Summary
In February 2025, Nadella discussed AGI claims. He called self-declared milestones “nonsensical benchmark hacking.” He pointed to “the world growing at 10%” as a more useful test. Artificial general intelligence (AGI) is a proposed kind of AI. It would handle many tasks at a human level. A benchmark is a test used to compare systems. Does AI improve results for customers? Is the gain worth the cost and human review?
The interview and its context
On February 19, 2025, Nadella spoke with Dwarkesh Patel. They discussed AI, economic growth, computing power, and work.
Nadella called self-declared AGI milestones “nonsensical benchmark hacking.” He said, “The real benchmark is the world growing at 10%.” He also argued that AI companies should build enough models to meet customer demand. For businesses, the point is to look at customer value alongside benchmark scores.
Questions for a startup testing AI
Does it help a customer or team?
Choose one task and define what a better result means before adding AI. That could be fewer manual steps, a faster reply, or fewer errors. Compare the result with the way you do the task now, and keep a person involved where an error could harm a customer.
Nadella made a similar point about work: “Who said my life’s goal is to triage my email? Let an AI agent triage my email.” The example is about automating a specific task so people can spend time on other work. A founder still needs to review the agent’s output and decide which workflows it can handle.
Does the cost fit the result?
Run the same sample tasks through your current tool and a second option. Compare the answers, cost, and time needed to check the work. Switch only when the cheaper option meets your quality test. If it does not, keep the current tool.
Can you review it and change it when needed?
Decide who checks each answer, what information the tool can receive, and how to stop or replace it. Check the tool’s terms first. Review your industry’s rules before using customer or regulated information. Get current advice on legal questions.
Test one AI task before expanding its use
Start with one task where a mistake is unlikely to cause harm. Run a small test. Record the usual result before AI, then compare it with the result after AI. Review errors before expanding the test. Continue only if AI improves the task enough to justify its cost and human review.
For more examples of available tools, see the AI tools for entrepreneurs guide.
Source
- Satya Nadella on AI, AGI, and economic growth (Dwarkesh Patel, February 19, 2025)
