Christoph Esslinger, Ph.D.

🚨 New “foundation model” for reasoning on text and time series by Stanford Biodesign Digital Health, Google and ETH Zürich. Spoiler alert: OpenTSLM lacks evidence to support its claims, does not show evidence of working beyond a narrow problem set, and is not more accurate than classical models. 👉 OpenTSLM is marketed as a breakthrough multimodal foundation model for temporal reasoning with order-of-magnitude gains in medical reasoning and beyond. ‼️But... no evidence to support those claims is to be found. The paper overstates results, omits key literature, and cherry-picks baselines, falling short of minimal standards for methodological rigor and transparency. Here are 🔟 points that I find problematic and share in the spirit of academic exchange. ⚠️ Core concerns: 1️⃣ Misleading framing: Tasks are mostly templated classification, not real temporal reasoning. No forecasting, imputation, or anomaly detection. 2️⃣ False claims: “order-of-magnitude gains in temporal reasoning” is simply false and misleading. 3️⃣ Missing literature: Complete omission of related work. E.g. Chronos by Amazon , TimesFM by Google, Moirai by Salesforce, Toto by Datadog, TabPFN by Prior Labs, TiRex by NXAI, Moment by Carnegie Mellon University... 4️⃣ Weak baselines (cherry-picking): self constructed datasets and omission of standard baselines. (E.g., Sleep-EDF results are far below models from the last decade) 5️⃣ Template dependence — Models were trained to solve one very particular task and don't show any signs of generalization. 6️⃣ Compute inefficiency — Enormous complexity and compute for subpar results. 7️⃣ No forecasting, classification or anomaly detection: Using ChatGPT-style captioning is not forecasting. The authors do include code to handle the Monash Dataset and The UCR Archive but completely omit any reference in the paper. 8️⃣ Missing results, logs, checkpoints, and relevant ablations — No breakdown of curriculum, rationale, or encoder impact. 9️⃣ Code irregularities — >122 files deleted before publishing; including evaluation notebooks and dataset generation script. Also the code history shows that some standard benchmarks (e.g. PAMAP2) were introduced, presumably evaluated and then replaced. 🔟 Suboptimal quality — The paper is signed, endorsed, and shared by researchers from from top institutions. Such an amazing set of authors can and has to produce better work. 🔍 Transparency, quality, and rigor matter — especially in fields like healthcare. 🎖️ If we want trustworthy foundation models we need to raise the bar, not lower. Patrick Langer, Thomas Kaar, Max Rosenblattl, Paul Schmiedmayer, Filipe Barata Welcome to the time series field, and #happyforecasting. 💬 Thoughts, evidence, or corrections welcome. If I’ve missed something, please point it out — I’ll happily adjust or retract. #AI #MachineLearning #TimeSeries #FoundationModels #Transparency #MLResearch #AcademicIntegrity

58

13 Kommentare

"EU AI Act - Germany: AI Ecosystem Brief", interesting report co-authored by Sean Musch, Michael Charles Borrelli, Charles Kerrigan, Dr. Martha Boeckenfeld, Martin Ebers, Valentino Halim, Mike Flache, Martin Heitmann, Dr. Nadia Morozova, Michael Alf, Sheila Beladinejad 🇨🇦, Dilem Kaya, & Prof. Dr. Patrick Glauner, and published now in September 2025 by AI & Partners. 👇 This report examines Germany’s evolving #AIgovernance #strategy within the European Union framework. The document outlines national instruments, sectoral initiatives, and international commitments guiding #responsibleAI adoption. It highlights institutional coordination, #ethical safeguards, and #strategic #investments. The text underscores Germany’s ambition to align #innovation, accountability, and global competitiveness in #AI. Link: https://lnkd.in/d2rYuQFp #artificialera #web3 #web4 #web5 #fintech #digitalera #digitaltransformation #ailaw #machinelearning #deeplearning #generativeAI #trustworthyAI #explainableAI #aiethics #artificialintelligence #explainableAI #responsibleAI #foundationmodels #GPAI #GAI #LLMs #NLPs #GenAI #frontierAI #narrowAI #AGI #ethicalAI #adversarialAI #RAG #agenticAI #AIagents

39