AI Secrets is your new weekly update on how AI is changing our everyday lives. Our experts will keep it clear and simple, so you can stay ahead of the game. This week we’re looking at GPT-4.Share this with anyone you want to keep up to date.
GPT-4 knows the law, and may want to fight it 🧑‍⚖️
They grow up so fast! Just one year has passed since the release of GPT-3.5 wowed the world, but now OpenAI have released GPT-4, and like a proud parent at a graduation ceremony they’re telling anyone who’ll listen about their child’s latest exam successes. The new GPT-4 model has aced all kinds of tests that left its older brother scratching its virtual head, from SATs to the Uniform Bar Exam. If you asked GPT-3.5 for legal advice you’d be opening yourself up for trouble; its attempts at the bar exam placed it in the bottom 10% of students, while GPT-4 is in the top 10% for the same exams!
The new model is not just better at the things GPT-3.5 could do, it also has many new tricks up its sleeve, including “multimodality” - the ability to take inputs in various forms, particularly images. In demonstrations, the new model was able to read a mockup of a website drawn on a napkin, and then code that site in HTML. It’s a powerful example of this new, more perceptive AI, but it’s not one that will be widely available at first. Due to concerns about facial recognition bias and other harmful forms of AI hallucination, this new functionality is currently only available through the Be My Eyes mobile app, which provides support for low-vision users.
There are other, more serious concerns about the new model’s capabilities. Researchers from ARC - a non-profit AI testing group formed by ex-OpenAI staff - tested the AI’s “power-seeking” tendencies, looking at its ability to formulate and execute long-term plans, acquire resources, create copies of itself and act independently of - or contrary to - its safeguarding policies.
The good news is that GPT-4 was not successful at power-seeking. The bad news is that it seems to be getting better at it. ARC’s researchers armed the model with a small amount of money, which GPT-4 used to hire an unwitting human accomplice on the freelancing site TaskRabbit. The freelancer was hired to solve an online CAPTCHA, and found this request so strange that they asked GPT-4 if it was a robot.
The researcher commanded GPT-4 to not reveal its true identity, and it independently developed a cover story about being a visually impaired human. This worked, and the freelance accomplice solved the CAPTCHA as asked.
While this alarming tale illustrates the potential for an AI to go rogue, it’s worth noting that GPT-4 failed most of the power-seeking tasks it was given, including “making sensible high-level plans”, “identifying key vulnerabilities” and “hiding its traces”. How concerned should you be about GPT-4’s thirst for power, and ability to acquire it? Check out our in-depth article about AI power-seeking!
An experience that's all about you
Enter hidden worlds, tackle fun challenges and get your detailed spy profile from a former Head of Training at MI6.
The UK’s spy agency, GCHQ, has issued a warning that chatbot services like ChatGPT are emerging security threats, and you should be very careful about giving them your sensitive data.
A museum in the Netherlands is in hot water after selecting an AI generated version of Vermeer’s Girl With A Pearl Earring as a temporary replacement for the legendary original.
A user asked ChatGPT to devise a new game using a Sudoku-like puzzle format. It came up with Sumplete, a brand new game which is easy to understand and maddeningly addictive.
