On October 8, 2026, MIT Statistics and Data Science Center Director Alexander (Sasha) Rakhlin argued that universities should prepare for AI to perform many intellectual tasks, rethink how they assess researchers, and invest in shared research systems. In an essay about AI’s effects on academia, especially mathematics, statistics, machine learning, and engineering, he focused on graduate research and education and described decisions departments and institutions need to make.

Rakhlin pointed to rapid advances in mathematics: a model reached gold-medal level at the International Mathematical Olympiad last year, and models are now producing research results, including a recently proposed solution to one of the Millennium Prize Problems. He said progress in a field depends partly on how quickly and reliably results can be verified. Formal proofs can be checked automatically and programs can be run, enabling systems to generate candidates, learn from outcomes, and improve. He added that AI can contribute to improving models, training procedures, and tools, creating a compounding cycle. These capabilities may let experienced researchers pursue technically demanding questions, while making it more important to distinguish finding a solution from understanding why it works, what generalizes, and what to ask next. Rakhlin said universities should prepare for AI to exceed human ability in many aspects of intellectual work.

He said polished papers are becoming a weaker signal of individual expertise in more fields. Departments should reconsider what they reward, giving greater recognition to asking good questions, replication, synthesis, informative negative results, and shared datasets. Hiring, promotion, and funding decisions should establish what each researcher contributed and accepts responsibility for, including work substantially performed by AI; expectations should be clear to current and incoming PhD students. Graduate training also needs care: routine calculations, coding, failed approaches, and small discoveries help students develop intuition and judgment, so delegating them can remove formative experiences even as AI enables more ambitious projects. Students should learn to formulate problems, audit model outputs, reproduce results, and defend their choices, while fundamentals may become more valuable.

Rakhlin said universities should pursue long-term questions, share results openly, and evaluate claims independently. He called industry partnerships essential but said commercial priorities may not cover all of science or remain aligned with it, making some technological independence necessary. He identified laboratories’ accumulated knowledge as a strategic asset: failed experiments, abandoned directions, and reasons approaches did not work often go unpublished, while scientists and engineers gain tacit knowledge through experience. That missing context may help explain why models struggle in some domains, especially empirical sciences, to anticipate consequences clear to experts. Capturing negative results and researchers’ interpretations could improve models’ scientific exploration, though current limitations may be temporary.

As a proposed direction, Rakhlin described MIT laboratories connected through shared AI research infrastructure. In his example, an AI agent could connect a neuroscience lab seeking neuron segmentation methods for microscopy images with a computer vision group’s relevant advance, propose benchmarks, support iteration, and share unresolved questions and lessons among students, faculty, and labs. He said workflows should capture hypotheses, interventions, outcomes, failures, and interpretations, with shared systems enabling agents to use cross-laboratory tools and information under appropriate permissions. Building this would require substantial public and institutional investment now in compute, secure data systems, and expertise adapting and post-training AI models. Tracing research could preserve the lineage of ideas and make graduate student contributions easier to recognize; agreed rules for consent and credit could support more open collaboration. Rakhlin said AI can already synthesize and reason over more sources than one researcher could absorb, and urged universities to invest so researchers can use it to connect and work together on difficult scientific and engineering problems.

On October 8, 2026, MIT Statistics and Data Science Center Director Alexander (Sasha) Rakhlin argued that universities should prepare for AI to perform many intellectual tasks, rethink how they assess researchers, and invest in shared research systems. In an essay about AI’s effects on academia, especially mathematics, statistics, machine learning, and engineering, he focused on graduate research and education and described decisions departments and institutions need to make.

Rakhlin pointed to rapid advances in mathematics: a model reached gold-medal level at the International Mathematical Olympiad last year, and models are now producing research results, including a recently proposed solution to one of the Millennium Prize Problems. He said progress in a field depends partly on how quickly and reliably results can be verified. Formal proofs can be checked automatically and programs can be run, enabling systems to generate candidates, learn from outcomes, and improve. He added that AI can contribute to improving models, training procedures, and tools, creating a compounding cycle. These capabilities may let experienced researchers pursue technically demanding questions, while making it more important to distinguish finding a solution from understanding why it works, what generalizes, and what to ask next. Rakhlin said universities should prepare for AI to exceed human ability in many aspects of intellectual work.

He said polished papers are becoming a weaker signal of individual expertise in more fields. Departments should reconsider what they reward, giving greater recognition to asking good questions, replication, synthesis, informative negative results, and shared datasets. Hiring, promotion, and funding decisions should establish what each researcher contributed and accepts responsibility for, including work substantially performed by AI; expectations should be clear to current and incoming PhD students. Graduate training also needs care: routine calculations, coding, failed approaches, and small discoveries help students develop intuition and judgment, so delegating them can remove formative experiences even as AI enables more ambitious projects. Students should learn to formulate problems, audit model outputs, reproduce results, and defend their choices, while fundamentals may become more valuable.

Rakhlin said universities should pursue long-term questions, share results openly, and evaluate claims independently. He called industry partnerships essential but said commercial priorities may not cover all of science or remain aligned with it, making some technological independence necessary. He identified laboratories’ accumulated knowledge as a strategic asset: failed experiments, abandoned directions, and reasons approaches did not work often go unpublished, while scientists and engineers gain tacit knowledge through experience. That missing context may help explain why models struggle in some domains, especially empirical sciences, to anticipate consequences clear to experts. Capturing negative results and researchers’ interpretations could improve models’ scientific exploration, though current limitations may be temporary.

As a proposed direction, Rakhlin described MIT laboratories connected through shared AI research infrastructure. In his example, an AI agent could connect a neuroscience lab seeking neuron segmentation methods for microscopy images with a computer vision group’s relevant advance, propose benchmarks, support iteration, and share unresolved questions and lessons among students, faculty, and labs. He said workflows should capture hypotheses, interventions, outcomes, failures, and interpretations, with shared systems enabling agents to use cross-laboratory tools and information under appropriate permissions. Building this would require substantial public and institutional investment now in compute, secure data systems, and expertise adapting and post-training AI models. Tracing research could preserve the lineage of ideas and make graduate student contributions easier to recognize; agreed rules for consent and credit could support more open collaboration. Rakhlin said AI can already synthesize and reason over more sources than one researcher could absorb, and urged universities to invest so researchers can use it to connect and work together on difficult scientific and engineering problems.