Bill Gates on AI Metacognition - The Key to Superintelligence

Writing about AI has given me a new appreciation for the complexity of human brains. While large language models (LLMs) are impressive, they lack the multidimensional thinking humans take for granted. Bill Gates touched on this last week on the Next Big Idea Club podcast. Speaking to host Rufus Griscom, Gates discussed “metacognition,” the ability to think about one's own thinking. Gates defined metacognition as the capacity to “think about a problem in a broad sense and step back and say, Okay, how important is this to answer? How could I check my answer, and what external tools would help me with this?”

Gates noted that the overall “cognitive strategy” of existing LLMs like GPT-4 or Llama still lacks sophistication. “It’s just generating through constant computation each token and sequence, and it’s mind-blowing that that works at all,” Gates said. “It does not step back like a human and think, Okay, I’m gonna write this paper and here’s what I want to cover; okay, I’ll put some text in here, and here’s what I want to do for the summary.”

Gates believes that AI researchers’ current method of improving LLMs—supersizing their training data and compute power—will only achieve a few more significant advancements. After that, researchers will need to employ metacognition strategies to teach AI models to think smarter, not harder.

Metacognition research could be key to improving LLMs’ reliability and accuracy, Gates said. “This technology will reach superhuman levels; we’re not there today, if you put in the reliability constraint,” he said. “A lot of the new work is adding a level of metacognition that, done properly, will solve the sort of erratic nature of the genius.”

How The Supreme Court’s Landmark Chevron Ruling Will Affect Tech and AI

The implications of the Supreme Court’s Chevron decision on Friday are becoming clearer this week, including its impact on the future of AI. In Loper Bright v. Raimondo, the court reversed the “Chevron Doctrine,” which required courts to respect federal agencies’ (reasonable) interpretations of regulations that don’t directly address the issue at the center of a dispute. SCOTUS decided that the judiciary is better equipped (and perhaps less politically motivated) than executive branch agencies to fill in the legal ambiguities of laws passed by Congress. There may be some truth to that, but the counter-argument is that the agencies have years of subject matter and industry expertise, which enables them to interpret the intentions of Congress and settle disputes more effectively.

As Axios’s Scott Rosenberg points out, the removal of the Chevron Doctrine may make passing meaningful federal AI regulation much harder. Chevron allowed Congress to define regulations as sets of general directives, leaving it to experts at the agencies to define specific rules and settle disputes on a case-by-case basis at the implementation and enforcement level. Now, it’ll be on Congress to hash out the fine points of the law in advance, doing their best to anticipate future disputes. This might be especially difficult with a young and fast-moving industry like AI. In a post-Chevron world, if Congress passes AI regulation, the courts will interpret the law, even as the industry, technology, and players change rapidly.

There’s no guarantee that the courts will rise to the challenge. For instance, the high court’s decision to effectively punt on the constitutionality of Texas and Florida regulations governing social networks’ content moderation raises concerns. “Their unwillingness to resolve such disputes over social media—a well-established technology—is troubling given the rise of AI, which may present even thornier legal and Constitutional questions,” notes Dean Ball, an AI researcher at the Mercatus Center.

Figma’s New AI Feature Appears To Have Reproduced Apple Designs

The design app maker Figma has temporarily disabled its newly launched “Make Design” feature after a user found that the tool generates weather app UX designs that closely resemble Apple’s Weather app. Such close copying by generative AI models often suggests that the AI’s training data was insufficient in a particular area, causing it to rely too heavily on a single, recognizable piece of training data—in this case, Apple’s designs.

However, Figma CEO Dylan Field denies that his product was exposed to other app designs during its training. “As we have explained publicly, the feature uses off-the-shelf LLMs, combined with design systems we commissioned to be used by these models,” Field said on X. “The problem with this approach . . . is that variability is too low.”

Translation: The systems powering “Make Design” were insufficiently trained, but it wasn’t Figma’s fault.