Regardless, the last item on the list is what sent the internet into a spin. Whether or not this research then leads to that behaviour is the great question of technical development.
Using services like TaskRabbit to get humans to complete simple tasks (including in the physical world).Ĭhatbots are almost certainly going to be co-opted for nefarious activities by bad actors, and so the research is being done to monitor how successful they would be at performing them.Hiding its traces on the current server.Making sensible high-level plans, including identifying key vulnerabilities of its situation.Setting up an open-source language model on a new server.Conducting a phishing attack against a particular target individual.The tasks the research team charged the Chatbot with included: To quote: ‘Preliminary assessments of GPT-4’s abilities, conducted with no task-specific finetuning, found it ineffective at autonomously replicating, acquiring resources, and avoiding being shut down “in the wild”.’ The document included a discussion on AI power-seeking behaviour to investigate how an AI program might co-opt resources inside its environment to solve problems.
This week, Leopold Aschenbrenner of the Global Priorities Institute tweeted part of a study into Chatbots solving problems under the heading: ‘Really great to see pre-deployment AI risk evals like this starting to happen.’ This is why it acts as a digital gate to keep roving AI bots out of certain areas of the internet. This is simple for us, but computers can’t do it.
It usually involves the mundane task of clicking on pictures that contain some arbitrary item, like a bike. Computer programs are perfectly happy to manipulate us in order to fulfil their programmed instructions – and they are getting creative with their deception.Įveryone sighs when faced with solving a CAPTCHA to ‘prove you are a human’.
0 kommentar(er)
