About this video

• Video Title: Why language models hallucinate, revisiting Amodei’s code prediction and AI in the job market
• Channel: IBM Technology
• Speakers: Tim Hwang, Skyler Speakman, Chris Hay, Kate Sol, Ayli McConnen
• Duration: 00:42:25

Overview

This video discusses the phenomenon of "hallucinations" in language models, examining the reasons behind them and potential solutions. It also revisits a past prediction about AI's role in code generation and explores the impact of AI on the job market, particularly in recruiting. Finally, it touches upon the potential for micro-scale AI models.

Key takeaways

• Hallucinations in LLMs: The paper "Why Language Models Hallucinate" suggests that hallucinations may stem from the models' training, where they are incentivized to guess and are rewarded for accuracy rather than admitting uncertainty. This leads to a need for better calibration between accuracy and uncertainty in model training.
• AI and Code Prediction: Dario Amodei's prediction that AI would write 90% of code within six months is revisited. While AI is assisting in code generation, the focus has shifted towards augmentation rather than complete automation, and societal adoption of these tools is still catching up.
• AI in the Job Market: The increasing use of AI in recruiting and job applications has created a challenging environment for job seekers. The article "The Job Market is Hell" highlights an "echo chamber" effect of AI inputs and outputs, emphasizing the importance of personal networks and genuine demonstration of skills.
• Micro-scale AI Models: Advancements in running LLMs on small devices, like a business card-sized circuit board, open possibilities for "micro micro" AI models. These could be integrated into everyday objects, offering specialized functions without constant internet connectivity, although widespread cloud connectivity might limit the need for disposable, localized models.
• Re-evaluating AI's Persona: There's a call to move away from treating AI as "little people" and to focus on their capabilities as computational tools with clear inputs and outputs, leading to more practical and scalable enterprise implementations.

Language models are incentivized to guess because their training often rewards them more for attempting an answer, even if it's incorrect, rather than saying "I don't know." This is because in some evaluation metrics and reward functions used during training, admitting uncertainty guarantees zero points, while guessing carries a chance of receiving points if the answer is correct.