Most companies still picture a developer hunched over a keyboard, typing every line. This breakdown flips that picture on its head, and the numbers behind it are hard to ignore.
I stumbled onto this deep-dive where the creator walks through a new Anthropic essay called “When AI builds itself.” The original poster does a great job translating dense research into plain talk, and I was genuinely stunned by a few of the stats. So let me share what stood out.
🤖 The old way vs the new way
Here’s the contrast that hit me. A few years ago, Anthropic worked like any tech company: engineers wrote code, code shipped. Then came chatbots. Then coding agents. Then agents spinning up sub-agents. Each step pushed the human further from the actual code.
The creator highlights the wild part. As of May 2026, the expert reports that more than 80% of code merged into Anthropic’s codebase was authored by Claude itself. Before Claude Code launched in early 2025, that number sat in low single digits. That’s roughly a year of change.
📈 The acceleration is speeding up
The author points to task length as the real signal. Watch the jump the original poster lays out:
- March 2024: Claude could handle tasks taking humans about 4 minutes
- One year later: about 90 minutes
- 2026: roughly 12-hour tasks
- The trend the expert cites suggests week-long tasks could land in 2027
And reproducing AI research papers? The contributor notes models went from a 20% success rate to nearly 100% in about 15 months.
⚠️ The honest catch
This is where I respect the breakdown. The creator flags that engineers now produce 8x more lines of code, but Anthropic staff only report about 4x more actual output. Translation: a lot of that AI-written code is, for now, lower quality than human code. Lines written is a weak measure, and the original poster is refreshingly clear about that.
The person who shared it also points out the missing ingredient: taste. Picking which experiments matter, spotting a dead end, deciding what to build next. Models still lean on humans for that judgment call.
🛠️ What you can actually take from this
The creator pulls out a few practical patterns worth copying:
- Hand AI underspecified problems. The expert admits his own prompts are often just a screenshot of an error plus “fix it,” and the models cope fine.
- Move up the ladder. From “the export button isn’t working, fix it” to “investigate why the network slows under heavy load.” Bigger problems, same short prompts.
- Keep your understanding. The contributor shares a line I keep thinking about: you can outsource your thinking, but you cannot outsource your understanding.
- Think net-new. Don’t just automate old tasks. Ask what work you’d never have attempted before because it was too expensive.
The one part that made me raise an eyebrow? Anthropic argues we should slow down frontier AI development. As the creator points out, that’s a convenient thing to say when you’re the one in first place.
There’s a lot more in the full video, including the “permanent underclass” idea and why human review might become the next bottleneck. Go watch the whole breakdown for the details, the charts, and the creator’s take on what comes next.