Software Engineering with LLMs in 2025: Reality Check (Category AI)
I watched an interesting talk by Gergely Orosz, engineer, author of The Pragmatic Engineer newsletter on Substack and author of The Software Engineer’s Guidebook. He gave it at LDX3 (LeadDev London) on 16 June 2025, the largest engineering-leadership festival, with 2000 attendees. I recently covered LeadDev’s Engineering Leadership Report 2025 too. Returning to the talk, here are its key ideas.
1. Orosz conducted a qualitative study to understand the reality of generative AI. He saw a large gap between optimistic claims from major-company CEOs—Microsoft’s CEO says AI writes 30% of code; Anthropic’s CEO predicts 90% within a year, in a talk I covered—and practical problems, including expensive AI-tool mistakes and failed public demos. Orosz therefore interviewed engineers personally across different settings.
Startups building AI tools
- Anthropic engineers say 90% of Claude Code was written using Claude Code itself.
- Windsurf engineers report 95% of code written with their tools’ assistance.
- Cursor engineers give more cautious estimates, around a 40-50% success rate.
Big Tech companies
- Google uses its own Cider IDE with integrated LLM tools. Its SREs are preparing for a 10-fold increase in production lines of code by expanding infrastructure and review tooling.
- Almost all Amazon developers use Amazon Q Developer Pro, especially for AWS work. Amazon is becoming an MCP pioneer, with most internal tools already supporting it and enabling many workflows to be automated.
AI startups that do not build the tooling itself The results are mixed. For example, incident.io actively uses AI to speed up teamwork and shares practices in Slack. A biotech startup, however, reported that LLMs had not caught on: writing correct code themselves was faster than checking and fixing AI-generated code.
Experienced independent software engineers
- Armin Ronacher, creator of Flask, feels excited again after 17 years of coding thanks to Claude Code, and believes “AI changes everything.”
- Peter Steinberger, creator of PSPDFKit, says languages and frameworks matter less thanks to AI tools.
- Simon Willison, a creator of Django, confirms that coding agents work and that improvements in models over the past 6 months have been a turning point.
Industry veterans’ views
- Martin Fowler compares LLMs’ arrival to the move from assembly to high-level languages: a similarly revolutionary productivity gain, but this time, for the first time, with nondeterministic tools.
- Kent Beck says he enjoys programming more than at any other time in 52 years. He compares LLMs’ impact with microprocessors in the 1970s, the internet in the 2000s and smartphones in the 2010s.
For survey statistics, DX—whose Measuring AI Code Assistants and Agents report I recently discussed—finds:
- On average, 50% of developers in organizations use AI tools weekly, rising to 60% in leading companies.
- Time savings are 3-5 hours per week, well short of claims of 10-20x improvements.
These survey findings differ from the enthusiasm of major-company CEOs. Experienced independent engineers do seem excited about the tools, though. Orosz believes current AI tools work better for individual developers than for teams inside large companies.
He closes by discussing major changes in software development and argues that the industry should experiment more, as startups do. His central recommendation: try what works, understand what has become cheap and what remains expensive, and adapt to the changing technology landscape.
#AI #ML #PlatformEngineering #Software #Architecture #Processes #DevEx #Devops