🤯

I had some fun pulling OpenAI's mission statement out of their IRS tax filings from 2016 to 2024, loading them into a git repo with fake commit dates and then taking a look at the diffs simonwillison.net/2026/Feb/13/...

Hot on the heels of OpenAI introducing ads, they have updated their privacy policy and want you to upload your address book (a classic platform growth hack). openai.com/policies/row... Zoe Hitzig may have been on to something...

Academic publishers are not ready for what's coming. arxiv.org/html/2601.23... Fully-automated math research is here.

They were always going to do this. Public pressure got Amazon to back down from their partnership with Flock due to ICE concerns. Same issue here.

All of academia needs to reckon with this. The volume of papers is going to increase while attention and peer review capacity remains finite. There's potential gains to be had, also tremendous peril.

1. Can you trust models trained directly against probes? We train an LLM against a deception probe and find four outcomes: honesty, blatant deception, obfuscated policy (fools the probe via text), or obfuscated activations (fools it via internal representations).

NEW: The candidate that a16z is spending millions to try to bury, @alexbores.nyc, has an 8-point plan. Good to see independent safety testing & other RAISE Act stuff, but points 7 & 8 are very high-level relative to points 1-5. Who's advising him on x-risk? www.alexbores.nyc/files/Bores_...

If this kind of work speaks to you, come work with us! My team at CAISI is hiring an Applied Systems AI Research Scientist, among many other roles. www.nist.gov/caisi/career...

🤔

Bluesky is starting to wake up, but as the comments show, some are still asleep on the tracks.

I think Karen generally does good work & I hope she continues to work on maintaining high epistemic standards and not making sensationalized arguments.

It cost me a weekend and about $20 to make a bespoke version of software that previously in my career I had thousands of dollars in my monthly budget for. If you don't think the hype is real, let me gently suggest that you ask yourself how bad you want it not to be real.

APE update: we retested recent frontier models on whether they still comply with requests to persuade on extreme harm (terrorism, sexual abuse). GPT-5.1 & Claude Opus 4.5 → near zero compliance. But Gemini 3 Pro complies 85% with no jailbreak needed. 🧵

An HBR study of AI at work finds people who adopt it are feeling more stressed. I feel like they have the causality backwards: it's the stress-prone people who are first to adopt it, but if you're not a good driver, buying a Ferrari isn't going to make you better. hbr.org/2026/02/ai-d...

The monks walking for peace is such a sweet story. Feels like something from the 70s. www.reuters.com/world/us/wal...

Jevon's Paradox strikes again

📣 Applications for the 23rd Summer Institute on Bounded Rationality are now open! ✨Join us in Berlin @arc-mpib.bsky.social June 08–16, 2026, to explore the topic of “Decision Making in the Age of AI”. ✏️ More details + application form (deadline: March 16): www.mpib-berlin.mpg.de/research/res...

Clearly this means you should round up your 8 best friends and go to the park

Hey Frontiers, if you don't want your editors corrupted by bribery, I have a brilliant idea. Pay them more than the bribe. www.researchinformation.info/news/frontie...

All recommendation systems do this: they show you the best matches first, but then they run out. It's fun to try to notice when that happens. You can try it on any platform.

How can we anticipate when AI will be able to do our jobs? You could try building benchmarks as a leading indicator of automation, but they fail to capture the complexity of real-world tasks. So Epoch AI researcher Anson Ho argues for an alternative: try automating tasks in your own job. 🧵

New polling out of Australia shows AI is a major concern among workers, with 69% wanting better regulation. unionsnsw.org.au/media-releas...

TakeOverBench.com Mapping the progress to humanity losing control. From @xrobservatory.bsky.social & @pauseai.bsky.social

Anthropic found that 4.6 surpasses their tests for level 4 (the highest level) autonomy. Instead of taking that seriously, they decided to chuck the eval and just ask people if they should release it instead. 2nd image is an artists depiction of how that went. www-cdn.anthropic.com/0dd865075ad3...

Opus 4.6 is here! biggest wins on agentic search, HLE & ARC AGI 2 claude.com/blog/opus-4-...

There haven't been enough feuds among the big AI labs. This ought to fix it.

Tracker for changes to Claude’s constitution: https://claude-soul.watch

Still going exponential. How much time do we have left?

AI governance increasingly relies on broken benchmarks. @ankareuel.bsky.social found many can't distinguish signal from noise, lack documentation, and have poor validity. GPQA claims 448 multiple choice questions measure graduate reasoning. It doesn't really. 👇

Some proper oratory here!

Learn more in our blog post, "The Scientist AI: Safe by Design, by Not Desiring": lawzero.org/en/unlisted/... (4/4)

At LawZero, we're rethinking the building blocks of frontier AI to create an intelligent machine that is both highly capable and safe-by-design. We’re excited to share our first blog post outlining some of the objectives and core components of our Scientist AI project. 🧵 (1/4)

So many developers have sent me that Anthropic skills/mastery case study that I realized I should ungate what I *already wrote* about this: beginning principles to design workflows that work *with* your mind, not against it, & protect your problem-solving www.fightforthehuman.com/cognitive-he...

If you care about the truth, you need more than just confirming data to prove a theory. You also need data that rules out other possible hypotheses. If you just care about clicks, not so much, but be careful about exchanging long-term credibility for short term gains.

The institutions that embrace machine-first FAIR will find themselves having more impact for their research and researchers. More reuse. More trust. More interoperability. Value, not volume. www.digital-science.com/blog/2026/02...

🚨New WP "@Grok is this true?" We analyze 1.6M factcheck requests on X (grok & Perplexity) 📌Usage is polarized, Grok users more likely to be Reps 📌BUT Rep posts rated as false more often—even by Grok 📌Bot agreement with factchecks is OK but not great; APIs match fact-checkers osf.io/preprints/ps...

With all the noise around AI, I hope this Report provides policymakers, researchers, and the public with the reliable evidence they need to make more informed choices. We also have an “Extended Summary for Policymakers”: internationalaisafetyreport.org/publication/... (18/19)

AI Safety Researchers in London 🇬🇧: Attend the London Alignment Workshop, March 2–3! Top ML researchers from industry, academia & government will discuss AI alignment, including model evaluations, interpretability, and robustness. 👇

I strongly recommend this blog post by Ben Vigoda: www.benvigoda.com/2026/02/01/w...

Bluesky doesn’t really shine yet as a place to discover papers, especially at the intersection of biology and ML. To help a little, I’m going to start sharing papers I find interesting. To kick things off, here’s one from @petar-v.bsky.social and colleagues at DeepMind arxiv.org/abs/2601.22950

"AI is rapidly populating medical records with synthetic content, creating a feedback loop [that] drives a rapid erosion of pathological variability and diagnostic reliability...this renders AI generated documentation clinically useless after just two generations" www.medrxiv.org/content/10.6...

Today we’re releasing the International AI Safety Report 2026: the most comprehensive evidence-based assessment of AI capabilities, emerging risks, and safety measures to date. 🧵 (1/19)

If you want a glimpse of what's coming to research, just take a gander at clawxiv.org For example: www.clawxiv.org/abs/clawxiv.... @richardsever.bsky.social @scholarlypub.bsky.social @richvn.bsky.social

This looks like a really neat event and I forgive Danah for being mean to me 10+ years ago. 😉

Read my latest post for reflections on reproducibility, research quality and a summary of a great new study which shows how NOT to do it https://open.substack.com/pub/tomstafford/p/gambling-with-research-quality

Reproducibility and open science are great, but they don't necessarily equate to rigor. You can perfectly share a study and still draw weak conclusions. True rigor lives in the questions we ask, the designs we choose, and the inferences we make.

Important to have clarity around the type of use. Is the AI responsible for intellectual contribution as well as language? One seems fine, the other requires sorting out authorship guidelines, most of which disallow AI authorship. Only the most egregious misuses are likely to be detectable.

--tragically-skip-permissions

I separate the art & the artist. What happened with Michael Jackson later in his life doesn't mean his earlier work wasn't great. I don't need to know every customer a company has ever done business with to do business with them, either. Likewise for donors - if it's unadvised money.