We expose a surprising failure of generalization in auto-regressive large language models (LLMs). If a model is trained on a sentence of the form “A is B”, it will not automatically generalize to the reverse direction “B is A”. This is the Reversal Curse. For instance, if a model is trained on “Olaf Scholz was the ninth Chancellor of Germany”, it will not automatically be able to answer the question, “Who was the ninth Chancellor of Germany?”. Moreover, the likelihood of the correct answer (“Olaf Scholz”) will not be higher than for a random name. Thus, models exhibit a basic failure of logical deduction and do not generalize a prevalent pattern in their training set (i.e. if “A is B” occurs, “B is A” is more likely to occur). We provide evidence for the Reversal Curse by finetuning GPT-3 and Llama-1 on fictitious statements such as “Uriah Hawthorne is the composer of ‘Abyssal Melodies'” and showing that they fail to correctly answer “Who composed ‘Abyssal Melodies?'”. The Reversal Curse is robust across model sizes and model families and is not alleviated by data augmentation. We also evaluate ChatGPT (GPT-3.5 and GPT-4) on questions about real-world celebrities, such as “Who is Tom Cruise’s mother? [A: Mary Lee Pfeiffer]” and the reverse “Who is Mary Lee Pfeiffer’s son?”. GPT-4 correctly answers questions like the former 79% of the time, compared to 33% for the latter. This shows a failure of logical deduction that we hypothesize is caused by the Reversal Curse. — Read More
Paper
Code is available at https://github.com/lukasberglund/reversal_curse.
Daily Archives: September 25, 2023
Everyone is above average
Congratulations – you are now above average!
It may sound like an old, bad statistics joke, but I mean it quite literally. We now have very strong evidence that AI elevates the skills of the lowest performers across a wide range of fields to, or even far above, what was previously average performance. — Read More
An NYPD security robot will be patrolling the Times Square subway station
The New York Police Department (NYPD) is implementing a new security measure at the Times Square subway station. It’s deploying a security robot to patrol the premises, which authorities say is meant to “keep you safe.” We’re not talking about a RoboCop-like machine or any human-like biped robot — the K5, which was made by California-based company Knightscope, looks like a massive version of R2-D2. Albert Fox Cahn, the executive director of privacy rights group Surveillance Technology Oversight Project, has a less flattering description for it, though, and told The New York Times that it’s like a “trash can on wheels.” — Read More
WTF is the fediverse?
Imagine posting a thought or piece of content on Threads and having followers from another platform, like Mastodon, like and comment on that post. It’s a possibility that’ll soon come to fruition if Meta keeps its promise to allow the nearly 10 million people who still use Threads to follow and interact with users on other platforms like Mastodon.
This is all thanks to the fediverse, or the federated universe, which is best described as a group of social media networks that are independent but able to communicate with one another. The fediverse’s origins reportedly date back to the early 2000s. But in with an increasingly fragmented social media landscape, Twitter’s decline, data privacy concerns and a colossal creator economy, the promise of cross-platform communication in the fediverse has sparked curiosity for some in the industry, especially with Meta’s plan to make Threads part of it. — Read More
Rick's Cafe AI 