Key to R1's success was distillation, a technique that makes AI models more efficient. This works by having a larger model train a smaller model: you run the teacher model on a bunch of examples and record the answers, and reward the student model for copying those answers as closely as possible so that it gets a condensed version of the teacher's knowledge. — Caiwei Chen
10. Sycophancy
As people around the world spend more and more time interaction with chatbots As with ChatGPT, chatbot manufacturers are struggling to determine what tone and “personality” models should adopt. Back in April, OpenAI admitted that it had struck the wrong balance between usefulness and whining, saying that a new update had rendered GPT-4o. too sycophantic. Not only is it annoying if you don't like it, it can mislead users, reinforcing their wrong beliefs and spreading misinformation. So consider this your reminder to take everything—yes, everything—that LLMs produce with a grain of salt. —Rhiannon Williams
11. Slop
If there's one AI-related term that has completely broken out of the nerd fence and entered the public consciousness, it's “sucks.” The word itself is obsolete (think pig feed), but “slop” is now commonly used to refer to low-effort, mass-produced content generated by artificial intelligence and often optimized for online traffic. Many people even use it as a shorthand for any AI-generated content. Last year it seemed inevitable: we marinated in it with fake biographies To shrimp jesus images for surreal human-animal hybrid video.
But people also enjoy it. The term's sardonic flexibility has made it easy for Internet users to use it in a wide variety of words as a suffix to describe something devoid of substance and absurdly mediocre: for example, “worker's slop” or “druze slop.” As the hype cycle resumes, “sucks” means a cultural reckoning about what we trust, what we value as creative work, and what it means to be surrounded by things designed for interaction rather than self-expression. — Caiwei Chen
12. Physical intelligence
Have you encountered a hypnotic video from earlier this year, when a humanoid robot was clearing away dishes in a dull, gray kitchen? This pretty much embodies the idea of physical intelligence: the idea that advances in artificial intelligence can help robots move better in the physical world.
It is true that robots have been able to learn new tasks. Faster than ever before, everywhere from operating rooms To warehouses. Self-driving car companies have seen improvements in how they… simulate roads too. However, it is still reasonable to be skeptical that AI has revolutionized the field. Suppose, for example, that many robots advertised as butlers in your home perform most of their tasks thanks to remote operators in the Philippines.
The path to developing physical intelligence is also likely to be strange. Large language models learn from text, which is abundant on the Internet, but robots learn more from videos of people doing things. That's why robot company Fig suggested in September that it pay people film themselves in their apartments doing household chores. Would you subscribe? —James O'Donnell
13. Fair Use
AI models are trained by studying millions of words and images on the Internet, including copyrighted works by artists and writers. AI companies say it's “fair use,” a legal doctrine that allows you to use copyrighted material without permission as long as you transform it into something new that doesn't compete with the original. The courts are starting to do their part. In June, Anthropic's training of its artificial intelligence model Claude on a library of books was found to be fair because the technology was “extremely transformative”
That same month Meta dialed similar victorybut only because the authors could not prove that the company's literary buffet was cutting into their salaries. As copyright battles brew, some authors are cashing in on the holiday. In December, Disney signed big deal with OpenAI to allow users of Sora, an AI video platform, to create videos featuring more than 200 characters from Disney franchises. Meanwhile, governments around the world rewriting copyright rules for content-eating machines. Is teaching an AI to work on copyrighted work a fair use? As with any billion dollar legal matter, it depends. — Michelle Kim
14. GEO
Just a few years ago, an entire industry was built around improving the ranking of websites in search results (well, only in Google). Now, search engine optimization (SEO) is giving way to GEO—generative engine optimization—as the AI boom has brands and companies scrambling to maximize their visibility through AI, be it in AI-enhanced search results like Google. AI Reviews or in responses from LLM. No wonder they are scared. We already know what news companies have been through. colossal drop V search web trafficand artificial intelligence companies are working on ways to cut out the middlemen and allow users to visit websites directly from their platforms. It's time to adapt or die. — Rhiannon Williams
