We frequently get requests to "explain" incidents. Why they happened. What you can do to prevent them. Towards this, we are testing out a YouTube channel. Please let us know what you think! https://t.co/Fzn3wbS1JG

Independent AI assessment is more important than ever. At #NeurIPS2025, Transluce will help launch the AI Evaluator Forum, a new coalition of leading independent AI research organizations working in the public interest. Come learn more on Thurs 12/4 👇 https://t.co/5Nzf9E2SPV

I've noticed a trend of researchers who need other researchers have adopted the Klarna strategy of "agree to study now, execute study over time." Me: "Sure, you are only asking for 1 hour a week for the next 26 weeks, that's practically free!"

For the old school RL researchers. #MountainCar #ToyDomains https://t.co/oIJYNvj69O

Insurance is an underrated way to unlock secure AI progress. Insurers are incentivized to truthfully quantify and track risks: if they overstate risks, they get outcompeted; if they understate risks, their payouts bankrupt them. 1/9

Remember when public cell phone ringing was a constant annoyance? It took ~10 years for social norms, technology settings, and human operation to reasonably solve "ringing." AI now brings thousands of different forms of "ringing," each requiring cross-sectional effort.

Calling all AI incident researchers (and AI deployment researchers more generally)! Spend a few days in Singapore exploring the practicalities of real world AI deployment. https://t.co/fskQ5tFJrU

Britain and Canada supplied crucial elements of the early US nuclear weapons program. "...in July 1945 British permission required by the agreement was given for the use of nuclear weapons against Japan." https://t.co/K5bOFwZxqu #MiningWikipedia

Someone needs to run a study of whether klaxons make people safer depending on when they are received relative to local shaking. I jumped under my desk, but My phone gave the emergency earthquake alert right when the shaking started and it was a bit distracting.

Signal is excellent software, but people that use it can see how the wrong person could be added.

An instance in which software UX caused problems surrounding a (supposed) missile strike. https://t.co/Fps4vNMytS

As far as I can tell, people are applying the term "agentic AI" as a fancy way of saying "AI that does stuff." Everyone imagines that to be something slightly (or vastly) different in technological realization and capability.

Join us at IAAI-25 for an exciting program in Philadelphia on Feb 27th! The Thirty-Seventh Annual Conference on Innovative Applications of Artificial Intelligence (IAAI-25) will showcase successful applications and novel uses of AI, highlighting cutting-edge innovations.

Strange feeling: seeing a ~trillion dollars in market cap get wiped out because the wider world figured out things that were already known within your research community.

Health insurers know when you see a new specialist after appointments within a specialty. Dataset I want: medical specialty doctor churn. It would really help me figure out which podiatrists are good...

The models are answering a different question more common in training data. Add "total" to these queries and performance will improve by resolving the ambiguity. But agreed, models are not mechanically counting letters in the manner humans do. https://t.co/9iFYcHKCsQ

Defending the capacity for an anonymous human-centered internet from the deluge of bots.

I was just visiting NYC and searched* Google for the cocktails named for the city's 5 boroughs. I am enjoying the implication that "Cape Code" and "BLT" are city boroughs. *I was **searching** not **asking**. But I still got this fun answer.

This explains in simple terms my increasing alarm for how we are building the future.

I have a huge podcast backlog that Google is helpfully eliminating by sunsetting Google Podcasts today. Definitely not porting my listen list over to YouTube as suggested.