The Archive

Link: nytimes.com

TabPFN-3 was released today, the next iteration of the tabular foundation model, originally published in Nature. Quick recap for anyone new to TabPFN: TabPFN predicts on tabular data in a single forward pass - no training, no hyperparameter search, no tuning. Built on TabPFN-2.5 (Nov 2025) and TabPFNv2 (Nature, Jan 2025), which together crossed 3M downloads and 200+ published applications. What's new: * Scale: 1M rows on a single H100 (10x larger than 2.5).A reduced KV cache (\~8GB per million rows per estimator) and row-chunked inference make this practical on a single GPU * Speed: 10x-10...

Reddit user describes friction in Claude's UI for retrieving specific answers from long conversations; suggests workaround of manual copying.

Reddit user complaint about Claude API plan limits and rate-cap implementation, claims marketing misrepresented compute capacity gains.

Workers are using an internal AI tool to automate non-essential tasks.

Benchmark comparing Gemma 4 multi-token prediction vs. DFlash speculative decoding on H100 using vLLM and SPEED-Bench dataset.

A new startup called Dessn has raised $6M to build AI-powered design tools that work directly with production codebases.

llama.cpp adds llama-eval benchmarking tool supporting AIME, GSM8K, GPQA for local quantized model evaluation.

Been using Claude Code heavily and kept running into the same thing everyone here talks about: the model ignores your rules. You tell it to write tests first, it writes the implementation. You give it coding standards, it cherry-picks which ones to follow. And as your rulebook grows, you're burning more and more tokens stuffing everything into context when only a handful of rules are relevant to what you're working on. So I built Writ. Two pieces: A retrieval engine that picks only the relevant rules and skills for the current task. It runs a five stage pipeline over a Neo4j knowledge graph...

Transfer learning evaluation across 11 pre-trained models for image classification; standard practice study without novel methodology.

Random-Set GNNs introduce uncertainty quantification for graph neural networks via epistemic and aleatoric uncertainty modeling.

Evaluates LLM ability to generate Entity-Relationship diagrams from natural language; identifies systematic limitations in database modeling tasks.

QDSB combines quantization with Schrödinger bridges for generative modeling from unpaired samples; highly specialized technical approach.

Applies Bayesian optimization and DRL to aerodynamic control of high-lift wings; domain-specific application outside AI research scope.

RuDE framework predicts post-training potential of LLMs before fine-tuning; addresses model selection gap using rubric-based evaluation.

RTX 4090 power optimization for llama.cpp: reduce consumption 40% via power limits without performance loss.

Reach-Avoid Probability Certificates (RAPCs) enforce probabilistic safety constraints in stochastic RL while minimizing cost.

Dual Group Advantage Optimization mitigates order bias in LLMs to improve RAG and in-context learning fairness.

Vapi says its enterprise business has grown 10-fold since early 2025 as companies shift customer support and sales calls to AI agents.

Humanoid robot as V2X complement for intersection collision avoidance; autonomous systems application, not frontier AI.

NOFE enables continuous dimensionality reduction via function-to-function mappings using Graph Kernel Operators.

GAN-based stochastic cloud subcolumn generator for Earth System Models improves representation of subgrid cloud variability.

Reddit discussion alleging Anthropic incorporates open-source features (MCPs, memory, goals) without attribution to original developers.

Target-guided dialogue system uses scenario modeling and intent-keyword bridging to steer conversations toward predefined topics.

Google reports hackers using AI to automate computer intrusions, escalating security concerns for enterprise infrastructure.

ClipSum leverages frozen CLIP features with temporal modeling for multimodal abstractive summarization of instructional videos.

Study reveals LLM miscalibration in social science measurement tasks; confidence filtering can bias downstream empirical estimates.

Counterfactual Trace Auditing framework measures how agent skills change LLM behavior via structured Skill Influence Pattern annotations.

Random soft prompts without training reach comparable reasoning accuracy to optimized prompts, suggesting injection itself aids LLM reasoning.