Microsoft’s CEO said at Meta’s Llamacon that up to 30% of the company’s software code is currently generated by artificial intelligence tools.
Meta bets appear to be scale first and optimize later. The company plans to invest between $6-65 billion in 2025 to strengthen its AI infrastructure. As mentioned in our 2025 revenue release, we have increased our capital expression guidance for 2025 from $645 billion to $645 billion. Zuckerberg did not offer point-in-time metrics comparable to Microsoft, but predicted a rapid increase in AI contributions for Meta’s own AI development.
As Zuckerberg detailed in a recent interview with Dwarkesh Patel, this stems from a significant internal effort. “We have important coding efforts and develop coding agents within the meta… Specifically, coding and AI research agents have gone through Llama’s research and have been fully integrated into the toolchain. The average very good guy on the team.”
An update to Copilot Telemetry in February 2023 showed that developers would accept AI proposals of 46% of all lines of code between languages and 61% in Java. This trend continues to accelerate, with GitLab’s 2024 DevSecops survey revealing that 78% of 5,000 developers will use or plan to use AI in their code within two years. Github Blog, February 2023
In a related note, Salesforce CEO Marc Benioff told analysts in February that the AI tools had increased engineering output by about 30%. Results: There will be no new software engineers hire in 2025. Gartner currently forecasts that 75% of enterprise software engineers will use AI coding assistants by 2028, up from under 10% in early 2024.
Evaluating the true capabilities of these models remains difficult. Public benchmarks can be misleading, as highlighted when Meta caused a mini-backlash after submitting a hand-tuned “Maverick-03-26 Experimental” build. The company later opened sourced different, less-performing sibling models. This led to the benchmark maintainer (LMSYS) closing the submission rules as stated in the register. Zuckerberg acknowledges the difficulties, saying that inside Meta, “Northstar is the product value that users report.” It’s so optimised for benchmarks like Chatbot Arena that “can lead us to be confused by the quality of the product, and such benchmarks “skew to specific use cases that are not typical of the average user.”
In a recent podcast, Microsoft CTO Kevin Scott said that “95%” code could be generated by AI by 2030, and that humans could move towards architecture and intention. However, Scott argued that instead of replacing engineers, he would raise the role of human engineers to a higher level of abstraction (author, design). Despite this optimism, Microsoft’s research study (as well as the discovery of the joint Microsoft-Meta Paper MSR Paper) solved less than 50% of the debug tasks in the SWE Bench Lite benchmark, highlighting the remaining gaps compared to human engineers.
“Even if the code is writing itself, you still need to pour concrete and string fiber. Building a gigawatt scale data center takes time.” – Mark Zuckerberg Dwarkesh Patel Podcast
Scott also pointed out that AI forces tech companies to rethink their products, bringing out similarities with early and disruptive technologies. “So I think it was very confused in the early days of the internet, and in the early days of mobile, when everyone had ideas about what was worth, and most of those ideas were actually durable,” he said. Meanwhile, Amazon’s Q Developer Agent has migrated 30,000 Java applications in-house. This is a move that CEO Andy Jassy will save $260 million a year.
If you have an entrepreneurial spirit, this is the perfect time to live. – Scott
It’s not accurate how Google and Microsoft CEOs are counted, but both say that around 30% of the current code is writing AI. Chart data raised from TechCrunch, April 29, 2025, Alphabet Revenue Call, April 24, 2025 (Business Insider), Github Blog, February 2023, Stack Overflow Dev Survey 2024
