Dario Amodei stood on stage at the World Economic Forum in Davos on January 20, telling an audience of global leaders that AI is already reshaping software development. He's been warning about this for nearly a year now. Back in March 2025, speaking to the Council on Foreign Relations, he predicted AI would be writing 90% of code within three to six months. "And then in 12 months, we may be in a world where AI is writing essentially all of the code."
We're past that six-month mark. The industry-wide transformation Amodei envisioned hasn't materialized in quite the way he suggested.
The Prediction, Repeated
This isn't Amodei being careless with timelines. Anthropic's CEO has been remarkably consistent about his view that software engineering sits at the frontier of AI capability. At Davos this week, he reiterated concerns about entry-level white-collar jobs, noting that while there wasn't a massive AI impact on the labor market right now, he is seeing changes in the coding industry.
And he's not alone. His company's Chief Product Officer, Mike Krieger (the Instagram co-founder), has echoed the sentiment that developers would soon double-check AI-generated code rather than write it themselves. Amazon Web Services CEO Matt Garman made similar claims. Mark Zuckerberg told Joe Rogan he believes AI will replace engineers.
The consistent drumbeat from executives at companies selling AI tools is that traditional programming is on borrowed time. Color me skeptical, or at least curious about the evidence.
What the Data Actually Shows
The evidence is messier than the predictions suggest.
A Stanford University study found employment among software developers aged 22 to 25 fell nearly 20% between 2022 and 2025, coinciding with the rise of AI-powered coding tools. That's a real number, and it's concerning. But correlation isn't causation, and Indeed's chief economist Svenja Gudell points to post-pandemic right-sizing as a more likely culprit.
The productivity claims? Also murky. Early studies from GitHub, Google, and Microsoft (all vendors of AI tools, notably) found developers completing tasks 20% to 55% faster. But a September report from the consultancy Bain & Company described real-world savings as "unremarkable."
Here's the number that caught my attention: a randomized controlled trial by the nonprofit METR found that when developers use AI tools, they take 19% longer than without. Not faster. Slower. And yet the same developers believed they were working 20% faster with AI. The gap between perception and reality is striking.
The Vibe Coding Phenomenon
Andrej Karpathy coined the term "vibe coding" in early 2025, describing an approach where developers describe software in natural language and let AI handle the actual code generation. The practice has exploded. According to Stack Overflow's 2025 Developer Survey, 65% of developers now use AI coding tools at least weekly.
But widespread adoption doesn't mean universal success. It's unclear at this point how viable vibe coding will be in the long term, with issues around code quality and code maintainability leaving plenty of room for skepticism.
The irony? Karpathy himself apparently doesn't trust the technology enough, reportedly hand-coding his own project rather than relying on AI assistance.
Inside the AI Labs
Where Amodei's prediction does seem to hold is within AI companies themselves. Gergely Orosz at The Pragmatic Engineer reported that Boris Cherny, creator of Claude Code, claims 100% of his code contributed to Claude Code was AI-written as of December 2025.
That's a notable data point, though it comes with obvious caveats. Claude Code is closed source, so the claim is hard to verify independently. And the person making it has every incentive to showcase his product's capabilities.
Still, Orosz's own experience aligns with this: "I let Claude Code generate all the code I end up committing. When the code is not how I want it, I do more prompting to get the LLM to fix it."
The November and December 2025 model releases (Opus 4.5, GPT-5.2, Gemini 3) appear to have been a genuine inflection point, at least for some developers.
The Counterarguments
Not everyone's buying it. IBM CEO Arvind Krishna directly pushed back on Amodei's March prediction, saying "I think the number is going to be more like 20-30% of the code could get written by AI—not 90%."
Grady Booch, the IBM Fellow who co-created UML and is one of the most respected voices in software architecture, has consistently argued that LLMs represent another abstraction layer rather than a replacement for engineering fundamentals. He brought two examples that he felt was even more disruptive than LLMs: the shift from monoliths to distributed systems in the 80s, and the rise of GPUs. Software engineers adapted to those changes. They'll adapt to this one.
Google DeepMind CEO Demis Hassabis, sharing the Davos stage with Amodei in a panel called "The Day After AGI," took a more measured view. He said he expected "new, more meaningful jobs being created" and advised undergraduates to become "unbelievably proficient" with AI tools rather than fearing them.
The Real Problem
Here's what I keep coming back to: Amodei made a specific, falsifiable prediction in March 2025. He said 90% of code would be AI-written within three to six months. We're now past that window, and Amodei's bold prediction hasn't quite come to fruition.
Big tech companies have highlighted gains. Google claims more than 25% of internal code is AI-generated. But 25% isn't 90%, and internal AI company figures aren't the same as industry-wide transformation.
The executives making these predictions have a structural incentive to hype their own products. When Amodei says AI will write all code in a year, he's simultaneously promoting Claude Code as the tool that will do it. That doesn't make him wrong, but it should inform how we weight the claim.
What Happens Now
Amodei isn't backing off. At Davos, he predicted AI models could achieve Nobel-level breakthroughs by 2026 or 2027, though he acknowledged bottlenecks like chip production might slow things down. He called the Trump administration's move to ease restrictions on AI chip exports to China "a big mistake" with "incredible national security implications," comparing AI chips to nuclear weapons.
The timeline for AI-dominated software engineering keeps shifting forward. Last year's "three to six months" becomes this year's "twelve months." At some point, the prediction either lands or it doesn't.
For now, the developers I talk to are using AI tools heavily but still writing and reviewing plenty of code themselves. The junior developer job market is brutal, but senior engineers who know how to architect systems and debug AI-generated slop remain in demand. The revolution is arriving unevenly, as revolutions do.
Whether Amodei's latest timeline holds is anyone's guess. But his track record on the previous one should temper expectations.
