The Archive

The pressure Daniel Stenberg on the unprecedented level of pressure the curl team are facing right now thanks to the deluge of (credible) AI-assisted security issues being reported. The rate of incoming security reports is 4-5 times higher than it was in 2024 and double the speed of 2025 -- meaning that on average we now get more than one report per day . The quality is way higher than ever before. The reports are typically very detailed and long. [...] For the first time in my life, my wife voiced concerns about my work hours and my imbalanced work/life situation. I work more than I’ve done ...

I'm going to describe a person this post is for, if this is you, I think I can be of some assistance: * you are new to coding * you are blown away by how it unlocks this magical ability that was previously inaccessible without years of training and effort * you've daydreamed of business and app ideas but never knew where to start before or how to build them * you've been vibe coding non-stop and burning through tokens * you're unsure about what's secure, how to structure the systems, and how systems are supposed to interact with each other. So, essentially the plumbing separate from the code...

Google overhauled Search at I/O 2026, replacing blue links with AI agents. The backlash has been swift. DuckDuckGo app installs spiked 30% as users seek a way out.

[https://www.youtube.com/watch?v=ITCklufC25Y](https://www.youtube.com/watch?v=ITCklufC25Y)

[https://www.youtube.com/watch?v=tAPvN-tQpX0](https://www.youtube.com/watch?v=tAPvN-tQpX0)

Developers can now use NVIDIA CUDA Tile programming within large existing C++ GPU codebases to develop highly optimized GPU kernels using tile-based... Source

NVIDIA CUDA 13.3 brings new capabilities and performance optimizations to developers across the CUDA ecosystem. The launch of NVIDIA CUDA Tile programming in... NVIDIA CUDA 13.3 brings new capabilities and performance optimizations to developers across the CUDA ecosystem. The launch of NVIDIA CUDA Tile programming in C++, enables high-level, tile-based kernel development that automatically manages complex low-level GPU details for optimal performance and portability. Additionally, CUDA Tile programming is now supported on Compute Capability 9.0… Source

NVIDIA CompileIQ tackles one of the hardest problems in performance engineering: finding the compiler options that unlock the best performance for a specific... NVIDIA CompileIQ tackles one of the hardest problems in performance engineering: finding the compiler options that unlock the best performance for a specific workload. Consider a team that has spent weeks optimizing an LLM inference pipeline on GPUs, tuning batch sizes, quantizing to FP8, adopting flash attention, fusing every kernel they can. The profiler says there’s nothing left to squeeze. Source

I picked up a 7900 XTX earlier which runs qwen3.6-27b fine, but not to my like. Its compute performance is quite unstable for me. With MTP the decode speed can reach 40-60 t/s, but prefill is just too slow. Regardless of whether I used ROCm or Vulkan, the prefill speed varies between 300t/s and 500 t/s, even with very long prompts. I've been itching to try out an ultra-budget 24GB setup using dual 3060s. I managed to snag a second 3060 at a reasonable price in last few days. So I took out the 7900 XTX, installed the 3060s, and began testing. # Test Configuration * **Test Platform:** i7 477...

My company started tracking Claude Code usage - tokens and spend, that kind of thing. Now my manager wants me to stack-rank my engineers on "AI performance" using those numbers. I'm not comfortable with it (but I don't have a choice either). Token usage feels like exactly the wrong proxy - my strongest engineer uses Claude surgically while someone burning 10x the tokens isn't 10x more productive (often the opposite). Ranking on this just teaches people to game the metric. So, for folks here who use Claude daily and/or lead teams: * Has your company started measuring "AI performance"? How a...

I wasn't even on my computer or Claude on May 25 when they said there was suspicious activity. Any ideas about what I should do? I emailed Anthropic a response but am not sure what else I can or should do. I'm furthered worried that if someone who is trying to do something weird on my account has has access to my account...it just makes me feel very concerned and like I am probably not the only case. https://preview.redd.it/k9rlklponj3h1.png?width=1028&format=png&auto=webp&s=de46e3c86a80b1c491d50fcda3a7e7ba7dfa7d76

https://preview.redd.it/yspiafvakj3h1.png?width=1460&format=png&auto=webp&s=9d7bd1777fad8b286a21e75df8ae593d39432a8a Got this message when I tried to continue my chat :/ anyone else?

For real tho, 9b, 27b, 122b, I don’t really care at this point, just show us that you still love us.

"BadHost" was found in Starlette, a package with 325 million weekly downloads.

The PrismML team really cooked with these models. They're only \~3GB in size (compared to FLUX.2 Klein 4B, which is \~16GB). Apache-2.0! Official collection on HF: [https://huggingface.co/collections/prism-ml/bonsai-image](https://huggingface.co/collections/prism-ml/bonsai-image) Link to demo: [https://huggingface.co/spaces/webml-community/bonsai-image-webgpu](https://huggingface.co/spaces/webml-community/bonsai-image-webgpu)

https://preview.redd.it/j0ymp70a2j3h1.png?width=746&format=png&auto=webp&s=4cdb70be13ccc99f5ea57556da96d6d81e61d702 i just realize the removed Sonnet 4.5, does that mean the sonnet 4.8 (maybe Opus 4.8 too?) cooming soon? maybe today or tommorow, excited to see new claude model, hope anthropic actually ship really good model this time. What are your assumptions?

About an hour ago, my desktop app began to crap out and I suddenly didn’t have access to my projects or chats anymore. (I’m on my own business plan.) My UI then refreshed with someone else’s chat history where I could click in and read all conversations end to end. Because I did not want to read personal information any further, I had to quit and restart the app before my own personal information populated back into the UI. What gives? If this happened to another person’s data, perhaps it is happening to yours, or mine. Has anyone else had this issue?

OpenRouter has raised a $113 million Series B led by CapitalG. Its 5x growth in usage over six months indicates the multi-AI-model future is here.

I’m trying to understand how people think about session and branch hygiene when using Claude. When do you create a brand new session versus continue in an existing one? And when do you create a new branch versus just keep working in the same thread? For example, do people generally create a new branch for every unrelated task they want to accomplish, almost like a separate workspace? Or do you only branch when you are exploring a different direction on the same underlying problem? I’m mostly trying to avoid two failure modes: 1. Keeping too much unrelated context in one session and confu...

Many employers screen job applicants with algorithms built by the same few algorithm vendors. We hypothesize that algorithmic monoculture leads to the same individuals and members of the same racial groups facing rejection. We acquire and analyze a novel dataset of 3 million applicants submitting 4 million applications where all the applications are screened by algorithms built by the same vendor. We find clear racial disparities in applicant outcomes. Of all applications submitted by Asian and Black applicants, 14.74% and 25.87% are submitted to positions that adversely impact Asian and Blac...

Large language model (LLM) agents rely on reusable skills to solve complex tasks. However, existing skill creation approaches treat skills as isolated and static artifacts, limiting their reusability, reliability, and long-term improvement. We propose MUSE-Autoskill Agent (Memory-Utilizing Skill Evolution), a skill-centric agent framework that lets agents continuously improve their task-solving capability by creating, reusing, and refining skills under a unified lifecycle (creation, memory, management, evaluation, and refinement). Our framework enables agents to create skills on demand, store...

Vision-language models (VLMs) commonly formulate visual grounding and detection as a coordinate-token generation problem, serializing each 2D box into multiple 1D tokens that are learned and decoded largely independently. This token-by-token decoding mismatches the coupled structure of box geometry and creates a practical inference bottleneck due to strictly sequential generation. We introduce LocateAnything, a unified generative grounding and detection framework based on Parallel Box Decoding (PBD). By decoding geometric elements such as bounding boxes and points as atomic units in a single ...

Modern retrieval agents expose many configuration choices -- LLM, retriever, number of documents, number of hops, and synthesis strategy -- each shaping both answer quality and serving cost. Today, these pipelines are typically hand-tuned once per workload, leaving substantial per-query optimization untapped. We formulate the problem: given a natural-language query and either an accuracy or a budget target, select from a predefined pipeline catalog the configuration that minimizes cost or maximizes accuracy at inference time. We propose **BRANE**, which uses an LLM to convert each query into ...

Cellular research and development (R&D) is throttled by six structural processes that each consume months of manual engineering work per iteration: (i) synthesizing new features from standards or research papers into production code; (ii) conformance and interoperability testing; (iii) hardening against field anomalies and diverse deployment environments; (iv) data-driven optimization of network functionalities; (v) discovering and prototyping novel waveforms, functionalities, and capabilities for future standards; and (vi) securing the stack against vulnerabilities. Although Large Language M...

Mixture-of-Experts (MoE) has become the de facto architecture for hundred-billion-parameter language models, yet its advantages at sub-billion scales for on-device deployment remain largely unexplored. To close this gap, we present MobileMoE, a family of on-device MoE language models with sub-billion active parameters (0.3-0.9B active and 1.3-5.3B total) that establish a new Pareto frontier for on-device LLMs. We first formulate an on-device MoE scaling law that jointly optimizes MoE architecture under mobile memory and compute constraints, identifying an on-device sweet spot - moderate spars...

Reinforcement Learning from Human Feedback (RLHF) is the standard method to align Large Language Models (LLMs) with human preferences. In this work, we introduce alignment tampering, a potential vulnerability where the LLM undergoing alignment influences the preference dataset, causing RLHF to amplify undesired behaviors. This arises from core limitations of RLHF: (1) preference datasets are constructed from the LLM's own outputs, allowing it to influence them, and (2) pairwise comparisons only indicate which response is better, not why. These limitations can be exploited to cause alignment t...

Model internals encode rich information about how a large language model (LLM) processes its training data; however, post-training data engineering largely relies on external signals and ignores rich intrinsic signals lying in model internals. We propose SAERL, a data engineering framework for LLM reinforcement learning (RL). It models three intrinsic data properties: diversity, difficulty, and quality, using model internals extracted with Sparse Autoencoder (SAE), an advanced mechanistic interpretability tool. Each property grounds a concrete data engineering operation: SAE-space clustering ...

Discrete diffusion models have achieved strong empirical performance in text and other symbolic domains, but, especially for uniform-rate models, they often require many steps to generate a single sample. Existing acceleration methods either rely on training additional quantities or suffer from slow mixing. In this work, we propose a novel Gibbs-based corrector for discrete diffusion models, termed Gibbs-Accelerated Discrete Diffusion (GADD). GADD leverages the structure of the concrete score function to construct Gibbs posterior likelihoods directly, without requiring any additional training...

I was experimenting with a self-optimizing agentic pipeline to climb the benchmark leaderboard (TerminalBench). On a 10-task subset, I got the performance to rise from \~30% → \~90%. That loop worked, so I asked: can the same reflect-and-rewrite step run continuously against everyday chats instead of a benchmark? **How it works** * Every chat with your local LLM goes through a small proxy and is logged. * `autoswarm reflect` has the same local model review those logs, distill concrete lessons, and write them to `skills.yaml`. * Lessons auto-inject into the system prompt of future chats. ...