Machine Learning

Yuandong Tian discusses his research on computer Go, from DarkForest to OpenGo, and the role of action-space design in applying reinforcement learning. The conversation covers gradient-free optimization, neural architecture search, Coconut’s approach to reasoning in latent space, representation learning and grokking. It then examines AI-assisted research, recursive self-improvement, coding-agent limitations, and whether alternative architectures can outperform transformers. The closing discussion considers data efficiency, robotics, restrictions on self-improving systems, and open-source models.

Ming-Yu Liu, NVIDIA vice president of research and head of the Cosmos Lab, joins Practical AI hosts Daniel Whitenack and Chris Benson to discuss AI systems that interact with the physical world. The conversation examines open models, world models and simulation as foundations for robotics and autonomous vehicles, and how research and deployment can advance physical AI. Published 1 October 2026; publisher runtime 47 minutes 19 seconds.

Sam Charrington speaks with Greg Burnham, who leads AI capabilities research at Epoch AI, about the progression from elementary mathematical tasks to difficult research problems, including Navier–Stokes. They examine how advanced systems solve mathematical problems, the roles of persistence and existing human work, and the evidence for new ideas. The conversation also considers how to measure progress when traditional benchmarks become less informative, why improvements appear steady across successive model generations, and the remaining weaknesses in open-ended research, learning from experience and choosing productive scientific questions.

Stanford linguist and computer scientist Chris Manning discusses what linguistics contributed to machine learning and why pragmatics and dialogue remain open problems for language models. The conversation examines early abstraction of verb categories in small transformers, how distributed representations shape learning, and whether language alone can support meaningful representations. It also considers language in world models, diffusion language models, and representation finetuning (ReFT), which steers frozen models through their hidden states. The final discussion asks where knowledge resides and whether concepts occupy linear subspaces.

NSF support for interdisciplinary collaborations developing mathematical and theoretical foundations for explainable, reliable, sustainable and trustworthy artificial intelligence.

Two-year, full-time postdoctoral appointment investigating commonsense reasoning in multimodal world models. Research combines neural and symbolic methods to represent abstractions, causality, space and time, with the aim of improving human-centred AI. The project involves collaboration across AI, computer vision and reasoning, alongside limited teaching and student supervision. The preferred start is in the first quarter of 2027.

EMBL-EBI seeks a scientific developer for the Open Targets Safety 2.0 project. The role builds machine-learning and data-processing pipelines that extract drug side-effect evidence from publications, clinical trials and drug labels, map it to biomedical ontologies, and integrate results with ChEMBL and the Open Targets Platform. The researcher will help curate training data, collaborate with pharmaceutical partners and communicate methods and results. This is a full-time hybrid role based in Hinxton, UK. The advert gives a closing date of 11 October 2026 at 23:59 CET.

The BrAIN group recruits a postdoctoral researcher for the ERC-funded BrainAlign project, investigating whether human brain activity can improve language models’ long-timescale understanding. The researcher leads the computational neuroscience work, integrates heterogeneous fMRI datasets and develops models that learn from naturalistic language comprehension. The project also uses these models to investigate language processing in the brain. Candidates need a relevant PhD, strong computational or machine-learning skills and programming experience; fMRI expertise is particularly useful. Apply through the Max Planck Postdoc Program with a CV and publications, research summary, statement of interest and up to three referees. The project advert describes rolling review; the programme’s September 2026 application window closes on 13 October at 12:00 CEST.

Benoit Assi examines limits on reliable theoretical predictions for present and future particle colliders. Low-order simulations of QCD radiation carry large uncertainties, hadronization is commonly modeled rather than derived, and effective theories contain more operators than measurements can constrain. Information theory and machine learning can incorporate improved calculations into simulations, quantify uncertainty, choose informative observables and advance hadronization theory. Geometry of effective-theory field space combines infinite operator families into finite physical quantities, while information measures identify the combinations experiments can resolve. The talk develops these complementary approaches to extracting the available information from LHC and future-collider data.

Nearest-neighbor indexes usually rely on a single distance function, but accurate comparisons can be expensive. This talk proposes a bi-metric framework: a cheap proxy metric builds the index, while the query procedure uses a limited number of evaluations of both the proxy and an expensive ground-truth metric. The theory applies to DiskANN and Cover Tree. When the proxy approximates the ground-truth metric within a bounded factor, the resulting structure can achieve arbitrarily good approximation guarantees under the accurate metric. Experiments on text retrieval using models with very different computational costs show improved accuracy-efficiency tradeoffs on almost all MTEB datasets compared with alternatives such as reranking. Joint work with Haike Xu and Sandeep Silwal.

Robert Gower introduces Polar Express for the polar decomposition and matrix sign function, motivated by Muon neural-network training. Using only matrix-matrix products makes the method suited to high-throughput GPUs. Each iteration adapts its polynomial update through minimax optimization, building on Chen and Chow and Nakatsukasa and Freund. Worst-case error minimization gives rapid initial and asymptotic convergence. The talk addresses finite-precision implementation in bfloat16 and reports improved validation loss when training GPT-2 on one billion FineWeb tokens across several learning rates.

Chris Musco studies how to approximate an unknown matrix by a structured one using a limited, adaptively chosen sequence of matrix-vector products. This models operator learning in scientific machine learning as well as computational algorithms. Randomized SVD provides strong guarantees for low-rank targets; analogous results for sparse and hierarchical structures are less developed. The talk presents progress on efficient algorithms for these classes and a broader complexity theory. Joint work with Noah Amsel, Pratyush Avi, Tyler Chen, Prathamesh Dharangutte, Chinmay Hegde, Feyza Duman Keles, Diana Halikias, Cameron Musco, and David Persson.