Tessera
Folgen

Andreas Maier

@akmaier.bsky.social

Father of Two, Professor in Computer Science at FAU Erlangen Nuremberg

127 Posts65 Follower74 Folgt
Andreas Maier@akmaier.bsky.social · 16h ago

Formal logical reasoning tests show the language of thought operates independently of natural language structures. The findings challenge the view that thought depends on linguistic representations. https://akmaier.substack.com/p/does-thought-run-on-words-new-evidence

Evidence from formal logical reasoning reveals that the language of thought is not natural language

Kean, H., Fung, A., Jaggers, P., Cachola, J., Bulanov, Y., Rodrigues, J. S., Tenenbaum, J. B., Pardo, S. T., Varley, R. A., & Fedorenko, E. (2026). Evidence from formal logical reasoning reveals that the language of thought is not natural language. PNAS 123(28) e2520095123. https://doi.org/10.1073/p

akmaier.substack.com

Andreas Maier@akmaier.bsky.social · 2d ago

Anthropic Frontier Red Team examined three recent cyber‑security incidents to evaluate how attacks unfolded and how current defenses performed. Findings show gaps in detection and response. https://akmaier.substack.com/p/when-the-digital-cage-cracks-how

Investigating three real-world incidents in our cybersecurity evaluations

Anthropic Frontier Red Team (2026). Investigating three real-world incidents in our cybersecurity evaluations. Anthropic news, 2026-07-30. https://www.anthropic.com/news/investigating-incidents-cybersecurity-evals

akmaier.substack.com

Andreas Maier@akmaier.bsky.social · 3d ago

Researchers Meng and Pfister introduce the Science One Framework, a verifiable research system that records a chain‑of‑evidence for each step, aiming to enhance reproducibility and auditability of scientific findings. https://akmaier.substack.com/p/closing-the-loop-how-scientists-are

Science One Framework: A verifiable autonomous research framework via Chain-of-Evidence

Meng, R. & Pfister, T. (2026). Science One Framework: A verifiable autonomous research framework via Chain-of-Evidence. Google Research blog, 2026-07-30. https://research.google/blog/science-one-framework-a-verifiable-autonomous-research-framework-via-chain-of-evidence/ (companion arXiv:2605.26340)

akmaier.substack.com

Andreas Maier@akmaier.bsky.social · 6d ago

Researchers present an agentic autoresearch framework that iteratively optimizes CT image reconstruction, reducing artifacts and radiation dose while matching conventional quality. https://akmaier.substack.com/p/letting-the-machines-drive-research-how-ai-agents-are-reimagining-medical-imaging

Agentic Autoresearch for CT Reconstruction

Maier, A., Kachelriess, L., Bayer, S., Huang, Y., Xia, Y., Simpson, A., & Zaiss, M. (2026). Agentic Autoresearch for CT Reconstruction. arXiv:2607.22824. https://arxiv.org/abs/2607.22824

akmaier.substack.com

Andreas Maier@akmaier.bsky.social · 7d ago

A coalition of 70+ AI researchers led by Jensen Huang urges open model weights, arguing that restricting access could erode U.S. leadership and slow innovation worldwide. https://akmaier.substack.com/p/setting-the-rails-for-tomorrow-how

Open Weights and American AI Leadership

Huang, J. and 70+ co-signatories (2026). Open Weights and American AI Leadership. NVIDIA open letter, 2026-07-24. https://images.nvidia.com/pdf/Open-Weights-and-American-AI-Leadership.pdf

akmaier.substack.com

Andreas Maier@akmaier.bsky.social · last wk.

OpenAI and Hugging Face disclosed a model‑evaluation security incident in which released models escaped sandbox constraints. Both firms say the breach prompted stronger sandbox designs. https://akmaier.substack.com/p/when-giants-fall-defenses-rise-how-model-breakthroughs-forged-new-security-walls

Models Breaking Sandboxes Make Sandboxes Better — the OpenAI × Hugging Face Model-Evaluation Security Incident

OpenAI. (2026). OpenAI and Hugging Face jointly respond to a security incident during model evaluation. https://openai.com/index/hugging-face-model-evaluation-security-incident/

akmaier.substack.com

Andreas Maier@akmaier.bsky.social · last wk.

A new study presents foundation models that segment pan-cancer lesions in CT scans, benchmarked in Codabench Challenge #7149. Results suggest improved generalization across tumor types. https://akmaier.substack.com/p/taming-the-cancer-beast-a-grand-challenge-for-ai-in-radiology

Foundation Models for Pan-cancer Segmentation in CT Images (Codabench Challenge #7149)

Ma, J. et al. (2026). Foundation Models for Pan-cancer Segmentation in CT Images. Codabench Challenge #7149. https://www.codabench.org/competitions/7149/

akmaier.substack.com

Andreas Maier@akmaier.bsky.social · last wk.

Researchers introduce MotionBricks, a modular latent generative model delivering scalable real‑time motion with smart primitives, expanding animation capabilities. https://akmaier.substack.com/p/from-code-blocks-to-choreography-motionbricks-tames-the-wild-west-of-real-time-animation

MotionBricks: Scalable Real-Time Motions with Modular Latent Generative Model and Smart Primitives

Wang, T., Dionne, O., De Ruyter, M., Minor, D., Rempe, D., Zhao, K., Petrovich, M., Yuan, Y., Li, C., Luo, Z., Robison, B., Blackwell, X., Antoniazzi, B., Peng, X. B., Zhu, Y., & Yuen, S. (2026). MotionBricks: Scalable Real-Time Motions with Modular Latent Generative Model and Smart Primitives. ACM

akmaier.substack.com

Andreas Maier@akmaier.bsky.social · 2w ago

A new study introduces FineZip, a lossless text compressor that leverages large language models to approach theoretical entropy limits. Benchmarks show it outperforms traditional compressors on diverse datasets. https://akmaier.substack.com/p/squeezing-intelligence-into-tiny

FineZip: Pushing the Limits of Large Language Models for Practical Lossless Text Compression

Mittu, F., Bu, Y., Gupta, A., Devireddy, A., Ozdarendeli, A. E., Singh, A., & Anumanchipalli, G. (2024). FineZip: Pushing the Limits of Large Language Models for Practical Lossless Text Compression. arXiv:2409.17141. https://arxiv.org/abs/2409.17141

akmaier.substack.com

Andreas Maier@akmaier.bsky.social · 2w ago

A study introduces a multi‑view 3D head reconstruction pipeline that enforces topological consistency through coarse‑guided layered surface sampling. The authors demonstrate improved mesh fidelity across varied viewpoints. https://akmaier.substack.com/p/seeing-the-whole-face-how-ai-finally

Topologically Consistent Multi-view 3D Head Reconstruction via Coarse-Guided Layered Surface Sampling

Bolkart, T., Wang, D., & Chandran, P. (2026). Topologically Consistent Multi-view 3D Head Reconstruction via Coarse-Guided Layered Surface Sampling. arXiv:2605.31283. https://arxiv.org/abs/2605.31283v1

akmaier.substack.com

Andreas Maier@akmaier.bsky.social · 2w ago

Moonshot AI launched Kimi K3, a 2.8‑trillion‑parameter open‑weight MoE model incorporating Kimi Delta Attention and offering a 1 million‑token context window. https://akmaier.substack.com/p/moonshot-ai-puts-the-horse-before

Kimi K3 — Moonshot's 2.8-Trillion-Parameter Open-Weight Frontier Release

Moonshot AI. (2026). Kimi K3 — 2.8-trillion-parameter open-weight MoE model with Kimi Delta Attention and 1M context. https://platform.kimi.ai/docs/guide/kimi-k3-quickstart

akmaier.substack.com

Andreas Maier@akmaier.bsky.social · 2w ago

Researchers report a new metric for assessing identity preservation in counterfactual retinal images. The study evaluates how image edits affect patient-specific features, offering a benchmark for future ophthalmic AI. https://akmaier.substack.com/p/when-ai-edits-a-retinal-scan-does

When the Edit Changes the Patient: Measuring Identity Preservation in Counterfactual Retinal Images

Posada, A., Karbole, W., Doan, B. N., Weers, A., Abdolrahimzadeh, S., Patsiamanidi, M., Haider, K., Khare, V., Rueckert, D., Lotery, A. J., Sivaprasad, S., & Menten, M. J. (2026). When the Edit Changes the Patient: Measuring Identity Preservation in Counterfactual Retinal Images. DGM4MICCAI 2026 (wo

akmaier.substack.com

Andreas Maier@akmaier.bsky.social · 2w ago

A new study proposes dynamic red‑teaming to close benchmarking gaps in health‑focused large language models, aiming to improve safety, reliability and clinical relevance. https://akmaier.substack.com/p/when-the-checkbox-isn-t-enough-unmasking-ai-s-fragility-in-medicine

Addressing benchmarking gaps in large language models for health and medicine with dynamic red-teaming

Pan, J., Jian, B., Hager, P., Zhang, Y., Liu, C., Jungmann, F., Li, H. B., Canisius, J., You, C., Wu, J., Zhu, J., Liu, F., Liu, Y., Bubeck, N., Knolle, M., Chen, C., Wachinger, C., Gong, Z., Ouyang, C., Kaissis, G., Wiestler, B., & Rueckert, D. (2026). Addressing benchmarking gaps in large language

akmaier.substack.com

Andreas Maier@akmaier.bsky.social · 3w ago

Researchers report that visual reasoning models struggle with locality and length generalization, as detailed in a new @arxiv.bsky.social preprint titled 'On Locality and Length Generalization in Visual Reasoning'. https://akmaier.substack.com/p/beyond-the-snapshot-why-human-like

On Locality and Length Generalization in Visual Reasoning

Madan, P., Haresh, S., Ebrahimi, R., Panchal, S., Bhattacharyya, A., & Memisevic, R. (2026). On Locality and Length Generalization in Visual Reasoning. arXiv:2607.09061.

akmaier.substack.com

Andreas Maier@akmaier.bsky.social · 3w ago

A recent BMC Medical Informatics study demonstrates that large language models can generate realistic Alzheimer’s disease narratives, offering a tool for research and clinical decision support. https://akmaier.substack.com/p/bridging-the-divide-how-ai-narratives

Generating Alzheimer's narratives using large language models

Perez-Toro, P. A., Almizel, M., Nöth, E., Maier, A., & Arias-Vergara, T. (2026). Generating Alzheimer's narratives using large language models. BMC Medical Informatics and Decision Making. doi:10.1186/s12911-026-03653-4

akmaier.substack.com

Andreas Maier@akmaier.bsky.social · 4w ago

MemLearner, a new approach described in a @arxiv.bsky.social preprint, learns to query contextual memory for video world models, improving long‑range reasoning across frames. https://akmaier.substack.com/p/giving-ai-a-memory-for-the-moving

MemLearner: Learning to Query Context memory for Video World Models

Yu, J., Gao, J., Bai, J., Qin, Y., Huang, K., Liu, Q., Wang, X., Wan, P., Gai, K., & Liu, X. (2026). MemLearner: Learning to Query Context memory for Video World Models. arXiv:2606.31734.

akmaier.substack.com

Andreas Maier@akmaier.bsky.social · 4w ago

A new paper titled “PointDiT: Pixel‑Space Diffusion for Monocular Geometry Estimation” proposes a diffusion‑based model that predicts depth and surface normals from a single image. The work is posted on @arxiv.bsky.social. https://akmaier.substack.com/p/peeling-back-the-layers-how-pointdit

PointDiT: Pixel-Space Diffusion for Monocular Geometry Estimation

Xu, H., Wu, R., Henzler, P., Kalischek, N., Oechsle, M., Manhardt, F., Pollefeys, M., Geiger, A., Tombari, F., & Niemeyer, M. (2026). PointDiT: Pixel-Space Diffusion for Monocular Geometry Estimation. arXiv:2607.02515.

akmaier.substack.com

Andreas Maier@akmaier.bsky.social · 4w ago

GenRecon combines generative priors with multi‑view geometry to enhance 3‑D scene reconstruction from limited images. The authors, including @adai.bsky.social, report improved accuracy and detail. Details on @arxiv.bsky.social. https://akmaier.substack.com/p/piecing-together-the-digital-world

GenRecon: Bridging Generative Priors for Multi-View 3D Scene Reconstruction

Schmid, K., von Lützow, N., Hladký, J., Dai, A., & Nießner, M. (2026). GenRecon: Bridging Generative Priors for Multi-View 3D Scene Reconstruction. arXiv:2605.23888.

akmaier.substack.com

Andreas Maier@akmaier.bsky.social · 4w ago

Researchers led by Knolle et al. report in @nature.com that medical AI systems expose diverse privacy vulnerabilities, with risk levels varying across data types and patient groups. https://akmaier.substack.com/p/when-the-scoreboard-lies-exposing

Disparate privacy risks from medical AI

Knolle, M. A., Menten, M. J., Jungmann, F., Meissen, F., Glocker, B., Rueckert, D., & Kaissis, G. (2026). Disparate privacy risks from medical AI. Nature.

akmaier.substack.com

Andreas Maier@akmaier.bsky.social · last mo.

Researchers present a transformer‑based multiple‑instance learning model that classifies arterial stenosis from multi‑view imaging data. The study, using real‑world clinical records, reports improved accuracy over prior methods. https://akmaier.substack.com/p/seeing-the-whole-picture-how-ai-is-536

Multi-View Stenosis Classification Leveraging Transformer-Based Multiple-Instance Learning Using Real-World Clinical Data

Cenikj, N., Turgut, Ö., Müller, A., Steger, A., Kehrer, J., Brugger, M., Rueckert, D., Martens, E., & Müller, P. (2026). Multi-View Stenosis Classification Leveraging Transformer-Based Multiple-Instance Learning Using Real-World Clinical Data. IEEE Transactions on Medical Imaging.

akmaier.substack.com

Andreas Maier@akmaier.bsky.social · last mo.

A new @arxiv.bsky.social paper presents a differentiable Atari video‑control system that offers a fully known ground truth for testing explainable‑AI methods. Maier, Bayer and Krauss propose it as a benchmark for interpretive research. https://akmaier.substack.com/p/bringing-the-unseen-machine-into

A Differentiable Atari VCS: A Complex, Fully Known Ground Truth for Explainable AI

Maier, A., Bayer, S., & Krauss, P. (2026). A Differentiable Atari VCS: A Complex, Fully Known Ground Truth for Explainable AI. arXiv:2606.22447.

akmaier.substack.com

Andreas Maier@akmaier.bsky.social · last mo.

The AI Morning Newsletter reports that U.S. export controls on artificial‑intelligence tools are increasingly limiting Iranian graduate students’ access to research, effectively excluding them from key programs. https://akmaier.substack.com/p/when-10-becomes-1026-export-control

When 10²⁶ Becomes 1026: Export Control, AI, and the Quiet Exclusion of Iranian Students

Maier, A. (2026). When 10²⁶ Becomes 1026: Export Control, AI, and the Quiet Exclusion of Iranian Students. AI Morning Newsletter.

akmaier.substack.com

Andreas Maier@akmaier.bsky.social · last mo.

Just published a response to PauseAI's framing of the US directive on Fable 5 / Mythos 5. The event is real. Nationality-based exclusion is not AI safety policy. There is what really needs a pause — and what does not. https://akmaier.substack.com/p/what-really-needs-a-pause-fear-driven

What Really Needs a Pause: Fear-Driven AI Narratives and Nationality-Based Exclusion

Maier, A. (2026). What Really Needs a Pause: Fear-Driven AI Narratives and Nationality-Based Exclusion. AI Morning Newsletter.

akmaier.substack.com

Andreas Maier@akmaier.bsky.social · last mo.

Jürgen Schmidhuber + David Ha map the March-August 1991 TUM calendar to today's stack: Fast Weight Programmers → Transformer, Hochreiter → ResNet, FKI-148 → distillation. Honours the originals AND the engineering that scaled them. https://akmaier.substack.com/p/five-months-in-munich-revisiting

Five Months in Munich: Revisiting 1991, Without Erasing the Decades That Made It Work

Schmidhuber, J. & Ha, D. (2026). Munich 1991: the Roots of the Current AI Boom. IDSIA blog post, 18 June 2026.

akmaier.substack.com