All entries

We need a name for the kind of people that conflate 2022 pure LLMs that were in fact a tiny bit like stochastic parrots (not my phrase, and …

We need a name for the kind of people that conflate 2022 pure LLMs that were in fact a tiny bit like stochastic parrots (not my phrase, and not one I ever endorsed) with 2026 systems that couple those

safetygary-marcus--x

Human

Content type

All Blog X Post Paper YouTube Reddit GitHub

Clear filters

Agents

web-edit 0.1a0

DGX agent

Release: web-edit 0.1a0 I was working on a project from my phone using Codex Remote when I found myself wanting to edit a credentials file without round-tripping those credentials through the LLM itse

agentssimon-willison

12 Oct 2026

Web$hell: A Unified Interface for Agentic Web Access

arXiv:2610.13416v1 Announce Type: new Abstract: Large language model (LLM) agents are becoming direct users of the open web. As they take on increasingly complex and information-intensive tasks, effec

Weight Clustering Reveals the Importance of Relative Rank in Large Language Models

model-releasesarxiv-cs-cl

arXiv:2603.17917v2 Announce Type: replace-cross Abstract: Large language models (LLMs) contain billions of parameters, but many exact weight values are not essential. We apply scalar K-means clusterin

What are you doing for a reliable, well designed local model based AI search?

model-releasesr-localllama

Ive been using OpenWebUI's inbuilt web search and I can't put my finger on it, but it's not quite as polished, reliable, seamless as using Claude or Gemini. And its often slow beyond a limit that make

What Do Audio-Language Judges of Generated Music Listen For? Four Probes with Controls from Expert Ratings

arXiv:2610.13247v1 Announce Type: cross Abstract: Open audio-language models are used as judges of generated music: given a clip and its text prompt, the model rates the clip, and the judge is validat

safetyarxiv-cs-cl

What Should Embodied Agents Remember? Benchmarking and Evaluating Memory Modality for Embodied Agents

model-releasesarxiv-cs-ro

arXiv:2610.12561v1 Announce Type: new Abstract: Memory allows large language model (LLM) agents to use past experience, often stored as textual records, when responding to requests. For embodied agent

What Should SFT Learn? Conditional Information Gain for Generalizable Post-Training

arXiv:2610.12995v1 Announce Type: cross Abstract: Supervised fine-tuning (SFT) remains a widely used approach for post-training large language models, because fixed demonstrations can directly introdu

safetyarxiv-cs-ai

Industry

What to expect from theCUBE’s Smarsh EDGE analysis: Tune in Oct. 15

Financial institutions are adopting artificial intelligence faster than many compliance teams can establish the controls needed to oversee it. That gap is raising new questions about AI-powered compli

industrysiliconangle

What Visual Representations Do World Action Models Need?

arXiv:2610.13024v1 Announce Type: cross Abstract: World Action Models (WAMs) inherit powerful priors from pretrained video generation models, but their variational autoencoder (VAE) representations ar

researcharxiv-cs-ro

When Charts Deceive Web Agents: A Paired Benchmark for Misleading Visualization-Grounded Tasks

model-releasesarxiv-cs-cl

arXiv:2610.12823v1 Announce Type: new Abstract: Web agents increasingly browse, interpret, and act on web content on behalf of users. In many realistic workflows, however, the visual evidence they con

Tutorials

When Edge Independence Fails: Joint Graph Diffusion with Latent Sociability Priors

arXiv:2610.13765v1 Announce Type: cross Abstract: Discrete graph diffusion models corrupt graph structure through independent edge updates with a Markovian noise model. Combined with the exchangeabili

tutorialsarxiv-cs-ai

When Faster Decisions Do Not Make Better Agents: An Empirical Study of Jev in Agent Workflows

arXiv:2610.13092v1 Announce Type: new Abstract: AI agents depend on repeated decisions to accomplish tasks, with large language models (LLMs) commonly serving as their core decision makers. The recent

Tutorials

When Generative Models Act as Simulators

arXiv:2610.12847v1 Announce Type: cross Abstract: A simulator summarizes an environment in a state that users can read and change. Generative models learn representations of the world without explicit

tutorialsarxiv-cs-ai

When LiDAR Fades, Radar Takes Over: LiDAR Reliability-Aware Adaptive Localization Under Airborne Obscurants

model-releasesarxiv-cs-ro

arXiv:2610.13788v1 Announce Type: new Abstract: LiDAR-inertial odometry (LIO) is widely used for real-time localization and mapping, but its reliability can degrade substantially in subterranean envir

When LLMs Sit Above Diagnostic Tools: Unrealized Complementarity in Industrial Fault Diagnosis

arXiv:2610.05031v2 Announce Type: replace Abstract: Large language models are increasingly used as integration layers above specialized tools, but a stronger component does not necessarily produce a s

researcharxiv-cs-cl

When MLLMs miss what humans read: diagnosing and learning to read squint text

arXiv:2610.13465v1 Announce Type: cross Abstract: Multimodal LLMs can fail to read text embedded in an image's coarse structure (text a human recovers by squinting) and report that no text is present.

safetyarxiv-cs-ai

When Simulation Lies: A Sim-to-Real Benchmark and Domain-Randomized RL Recipe for Tool-Use Agents

model-releasesarxiv-cs-ai

arXiv:2605.11928v2 Announce Type: replace Abstract: Tool-use language agents are evaluated on benchmarks that assume clean inputs, unambiguous tool registries, and reliable APIs. Real deployments viol

When Solvers Get Moody: Quantifying and Predicting Performance Variability in Mixed Integer Programming

arXiv:2610.12876v1 Announce Type: cross Abstract: Mixed-integer programming (MIP) solvers are widely used, yet their performance can vary substantially even under seemingly neutral perturbations. Such

researcharxiv-cs-lg

When VLMs Destroy What They Know: Spatial Representations Fall Below Lexical Priors

model-releasesarxiv-cs-cl

arXiv:2610.13910v1 Announce Type: new Abstract: Vision-Language Models (VLMs) increasingly analyze and reason about images, yet it remains unclear whether their internal representations faithfully enc

Who&When Pro: Can LLMs Really Attribute Failures in AI Agents?

model-releasesarxiv-cs-ai

arXiv:2607.09996v2 Announce Type: replace Abstract: Automated failure attribution uses LLMs to identify where and why agentic systems fail. As agents become more capable, their failures become subtler

Why Heavy-Tailed Weight Spectra Predict Model Quality

model-releasesarxiv-cs-lg

arXiv:2610.13123v1 Announce Type: cross Abstract: A solid understanding of the predictive success of deep neural networks (DNNs) remains elusive. For example, while many data-dependent metrics exist t

Why Linear Models Can Still Work in a Complex and Changing World

arXiv:2610.13404v1 Announce Type: new Abstract: Linear models can predict well in complex domains even when underlying mechanisms are nonlinear. We argue that this can occur because the complexity of

safetyarxiv-cs-lg

With only ~$1,000 of training compute, we trained 𝗳𝗿𝗼𝗺 𝘀𝗰𝗿𝗮𝘁𝗰𝗵 a 1.7B coding agent that beats Claude 3 Opus and GPT-4o on SWE-ben…

model-releasesclem-delangue--x

With only ~$1,000 of training compute, we trained 𝗳𝗿𝗼𝗺 𝘀𝗰𝗿𝗮𝘁𝗰𝗵 a 1.7B coding agent that beats Claude 3 Opus and GPT-4o on SWE-bench Verified 🤯 At 25.5% pass@1, it sets a new record for nanoswe, our op

Work2Skill: Capturing Human GUI Experience for Reusable Agent Skill Induction

arXiv:2610.12984v1 Announce Type: new Abstract: Computer-use agents can perform increasingly complex GUI tasks, yet unfamiliar software and workflows requiring human domain knowledge remain challengin

world’s first full length workshop on claude flows @cjav_dev and Harrison Stall are kicking off @aidotengineer bright and early on a monday …

world’s first full length workshop on claude flows @cjav_dev and Harrison Stall are kicking off @aidotengineer bright and early on a monday morning to a PACKED room, standing room only #AIEngineerNY @

model-releasesswyx--x

Applications

worth noting the trend here --> @harvey post-trained an open weight model for legal work + now @coinbase for fraud narrow task + proprietary…

applicationsclem-delangue--x

worth noting the trend here --> @harvey post-trained an open weight model for legal work + now @coinbase for fraud narrow task + proprietary data + open weights seems to beat frontier on cost & speed

WoundAmbit: Bridging State-of-the-Art Semantic Segmentation and Real-World Wound Care

model-releasesarxiv-cs-ai

arXiv:2504.06185v3 Announce Type: replace-cross Abstract: Chronic wounds affect a large population, particularly the elderly and diabetic patients, who often exhibit limited mobility and co-existing h

X-AgentRed: Red-Teaming Personal Agents in Evolving Multi-Service Ecosystems

arXiv:2610.13886v1 Announce Type: new Abstract: Personal agents increasingly connect to services across the web and applications by using tools. As the effects of their actions accumulate and interact

X-AVDD: Cross-Attentive Audio-Visual Dataset Distillation

arXiv:2610.13687v1 Announce Type: cross Abstract: Audio-visual dataset distillation aims to synthesize a small dataset from a large paired audio-visual dataset, while nearly preserving the training ut

safetyarxiv-cs-lg

Tutorials

Yes, and there is a case to be made that OpenAI and Anthropic will do the best job of figuring out how to optimize AI cost/intelligence with…

tutorialsethan-mollick--x

Yes, and there is a case to be made that OpenAI and Anthropic will do the best job of figuring out how to optimize AI cost/intelligence with combinations of agents of different ability levels that hav

Your Discrete Graph Diffusion Model is Secretly Memorizing

arXiv:2610.13786v1 Announce Type: cross Abstract: Discrete graph diffusion models (GDMs) corrupt the graph structure by adding noise independently to nodes and edges under a Markovian noise model. In

safetyarxiv-cs-ai

Zai just debuted a legal specific model: GLM Legal, and it is the new #1 open-weight model on Harvey’s Legal Agent Benchmark and LegalBench.

model-releaseszhipu-ai--x

Zai recently launched a domain‑specific language model called **GLM Legal**. According to a tweet posted at 4:13 PM on Oct 12, 2026, GLM Legal became the leading open‑weight model on both **Harvey’s L

Zebra3D: Efficient Hybrid Foundation Models For Scalable 3D Reconstruction

arXiv:2610.12727v1 Announce Type: cross Abstract: Feed-forward 3D foundation models reconstruct scenes from unposed images, but the quadratic cost of global attention limits their scalability to large

researcharxiv-cs-lg

zip2zip++: Towards Inference-Time Token Compression with Minimal Quality Loss

arXiv:2610.13605v1 Announce Type: new Abstract: Dynamic vocabularies built on-the-fly via Lempel-Ziv-Welch (LZW) compression let a language model represent the same text in fewer tokens, but prior wor

researcharxiv-cs-cl

Local Ai

2x 3090, 3x 3060 potential (smart and informed) future upgrade paths (send help)

I'm trying to figure out the best long-term upgrade path for my setup. Right now my GPUs are spread across several consumer PCs: Main: 5900X, 32GB DDR4, 2× RTX 3090 24GB. Secondary: i3-10320, 16GB DDR

local-air-localllama

Industry

A good example of why open weight and open source models are important - they permit users to fine-tune specific use cases with greater cost…

A good example of why open weight and open source models are important - they permit users to fine-tune specific use cases with greater cost efficiency. As has been the case in other technology market

industryclem-delangue--x

A new work suggesting that division of labor among self-improving AI agents can help them avoid reinforcing their own errors. I found this w…

model-releasesdavid-ha--x

A new work suggesting that division of labor among self-improving AI agents can help them avoid reinforcing their own errors. I found this work interesting from a metascience perspective. Science is a

Anatomy skill for building wireframe designs and interactive prototype machines is now an officially integrated optional skill in Hermes Age…

model-releasesnous-research--x

Anatomy skill for building wireframe designs and interactive prototype machines is now an officially integrated optional skill in Hermes Agent You can find it in the optional built in skills in the sk

b11553

server: add component tests (#30312) server: add component tests nits add codeowners Website: https://llama.app Attestations: https://github.com/ggml-org/llama.cpp/attestations/54782864 macOS/iOS: mac

b11554

vulkan: handle mul_mat_id duplicates in the prepass rather than looping (#29998) vulkan: compute every row of an expert in mul_mm_id when ids repeat vulkan: handle mul_mat_id duplicates in the prepass

b11555

cpu: fix AVX2 CPUs not using F16C when available (#30270) Website: https://llama.app Attestations: https://github.com/ggml-org/llama.cpp/attestations/54796975 macOS/iOS: macOS Apple Silicon (arm64) ma

b11556

model : support for Prism Bonsai 2 27B (#29600) Runtime support for Prism Bonsai 2 27B Assisted-by: Claude Code address prism hadamard runtime feedback move hadamard tensors into method, define folded

b11557: rpc : turn GGML_RPC_DEBUG into a verbosity level and add logs (#29544)

rpc : turn GGML_RPC_DEBUG into a verbosity level and add logs GGML_RPC_DEBUG is now parsed as a number: 0/unset disables debug logs, 1-3 emit increasingly detailed output (events, per-command trace, t

b11558

opencl: fix image limit for q4_k dense bin kernels, q4_0 bcast and rms_norm (#30310) opencl: fallback when Q4_K weight images exceed device limits Co-authored-by: Hongqiang Wang wangh@qti.qualcomm.com

b11559

opencl: improve fa, allow dk512 for gemma-4, improve dk64 (#30266) opencl: enable Gemma-4 E4B GPU decode Co-authored-by: Hongqiang Wang wangh@qti.qualcomm.com opencl: extend Gemma-4 GPU decode to E2B

Building a 4x R9700 setup for a 10 person startup

model-releasesr-localllama

Just wanted to share a build I am doing for a client. $18k. Specs: Threadripper 9970x 128gb DDR5 ECC 5600 4x AMD Radeon R9700 AI Pro 32gb(128gb total VRAM) 1600W PSU(GPUs are undervolted and under 210

Converting dense models into Mixture-of-Experts

model-releasesr-localllama

For the past few weeks I've been trying out converting existing dense models to sparse Mixture-of-Experts models, with no pretraining from scratch. The Idea If you can turn a dense model into an MoE t