Retweeting a beautiful 1950s piece on speech research in Edinburgh (fixed links). More on the history: More about BBC's "eye on research": The program itself (well worth 30 minutes):

Image models struggle to create Go boards. Sadly ChatGPT can't currently copy one either, it seems conditioning the model is through too small a bottleneck. "Please copy this image exactly to create a new image. Every line and stone should appear in exactly the same places."

Join millions who have switched to Grok.

I'm really looking forward to having Slava in Edinburgh. There's an exciting opportunity to join his group as a machine learning PhD student:.

Some funding available to support probabilistic AI activity in the UK. For events, public engagement, visits, and 1-year bridging postdocs:

RT @KyleCranmer: Big congrats and thank you to the growing group of developers contributing to the SBI (simulation-based inference) package….

Five 3-year postdoc positions in probabilistic AI across the UK.

One could write a more complicated extension that lets redirects be configured and enabled after installation, and only asks for any new hosts permissions then. But I haven't found the time to do that so far.

Redirects can be used for evil, so still need a powerful "hosts" permission. If my developer account were hacked, the extension could be altered to steal your data on those websites. The extension settings page lists the "hosts" permissions, and you can disable any of them.

Here's why I had to update: The Chrome version of the extension now has no javascript, just some simple redirects encoded in a static .json file, but it seems I still have to ask for scary-sounding permissions.

Because Google, I had to rewrite my "redirectify" Chrome extension that redirects pdf links to html pages. It will auto-disable and you'll have to accept new permissions before being able to use it again. Sorry! (Boring details in 🧵).

RT @oisinmacaodha: Interested in starting a PhD in Machine Learning? Consider reaching out to Nikolay Malkin who will be joining the School….

Given observations of a function with variable+unknown noise, it's hard to know if the function is wiggly or the observations are just noisy. We need multiple observations at some inputs to disambiguate. This paper uses that intuition but goes a lot further. Looks interesting!.

RT @oisinmacaodha: Reminder about the open machine learning faculty position at the University of Edinburgh. Closing date is one week from….

RT @MaikeZufle: Last week, I presented my thesis at @GenBench #EMNLP2023 - the last step of my Masters journey at @EdinburghUni. I feel hon….

RT @tetraduzione: At @InfAtEd we are looking for the Director of our brand new center for #GenerativeAI #GAIL. come chat at #NeurIPS2023 if….

I'm sure Gatsby don't really need the promotion, but it would be an excellent place to be ML faculty. (I did my PhD there; it's a great environment.).

The University of Edinburgh is looking for someone to direct a new generative AI lab. It's a professor role, with lots of resources. More info: For earlier-career faculty jobs (and a communications role) see:

I got nerd-swiped by this nice multiple choice question by @yaroslavvb. Explanations of the distractors are in a follow-up thread:

Come join us in Edinburgh! This thread, from one of my fantastic colleagues, advertises our machine learning faculty position. (The photo is across the core atrium of our main research building, the Informatics Forum.).

Congratulations to @timudk! It's been a pleasure to regularly hear about his research over the last few years.

Does ChatGPT get this difficult case wrong too? Yes. If asked to be careful, or given hints, does ChatGPT give any indication that there's a problem? Usually not! . However, careful prompting and follow-up discussion, or asking in Chinese can work: