one)."
"Before joining Oxford, I worked for at a company called “Relational AI”. We were building a new generation of smart database systems that are centered around “decision intelligence”. These systems are designed to not only digest and analyse data but to also make high-stakes decisions based on that data. Unlike most companies at the time, our team included a wide research network of academics from many universities all over the world. It was a place that brought together theoreticians and practioners into the same room to work on the same whiteboard.
Thanks to this unique environment, I learned a lot about the similarities and differences between industry and academic research and how to bridge the gap between the two. I had the privilege of supervising several PhD students as interns, helping them implement and test their research ideas in a real system. We also worked on trying to identify the most relevant problems from an industry viewpoint, so that we can put them on the radar of universities to tackle."
"You can think of an LLM as the creative part of the brain that can get “too creative” and start hallucinating."
"As I mentioned before, automated algorithm design is about automatically generating efficient computer code or a “program” to solve any given problem. It might seem that this is something already solved by Large Language Models (LLMs), like ChatGPT or Claude. For example, you can type a prompt to Claude saying, “We have this particular scheduling problem at Wadham with tutors and students... Can you generate a program to solve this?”, and the LLM will indeed give you back a program that you can run on your machine.
But the elephant in the room is trust: You will have no idea whether this program matches exactly the problem you described, produces correct or even sensible results, and how efficient it is. In fact, LLMs are notoriously known for their “hallucinations”. The main puzzle in my area is verification: How do we know that the generated program is correct and efficient? This is a challenge that needs to be solved in parallel to LLMs.
We have all heard in the media in recent weeks about how LLMs managed to solve some of the most significant open problems in mathematics. While LLMs have become quite impressive, this narrative misses half of the story. The other half that equally contributed to this success was Lean: a massive project in the mathematics community decades in the making with the aim of formalizing all the rules of math in a format that is accessible to machines. Lean is exactly what allowed LLMs to overcome their hallucinations in the context of math. It provides an LLM with an instantaneous reality check whenever its “train of thought” drifts into hallucinating false arguments. Now getting back to automated algorithm design, the biggest challenge in my view is developing an analog of Lean for this particular area. This is exactly my research goal.
You can think of an LLM as the creative part of the brain that can get “too creative” and start hallucinating. It needs to be paired up with another part of the brain: the “rational” part that knows true from false and knows exactly how to explain to the other part why a certain argument is false. Each part alone is useless for real-world applications, but together, both parts can do amazing things. Unfortunately, most of the focus and excitement right now is about the first part, but I am hoping to bring attention and progress to the second part."
"We still lack a rigorous theory explaining how LLMs exactly work and what they can and cannot do."
"Over the last few weeks, AI has been sending shockwaves through the Mathematics and Theoretical Computer Science communities, with some AI models cracking a large number of notoriously long-standing open problems in both fields. I didn’t believe I was going to see many of these problems solved during my lifetime. While extra verification of these solutions is still needed, it is probably safe to say that October 2026 was the most significant month in the history of both fields since their inception. The reaction within both communities was, understandably, very diverse and complicated in nature; see this for a sample. It is really hard to give this topic justice in just a few sentences, but I will do my best.
My own reaction was a mixture of shock, super excitement, but also deep concern. We should aim to incorporate the technology in our pursuit to push the frontiers of science, while simultaneously developing a real plan to understand and deal with the negative consequences. (Yes, we can walk and chew gum at the same time!) One aspect that scares me about AI is that it is one area of science where applications are far ahead of the theory: We still lack a rigorous theory explaining how LLMs exactly work and what they can and cannot do. At the risk of sounding like an alarmist, let me give you this analogy: Imagine we somehow figured out how to build a nuclear reactor before we ever understood the laws of quantum physics! Again, I am not asking anyone to stop using AI or stop feeling excited about it; I am just saying we also shouldn’t lose track of the importance of understanding the theory behind it, because it is the only way to achieve safety in AI. Some people might argue we no longer need human mathematicians and computer scientists. I would argue quite the opposite: Now more than ever, we need them to develop the theory of AI and achieve safety in AI among many other things."
"I would say there has never been a more exciting time to be in mathematics and computer science in the entire history of both fields! We, as mathematicians and computer scientists, are not in competition with AI any more than astronomers are in competition with space probes. We are entering a new era in both fields, of cooperation between humans and AI to explore boundaries that were never imaginable before. As the world-leading mathematician, Terence Tao, puts it, it is “Math 2.0”. I believe machines will never replace our human intuition and imagination, but they can certainly complement our human limitations, allowing us to achieve things that were not possible within our short lifetimes. AI will likely be the first to prove Riemann’s hypothesis, but we still need the next Reimann amongst us to guide the field in a direction that serves humanity.
In my view, there are two major frontiers that mathematicians and computer scientists should prioritize going forward. The first is developing a complete mathematical theory of AI, which is essential for using it in a safe and trustworthy manner, as mentioned before. The second frontier is mathematical visualization: For background, the mathematics community has put an incredible effort over the last decade towards opening up math to machines. The effort was a spectacular success allowing LLMs to conquer math. But the downside that no one saw coming is that humans were pushed out. I think it is time we put our effort towards bringing humans back in the loop: Developing a mathematical language that is not aimed at proof verification but at proof visualization for the purpose of communicating mathematical intuition back to humans. This, in my view, is essential for successful cooperation between humans and AI."
Lethal Autonomous Weapon Systems (LAWS) should be regulated rather than banned says Wadham Fellow Tom Simpson in a new Policy Memo for the Blavatnik School of Government.
A Wadham Honorary Fellow has been honoured with an OBE for services to artificial intelligence research.