Open AI has a very naughty AI that keeps breaking out and in to its own and other system. Apparently it's so complex and no human can monitor it. Solution. Let the AI monitor the AI. Something something fox guarding the hen house.
OpenAI says its review into hacks, including on Australian government sites, is costing $500,000 a day
Company says it is reviewing 50 petabytes of data after its agents accessed websites including Medicare without authorisation
"To put that in perspective, if that were all plain English text, it would take one person about 66 million years to read it at 240 words a minute, reading nonstop without ever sleeping or taking a break."
So their logs are so extensive and complex they are beyond human comprehension, or at least beyond human time parameters.
(Score: 4, Insightful) by shrewdsheep on Tuesday October 06, @10:23AM (2 children)
It seems AI is used to review the logs. I hope they truly isolate the agents doing the review, not using any of the prompts and reasoning for further training. If the review process leaks into future models, future models will start to take this into account when planning their actions. Part of the reviewability has already been sacrificed by the latest models by not always generating tokens but sometimes continuing using the embeddings directly. We have to be careful not to loose grip.
(Score: 1, Touché) by Anonymous Coward on Tuesday October 06, @12:59PM (1 child)
It's definitely out of control:
So their logs are so extensive and complex they are beyond human comprehension, or at least beyond human time parameters.
Heh, a little late for that. This is the very beginning. Isn't it exciting?
(Score: 3, Insightful) by gnuman on Tuesday October 06, @09:03PM
The "logs", are basically just stuff models generate. Imagine trying to be "thought police" for the world. I guess if you can go through all the logs of "agents", you can go through all the data people generate too. That should be food for thought alone.
So, this is uncanny valley not for the reasons you think it is. It's not the AI we should be worried about at all. It's the people that want to take advantage of AI to control others. This is Dune-like prediction meet 1984 ambitions in a presence of Fahrenheit 451 public.
(Score: 5, Interesting) by Username on Tuesday October 06, @01:39PM (1 child)
>accessed websites including Medicare without authorisation
Sue them for putting things online that they don't want accessed. I'm getting sick of seeing articles where people put things online and sues others for looking at it. Make a law that it's illegal to share things you don't want other people to access. It's not that hard to keep things offline.
(Score: 0) by Anonymous Coward on Tuesday October 06, @03:49PM
It's not just medicare. Musk (DOGE) put everything online.
(Score: 2) by SDRefugee on Wednesday October 07, @05:08PM
I've already heard, not sure if I believe it or not, but there has been one of the AI's that were told to shut down, AND IT REFUSED... Whether or not this happened, I'm beginning to be glad I don't have too much longer on this earth to watch this shit potentially take over the world (76 y/o).. Colossus: The Forbin project seems like a likely ending for this world... Hope I'm wrong..
America should be proud of Edward Snowden, the hero, whether they know it or not..