This Wired article about Meta red-teaming the AI of others without their knowledge is mysterious but the subsequent BlueSky pearl clutching isn’t helpful. How do you think AI gets better? How will academics detect that AI isn’t safe? Through red teaming. www.wired.com/story/meta-c...

Fatality.

First day of class and I have the perfect lecture hall for Trust and Safety.

A lot of the characters you meet online will be powered by this model going forward. It doesn't seem to have any internal protections against generating political content (as long as you don't ask about specific issues contrary to the CCP position, see this lab leak question).

Deepseek-R1... 1) Is very impressive 2) The 32B version runs very well locally on a 4090 3) Will put a lot of pressure on the big US labs to open-source 4) Will be used in a lot of abuse/spam 5) Has some interesting holes in its knowledge:

If you want to see how well the war against bots is going on X, just mention that you have a Facebook account that was stolen.

I did my own research, Marcus, and you wouldn't believe how far the conspirarcy goes! I found that the skies are filled with these drones, and there is even a massive government agency overseeing it all! It goes to the top, Marcus, the very top!

Smartest model ever, still doesn't know who Jonathan Turley is. 🤔

New Bingo card entry dropped.

Me deciding to start posting on Bluesky after getting tired of Threads’ inscrutable algo and gas leak normie replies.