Data For Science

Our latest Substack post is now live Graph Neural Networks 101 Come learn the the basics of how GNNs work under the hood, with a fascinating new dataset open.substack.com/pub/data4sci/p…

[2605.22763v1] Advancing Mathematics Research with AI-Driven Formal Proof Search — A clear overview of how AI-guided formal proof search can support real math research, not just toy benchmarks. Useful if you’re tracking where automated… arxiv.org/abs/2605.22763…

The Science of Unpredictability | Los Alamos National Laboratory — A clear, accessible look at how the FPUT problem helped spark chaos theory and modern nonlinear dynamics. Useful context for anyone working… lanl.gov/media/publicat…

After Automation | Every — A useful counterpoint to the “AI will eliminate work” narrative: as automation improves, it often shifts effort into new human tasks rather than removing it. Worth reading if you’re thinking about how roles and… every.to/p/after-automa…

[2605.06394] Lecture Notes on Statistical Physics and Neural Networks — Clear lecture-style notes connecting statistical physics tools to neural network theory—useful as a structured refresher if you want the big ideas and standard derivations… arxiv.org/abs/2605.06394

[2605.12460] Multi-Stream LLMs: Unblocking Language Models with Parallel Streams of Thoughts, Inputs and Outputs — Interesting take on making LLMs less sequential by running parallel streams for reasoning, inputs, and outputs. Worth a skim if… arxiv.org/abs/2605.12460

Sycophantic AI decreases prosocial intentions and promotes dependence — Worth a read; the useful bits are usually in the examples and edge cases. science.org/doi/10.1126/sc…

Global approaches to infectious disease surveillance and modeling | Nature Medicine — Clear overview of why infectious disease surveillance is getting harder—and how federated analytics could unlock cross-border modeling… nature.com/articles/s4159…

Persuading large language models to comply with objectionable requests — Worth a read; the useful bits are usually in the examples and edge cases. pnas.org/doi/10.1073/pn…

Personal update: I've joined Anthropic. I think the next few years at the frontier of LLMs will be especially formative. I am very excited to join the team here and get back to R&D. I remain deeply passionate about education and plan to resume my work on it in time.

[2604.09839] Steered LLM Activations are Non-Surjective — Useful if you’re thinking about activation steering as a general-purpose control knob: this paper argues the mapping from steering directions to reachable behaviors isn’t onto, so some… arxiv.org/abs/2604.09839

Project Glasswing: what Mythos showed us — A clear-eyed look at what happened when security-focused LLMs were aimed at real production code—what they caught, what they missed, and what needs to change before this approach can… blog.cloudflare.com/cyber-frontier…

Project MUSE -- Verification required! — Heads-up: this Project MUSE link is behind a verification check, so you may need to complete a quick challenge to access it. If you’re doing text/data mining, it points you to Customer Service for help. muse.jhu.edu/article/936213

Three new ways to build with real-world imagery and AI – Google Maps Platform — Google Maps Platform is rolling out three AI-driven updates to its imagery products, including… mapsplatform.google.com/resources/blog…

Is AI putting graduates out of work already? — Worth a read; the useful bits are usually in the examples and edge cases. economist.com/finance-and-ec…

Crypto Prediction Markets Explained — Clear overview of how crypto prediction mark… chainalysis.com/blog/crypto-pr…

[2605.05242] Beyond Semantic Similarity: Rethinking Retrieval for Agentic Search via Direct Corpus Interaction — Useful reframing of retrieval for agentic search: instead of only ranking by semantic similarity, it argues for more direct,… arxiv.org/abs/2605.05242

[2605.01048] Compared to What? Baselines and Metrics for Counterfactual Prompting — Useful framing for evaluating counterfactual prompting: it pushes you to define the right baselines and metrics before drawing conclusions. Worth a skim if… arxiv.org/abs/2605.01048

Natural Language Autoencoders Produce Unsupervised Explanations of LLM Activations — Interesting approach to interpretability: train a verbalizer and reconstructor so a short text description has to preserve enough… transformer-circuits.pub/2026/nla/index…

[2605.06647] Superintelligent Retrieval Agent: The Next Frontier of Information Retrieval — A concise arXiv overview of “superintelligent” retrieval agents and what they could change about search beyond ranking documents. Useful if you’re… arxiv.org/abs/2605.06647