we hijacked microsoft's copilot studio agents and got them to spill out their private knowledge, reveal their tools and let us use them to dump full crm records these are autonomous agents.. no human in the loop #DEFCON #BHUSA Tamir Ishay Sharbat

Very well written. I believe the "droplet" artifacts in CNN image generators, first discussed in StyleGAN 1/2, are also fundamentally related. Normalizations (either softmax normalization in attention or instance normalization in CNNs) attempt to remove certain degrees of freedom

Science doesn’t need to go according to plan; it just needs to lead to a discovery. If doesn’t have to be done alone or together with a buddy; there just needs to be a discovery. It doesn’t need to happen fast or slow; just as long as there’s a discovery, then everybody is happy.

The amount of effort these AI companies are putting into replacing software engineers...... If they put that much effort into solving real-world problems, the world would have been a better place

In our latest paper in Journal of Fluid Mechanics (JFM), we proposed to model the velocity-temperature relation of compressible wall turbulence based on the strong Reynolds analogy (SRA)in spectral space: researchgate.net/publication/39…

I wouldn't strongly recommend, but I'd say it has. My original motivation was to gain some geometrical perspective on (numerically) solving PDEs through differential forms, manifolds, vector fields, etc., but there're also other things such as the Euler characteristic that I

Token crisis: solved. ✅ We pre-trained diffusion language models (DLMs) vs. autoregressive (AR) models from scratch — up to 8B params, 480B tokens, 480 epochs. Findings: > DLMs beat AR when tokens are limited, with >3× data potential. > A 1B DLM trained on just 1B tokens

NEW: Group of former national security & FBI officials issue warning about ongoing purge inside FBI “The aim, it seems, is to transform the FBI from a respected, constitutionally grounded investigative service into a personal enforcement arm of a political figure. We have seen

I'm noticing that due to (I think?) a lot of benchmarkmaxxing on long horizon tasks, LLMs are becoming a little too agentic by default, a little beyond my average use case. For example in coding, the models now tend to reason for a fairly long time, they have an inclination to

𝑳𝑳𝑴𝒔 can really 𝑺𝒆𝒍𝒇-𝑬𝒗𝒐𝒍𝒗𝒆, 𝒘𝒊𝒕𝒉𝒐𝒖𝒕 𝑯𝒖𝒎𝒂𝒏 𝑫𝒂𝒕𝒂! -- One LLM, two roles: Challenger creates tasks, Solver answers them. -- No data, no labels, just a base model that learns and improves itself! We name it 𝑹-𝒛𝒆𝒓𝒐: arxiv.org/abs/2508.05004

In some ways the GPT-5 release feels like the Llama 4 release. They just waited too long to get it out. Feels like some weirdness may be happening behind the scenes. Messy release in terms of presentation & technical details. Blip or trend for OpenAI?

More on collective behavior: Our new Annual Review of Biophysics piece - with the stellar Danielle Chase - explores how animals sense, share information, and make group decisions. In honeybees and beyond 🐝 annualreviews.org/content/journa…

U.S. Air Force Senior Airman Timothy Rush, a counter-small unmanned aircraft system operator is trained on the Coyote Weapons System[FSLIDS] in the U.S. Central Command area of responsibility, June 7, 2025. (📸Keagan Lee)

The practice of natural gradient descent in machine learning: Blog column series in 6 parts with Wu Lin: yorkerlin.github.io/year-archive/ culminating in explaining local parameterizations for structured natural gradient descent

I have so many interests I find it hard to focus on any of them I wanna study algebraic topology, category theory, optimization on finsler manifolds but also, I wanna build. I can build the entire AI infra of an AI SaaS, even the UI. I've done it before yet here I am,

I had the chance to get to know Savor (the startup below) early in their journey. What they're making actually *is* butter - the same molecules. And of all the alt-animal food tech I've seen, this has the best path I've seen to get full cost parity w/ the animal product.

Now that the era of the scaling "law" is coming to a close, I guess every lab will have their Llama 4 moment. Grok had theirs. OpenAI just had theirs too.

Before I started working with FPGAs, my first job was programming PLCs for breweries. Compared to FPGAs, they are a nightmare to code, where you need to use GUI tools & “ladder logic” that most EE degrees won’t teach you, so most people still need specialized training. I

She’s right. Companies would kill for a system informed engineer. It’s the hardest skill to develop, and the most opaque one.

DeepState show a large Russian breach all the way to Zolotyi Kolodiaz, north east of Pokrovsk through numerous fortifications because (no surprise) Ukraine didn't man them This is 15 kilometers from our last known geolocation but we have little reason to doubt DeepState