Software engineering with LLMs in 2025: reality check (AI column)
I saw something interesting. speech Gergely Orosz is an engineer and author of The Pragmatic Engineer on Substack and the author of The Software Engineer's Guidebook. He spoke at the LDX3 conference. (LeadDev London) 16 June 2025 The largest engineering leadership festival of the year 2000 participants. By the way, I've recently told reportEngineering Leadership Report 2025" from this company LeadDev. But if you go back to the report itself, then below the sammari of his key ideas.
1. Orosz conducted a qualitative study To analyze the real situation with Gen AI. He decided to do it because there is a big gap between the optimistic statements of CEOs of large companies: Microsoft CEO claims that the company is not a good company. 30Percentage of code written by AI, Anthropic CEO predicts 90Percentage of AI code in a year (me told speech) real problems in practice, including costly AI tool errors and failed public demonstrations. To do this, Orosz decided to interview engineers personally from different contexts.
Startups that make AI tools
- Anthropic engineers say that 90A percentage of their Claude Code product is written using the same code.
- Windsurf engineers claim 95Percentage of code written with them Cursor engineers give more cautious estimates - about 40-50% success rate
Big Tech companies
- Google: They use their own IDE Cider with integrated LLM tools. SRE inside preparing for 10Multiply the number of lines of code in production, expanding the infrastructure and review tooling.
- Amazon: Almost all developers use Amazon Q Developer Pro, especially for AWS development. Amazon is becoming a pioneer of MCP, and most internal tools already have MCP support, allowing you to automate many workflows.
AI startups (But not what tuling does.) Mixed results. For example, incident.io actively uses AI to speed up team work and shares practices in Slack. However, the biotech startup reported that LLMs didn’t catch on – it’s quicker to write correct code on your own than it is to check and fix AI code.
Experienced engineers (independent software engineers)
- Armin Ronacher (Flask-maker)after 17 Years of coding inspired again thanks to Claude Code, believes that "AI changes everything"
- Peter Steinberger (creator of PSPDFKit): notes that languages and frameworks have become less important thanks to the AI tool n
- Simon Willison (creator of Django): confirms that coding agents do work, and model improvements over the past 6 The months were a turning point.
Opinions of industry legends Martin Fowler compared the advent of LLM to the transition from assembler to high-level languages – a similarly revolutionary performance gain, but for the first time ever, with a high-level language. deterministic tools. Kent Beck: I enjoy programming more than ever. 52 year. He compares the effects of LLM to microprocessors. (1970s)Internet (2000s) smartphone (2010s).
If we talk about statistics that are collected through surveys, then according to surveys of the DX platform (recently handler their reportMeasuring AI code assistants and agents")
- On average. 50% of developers in organizations use AI tools weekly, in top companies – up to 60%.
- The time savings are 3-5 hours a week, which is far from the stated 10-20x improvements These survey results show results that are different from those of CEOs of large companies. But you can see that experienced independent engineers are excited about this tuling. Orosz He believes that current AI tools work better for individual developers than for teams within large companies.
At the end, Orosz is talking about significant changes in software development, and the industry needs to experiment more like startups. The key recommendation is to try what works, understand what has become cheap and what has become expensive, and adapt to the new technology landscape.
#AI #ML #PlatformEngineering #Software #Architecture #Processes #DevEx #Devops