Andreas Mueller

@GaelVaroquaux @OfirPress @fpedregosa @agramfort @scikit_learn Also, y'all still on Musk's echo chamber 🙃

@GaelVaroquaux @OfirPress @fpedregosa @agramfort @scikit_learn The actual implementation of a solution is not usually the bottleneck. We often write out 2 or 3 versions to discuss them before settling on what to do. And the agent won't have the context to decide which solutions are appropriate.

The SWE-bench paper (arxiv.org/pdf/2310.06770, one of the main LLM evals on code) uses an actual PR from the scikit-learn repo as illustration♥️

📢 We are excited to announce "#FMSD: 1st Workshop on Foundation Models for Structured Data" has been accepted to #ICML 2025! Call for Papers: icml-structured-fm-workshop.github.io/call-for-paper…

Open protocols like A2A and MCP are key to enabling the agentic web. With A2A support coming to Copilot Studio and Foundry, customers can build agentic systems that interoperate by design.

@howdataworks Integrating anomaly detection, root cause analysis and causal modeling seems like a promising approach, as in "Root Cause Analysis of Anomalies in Multivariate Time Series through Granger Causal Discovery"

@amuellerml @amuellerml, it's fascinating to see such insights into time series anomaly detection! Bridging research with practical applications is crucial for real-world impact. What emerging trends in this space excite you the most? 📈 #Innovation

New preprint arxiv.org/abs/2502.05392 Open Challenges in Time Series Anomaly Detection: An Industry Perspective This is a vision paper about what I think it missing from current research in time series anomaly detection, and how it could align better with practical applications.

@axnsantana @FrankRHutter I think they attack the problem from two different ends, and the ideal solution is somewhere in the middle. Right now, the computational cost of the two is really not on the same scale, but Carte can make use of sources of information that are unavailable to current tabpfn.

@axnsantana @FrankRHutter These are both super interesting but orthogonal. Carte explicitly addresses world knowledge in string categories and column names, while TabPFN doesn't use string content. Carte requires fine-tuning, which can be quite expensive, while TabPFN does ICL.

The data science revolution is getting closer. TabPFN v2 is published in Nature: nature.com/articles/s4158… On tabular classification with up to 10k data points & 500 features, in 2.8s TabPFN on average outperforms all other methods, even when tuning them for up to 4 hours🧵1/19

is that "agentic" enough🤣

Yes!

Today OpenAI announced o3, its next-gen reasoning model. We've worked with OpenAI to test it on ARC-AGI, and we believe it represents a significant breakthrough in getting AI to adapt to novel tasks. It scores 75.7% on the semi-private eval in low-compute mode (for $20 per task in compute ) and 87.5% in high-compute mode (thousands of $ per task). It's very expensive, but it's not just brute -- these capabilities are new territory and they demand serious scientific attention.

Congrats to my friends at @Microsoft on getting Python in Excel to GA! techcommunity.microsoft.com/t5/excel-blog/…

@cvondrick Btw I haven't looked into this closely but this method of calibration has been used in practice: Consensual Affine Transformations for Partial Valuation Aggregation. AAAI 2019: 2612-2619

@cvondrick Also maybe requiring reviewer training might be helpful, the slides have some evidence for that, and also demonstrate several of the reviewer biases that reviewers probably should be made aware of.

I'm pretty frustrated with the current review process in ML (both from an author, reviewer and meta-reviewer perspective). There's possible solutions or at least experiments and changes, but I feel like business as usual is no longer feasible.