My keynote at the Doctoral Consortium of ACM RecSys 2026 in Minneapolis
On 27 September 2026, I gave a keynote at the Doctoral Consortium of ACM RecSys 2026 in Minneapolis. The title was deliberately provocative: “When AI Does RecSys Experiments for $1: Why Spend Four Years on a PhD?”
The title sounds cynical, but my conclusion was optimistic. If AI takes over more of the mechanical work in research, a PhD may become more interesting. The role of a PhD student, however, will change substantially.
The consortium was organised by Markus Zanker, Xia Ning, and Alex Tuzhilin. Markus opened the event, followed by my keynote. Afterwards, the consortium split into three small tables, each with a few PhD students and several senior researchers. At my table, Christine Bauer and Alex Tuzhilin joined me in discussing the work of three PhD students. This is one of the things I like about doctoral consortia: students get enough time to discuss their research properly and to get to know senior people from the community outside the usual two-minute coffee-break conversation. If you are a PhD student and get the chance to join a doctoral consortium, take it. Unfortunately, I had to leave after the first sessions.
From SciGen to AI scientists
I started with SciGen, the 2005 system that generated nonsense computer-science papers. SciGen produced text that looked like research. Today’s AI-scientist systems increasingly carry out parts of the actual research process.
Sakana AI’s AI Scientist, for example, covers idea generation, novelty checking, coding, running experiments, analysing results, writing a manuscript, and even automated reviewing. This goes far beyond “ChatGPT helps me write”. The research loop itself is being automated. Although, in our recent evaluation, we concluded that Sakana’s AI Scientist performs at the level of an unmotivated undergraduate student rushing towards a deadline. Anyway, a few years ago, such a level of work was unthinkable.
I described this development as three generations of AI.
- Generation 1: Chat. Understand and communicate.
- Generation 2: Code. Use tools and write software.
- Generation 3: Research. Plan and execute substantial parts of scientific workflows.
Google, Meta, Anthropic, Microsoft, and several research groups now have systems aimed explicitly at scientific research or recommender-systems experimentation. Our survey on AI scientists documents 62 AI-scientist systems or capabilities by September 2026, with particularly rapid growth since 2023.
Research is becoming cheap
My central argument was simple: AI research is becoming fast, scalable, and cheap.
I showed examples from mathematics, biology, medicine, and computer science. Some require a lot of compute. Others cost surprisingly little. FutureHouse’s Robin reportedly performs a research run for about $10.76. Our AutoRecLab can turn a natural-language RecSys research prompt into executable code, run the experiment, produce plots, and analyse the results for around $1 per run. A comparable experiment may otherwise require roughly one to three researcher-days for implementation, debugging, execution, and plotting.
If this kind of experiment costs one dollar and a few minutes instead of a couple of days, you will not run three experiments. You will run thirty, or three hundred.
We currently spend a great deal of time discussing whether researchers may use an LLM to improve a sentence in a paper. Meanwhile, the marginal cost of substantial parts of empirical research may collapse.
Why RecSys is particularly affected
Recommender-systems research is unusually suitable for automation. Much of our experimental work follows an executable pipeline: choose datasets, preprocess them, select algorithms and baselines, define experimental conditions, run models, calculate metrics, analyse results, make plots, repeat. Current agents can already automate a substantial part of this loop.
The opening image of my talk therefore contained two huge buttons: “Start Research” and “Build / Improve Production System”. Today we already have a spectrum from classical AutoRecSys and hyperparameter optimisation to increasingly agentic research systems. The line between carrying out research and asking an autonomous system to carry it out may blur considerably.
Advice 1: Become the boss of AI agents
This was the part of the talk that mattered most to me because I was speaking to PhD students. My main advice was to get used to being the boss of AI agents.
One slide showed a PhD student as the conductor of an orchestra of agents. One agent performs a literature review. Five run experiments in parallel. Others prepare teaching material, maintain the website, arrange conference travel, draft emails, and handle outreach. On the student’s sweater, “PhD Student” has been changed to “PhD Conductor”. Next to him is a stack of books titled Delegation, Multitasking, Managing Change, Creative Thinking, and Leadership.
The image is exaggerated, but I think the direction is right. The important question shifts from “Can you do every step yourself?” to “Can you decide which steps matter, formulate good tasks, allocate resources, evaluate the output, notice when something is wrong, and combine the results into something meaningful?” These are very different skills.
Advice 2: Do more, do it really well, and sell it really well
AI will raise expectations. If everyone can run more experiments, write more code, and explore more ideas, the bar for a meaningful contribution rises. My summary slide said: “Do more, do it really well, and sell it really well!”
The accompanying images showed ACM TORS at the gym, in court, on the beach, and at a fashion show. Research does not end when an experiment finishes. Someone has to identify the important question, distinguish an interesting result from noise, explain why it matters, defend it against criticism, and make sure the right people know about it. When producing research gets cheaper, judgement and communication become more valuable. With AI, one can create “advertisement” that until recently only expensive marketing agencies could have created.
Here are some examples.
Submission numbers at the major AI conferences have grown rapidly, ICLR in particular. I would not attribute this simply to generative AI making papers easier to produce. AI itself has become one of the most active research fields in the world. Since ChatGPT, researchers, universities, and companies have rushed into the area, creating more groups, more projects, and consequently more papers. Cheaper AI-assisted research may add another acceleration on top of that. More output, however, does not automatically mean more knowledge.
Advice 3: Embrace change, be critical, be brave
Another slide simply said “Embrace Change”. Later I showed two versions of the same PhD student. In the first, the student conducts a research operation supported by agents. Experiments are running, literature is being synthesised, outreach is happening, and work that would previously have required several people is happening in parallel.
In the second picture, the agents sit idle while the student works alone at a desk. That slide was called “The Alternative”. The point is not that everybody should maximise paper counts. But deliberately refusing tools that greatly extend your capabilities seems unlikely to be a sustainable strategy.
At the same time, PhD students should remain critical. One slide said: “Be Critical and Brave — but don’t jeopardize your PhD.” Underneath was the line: “Realize, most of today’s AI policies are made by old men who could be your grandfathers.” I am, of course, much closer to that generation than the PhD students in the room. The underlying point is that many rules governing AI use in academia are currently being written by people who built their careers under very different technological conditions. The researchers who will spend the next decades working under these rules should have a strong voice in shaping them.
The downside: AI will not benefit everyone equally
I am personally extremely happy with where AI is today. I enjoy having ideas. I enjoy deciding which ideas are worth pursuing, shaping them, prioritising them, and thinking strategically about where they might lead. I do not particularly care whether I personally implement them. What I enjoy is seeing an idea become real, seeing whether it works, improving it, and eventually seeing other people use it, discuss it, or build upon it.
Today’s AI systems are almost perfectly suited to that way of working. Ideas that previously remained on my to-do list because neither I nor someone in my group had the time to implement them can now be explored. My group members and I can test more ideas, iterate on them faster, and shape them much further than we could before. For me, this is fantastic.
But other researchers enjoy very different parts of the process. Some people genuinely love coding. Others love writing. They may get the same satisfaction from spending a day implementing an elegant system or finding exactly the right formulation for a paragraph that I get from seeing one of my ideas become reality.
For those people, today’s development is much less obviously good. If the task you most enjoy is precisely the task that AI becomes increasingly good at doing, then “AI makes you more productive” is not necessarily an attractive proposition. Productivity is not the only reason we work. We also care about enjoying the work itself.
This may create a new divide. People who enjoy generating ideas, setting directions, prioritising, strategising, and coordinating may benefit greatly from AI agents. Someone who has many ideas but previously lacked the time or people to realise them suddenly gains considerable leverage. People whose strength and enjoyment lie primarily in implementation may see a larger part of what they value being automated.
The same applies to management. The future I described with the “PhD Conductor” requires people to delegate, coordinate, evaluate, and direct agents. Some people enjoy that. Others went into computer science precisely because they preferred programming to managing a team.
There is also a more traditional inequality of access. A PhD student with the best models, powerful agents, substantial compute, and a generous budget may soon have a very different research capacity from someone without those resources. AI could amplify existing inequalities between institutions, countries, and individuals.
The future isn’t 10 work-hours per week
Another slide said: “The future isn’t 10 work-hours per week. It’s 40+ or 0.” I do not mean this literally as a prediction about working hours. The point is that productivity improvements do not automatically turn into leisure.
If one researcher uses AI to accomplish five times as much in the same time, others will probably start using AI too. Expectations rise. What was exceptional becomes normal. Email did not give us more free time because letters became faster, and computers did not create a three-day academic work week because calculations became easier.
The “0” points to the other side of the same development. If some people or roles cannot make effective use of these systems, or if AI takes over much of the work that gave those roles their value, they may find that there is much less demand for what they used to do. The gap may therefore be between people whose work is strongly amplified by AI and people whose work is increasingly displaced by it.
The optimistic scenario is that we spend less time debugging mundane code, formatting plots, and performing repetitive literature searches, and more time thinking. The less comfortable scenario is that we simply do much more. Probably both will happen.
Generation 4?
Today’s agents leave me the parts of research I enjoy most. They can perform increasingly large parts of the research process, but someone still has to provide the goals, choose the ideas, set priorities, allocate resources, and decide what is worth pursuing. That is essentially what I called Generation 3.
For now, that arrangement makes me very happy. But Generation 4 would mean that AI also generates the ideas, chooses which ones are promising, develops the strategy, allocates the resources, decides which experiments should run, and evaluates what to do next.
Then the question becomes personal for me as well: if AI does not only implement my ideas but also has the ideas, develops the strategy, and decides what should be done, what exactly is left for me, or for anyone?
I do not know. My slide for this was a giant question mark disappearing into a black hole with the line: “Will you be needed? Who knows?” “Who knows?” is still my answer.
So why spend four years on a PhD?
At the end I returned to the title: “When AI Does RecSys Experiments for $1: Why Spend Four Years on a PhD?” My answer on the slide was: “What a Stupid Question!!!”
Cheap AI research does not mean that a PhD student should achieve less. It means that a PhD student may be able to achieve much more. So perhaps the better statement is: “Because AI does RecSys experiments for $1, you should spend four years on a PhD — and truly achieve something!”
Use these systems. Learn to direct them. Stay critical, be brave, develop judgement, and have ideas. Most importantly, use the additional capacity to attempt research that would previously have been unrealistic for a single PhD student.
0 Comments