MIT: Rakhlin Explains How Universities Should Prepare for AI Smarter Than Scientists Themselves

AI's Accelerating Progress in Science

Sasha Rakhlin, director of MIT's Center for Statistics and Data Science, has published an essay examining how universities should prepare for a near future in which artificial intelligence surpasses human intelligence across many aspects of intellectual work. This extensive reflection synthesizes ongoing discussions about how AI is transforming mathematics, statistics, machine learning, and engineering — especially at the graduate and research level.

The pace of AI progress is striking. Last year, a neural network model reached gold-medal level at the International Mathematical Olympiad in mathematics. Just a year later, models are already producing original research results, including a recently proposed solution to one of the seven Millennium Prize Problems. Tasks that once required advanced mathematical ability can now be performed automatically.

Rakhlin points to a key factor: the speed and reliability of verifying results. Formalized proofs can be checked automatically, and programs can be run and tested. Wherever evaluation is fast and reliable, systems can generate candidate solutions, learn from the outcomes, and improve. This creates a compounding effect: improved models participate in the next round of development, accelerating progress exponentially. The same logic applies to AI research itself — improving models, training procedures, and supporting tools.

Rethinking Criteria in the University System

For experienced researchers, this opens up the ability to tackle questions that previously required impossible technical resources. But the gap between obtaining a solution and understanding why it works and what it generalizes to will become even sharper.

The current university reward system needs rethinking. In a growing number of fields, a published paper has become a weak signal of individual expertise. Rakhlin warns against simply defining value as "everything AI still can't do." Instead, departments should place greater value on:

• Asking good questions • Replication and reproduction of results • Synthesizing ideas from different sources • Informative negative results • Open datasets

Evaluation must explicitly establish what exactly a researcher contributed and what they take intellectual responsibility for, including cases where significant portions of the work were performed by AI. These expectations should shape the logic of hiring, promotion, and funding — and be clearly articulated for current and future graduate students.

Retraining is especially difficult. Intellectual abilities develop through practice. Routine calculations, coding, failed attempts, and small discoveries all helped students build intuition and judgment. If this work is delegated to AI, students will lose these formative experiences, even though AI allows them to take on more ambitious projects. Universities must distinguish between unnecessary friction and work that builds genuine expertise. Students need to learn how to frame problems, verify model outputs, reproduce results, and defend their choices. Foundational knowledge may become even more valuable as the basis for using these tools competently.

Infrastructure for Scientific Independence

Rakhlin sees a real opportunity for universities: to preserve the ability to pursue questions on long-term horizons, share results openly, and independently verify claims. Industry partnerships are necessary, but commercial priorities won't cover the full breadth of science and may diverge from it over time. Universities need technological independence.

Academia's most important strategic asset is the accumulated knowledge and experience within labs. Published papers are only part of the picture. Failed experiments, abandoned research directions, and the reasons an approach didn't work remain unpublished. Scientists possess tacit knowledge about which interventions are likely to fail and why. This context explains why AI currently struggles in some fields, especially the empirical sciences, to anticipate consequences that are obvious to an expert. These limitations may be temporary; if negative results and researchers' interpretations are collected, models will become far better at scientific exploration.

Rakhlin proposes a vision: MIT's labs functioning as a single living scientific organism, connected through shared AI research infrastructure. Imagine a neuroscientist searching for neuron segmentation methods in microscopy images. An AI agent recognizes relevant progress happening in a computer vision group, connects the researchers, suggests benchmarks, and helps iterate. It could surface unsolved questions for students and faculty, and make results from one lab accessible to others.

This requires building workflows that capture hypotheses, interventions, outcomes, failures, and interpretations. Shared systems need to link these workflows, allowing agents to use tools and information across labs with proper permissions. This demands significant public and institutional investment in computing, secure data systems, and expertise in adapting and fine-tuning AI models.

Such tracking would preserve the lineage of ideas and make contributions — including those of graduate students — easier to recognize. With agreed-upon attribution rules, researchers could collaborate more openly. This would also help answer the question of how to recognize and reward intellectual contributions.

AI systems are already synthesizing information from sources that exceed what any single researcher could absorb, and their ability to make useful connections will keep growing. Universities should harness this power to rally around difficult scientific and engineering problems, helping researchers build on each other's strengths and ideas. Rakhlin sees this as a promising path forward — but universities must invest now to make it possible.