Movies are supposed to transport you places. At the end of last month, I was sitting in the Chinese Theater, one of the most iconic movie theaters in Hollywood, in the same complex where the Oscars are held. And as I was watching the movie, I found myself transported to the past, thinking about one of my biggest regrets. When I was in high school, I went to a theater to watch a screening of a movie one of my classmates had made. I was 14 years old, and I reviewed it for the school newspaper. I savaged the film’s special effects, which were done by hand with love and care by someone my own age, and were lightyears better than anything I could do. I had no idea what I was talking about, how special effects were made, or how to review a movie. The student who made the film rightfully hated me, and I have felt bad about what I wrote ever since.
So, 20 years later, I’m sitting in the Chinese Theater watching AI-generated movies in which the directors sometimes cannot make the characters consistently look the same, or make audio sync with lips in a natural-seeming way, and I am thinking about the emotions these films are giving me. The emotion that I feel most strongly is “guilt,” because I know there is no way to write about what I am watching without explaining that these are bad films, and I cannot believe that they are going to be imminently commercially released, and the people who made them are all sitting around me.
Then I remembered that I am not watching student films made with love by an enthusiastic high school student. I am watching films that were made for TCL, the largest TV manufacturer on Earth as part of a pilot program designed to normalize AI movies and TV shows for an audience that it plans to monetize explicitly with targeted advertising and whose internal data suggests that the people who watch its free television streaming network are too lazy to change the channel. I know this is the plan because TCL’s executives just told the audience that this is the plan. – Read More
Daily Archives: December 12, 2024
It’s Surprisingly Easy to Jailbreak LLM-Driven Robots
AI chatbots such as ChatGPT and other applications powered by large language models (LLMs) have exploded in popularity, leading a number of companies to explore LLM-driven robots. However, a new study now reveals an automated way to hack into such machines with 100 percent success. By circumventing safety guardrails, researchers could manipulate self-driving systems into colliding with pedestrians and robot dogs into hunting for harmful places to detonate bombs.
Essentially, LLMs are supercharged versions of the autocomplete feature that smartphones use to predict the rest of a word that a person is typing. LLMs trained to analyze to text, images, and audio can make personalized travel recommendations, devise recipes from a picture of a refrigerator’s contents, and help generate websites.
The extraordinary ability of LLMs to process text has spurred a number of companies to use the AI systems to help control robots through voice commands, translating prompts from users into code the robots can run. For instance, Boston Dynamics’ robot dog Spot, now integrated with OpenAI’s ChatGPT, can act as a tour guide. Figure’s humanoid robots and Unitree’s Go2 robot dog are similarly equipped with ChatGPT.
However, a group of scientists has recently identified a host of security vulnerabilities for LLMs. So-called jailbreaking attacks discover ways to develop prompts that can bypass LLM safeguards and fool the AI systems into generating unwanted content, such as instructions for building bombs, recipes for synthesizing illegal drugs, and guides for defrauding charities. — Read More
Banishing LLM Hallucinations Requires Rethinking Generalization
Despite their powerful chat, coding, and reasoning abilities, Large Language Models (LLMs) frequently hallucinate. Conventional wisdom suggests that hallucinations are a consequence of a balance between creativity and factuality, which can be mitigated, but not eliminated, by grounding the LLM in external knowledge sources. Through extensive systematic experiments, we show that these traditional approaches fail to explain why LLMs hallucinate in practice. Specifically, we show that LLMs augmented with a massive Mixture of Memory Experts (MoME) can easily memorize large datasets of random numbers. We corroborate these experimental findings with a theoretical construction showing that simple neural networks trained to predict the next token hallucinate when the training loss is above a threshold as it usually does in practice when training on internet scale data. We interpret our findings by comparing against traditional retrieval methods for mitigating hallucinations. We use our findings to design a first generation model for removing hallucinations — Lamini-1 — that stores facts in a massive mixture of millions of memory experts that are retrieved dynamically. — Read More
Rick's Cafe AI 