An interesting presentation at a London conference recently caught my attention. It explored the stark contrast between CEOs' glowing statements about AI development tools and the actual temperature on the ground in development teams. Let me break down the key insights from this eye-opening discussion.
CEO Optimism vs. Reality Check
The headlines around AI have been absolutely dazzling lately. Microsoft's CEO claimed that "30% of all code is written by AI," while Anthropic's CEO boldly stated that "all code will be AI-generated within a year." Google's Jeff Dean predicted that "AI will reach junior developer level within a year."
But what's the reality? One startup engineer shared how they lost $700 due to bugs from Devon, a $500/month autonomous AI agent. At Microsoft's Build conference, we witnessed a Copilot agent completely fail while trying to apply changes to a .NET codebase in front of a live audience.
What does this stark contrast tell us? The presenter decided to dig deeper by talking to actual developers to understand what's really happening on the ground.
The Real Story Behind AI Development Startups
Anthropic's Experience
Interviews with the Anthropic team revealed some fascinating insights. When they gave their engineers access to Claude Code, every single engineer started using it daily. Even more surprising? 90% of the Claude Code product itself was built using Claude Code.
Claude Code hasn't even been publicly available for a month, yet it saw a 40% usage increase from day one, with a 160% increase since launch.
Windsurf and Cursor's Reality
The Windsurf team claimed that 95% of their code is written using Windsurf. Cursor, on the other hand, was more candid—they admitted that while about 40-50% is effective, the rest isn't quite there yet.
Big Tech's AI Adoption Status
Google's Internal Situation
Google is all about custom everything. They use Borg instead of Kubernetes, their own repositories instead of GitHub, and Cider (a VS Code fork) as their IDE.
According to Google engineers, AI is truly integrated everywhere. Their Cider IDE has LLM integration, autocomplete, chat-based IDE features, and AI is even built into Critique, their code review tool.
What's particularly interesting is that Google's SRE team is strengthening infrastructure to handle 10x more lines of code going into production. Is Google seeing something we don't know about?
Amazon's Hidden Strength
While Amazon isn't well-known for AI, internally almost every engineer uses Amazon Q Developer Pro. It's particularly praised for AWS-related coding.
What's more intriguing is Amazon's API-first approach. Since 2002, they've required all teams to expose data and functionality through service interfaces, which has made it easy to build MCP (Model Context Protocol) servers. Most of Amazon's internal tools and websites already support MCP.
Experiences from Regular Startups and Independent Developers
Success Stories
Lawrence Jones from Incident.io mentioned that his team members actively use AI and share tips and tricks with each other. For well-defined tickets, agents creating the first pass is quite effective.
Armin Ronacher, creator of the Flask framework, recently published a piece titled "AI Changed Everything." He admitted that six months ago, he wouldn't have believed anyone who said they'd prefer being an engineering lead guiding a virtual programmer intern.
Failure Cases Exist Too
An engineer from a biotech AI startup honestly shared: "We've experimented with multiple LLMs but haven't settled on any." They found it faster to write correct code themselves and review LLM code to fix issues. Since they're building new software, they can't rely on existing patterns.
Veteran Developers' Perspectives
What's really fascinating is the reaction from experienced developers. Peter Steinberger, founder of PSPDF Kit, said he "hasn't been this excited and amazed by technology in a long time." As an iOS expert, he can now easily code in other languages like TypeScript.
Simon Willison, Django's creator, stated that "coding agents actually work" and that "the model improvements over the past six months have been some kind of turning point."
Most impressive was Ken Beck's comment—after 52 years of programming, he said "this is the most fun I've had in all 52 years." Thanks to LLMs, he can now pursue truly ambitious projects.
Remaining Questions
However, there are still unresolved questions:
1.Why are founders and CEOs much more enthusiastic than engineers?Even the founder of Warp, an AI tool company, complained that "senior engineers don't use AI much."
2.How mainstream has AI usage become?According to DX's survey of 38,000 developers, about 50% in medium-sized organizations use it once a week or more. Once a week, not daily.
3.How much time does it actually save?Peter mentioned 10-20x productivity improvements, but the DX survey showed about 3-5 hours saved per week.
4.Why does it work well for individuals but not at the team level?This is a common issue many people point out.
Conclusion: At the Starting Point of Change
Martin Fowler compared this change to the transition from assembly language to high-level programming languages. The difference this time is that it's non-deterministic for the first time.
Ken Beck compared it to changes on the level of microprocessors, the internet, and smartphones. He said, "The entire landscape of what's cheap and what's expensive has changed," and "things we assumed were expensive or difficult have become ridiculously cheap."
Looking at the current situation, AI development tools seem to be passing some kind of turning point. Between CEOs' exaggerated marketing and some failure cases, there's actually gradual but meaningful change happening.
What's particularly noteworthy is that experienced developers are responding positively to these tools. These are people who were skeptical about new technology, but now they're becoming the most enthusiastic users.
Ultimately, we're at a point where we need to figure out what works and what doesn't through experimentation, just like the startups are doing. The criteria for what's cheap and what's expensive, what's possible and what's impossible, are completely changing.
In this wave of change, what matters isn't blind acceptance or unconditional rejection, but careful experimentation and learning. While AI tools aren't perfect, they're definitely changing how we develop software. How we leverage this change is ultimately up to each of us.
The transformation is real, even if it's not as dramatic as some CEOs claim. We're witnessing a fundamental shift in software development, and the developers who embrace thoughtful experimentation with these tools will likely be the ones who benefit most from this new landscape.