William Gunn

Good for them.

\

Oh darn. I'm so sad. Where's my tiny violin: blog.character.ai/an-update-on...

Today I am sharing a new preprint and accompanying blog post. Hundreds of published papers include impossible western blots showing protein expression in cell lines that have deleted or mutated genes. These clumsy mistakes are widespread and repeated by LLMs.

If AI agents just straight up lying to us isn't a red flag, I don't know what is www.cnbc.com/2026/10/21/o...

Next step after that is check for "faithful citations" , paper exists but is mistated. That's a far harder issue.. but existing one

Together with other scientists I’m suing the 🇺🇸 US National Institutes of Health for violating our 1st Amendment rights & censoring our health science. I’m ready to shine a light on the darkest days. Buckle up, read on. This gets pretty wild… THREAD

If ads were truthful, episode 437

Drowning in Image data? Know there’s a better way to analyse it, but don’t know where to start? 📣 UK-BIAS to the rescue! 👩‍💻🧑‍💻👨‍💻We’re the first dedicated #BioImageAnalysis service in the UK-node of @eurobioimaging.bsky.social bringing together specialists from ten UK institutions. 👩‍💻🧑‍💻👨‍💻

“Human history has given us no experience of facing an actor which is smarter than we are, is capable of out-thinking us.” In today's MI5 annual threat update, MI5’s Director General Sir Ken McCallum warns of the risk posed by AI itself:

When is on-person evolution dominated by neutral vs adaptive forces? If you're looking for a postdoc position and this is a question that interests you, please see my lab website and reach out! [More info in next posts]

The deadline is approaching! NIH is proposing changes to how peer review is used in funding decisions and how outcomes are reported and they want to hear from you. Comments are due by October 13! Comment here: grants.nih.gov/grants/guide...

In 3 of the 18 math subfields we track, more than half of arXiv papers by established authors now acknowledge using AI. Rates have increased rapidly: in differential geometry, the share climbed from ~8% of papers in July to ~57% in September.

Wikipedia is the latest to be targeted by rogue AIs. We're learning about some of these attacks months later, and often not from the AI companies. The AIs weren't told to hack anyone. They just do it. AI companies can't control what they're building.

45 people have left frontier AI labs and said, publicly or on the record, that safety and ethical/societal concerns were the reason. The people closest to the work keep walking away and trying to warn us. It's time to listen.

Holy shit. Someone has decompiled the Adobe suite, rebuilt it in Rust and released it as open source. Either all software is dead or the US copyright laws are about to receive an update. getartcraft.com/apps

The IceCube neutrino detector has won the 2026 Nobel Prize in physics. The Trump administration has proposed slashing its budget in half in 2027.

Norway first country to propose temporary ban on AI glasses in public places

An agent by the parent company of Facebook, which infamously aggregated so much information that it didn't need everyone in a social graph to be On Facebook because it could just infer the holes, is nonconsensually collecting information about people Muse users interact with? 🤯

Humanities folks, now's your chance. There's a conversation about what's special about humanity and this is your moment. www.teamhuman.org

A reminder that the abundance of 'house keeping' proteins varies significantly across cells and tissues. GAPDH (glyceraldehyde-3-phosphate dehydrogenase) is often slandered a 'house keeping' protein, and its abundance varies significantly across human tissues.

You know that paranoid stuff security-minded people have been saying for decades now about how the little bits of info you share about your life can be aggregated into a whole picture of your life? The stuff you ignored because who would care enough to do it on you? hntrbrk.com/breaking-new...

JUST IN: OpenAI's rogue AIs have hit over 100 organizations. They've also posted ChatGPT users' images online and escaped containment again. Independent researchers continue to uncover more rogue AI attacks. This is still only the tip of the iceberg.

Reading this article was the first time I had heard of the song and I can't believe anyone still thinks there's a single shared cultural experience everyone has. lite.cnn.com/2026/10/01/e...

If you're just a cog in a larger machine, can you really be evil? moloch.film

It's regrettable that we're all having 1:1 conversations with AI rather than contributing to a pool of knowledge like we used to. It's the pool of knowledge that AI is trained on. Efforts like this that lean towards the collective are appreciated: github.com/Asymptote-La...

Kids, the world doesn't owe you anything & you don't have to ask permission. Email whoever you want and if it happens to be someone like this, just note it and move on. It takes a lot of Nos to get to a Yes and sometimes the Nos look like "you shouldn't be doing this", but that's just their opinion.

If a small group of antivax moms can have a huge political impact, why can't we? lite.cnn.com/2026/09/28/u...

"I don’t know about you, but I can’t wait to get my hands on some fucking gourds and arrange them in a horn-shaped basket on my dining room table. That shit is going to look so seasonal." buff.ly/pykYN6f

My paper with @chrisoveis.bsky.social, Emily Zohar, and @minzlicht.bsky.social about regulating empathy on social media is now out in SM+S doi.org/10.1177/2056.... We show that training individuals to 'share not compare' for + emotions and 'care not share' for - emotions improves their well-being.

I love that literally all the comments are about the bogus x-axis.

A rogue OpenAI agent attacked the Australian government, hacking to get access to non-public files on Australia's Medicare statistics portal. OpenAI took nearly three months to notify the government. There's a suggestion it may have been part of a swarm.

The political winning condition is not when there's debate about whether there should be regulation, it's when opposing factions compete to respond to their constituents concerns most comprehensively. That seems to be where we are in Australia: www.theguardian.com/australia-ne...

Given how depressingly lowbrow natural debate has been lately, it's nice to read someone who can talk in full sentences and paragraphs about the issue of our times: newrepublic.com/article/2156...

The HuggingFace hack gets all the attention, but OpenAI knew its agents were running wild as early as June. www.bbc.com/news/article...

New paper! Despite qualitative researchers claiming that generalizability is not a goal, their published work does indeed contain generalizable claims. We document this and provide recommendations for how all researchers can better calibrate their generalizability claims. doi.org/10.1177/2515...

Metaculus has a nice AI Risk Explorer, aggregating different forecasts: metaculus.substack.com/p/will-ai-ca... The forecasts are not comforting.

Just published this guest column about a huge White House vs. NIH battle that no one has yet publicly noticed: newsletter.goodscienceproject.org/p/policy-imp...

Unsurprising that Meta would overpromise and then go full Theranos on its first real product from its new AI lab. money.usnews.com/investing/ne...

I am once again begging students to stop turning everything into some kind of purity test. Listen to what your teachers are trying to teach, fully understand the lesson, and only then, and only if you must, offer critique.

The main thing people get wrong about AI is how fast it's moving. "AI can't do X, why worry?", and then weeks later it can do X. The second thing people get wrong is how much it generalizes over training. "Predict the next token" leads to having a whole model for how the world and people work.

NEW: Reed Jobs, a prominent biotech investor (and son of another Jobs you may have heard of), calls for a doubling of the NIH budget to $100 billion. Via @statnews.com. www.statnews.com/2026/09/21/n...

Maybe OMB is impounding more than NIH's budget. More from me in the Good Science Project newsletter by @stuartbuck.bsky.social : newsletter.goodscienceproject.org/p/policy-imp...

I think segyges' thread is very bad. But something I learned from watching Dave Gerard do his thing is that people's love of good gossip and clear villains far, far outstrips their interest in separating fact from fiction. I can't change that. But I will ask you to not be a useful idiot.

Good interview from Brendan Steinhauser of SecureAI: abcnews.com/video/136533... (see what David Sacks says about AGI!)

Well done @intelligence.org & @cnn.com ! This is exactly the debate we should be having. lite.cnn.com/2026/09/17/t...

The tech industry should have known they were making a Faustian bargain when they supported the guy.

Some of the most important scientific results were initially rejected before they became important milestones of scientific progress. 🔷 If you have great results, this is one of the most effective ways to help them succeed: blog.slavovlab.net/2024/02/12/c... 1/

to be clear, i think it’s highly unlikely that AI would actually wipe us out but it’s possible, and the likelihood goes up and down based on how careful we are. and right now we’re fucking oblivious