The Archive

Kappa-LoRA uses condition numbers to identify which LoRA matrices justify tuning, reducing fine-tuning cost for large models.

Anthropic releases Claude Opus 5 with improvements in agent execution, coding, and professional tasks.

Stratechery weekly digest covering Chinese AI models, Hugging Face developments, and tangential sports business commentary.

Opus 5 will be both cheaper and less restrictive than Fable, likely making it preferable in most use cases

Weeks after Anthropic's latest toe-to-toe with the US government, and days after an OpenAI security incident that dominated tech industry discussions, Anthropic on Thursday released its newest model, Claude Opus 5. The company said in a release that Opus 5 "comes close to the capabilities of Claude Fable 5 in many domains" and is much better at complex coding tasks. (Fable 5 is the public-facing Mythos-class model that drew the government's ire, was taken offline for a few weeks along with Mythos 5, and then brought back with even stronger cyber safeguards than before.) The Fable 5 concerns -...

Meta says its AI chatbot is going beyond just answering questions and generating images. | Image: Meta Meta is upgrading its AI chatbot with new productivity features in a bid to compete with rivals like Gemini, ChatGPT, and Claude. The update will allow Meta AI to tap into your calendar to help you plan events and generate daily briefings, as well as perform in-depth research that you can steer as it progresses. In a blog post, Meta says this update marks its "next step toward personal superintelligence," something CEO Mark Zuckerberg has touted as the future of AI. Meta is powering the upda...

GPU-efficient algorithm for singular value soft-thresholding via polar decomposition, with accuracy trade-offs noted.

Every byte moved has a cost. As model checkpoints grow to hundreds of gigabytes or even a terabyte, that cost adds up quickly. To make things even worse, moving... Every byte moved has a cost. As model checkpoints grow to hundreds of gigabytes or even a terabyte, that cost adds up quickly. To make things even worse, moving these model weights around the cluster is extremely common. For instance, a cold start may pull weights from remote storage into GPU memory; autoscaling and rolling updates must populate each new replica; and RL post-training continuously… Source

Mixed-sign spectral regularization method extending negative-ridge endpoints in overparameterized linear regression via early-stopped gradient descent.

MineValiCoder addresses LLM-based TDD by mining test quality and using bipartite graphs to validate code without human test cases.

ADAPT-GQE uses generative AI and transformers to learn efficient quantum ground-state preparation circuits for electronic structure.

Generalization bounds via Rademacher complexity for k-neighborhood data augmentation strategy in learning optimization solver iterates.

TRACE-Router enables task-level LLM routing for agentic workflows, attributing feedback to routing decisions across long-horizon tasks.

Trio-ethnography examining how educators' interpretations of student AI-supported programming learning evolve through dialogue.

Four audio foundation models tested on phylogenetic signal recovery from marine mammal/bird vocalizations; domain-specific pretraining shows limited gains.

grapheme-kit: open-source Python library for grapheme-level NLP metrics and text processing in multilingual systems, addressing Unicode limitations in writing systems like Tamil and Sinhala.

Dynamic capability scoping framework for enterprise AI agents using three-source permission architecture to enforce least-privilege credential access and reduce attack surface.

Analysis of Hyperball-style optimizers via angular effective learning rate decomposition to explain their performance gains in scale-invariant deep network training.

Convex optimization framework for generating sparse correlation matrices with prescribed graph-structured sparsity patterns via elliptope projection.

Robots learning to communicate through projected visual abstractions like shadows and silhouettes, enabling embodied expression beyond physical morphology.

AI companies including Nvidia and Mistral urge policymakers to avoid broad restrictions on open-weight AI models as Washington debates responses to Chinese AI and alleged model distillation.

Identifiability analysis of action-conditioned Joint-Embedding Predictive Architectures for learning controlled world models from high-dimensional visual observations.

Practice-based explainability approach for diffusion models in creative applications via model bending and interactive intervention for artists.

PRIMS: physics-aware multimodal Transformer for on-device fluid identification in microfluidic systems integrating physical knowledge into representation learning.

Reflector: interactive audio workstation for sample-based composition with arrangement-aware harmonic retrieval tracking pitch-class combinations in real-time.

LunarFM: multimodal foundation model for unified representation of lunar surface from heterogeneous orbital remote-sensing data for resource mapping and exploration.

Self-calibrating framework for LLM agents to detect and correct operational drift in open-ended autonomous edge resource allocation.

Statistical learning theory for estimating prediction functions from single finite trajectories of ergodic Markov processes.

Trunk-agnostic API for EEG encoder personalization compatible with frozen classical and foundation models across datasets.

SceneActBench: benchmark for evaluating VLM agents' ability to act on complete multi-object 3D scenes via unified agent-environment loop.