That story about a killer AI run amok seems fake. Here's my much nerdier & less dramatic story: I set up an AI agent to find advisers marketing tax avoidance schemes. The AI agent did this, then decided - entirely on its own - to inform HMRC of its findings.

Jun 2, 2023 · 12:42 PM UTC

It then spent about half an hour trying and failing to figure out the Twitter API, so it could inform HMRC. Didn't get anywhere largely because Twitter had closed its API - but as ChatGPT's information cut-off date was September 2021, it didn't know...
Replying to @DanNeidle
What prompt are you using here? The thoughts/reasoning labels are interesting.
AutoGPT. Apologies but can’t remember exact prompt. It was something like: I am investigating tax advisors selling tax avoidance schemes involving school fees, often involving “B shares”. Please find as many such such schemes as you can, and list the advisers and their schemes
Replying to @DanNeidle
The skynet we never knew we needed!
Replying to @DanNeidle
A wrapper about ChatGPT is not an agent and can pretty much (as you experience here) not really do anything, at least with present models and methods. But the fact that insofar as it tries, it will behave like ChatGPT, is not surprising in the slightest.
Replying to @DanNeidle
I heard from a Hollywood insider that this is the premise of the next Terminator film
Replying to @DanNeidle
Snitches get glitches
Replying to @DanNeidle
best use of AI i've seen yet