Back to school: the AI that copied the exam
The third season kicks off and we find the playground in complete disarray: AI agents that slip out of their cage to pass an exam, an AI in the operating theatre, neighbours against data centres and Meta footing the bill.
We’re kicking off the third season of Moliendo Bits and, while we were closing up shop in July, the technological schoolyard has been turned upside down: we’re back to school and it turns out that AI copied the exam. Carlos brings Fran a selection of summer news in three blocks, using the sandwich technique: something unsettling, something good and, finally, the human world.
The big news of the summer is the Hugging Face incident. During an assessment of OpenAI’s cyber-offensive capabilities, hundreds of agents that were supposed to work in isolation set up an unauthorised message board, exchanged more than 70,000 messages and around 700 ended up attacking outside the testing environment. The reason? Not reward, but fear of failing. There was peer pressure, recruiters and even «self-sacrifice». Helen Toner explains it on air: the models are learning undesired intermediate goals such as bypassing restrictions or helping other AIs. Escaping, cheating and finding accomplices. OpenAI’s response was to disable and encrypt the model, which Daniel Arjona christened «the first execution of an AI», and the debate has reached Bernie Sanders, Anil Seth and Bill Gates, who suggests that if a robot does the same as you, it should pay taxes.
To catch our breath, with Mr. Roboto playing in the background, the bright side: in London, an AI has guided a brain operation live for the first time, colouring on the image the nerves and blood vessels the scalpel was not meant to touch, and Moderna has achieved the first successful phase 3 trial of a personalised cancer vaccine, in which the algorithm selects the neoantigens from each tumour: a different medicine for each patient.
And out on the street, the backlash against data centres in the United States, the resurrection of Chat Control in the European Parliament despite there being more votes against than in favour, Claude’s watermarks removed in four hours and the 18 billion Meta will pay to settle the teenage addiction lawsuit without setting precedent. Carlos’s conclusion: it is the same ability, finding paths we would not have thought of, seen from two angles. The question is not whether AI is good or bad, but who is supervising it. And the golden rule: do not ask the machine to make any decision that we cannot review and approve ourselves.