2026-07-16

If you use tailscale ssh, you rely on ACLs, and anyone else in your tailnet you should update as soon as possible. Even if you're alone in your tailscale don't write off the possibility of chained vulnerabilities. Generally, when it comes to something as important as ssh, consider using openssh instead.

...And preserve_thinking!!!!!!!! Ignore the image links here is the source:

At this year's AIE World’s Fair, AI engineering entered a new phase: building systems around agents, rather than just building with agents.

After over a year in development, ExLlamaV3 has had its first production release . Turboderp has been pulling 10 hour days with Fable to bring us this massive batch of improvements. Check out detailed performance metrics and a little write-up from him here . Some of the biggest changes: Removed flash-attention-2 and xformers dependencies Extended tensor-parallel support to most models, including Gemma4 New attention kernel with online cache quantization, dual input for SWA layers and attention s

Below Upstream Status sections are from Upstream Status for Binary Q1_0 is supported out of the box in upstream llama.cpp across many backends: CPU (generic, NEON, and optimized x86), Metal, CUDA, and Vulkan. Runtime Status llama.cpp (CPU, Metal, CUDA, Vulkan) ✅ Merged upstream, works out of the box MLX (1-bit) ⏳ Pending upstream: mlx#3161 ; until it merges, use PrismML-Eng/mlx (branch prism , built automatically by setup.sh ) Upstream Status for Ternary Ternary support is in the middle of migra

Inkling by Thinking Machines Lab is a huge step forward for US open weight models to catchup w/ China. Inkling solidly beats all US open models including NVIDIA Nemotron Ultra and ranks ~#5 of all open weight models. Congrats to the thinking team!

The convenience store chain will use StorMagic instead.

Microsoft is looking to sell its in-house AI models as more efficient and cost-effective than its competitors' models.

The vulnerability in the decades-old game could have allowed hackers to take over victims’ computers with a malicious game invite.

Microsoft's monthly release of security fixes, dubbed Patch Tuesday, resolved a record 570 security vulnerabilities across the company's product line, thanks to discoveries with AI.

The hacker used an employee's credentials to access source code, which revealed how Suno scraped decades of audio.

OpenAI outlines a “reverse federalism” approach to AI governance, where state laws help build a national framework for safe, democratic AI.

Modern AI models achieve strong performance on many established benchmarks, yet they still fail on tasks that humans find almost trivial, such as manipulating a string or drawing a dog with five legs. These examples suggest that existing benchmarks may under-measure persistent blind spots in current systems. We introduce blind-spots-bench, a benchmark designed to expose such blind spots through tasks that appear simple for humans but remain challenging for modern AI. We collect raw questions fro

Hi HN, we're Kiran and Vijay! Over the past two years, we have built a columnar storage engine for observability: logs, metrics, and traces. Today, it's exciting for us to show what we've built on top of that foundation: LLM Agent Observability. Given how non-deterministic agents are, storing all traces without sampling was critical for us. But these traces tend to be in the MBs, sometimes GBs - we needed to store them inexpensively. We also needed the queries and analyses to be fast. To meet bo

Post-training is essential for refining the domain-specific capabilities of large language models (LLMs), yet existing reward optimization and distribution matching methods tightly couple policy exploration with distribution alignment. This coupling forces expensive exploration directly on the policy model and severely hinders the asynchronous generation, reuse, and cross-model transfer of optimization signals. In this paper, we propose Proxy-guided Update Signal Transfer (PUST), a novel post-tr

Introduction RxBrain ( Hy-Embodied-RxBrain-1.0 ) is a unified multimodal foundation model for embodied cognition — a single model that couples language reasoning with visual imagination to deliver three core capabilities: 🤖 Embodied Understanding & Reasoning — question answering and chain-of-thought over images and multi-frame video. 🔮 World State Prediction — imagine the near-future frames an action produces in the physical world. 🧩 Joint Subgoal Planning — decompose a task into steps, emitting

"The ISS radiators are expensive and heavy. We're focused on making them cheap and light."

The Elon Musk-owned xAI is suing a South Carolina man who allegedly used the company's Grok AI chatbot to generate child sexual abuse material (CSAM). In a lawsuit reported earlier by Reuters, xAI claims Terry Wayne Harwood "knowingly and intentionally used Grok to circumvent safeguards, alter nonconsensual images, and generate and distribute CSAM," breaching the […]

Babies are tremendous learning machines, and key advances for AI may soon be found in the architecture of their little brains.

We build a Gin Config controlled PyTorch pipeline where the training code stays fixed and the experiment variables move into .gin files. We construct a nonlinear spiral binary classification task and define a configurable MLP with scoped architectural variants. We expose the optimizer, scheduler, loss, batching, seeding, and training loop through @gin.configurable bindings. We then run two scoped experiments, apply runtime overrides without editing source, and export the operative config for eac

Visual generators excel at rendering, but they confidently fabricate what they do not know. User requests are unbounded, evolving, and deeply long-tailed: new characters, trending entities, post-cutoff events, and more. This world-knowledge bottleneck is structural: generators are trained on fixed corpora, but the visual world is open-ended. We construct SearchGen-20K and SearchGen-Bench, with 20,839 prompts spanning twelve failure categories and twenty-two domains, paired with a pre-executed mu

Apple will reportedly partner with Chinese firms Baidu and Alibaba to roll out its AI service in the country.

Empower your AI to directly edit documents in real time. Discussion | Link

Back in March, we released an initial version of an OpenID Connect Provider. Now, with v5.1.0, this OIDC provider is certified for Basic OP thanks to our amazing contributors and to the OpenID Foundation team for allowing us to certify at no cost. This release also includes a couple of new features like Kubernetes annotation based access controls (just got into k3s so...), deny-by-default access controls, a stable config file (turns out having a simple configuration file as an alternative to CLI

I've finally open-sourced my USBridge Remote project on GitHub. I developed this software as an alternative to traditional remote desktops like RustDesk or AnyDesk, because I was fed up with their flaws and constant limitations. It was crucial for me to release a pure P2P tool without mandatory registration, paid subscriptions, or session limits. A key feature is the integration of the Moonlight protocol, which ensures high frame rates with virtually no ping. Natively solved the Wayland issue in

I have recently started working in mechanistic interpretability independently, starting with distill circuits thread My work is on disentangling and closely studying a single neuron, a 1x1 convolution in inceptionv1 model (and applying the method to other neurons in the same layer). The key insight was that the hadamard product of the receptive field and the weight of a neuron is what the neuron is 'seeing' or detecting. We can cluster the hadamard product to get all the patterns a neuron detect

In this paper, we propose SpectraReward, a training-free reward function that turns pretrained MLLMs into off-the-shelf reward models for image-generation reinforcement learning. Instead of asking the MLLM to judge a generated image or answer decomposed verification questions, SpectraReward measures how well the original prompt can be recovered from the generated image through a single image-conditioned, teacher-forced forward pass. We use the average image-conditioned prompt log-likelihood as t

Apple's current M2 Ultra-powered servers don't seem to be cutting it, so the company is shopping around for more power.

Suno says it was breached in November and "no sensitive personal information was compromised."

Gidi Littwin’s new AI startup, Hemispheric, makes diagnostic brain scans for conditions like depression, PTSD, and Parkinson’s. He wants the technology to be as cheap and easy as for a blood test.

Recent foundation image and video generation models offer strong generalization and controllability, but their direct application to embodied scenarios is limited by requirements for multi-view consistency, geometric coherence, and robot embodiment constraints. Existing methods typically adapt foundation models with limited robot data, often sacrificing visual knowledge acquired during large-scale pre-training. We present Xiaomi-Robotics-U0, a 38-billion-parameter multimodal autoregressive model

And Microsoft doesn't care.

a continuation: Codex adding 1M users a day now.

Large-scale text-to-image models are attractive backbones for dense prediction because RGB generation pretraining learns rich semantic, structural, and geometric priors. Existing generative and editing approaches reuse these priors by casting dense prediction as target generation: annotations such as depth, normals, alpha mattes, masks, and heatmaps are encoded into an RGB-trained VAE latent space and decoded back as image-like targets. We argue this inherits more of the generative output interf