
Lessons from Kyunghyun Cho
Kyunghyun Cho co-directs NYU’s Global AI Frontier Lab with Yann LeCun. The lab spans fundamental AI, responsible AI and health care. These lessons examine his research and public discussions. — Cho — Global AI Frontier Lab at NYU.
Part 1: The Attention Mechanism and Sequence Modeling
- On the fixed-length vector bottleneck: The attention paper identifies compression into one fixed-length vector as a potential bottleneck for translating long sentences. — Bahdanau, Cho and Bengio — Learning to Align and Translate.
- On the core idea of attention: Let the decoder select relevant source annotations while generating each target word, rather than depending only on one compressed sentence representation. — Bahdanau, Cho and Bengio — Learning to Align and Translate.
- On early skepticism of deep learning: Cho used a provocative remark from his friend Felix to open talks and challenge audiences skeptical of neural approaches. — Kyunghyun Cho — Bye, Felix.
- On the Gated Recurrent Unit (GRU): The encoder–decoder paper introduces a recurrent unit with reset and update gates to retain useful information over different timescales, using a simpler design than LSTM. — Cho et al. — RNN Encoder–Decoder.
- On the transition from statistical translation: The attention paper contrasts separately tuned translation components with a neural model whose components are trained jointly to generate translations. — Bahdanau, Cho and Bengio — Learning to Align and Translate.
- On sequence-to-sequence learning: An encoder–decoder can map a variable-length input sequence to a variable-length output sequence through a learned representation. — Cho et al. — RNN Encoder–Decoder.
- On long sentences in translation: In the paper’s English–French experiments, attention-based models handled longer sentences better than the tested fixed-vector encoder–decoder baseline. — Bahdanau, Cho and Bengio — Learning to Align and Translate.
- On bridging architectures: Attention first worked alongside recurrent translation models; later, the Transformer dispensed with recurrence. The Transformer paper cites the earlier Bahdanau, Cho and Bengio work, but Cho was not a Transformer coauthor. — Vaswani and Coauthors — Attention Is All You Need.
Part 2: The Future of Deep Learning and World Models
- On the limits of passive data: Cho argues that modern models already extract much of the correlation available in passively supplied data; he sees choosing more informative data as the more promising sample-efficiency problem. — SuperDataScience — Attention, World Models and the Future of AI.
- On the next frontier of learning: Build learning systems that deliberately seek informative observations rather than only consuming whatever data they receive. — SuperDataScience — Attention, World Models and the Future of AI.
- On world models: Cho treats high-fidelity simulation and higher-level latent models as competing research approaches, not a settled choice. — SuperDataScience — Attention, World Models and the Future of AI.
- On chasing scale: Cho suggests that gains from making models larger may partly compensate for limitations of current optimization methods. Better algorithms could change those scaling constraints; he presents this as his view, not an established explanation. — Cohere For AI — A Holistic View of Foundation Models.
- On emergent communication: Agents with different information can learn to communicate toward a shared goal. Cho’s presentation studies a visual sender and a text-reading receiver that exchange messages to resolve an object’s identity. — Cho — A Multi-Modal, Multi-Agent System.
- On foundation models: Strong answer accuracy does not establish consistent reasoning. Cho’s evaluation slides examine language models that can answer differently when equivalent intermediate expressions are substituted. — Cho — Beyond Test Accuracies.
- On latent planning: A useful plan may work with high-level states instead of imagining every intermediate step. Cho uses planning a trip to illustrate this motivation for latent dynamics. — SuperDataScience — Attention, World Models and the Future of AI.
- On retrieval-augmented generation: Adding retrieval does not by itself explain why a language model will use retrieved passages rather than information in its parameters. Cho raises this as a training-design problem and discusses trade-offs in separating training and retrieval corpora. — Cho — A Random Thought on Retrieval-Augmented Generation.
- On uncertainty-aware prediction: When a problem has inherent uncertainty, consider a set of plausible answers rather than reporting only the single best prediction. Cho uses multiple acceptable translations as an example. — Cho — Uncertainty and Calibration.
Part 3: Pragmatic AI in Healthcare and Science
- On clinical operations: Cho reframed health-care AI around improving care under real resource constraints, rather than focusing only on rare diagnostic puzzles. — NEJM AI Grand Rounds — AI’s Next Frontier.
- On hospital deployments: Medical-image models can exploit hospital-specific patterns and then fail to generalize to another hospital. Cho and coauthors use this problem to motivate separating target information from experimental-condition effects. — Makino and Coauthors — Supervised Contrastive Block Disentanglement.
- On drug discovery: Cho proposes sharing information across symptoms, therapies and populations so each experiment informs later drug-discovery decisions. He presents this as an end-to-end research direction, not an established clinical result. — Cho — Drug Discovery May Be in the Cold War Era.
- On medical records: Patients should be able to assemble and use their own medical records across providers. Cho describes building a tool to make his fragmented records easier to access and understand. — NEJM AI Grand Rounds — AI’s Next Frontier.
- On a learning-health-system proposal: Cho sketches a speculative learning-health-system proposal: record clinicians’ reasonable alternatives and randomized choices to continually update treatment-effect estimates. He acknowledges remaining bias and the continued need for trials of new treatments; this is a proposal, not established clinical practice. — Cho — A Society-Level Forever Clinical Trial.
- On mechanism and practical progress: Practical progress need not wait for a complete mechanistic explanation. Cho argues that useful drug-discovery approaches can combine partial biological knowledge, data and experiments. — Cho — Drug Discovery May Be in the Cold War Era.
- On clinical model adaptation: General-purpose models may need clinically relevant training and task-specific adaptation for hospital operations. Cho describes testing these choices rather than assuming an off-the-shelf model is sufficient. — NEJM AI Grand Rounds — AI’s Next Frontier.
- On operational prediction: Predicting readmission can help hospitals identify where preventive interventions may be useful. Cho came to see this seemingly ordinary task as important because readmissions consume limited beds and staff time. — NEJM AI Grand Rounds — AI’s Next Frontier.
- On healthcare data: Clinically relevant training matters when labeled examples are limited. Cho reports that his team’s clinical language-model work performed particularly well in this setting. — NEJM AI Grand Rounds — AI’s Next Frontier.
- On interdisciplinary collaboration: Sustained conversations with clinicians changed Cho’s understanding of which health-care problems mattered. The NYUtron collaboration combined clinical priorities with language-model research. — NEJM AI Grand Rounds — AI’s Next Frontier.
Part 4: Doing Good Science and Research Philosophy
- On chasing state-of-the-art: Research can be valuable before it reaches state-of-the-art performance. Cho’s keynote traces promising translation work whose early results fell short of leading systems. — Cho — A SOTA-Less, Novelty-Less Journey.
- On novelty: An idea can become useful when conditions and execution change, even if its basic form is old. Cho connects neural translation to earlier work rather than presenting it as invention without precedent. — Cho — A SOTA-Less, Novelty-Less Journey.
- On the hero scientist narrative: Cho rejects treating prominent researchers as unquestionable authorities on AI risk. — VentureBeat — Interview with Cho.
- On failure in research: Cho recalls that most research ideas in his experience do not work, while the early neural-translation period was unusually productive. Treat failure as part of research rather than expecting every idea to succeed. — NEJM AI Grand Rounds — AI’s Next Frontier.
- On peer review: Reviewer scores need not capture all signals of a paper’s future influence. Cho and coauthors found that authors’ rankings of their own submissions added predictive information about later citations. — Su and Coauthors — Self-Rankings Beyond Peer Review.
- On academic freedom: Cho values academia for the freedom to learn and move into fields beyond the career paths he already knows. — NEJM AI Grand Rounds — AI’s Next Frontier.
- On community consensus: Research incentives can pull academics toward fashionable topics. Cho argues that universities should preserve curiosity-driven scientific work rather than simply copying the product priorities of industry. — Technically Optimistic — AI and Accountability.
Part 5: Causality and Deep Learning
- On learned causal estimation: Cho explores training neural networks on simulated causal problems to learn estimators, rather than hand-designing a new estimator for every setup. Identifiability remains a constraint. — Cho — Causality and Deep Learning.
- On correlational models: A predictor can fail when a spurious relationship changes between training and deployment. Specify which distribution changes rather than treating every shift as the same problem. — Cho — Beyond Test Accuracies.
- On unobserved confounders: Unobserved confounding complicates causal analysis. Cho’s slides show that stronger hidden confounding makes one mixed observational–interventional inference setup more challenging. — Cho — Causality and Deep Learning.
- On deep learning as a causal tool: Cho explores learning causal-discovery algorithms from simulated examples, including identifying causes of a target outcome rather than reconstructing an entire graph. He emphasizes that statistical guarantees and real-world validation remain open research challenges. — Cho — Causality and Deep Learning.
- On estimating counterfactual effects: Counterfactual effects can be harder to estimate than average treatment effects. Cho’s causal-inference slides show an example where more observations do not necessarily improve counterfactual estimation. — Cho — Causality and Deep Learning.
- On experimental design: Choose experiments using feedback from earlier results. Cho illustrates this with selecting successive batches of proteins to test, rather than attempting every possible candidate. — SuperDataScience — Attention, World Models and the Future of AI.
- On causal representations: Cho and coauthors study separating information about a target from variation caused by experimental conditions. Their block-disentanglement method learns distinct embeddings rather than assuming every latent dimension is an independent real-world cause. — Makino and Coauthors — Supervised Contrastive Block Disentanglement.
- On the intersection of fields: Cho’s interest in causality is tied to control: deciding which learned associations to use and which to suppress when acting through a system. He describes experiments and learning algorithms as possible tools for that research. — Cohere For AI — A Holistic View of Foundation Models.
- On limits of observational data: More data do not automatically make a causal quantity identifiable. Cho separates whether an effect can be recovered from the available observations from how accurately an estimator can compute it. — Cho — Causality and Deep Learning.
Part 6: Rethinking Evaluation and Test Accuracies
- On test-set accuracy: Test accuracy alone leaves important questions unanswered: how modeling assumptions affect results, how optimization reaches a model, and how the model arrives at an answer. — Cho — Beyond Test Accuracies.
- On limits of evaluation protocols: An evaluation protocol can stop being informative as models change. Cho notes that a test requiring models to distinguish their own answers becomes unhelpful when different models give essentially the same answers. — Cho — Beyond Test Accuracies.
- On evaluating open-ended generation: A single reference can be inadequate when many outputs are acceptable. Cho and coauthors’ QAGS paper also shows why high word overlap can hide a factual error. — Wang, Cho and Lewis — Evaluating Factual Consistency.
- On learned evaluation: A learned evaluator needs evaluation of its own. The QAGS study tests agreement with human judgments and varies the underlying question-generation and answering models. — Wang, Cho and Lewis — Evaluating Factual Consistency.
- On qualitative analysis: Inspect generated examples as well as scores. Cho’s keynote uses translation outputs and repetitive generated text to expose behavior that aggregate performance alone does not explain. — Cho — A SOTA-Less, Novelty-Less Journey.
- On memorization vs generalization: A useful language model can give probability to configurations absent from its training examples. Cho’s compression discussion frames this departure from the observed corpus as a basis for generalization. — Cho — Are JPEG and Language Models Similar?.
- On the cost of evaluation: Human checks of summary factual consistency can be slow, costly and difficult. This motivates automatic assistance, but human judgments remain the reference for testing QAGS. — Wang, Cho and Lewis — Evaluating Factual Consistency.
- On calibration: Evaluate the uncertainty attached to predictions, not only the most likely answer. Cho discusses calibration as checking whether stated confidence matches observed correctness. — Cho — Uncertainty and Calibration.
Part 7: AI Hype, Existential Risk, and Real World Issues
- On existential risk narratives: Cho worries that speculative existential-risk debate can crowd out the concerns of people already subject to AI systems. He argues that their experiences should be part of safety discussions. — NEJM AI Grand Rounds — AI’s Next Frontier.
- On immediate harms: AI can amplify existing social disadvantages. Cho argues that providers must examine data collection and communicate how deployed systems work and affect people. — Cho — Social Impacts and Bias of AI.
- On the risk agenda: Include the people affected by AI systems in discussions about their safety, especially those who cannot realistically opt out. Cho compares their missing representation with patient advocates in medicine. — NEJM AI Grand Rounds — AI’s Next Frontier.
- On AI hallucinations: Cho sees potential in generative models proposing hypotheses that people would not consider themselves. Such suggestions still need experimental verification; fluency alone does not make a generated claim true. — Technically Optimistic — Language, Culture and AI.
- On marginalized languages: Do not assume a large web crawl will adequately represent every language. Cho warns that languages with little online text or no writing system can be left behind, undermining machine translation’s goal of equitable information access. — Technically Optimistic — Language, Culture and AI.
- On open research: Cho values the period when useful AI technologies could be developed openly in university research, allowing others to benefit. He contrasts that openness with product competition and trade secrets. — NEJM AI Grand Rounds — AI’s Next Frontier.
- On regulating AI: Cho proposes accountability for deployed AI products through safety reporting and involvement of affected users. He distinguishes a reporting framework from prescribing every company’s development process. — NEJM AI Grand Rounds — AI’s Next Frontier.
- On AGI timelines: When asked whether AGI would be achieved, Cho declines to offer an authoritative forecast and instead cites Yann LeCun’s skepticism about the concept. — Cho — Questions from Deep Learning.
- On beneficial applications: He favors tangible medical and scientific benefits over sensational AI narratives. — VentureBeat — Interview with Cho.
Part 8: Coding Agents, Education, and the Next Generation
- On classroom adoption: Cho recalls that roughly four-fifths of his approximately 200 students had no coding-agent environment installed. He therefore began with setup and access rather than assuming prior adoption. — SuperDataScience — Attention, World Models and the Future of AI.
- On teaching fundamentals: Teach mathematical and statistical foundations alongside coding assistants. Cho’s course paired application-building with labs on the internal workings of learning algorithms. — Cho — Teaching Fundamentals of Machine Learning.
- On adapting curricula: Cho redesigned his machine-learning class around algorithm explanations, demonstrations of working with coding agents, and students building applications themselves. He described this as an experiment still underway. — SuperDataScience — Attention, World Models and the Future of AI.
- On AI assistance in research: Cho sees potential for research systems that retrieve relevant literature and connect information across fields, extending what one researcher can read and know. — SuperDataScience — Attention, World Models and the Future of AI.
- On testing student projects: Evaluate what students can make work, not only fragments of code. Cho asked teams to build an application from scratch and tested it live after their presentation. — Cho — Teaching Fundamentals of Machine Learning.
- On mentorship: Mentorship involves balancing ambitious long-term research with students’ near-term career options. Cho describes discussing that compromise with students approaching graduation rather than pretending every project can ignore employment realities. — Cohere For AI — A Holistic View of Foundation Models.
- On computer science as problem-solving: Cho frames much of computer science as solving problems, with its foundations evolving alongside those problems. He contrasts this with discovering an independently existing natural universe. — Cohere For AI — A Holistic View of Foundation Models.