Advanced AI gets more unpredictable

Catch up quick: OpenAI released the o3 and smaller o4-mini models a week ago and called them "the smartest models we've released to date."

• The company and early testers lauded o3 for its overall reasoning prowess — its ability to respond to a user prompt by planning, executing and explaining a series of planned steps.
• They also highlighted o3's reliability in conducting web searches and using other digital tools without constant user supervision or intervention.

o3 won praise from reviewers not only for bread-and-butter AI work like writing, drawing, calculating and coding but also for advances in vision capabilities.

• One popular — and, for privacy experts, potentially alarming — trick that went viral: Using o3 to look at virtually any digital photo and identify where it was taken.

What they're saying: "These models can run searches as part of the chain-of-thought reasoning process they use before producing their final answer. This turns out to be a huge deal," the developer Simon Willison wrote.

• "This is the biggest 'wow' moment I've had with a new OpenAI model since GPT-4," Every's Dan Shipper reported.
• Economist-blogger Tyler Cowen declared that o3 heralded the advent of AGI: "I think it is AGI, seriously. ... Benchmarks, benchmarks, blah blah blah. Maybe AGI is like porn — I know it when I see it. And I've seen it."

Yes, but: Plenty of reviewers found reasons to criticize o3, including math errors and deceptions.

• A study of models' performance in financial analysis placed o3 at the top of the heap — but it still only provided accurate results 48.3% of the time, and its cost-per-query was the highest by far, at $3.69. (The Washington Post has more on the study.)

Between the lines: Intriguingly, OpenAI notes that despite o3's impressive capabilities, it's actually regressing in some areas — like its tendency to "hallucinate," or make up incorrect answers.

• In one widely used accuracy benchmark test, OpenAI found that o3 hallucinates at more than twice the rate of its predecessor, o1.

• o3 also answers more questions — and gets more of them right — than o1. "More research is needed" to understand why o3's error rate jumped, OpenAI says.

Zoom out: AI analyst Ethan Mollick describes o3's impressive but scattershot performance as an example of "the jagged frontier": "In some tasks, AI is unreliable. In others, it is superhuman."

• Mollick argues that "the latest models represent something qualitatively different from what came before, whether or not we call it AGI. Their agentic properties, combined with their jagged capabilities, create a genuinely novel situation with few clear analogues."

Our thought bubble: Software makers and programmers have spent decades trying to make their work more reliable, scalable and flexible, and they've made plenty of progress.

• Making AI is newer, stranger and so far not well enough understood to be turned into a predictable discipline.

The bottom line: Designing, building and training AI models remains stubbornly resistant to developers' efforts to impose scientific rigor on their field or duplicate their results.

• Apparently, this process is still more like raising a kid than building a bridge.
• That adds to the sense of mystery and possibility surrounding AI development — but also frustrates efforts to domesticate it or harness it for economic advantage.