The New Era of Programming: When AI Writes Its Own Tools
There’s something happening inside the world of programming right now, and if you’re not paying attention, you’re already behind. It started innocently enough—a post by Andrej Karpathy that’s racked up over 15 million views. His confession? He feels completely behind as a programmer, and he’s one of the best in the business.
Let me explain what’s really going on here. The way programming has been done for decades is being fundamentally refactored. Less code is being written by humans. Less code is being reviewed by humans. Yet somehow, we’re building more incredible applications than ever before.
When the Best Feel Left Behind
Karpathy’s post hit a nerve because it articulated what many developers are feeling right now. According to his assessment, there’s a new programmable layer of abstraction involving agents, sub-agents, prompts, context, memory, modes, permissions, tools, plugins, skills, hooks, MCP, LSP, slash commands, workflows, and IDE integrations. It’s like trying to learn guitar when the instrument keeps sprouting new strings every week.
Here’s the thing though—he nails the core issue. The model capabilities are already there. What we’re seeing now is the scaffolding catching up. The tools built around these models are becoming increasingly powerful, and it’s happening at a pace that makes Moore’s Law look leisurely.
His key takeaway? “Roll up your sleeves to not fall behind.” If you’re a programmer or want to get into programming, don’t be afraid. Try the tools, tinker with them, experiment, find out what works, find out where those edges are.
Even the Builders Can’t Keep Up
Think about it this way: Boris Cherny, one of the leads on the Claude Code team, replied to Karpathy’s post with “I feel this way most weeks, to be honest.” Let that sink in. The guy literally building the tools that are transforming programming feels like he’s struggling to keep up with his own creation.
Recently, Cherny’s team was debugging a memory leak in Claude Code. He started approaching it the old-fashioned way—connecting a profiler, using the app, pausing the profiler, manually looking through heap allocations. Then a coworker just pointed Claude Code at the problem and said “go look.” Claude Code one-shotted a PR with a fix.
To be fair, this is like watching the Band-Aid company struggle to keep up with their own product improvements. When your tools are evolving faster than you can master them, we’ve entered uncharted territory.
The Opus 4.5 Moment: A Vibe Shift in Real-Time
Igor Babushkin, one of the founders of xAI, chimed in on the discussion. Karpathy’s response was telling: “It’s very good. People who aren’t keeping up, even over the last 30 days, already have a deprecated worldview on this topic.”
Thirty days. That’s all it takes for your understanding of programming capabilities to become outdated. The reality is that we’re experiencing what feels like a phase change in software development, and it’s centered around the Claude family of models wrapped with Claude Code.
Toby Lütke, CEO of Shopify and prominent in the AI community, put it simply: “Opus 4.5 feels very different for coding than anything else that came before. It’s kind of stunning what it can do.”
When Code Writes Itself: The Self-Improvement Loop
Here’s where things get really interesting—and slightly unsettling if you’re being honest about it. Boris Cherny dropped a bombshell revelation: 100% of Claude Code’s code in the last 30 days was written by Claude Code itself. That’s not just AI-assisted development. That’s genuine self-improvement.
In the last 30 days, he landed 259 PRs, 497 commits, 40,000 lines of code added, and 38,000 lines of code removed. Every. Single. Line. Written by Claude Code plus Opus 4.5. Claude now consistently runs for minutes, hours, and days at a time using stop hooks.
What most people miss is that this isn’t just about raw productivity metrics. It’s about a fundamental shift in how software development works. We’re entering a new period in coding history, and like the shift from assembly to high-level languages, there’s no going back.
How Elite Teams Actually Code Now
A prominent coding leader recently talked with the Claude Code team to understand their workflow. His takeaway? “It feels like I get a peek into the future and I get why Dario said six months ago that 90% of code will be written by AI.”
Here’s how this team operates differently than any traditional engineering organization:
The first code review is always done by Claude Code. When you write a PR, the code is immediately reviewed by AI. The test suite? Almost 100% written by Claude Code. Now, there still needs to be a human somewhere in the loop because when you have specifications, the human is the one defining those specs. That’s where human taste comes in.
They ship about five releases per engineer per day. That’s insane velocity. Even the best engineering teams a few years ago couldn’t achieve that kind of speed. Prototyping happens surprisingly quickly—they go through 10+ actual prototypes for a new feature.
They use sub-agents extensively. It’s not just a single agent with one human overseer. It’s humans overseeing agents overseeing teams of sub-agents. The organizational chart looks more like Inception than a traditional dev team structure.
Confession: I Ship Code I Never Read
Peter Steinberger, another well-known developer, made a confession that would have been career-ending heresy just a year ago: “I ship code I never read.” His 2025 workflow is fundamentally different from traditional development practices.
He watches the stream of generated code, sometimes looks at key parts, but admits most code he doesn’t actually read. What he does know is where components are, how things are structured, and how the overall system is designed. Watch for this becoming the norm rather than the exception.
For those of you who have been pessimistic about “vibe coding,” the best in the world are now adopting it. It’s more than okay to ship code you haven’t read line by line. The skill isn’t in reading every character—it’s in understanding architecture, design patterns, and system behavior.
The End of Me: A 55-Year Career Automated
Google’s ex-CEO Eric Schmidt recently reflected on his five-decade programming career with striking candor. According to Schmidt, everything he did in his 20s that got him to where he is has now been completely automated. Every aspect of the programming, every aspect of the design is now done by computers.
He recently had AI write a whole program for him. As he watched it generate classes and interaction details, his reaction was visceral: “Holy crap, the end of me.” Consider this—watching your entire profession start, evolve, and potentially end within your own lifetime is profound in ways we’re still processing.
The trick is understanding what this means for the next generation. Each person now has a supercomputer and a super programmer in their pocket. Computer science isn’t going away, but computer scientists will be supervising these systems rather than doing the grunt work.
Six to Twelve Months from Solving Software?
McKay Wrigley, who teaches AI coding, made a bold claim: “The more I code with Opus 4.5, the more I think we’re six to twelve months away from solving software.”
What he means is worth unpacking. He now builds three versions of an app in a few hours just to explore options—something that would have taken him one to two weeks less than a year ago. The model is pretty much there. The capability exists today.
Don’t get me wrong, but “solving software” doesn’t mean we’re done building things. It means the bottleneck has fundamentally shifted.
The Real Impact: Human Taste Becomes the Bottleneck
Here’s what’s interesting about this transformation: If intelligence and the creation of software becomes essentially free and infinite, the only thing that’s going to matter is the experience.
What does the application look like? How does it feel? How does it interact with the user? These are all decided by human taste. As the signal-to-noise ratio becomes flooded with noise—and it will—the ability to pick out signal from noise becomes the only skill that matters.
The bottom line is this: Human taste is going to become not only much more valuable, but potentially the biggest bottleneck in software development. Not coding ability. Not technical chops. Taste.
What This Means for You
Coding is vastly different than it was just a year ago. Imagine what it’s going to be like a year from now, or five years from now. The profession isn’t dying—it’s evolving faster than we can comfortably track.
The notion of what makes a “good programmer” is being redefined in real-time. It’s no longer primarily about syntax mastery or algorithmic problem-solving. It’s about understanding systems, directing AI agents, making architectural decisions, and most importantly, having the taste to know what’s worth building in the first place.
You might be wondering how to stay relevant in this new landscape. The answer isn’t to resist the change or pretend it isn’t happening. The answer is to experiment aggressively, fail fast, and develop a deep intuition for what these tools can and can’t do. That intuition—that sense of possibility and limitation—is the new core competency.
The staircases have started moving at Hogwarts, and standing still isn’t an option anymore.