Michael Armstrong

@medialator.bsky.social

Continually in transition, so just as soon as I get round to writing this description it will be out of date. Currently wrangling python code to monitor TV subtitles. Also to be found at @medialator@mastodon.social

A message to trans + people in the UK. Carry on using the facilities you feel most comfortable in from tomorrow. Your comfort matters. Your right to exist in public life matters. The guidance is completely unworkable and it will fail!!

Haters focusing on Liddle's violence toward his pregnant partner, his galloping racism, or his repeated defences of paedophilia, should remember to mention he was also one of the worst writers to ever do it. A calorie-free hack of the first water. Just absolute dogshit, back to front.

I remember a paper from several years ago about classical RL agents learning to exploit bugs/limits in their training gyms to “solve” the problem. One was a little physics sim, the agent learned to exploit the sim’s floating-point error to fly across the world at impossible speed to meet the goal.

Timnit Gebru@timnitgebru.blacksky.app · 5d ago

We're in the era of incompetence & cybercrimes headlined as "unprecedented model capabilities gone rogue." So OpenAI & Anthropic are trying to one up each other with such incompetence because the "press release as a service" performing media & clueless politicians parrot pre IPO CEO talking points.

In fact, one of the hard things about training & evaluating ML systems, even without an LLM or agent in sight, is making sure the system is learning to solve the real problem, not exploiting quirks of your data set and/or evaluation. There are so many ways to open the door to “cheating”.