Hark is scarily proactive. It noticed I'd recently registered for #NeurIPS2026# in Atlanta and offered to pull the schedule, which had just gone up on the website. Instead, I asked if it could build me a web app to explore the papers. A few minutes later, I had a working one. It covers all 1,829 papers being presented in Atlanta, with the full poster-session schedule, search across titles, authors and abstracts, and filters for orals and spotlights. My favorite part: there's a star button to build my own watch list, and it labels the papers that match my research interests (which I think it picked up from our past conversations). I asked for a schedule and got a personalized conference planner.

Which review complaints get ICLR papers rejected? I used @adaption_ai's Visualize to cluster the harshest review of 5,000 ICLR 2026 papers. Complaints about weak baselines or novelty drop acceptance to ~23%, while those about compute cost or real-world validity leave it at 36%.

Kernels encode the inductive biases of a wide range of ML models. Yet most kernel design approaches are limited to a fixed library of base kernels and operators. What if we let coding agents write kernels as programs instead? We introduce Kernel Autoresearch (Kernaut 👨‍🚀), which treats kernel design as open-ended program synthesis. Coding agents write kernel components as programs, and construction contracts ensure that accepted kernels are valid. 🧵

We're excited to have @sirbayes in our next virtual seminar! He will talk about his recent work on the Model Discovery Agent (MDA), which learns mechanistic world models for AI-driven scientific discovery. You don't want to miss this!

I got early access to @Alibaba_Qwen's new image model, Qwen-Image 2.1. I ran 5 prompts (one-shot generation) covering a dense infographic, a multi-panel comic, a macro photo, a chalkboard menu, and a movie poster. Thread below 🧵

In today's agents, we have many classification steps that still burn a full LLM call. This is a great write up on how we can utilize @typesafeai's Jev for those: model routing (cheap model for simple lookups, stronger one for hard debug) and auto mode (flag risky tool calls before they run).

🚀 Meet Qwen3.8-Omni-Flash, Qwen's first omni-modal model built around agentic capabilities! Native audio-video understanding, reasoning, and tool use come together in one model: understand the content, plan the task, execute with tools, and deliver the result. Highlights: 🥳 - Audio-video intelligence that gets things done: jointly reason over what's seen and heard, and orchestrate tools across long workflows to auto-edit vlogs, translate short videos, and turn movies into recaps. - A major leap: approaching Gemini 3.8 Flash in audio-video capabilities; +19.5 points on average in agent performance across WildClawBench-MM & UniClawBench. - 1M-token context with agentic perception: actively explore long videos and locate key moments with higher accuracy, using 51.8% fewer tokens than static understanding on OmniVideoBench. Video input costs are reduced by about 89% compared with Qwen3.5-Omni-Plus, making long-form audio-video understanding and agentic workflows more affordable than ever. To help you build apps around Omni, we're also open-sourcing Qwen-MM-Plugins and Qwen-Live Harness! 🛠️ We can't wait to see what you build with Qwen3.8-Omni-Flash! 👀 - Blog: - Qwencloud: - Qwen Studio: - API: - Qwen-MM-Plugins: - Qwen-Live Harness: coming soon

Excited to share PlugBO 🔌, a modular framework for agentic Bayesian optimization (BO) In a standard BO routine, we usually declare the surrogate family, the acquisition function, and the search bounds before the first evaluation, and we keep them fixed after that. What if, instead, we treat those choices as live slots that an LLM agent can inspect and swap?