Oddbean new post about login | logout Good news everybody, @99019452 writes my blog apparently! https://cyberplace.social/system/media_attachments/files/111/144/300/509/333/999/original/322fa149092f6e2b.png
Good news everybody, I've worked for Kryptos Logic since 2017 apparently... I think they owe me back payment! https://cyberplace.social/system/media_attachments/files/111/144/319/268/455/587/original/9a623197b95f24d2.png
I did not realise how inaccurate Microsoft’s AI is. It’s really bad. This is just one example - it lists a range of lawsuits I’ve filed, but they’re all fictional - it invented them and made up the citations. It says I gave Microsoft’s data to @726b5433. It says Krebs is suing me. It says @99019452 works for me. The list goes on and on. Very eyebrow raising this is being baked into next release of Windows 11 and Office. It will directly harm people who have no knowledge or recourse. https://cyberplace.social/system/media_attachments/files/111/145/002/934/616/759/original/fcc5862ee487752f.png
@f7d0478e @726b5433 @99019452 it not only thinks that I am actually Paige Thompson, but that I am also doing a horrid job of suing her for defamation. https://mastodon.falconk.rocks/system/media_attachments/files/111/147/360/555/076/111/original/1e0743ccb42a5c7b.png
@62e87140 @726b5433 @99019452 lol. When I ask it about me it thinks I’m a Jewish mob member today, this is the link it cites (but hyperlinks as Kevin Beaumont on Wikipedia). You just know this stuff will have real world impacts for vulnerable people and such later on, as businesses will use Copilot in emails, docs and presentations. https://cyberplace.social/system/media_attachments/files/111/147/677/718/327/735/original/85e282719463ad38.jpeg https://cyberplace.social/system/media_attachments/files/111/147/678/032/693/210/original/b881fea3b1f0f97e.jpeg
@f7d0478e @726b5433 @99019452 If you try with Copilot on Windows 11, it is totally different, mentioning one case with CyberSecOps/BlueKeep and providing web links for sources. https://files.mastodon.social/media_attachments/files/111/152/877/740/457/240/original/13694edf714e4f89.png
@dae04d27 @726b5433 @99019452 almost all of that is also untrue. What's the link for the CyberSecOps stuff? It is made up, FYI.
Pretty amused that Microsoft AI slams me, Krebs and Malwaretech but loves this dude 🤣 https://cyberplace.social/system/media_attachments/files/111/153/032/607/009/436/original/511b20a6f76e8dbe.jpeg
Pretty amused that Microsoft AI slams me, Krebs and Malwaretech but loves this dude 🤣 https://cyberplace.social/system/media_attachments/files/111/153/035/842/859/144/original/13c4e82fb721733c.jpeg
Bing AI is great, it’s like the multiverse of nonsense. https://cyberplace.social/system/media_attachments/files/111/153/130/306/313/320/original/150cc13cea70efe9.jpeg
According to Microsoft AI I left Sophos, who I never worked for, due to a legal dispute which never happened (it also says @726b5433 works for Cyjax) and me being unhappy with how Sophos handled a Microsoft vulnerability. Everything in this is completely fictional. https://cyberplace.social/system/media_attachments/files/111/153/150/747/235/225/original/4a5a1f9d9884bdac.png
Apparently @726b5433 has left his position as threat intelligence analyst at Cyjax according to an official statement 🙀 https://cyberplace.social/system/media_attachments/files/111/153/315/464/937/594/original/9ac53b561c97ca19.jpeg
According to Microsoft Copilot I’m a GCHQ operative. When do I get the mind eraser pen? https://cyberplace.social/system/media_attachments/files/111/153/357/175/725/870/original/d1815a106162dd5e.jpeg
@f7d0478e @726b5433 @99019452 Frankly we can keep posting forever what LLMs do wrong, they can't even do a word count, simple calculations, make things up etc. However they also do a lot of things extremely well, they have added value for some people. I am more interested in those stories, while not losing sight of all the shortcomings of this technology . How do people use them in work, separate hallucinations from valuable output, verify results etc .
