this finding matches my experience: the valuable thing that knowledgeable human devs can bring to a project, that the agents aren't yet good at, is ontology creation.

Think of this as a pattern for any LLM integrated system that claims to provide guarantees. Agents propose, domain verifiers validate, approved proposals are committed and every step and decision is logged. Yes, most domain specific verifiers can be hard. So is building most deterministic system.

Someone just wrote me to ask if I could be their Arxiv endorser, meaning someone who vouches for the author uploading non-slop. I declined, because I am not clear on the recent submission rules. It appears publishing a preprint yesterday put me on a list of verified authors.

New from me. arxiv.org/abs/2606.04903

Great thread! Empirical evidence about LLM use in scientific articles.

9. Here's a surprise: controlling for the other covariates (correct me if I have that wrong, Kyle), we see the *most* LLM use in the high impact journals, not low impact journals.

While the curl project does not ban the use of AI tools - recognizing they can enhance development - AIs are merely tools. Humans must always drive the process, taking full responsibility for presenting, reviewing, and understanding every change.

Having been a target of a social media-driven OSS pile on, the only thing you can do is continue without taking their written slop into consideration and locking the thread. In my case it was proceeding with a Code of Conduct and daring the assholes to fork. They gave up really quickly

The bitter math lesson: you can essentially solve all of math by just doing more matrix multiplication

one of my favorite cheeky papers, from 1999, is this one where the authors argue that sometimes it is better to simply wait and do nothing, because your astrophysical simulations will complete faster if you simply wait for the next epoch of compute arxiv.org/pdf/astro-ph...

The leaders running this initiative at Stanford really hope their idea spreads to other health systems — eventually making patients’ voice a critical feedback channel that shapes how AI tools are implemented, @brittanytrang.com reports: www.statnews.com/2026/05/27/s... via @statnews.com

This is a continuation of the trend away from music as a primary source of entertainment, and toward music as a background experience. See the rise in searches for "coding" or "lo-fi" or "asmr" or "white noise" or "focus" or "healing," instead of searching for artists or music genres.

When I woke up this morning I didn't think I'd be spending a bunch of time today getting familiar with Catholic theology, but here we are. Notes on Pope Leo XIV's encyclical on AI. simonwillison.net/2026/May/25/...

Short and very much worth the read.

Best joke I've seen a Claude model create.

At @arxiv.bsky.social, we are receiving a new type of paper that I call an "I did this experiment" paper. These papers typically report some experiment with an LLM or LLM "agentic" workflow. They are the kind of experiments an "insider" engineer would run to optimize a system. 1/

my take for the last year or so on this kind of stuff is that even if AI doesnt technically get any better from this point on, we're still years away from optimizing the tooling to really get reach the full power of the things

I use this agentic AI workflow in a new Health AI post to dispel the myth that AI can't be trusted because it hallucinates. Agents can improve trust by breaking tasks into pieces with checks that reduce hallucinations caused by LLMs healthaiinsights.substack.com/p/myth-vs-re... #ai #agenticAI #llms

This thread is incredible and everyone with even a passing interest in AI consciousness should read it.

The first Captain Disillusion debunk I've seen on Bluesky, and it's a good one.

When I heard Carlini predict that eventually everything would be written in memory-safe languages, I envisioned mass migration off of C. I didn't expect the C-family to make its own languages more memory safe, but here we are.

TBH I was pretty torn about contributing to this. In the end I decided that writing something restrained was better than writing nothing.

"No way to prevent this" say users of only language where this regularly happens https://xeiaso.net/shitposts/no-way-to-prevent-this/CVE-2026-45250/

OpenAI's claim that this is a central conjecture in discrete geometry is not an exaggeration. This will I think be looked back on as the first time that AI solved a major mathematics problem (defined as a problem that all experts in some subfield had thought about). openai.com/index/model-...

It's no mandate of heaven, but having "Read the proof↗️" has a few mandate particles on it. Imagine what math-mythos could find out

Claude Opus 4.7 just created a project memo with a section titled, "Open questions for the human before starting."

Maintainer of curl, one of the most-used, and most secure, services on the internet. Mythos only found one vulnerability in a recent scan.

Hello Bluesky. It's Playboy. For our first, and timely post, we share our latest investigation into OpenAI's disastrous plans to become x-rated. "Altman’s idea of an “erotica” feature seemed riddled in uncertainty." Read our piece "Why ChatGPT Can't Be Sexy" here: www.playboy.com/read/politic...

6/ Research culture + references: • Citation hygiene problems (Suflaky): no central bibliography source; metadata disagreements. https://x.com/Suflaky/status/2056388796938179019 • Jean-Pierre Serre on Quora (datagenproc): https://x.com/datagenproc/status/2056476859534295235 • Score-Difference Flo...

Conclusion: agents can already help scientists with tedious data-reuse work, but they are not reliable enough to run fully autonomously. Careful human-in-the-loop review is still necessary. 10/10

Extremely strong agree. This post is about adding an Edit button, but imo it is true about any possible new feature, even the most innocuous. Attack it first, if it survives, float it to the community as a possibility.

A security expert using Mythos compromised MacOS. www.wsj.com/tech/ai/anth...

This is fantastic. And beautifully presented, proving the talk's correctness.

Ultimately, the continued losses on the direct path are the context for Texas v. Epic. If you can't get the records through HIPAA subpoenas, you move upstream to EHR vendor proxy-access configurations and let parental-rights litigation do the work.

Terence Tao on New Mathematical Workflows. Includes things he learned from several crowdsourced math projects. youtu.be/Uc2zt198U_U

😱 1 year ban from arXiv and no more tech reports anymore, must be papers accepted at a reputable venue. Wow, talk about taking action against AI slop science...

Hey folks, I'd like to hear from folks who have deployed their own LLM on their own machines, instead of using Open AI or Claude. What tutorials did you use? What is working well? How did you train it? How did you keep it secure? Ethical? Would love your thoughts!

I'm on a mission to make Claude the best MCP client. Diving deep into every MCP issue in our public repos. Having MCP problems with Anthropic products? File here: ✨ github.com/anthropics/c... ✨ github.com/anthropics/c... Blocked or no response? Tag me, we'll fix it.

honestly haven't felt this hopeful about climate in a *while*

"I was gone four whole hours, and the family had no idea where I was – they even called the police." When she was four, Dr. Jane wanted to learn about hens and find out how they laid eggs – and so she went off, without letting anyone know where she was going. Video: National Geographic

I'm on it. There is a Mythos scanning #curl blog post pending.

Why was the web designed so a web site gets so much information about its visitors? Here is a page that tells you what information your browser shares with it. sinceyouarrived.world/taken

as technologists, we are in fact responsible for the consequences of what we create and to close your eyes to this is hubris, carelessness, or malice often it's whatever but this time the stakes are too high

this from the end of the blog post feels like somewhat timeless advice as it applies to tools broadly

To add some fuel to the AI coding fire, now that these numbers are public in the shareholder letter Block has: - 15% of PRs are entirely automatic (human writes ticket -> bot picks up -> human approves PR) - SEVs from code changes down 40% - PR rate up 2.5X s29.q4cdn.com/628966176/fi...

the quality of the models matters enormously — better models are less vulnerable to injection and less likely to make weird mistakes — which effectively makes AI safety a luxury good

I love all these people who are inveighing against the 24-hour news cycle and its evils. You know you're on Bluesky, right? Don't pretend this is anything but a group home for those who are addicted to the drip-drip-drip of information and discourse.

I'm following this, and Simon is doing an outstanding job imo.

uh, so i think i can claim that one of my open-strix agents is doing fully autonomous software engineering kinda want to hedge that, because it’s a bold claim — i feed it new directions, and it really only comes back to me for issues that are truly ambiguous and need deeper discussion