0:00 / 0:00
research
Every AI output is a hallucination
Weights & Biases19 June 2026Watch on YouTube
Description
"My mission is to build technologies that can't lie." That's Dan Klein on this week's Gradient Dissent. Dan is a Berkeley computer science professor now building Scaled Cognition, a company designed around one bet: that reliability is the part of intelligence that's been left behind. Lukas and Dan get into why ChatGPT is always confident even when it's wrong and how Dan is working to fix this at the foundation.
What you'll learn
- AI hallucinations result from fundamental design problems in current AI systems, not technical glitches
- Models lack reliability because they are inherently overconfident in their answers regardless of actual accuracy
- Reliability is a core component of intelligence that has been neglected in foundation model development
Frequently asked questions
Why does ChatGPT always give confident answers, even when it's wrong?
ChatGPT and similar models are designed in ways that cause them to always respond with confidence, even when hallucinating or being inaccurate. This is an inherent characteristic of how these systems work, not something easily fixed afterward.
What is Scaled Cognition and what does the company aim to achieve?
Scaled Cognition is a company founded by Dan Klein focused on building AI technologies that are reliable and cannot lie. The company is built on the assumption that reliability is the key component of intelligence that has been overlooked so far.
How can AI systems become more reliable and less overconfident?
Dan Klein is working on approaches that address how AI models work at a fundamental level. The goal is to build systems that can express their uncertainty and are not automatically overconfident in their responses.