'People use AI to avoid paying for school photos, but they don't realise the consequences' Image caption, Photographer Ian Gilbert says he's heard parents boasting about how they use AI to take the watermark off to avoid paying for school photos A photographer is warning parents who use artificial intelligence to take the watermark off school pictures about the dangers of how companies can use images of their children. Ian Gilbert, founder of Emotion Picture Photography, said he had heard parents boasting at the school gates about how they use the tool to avoid paying for photos. The scary thing is when you submit a picture to AI by default it is allowed to do whatever it wants and I don't think a lot of parents realise that, he said. An AI safety expert said he understood why people did it, and that it may seem harmless, but uploading any identifiable information, including photographs, was not worth the risk. Gilbert, from Pontypool, Torfaen, urged parents to think carefully before uploading their child's photograph to an AI tool.

Reinforcement Learning from Human Feedback (RLHF) (Christiano et al., 2017) has rapidly become a standard method to align Large Language Models (LLMs). One of the main practical issues that all the prominent existing RLHF methods (offline or online) (Ouyang et al., 2022; Rafailov et al., 2023; Azar et al., 2023; Zhao et al., 2023b; Ahmadian et al., 2024) encounter is that their optimal solution heavily depends on the training task in terms of the distribution used to generate the preference data (behavior policy) (Munos et al., 2023; Azar et al., 2023). This makes the existing RLHF methods prone to out-of-distribution (OOD) tasks (Li et al., 2024; Kirk et al., 2024) where the evaluation distribution is significantly different from that of the behavior policy. Also, whenever the base/SFT models significantly differ from the behavior policy, the dependency of the RLHF solutions on the behavior policy makes the preference dataset and reward model less useful (Gao et al., 2022) as RLHF may undo the SFT/pretraining. To address this challenge, we introduce an alternative approach for aligning LLMs from human preferences based on more principled and robust foundations. Our goal is to find a solution that is robust to the changes in the preference dataset, meaning that changes in the distribution from which the completions are sampled do not affect the final outcome of learning significantly. To achieve this goal, we exploit the concept of self-improving (Huang et al., 2022; Bai et al., 2022) language models. By self-improving LLM we refer to a model capable of enhancing its outputs recursively with each inference iteration.

Mark Carney knows how to illustrate economic trends through the use of creative language. And when he talks, people tend to listen. "The massacre of the Dilberts" was how the governor of the Bank of England encapsulated the fear that middle-management jobs would be wiped out by automation – for people unfamiliar with American cartoon strips, Dilbert is a white collar office worker and the strip mocks the absurdities of office life. In his native Canada this week, Carney made a number of points in a speech on automation. Most obviously, many office jobs done by people would be done by computers, a process that was already well advanced. "When I look back 30 years ago, what I used to do in the City of London when I worked at an investment bank, probably about three-quarters of what I did is now done by machine," he said.