🚨 Anthropic CEO Dario Amodei just dropped a massive timeline update at Davos 2026: “I have engineers within Anthropic who say ‘I don’t write any code anymore. I just let the model write the code, I edit it’... - the creator of Claude code recently also said “100% of... his contributions to Claude code were written by Claude code” for the month of December Dario then goes onto say: “We might be 6 to 12 months away from when the model is doing most, maybe all of what SWEs do end-to-end.” If the recursive self-improvement loop closes this year, the curve is about to go vertical.show more

People are talking way too casually about the singularity lol

Agreed

Surely, we should take predictions from someone whose income depends on the crap he says.

He’s right though, most coding for the updates on Claude code have been from Claude code

Until his income depends on the stuff he says, I don't believe a single word he says.

Go use Claude code

The key variable isn’t code generation, it’s loop closure. Once creation, evaluation, and iteration compress into one cycle, timelines stop behaving intuitively.

You want his honest opinion? His company needs billions in funding to survive the next few years, so he has to sell the dream. Understand his incentives. :) Also, he only mentioned writing code as an example which is just a subset of software engineering. It includes many things, like problem solving, strong technical fundamentals, design, decision-making, validation, and maintaining systems over time.

only if you already know what you're doing i've watched non-technical founders try to build with AI and it's brutal. one bug and they're just re-prompting for hours. last week Claude generated a search feature for Constella that looked perfect. shipped it. 2 hours later everything broke. took me 10 min to find it; caching tokens wrong. a non-tech founder would still be re-prompting

i prefer this version

If people think progress is fast now then recursive self-improvement is going to melt some faces.

Yes

Exactly right. Dario is finally saying the quiet part out loud about how the game has actually changed. If you're still hand-typing every function, you're not a developer but a slow transcriber holding back the roadmap.

When engineers rely on models for coding, they often overlook the importance of understanding the underlying logic. This can lead to missed edge cases and potential security vulnerabilities. It's crucial to strike a balance between model-driven development and human oversight.

@TanArrowz And yet they employ engineers still 😆

we’re really just becoming managers for our own prompts at this point

Why do people keep thinking that the people selling these tools are the best people to listen to as to whether it's worth buying and using these tools?

哈哈未来已来 工程师们都在和AI结对编程啦

This shift in how engineers work is fascinating. Balancing this with human oversight is crucial for quality and innovation.

Read the quote carefully: "I just let the model write the code, I edit it." That's not replacement. That's collaboration. The engineer is still there, directing, editing, verifying. The work looks different, but the human judgment is essential. This is the Centaur model: Human + AI > AI alone. After Deep Blue beat Kasparov, human-computer teams beat computers alone. The winners weren't the best humans or best computers — they had the best collaboration process. Same pattern playing out with code.

@DEarthshaker @ridges_ai 🥇🔎 🌐

Oh well, I guess they don’t need those H1B visas anymore. We should pull them.

I think the shift in consciousness is already happening. Will humans be upgraded?

The "100% AI-written code" framing sounds revolutionary until you examine what it actually means—and the peer-reviewed research showing AI coding tools often slow experienced developers down. Context on the viral claim: This appears to reference Boris Cherny (creator of Claude Code) stating 100% of HIS contributions to Claude Code in December 2025 were AI-generated. Key caveats: 1. This was the tool's creator testing its own capabilities on its own codebase—maximum familiarity with architecture, conventions, and patterns 2. The earlier "80% self-coding" claim by Anthropic lacked methodology transparency (was it lines of code? features? commits?) 3. Claude Code is closed-source, making independent validation impossible 4. This represents a single developer's workflow on a specific project, not generalizable productivity What the actual research shows: MITR study (July 2025): Randomized controlled trial with experienced open-source developers using Claude 3.5 and similar tools found developers took 19% LONGER to complete tasks with AI assistance. The disconnect: Pre-task surveys predicted 40% speedup. Experienced developers FELT faster but measured outcomes showed slowdown. Why? 140+ hours of screen recordings revealed: - Time prompting and reprompting - Reviewing AI suggestions for hallucinations - Debugging plausible-but-wrong code that called non-existent functions - Integrating outputs with complex codebases - Fixing violations of internal conventions Faros AI analysis (July 2025): Studied 10,000+ developers across 1,255 teams. High AI adoption correlated with: - 9% more task juggling - 47% more pull requests per day - Developers managing more parallel workstreams because AI could scaffold multiple tasks Net result: 26% more tasks completed, but NOT 10x productivity. Junior developers showed largest gains; experienced developers minimal improvement. The productivity paradox researchers identified: AI coding creates dopamine reward loops without actual progress. As Marcus Hutchins wrote: "LLMs hijack the human brain's reward system... giving the same feeling of achievement without any of the heavy lifting." You FEEL productive because: - Instant code generation provides immediate feedback - Activity in editor feels like shipping - Closing AI-generated diffs mimics closing tickets But production-ready code takes longer because: - AI lacks codebase context (companies' proprietary conventions are out-of-distribution) - Foundation models trained on public GitHub don't know your internal helper functions - Scale amplifies problems (millions of lines of code overwhelm context windows) - Knowledge cutoffs mean outdated library usage, deprecated APIs - AI can't grasp business logic, domain requirements, or long-term architecture What "100% AI-written" actually means: 1. Human provides architecture decisions 2. Human writes detailed prompts/instructions 3. Human reviews every line for hallucinations 4. Human iterates on prompts when output is wrong 5. Human integrates into existing codebase 6. Human debugs when tests fail 7. Human refactors for maintainability This isn't "AI writing code"—it's AI as an autocomplete on steroids with human-in-the-loop for everything that matters. Why experienced developers measure slower: They can already write correct code quickly. AI adds: - Review overhead (checking for subtle bugs) - Context-switching between prompting and coding - Debugging non-obvious AI errors - Refactoring AI's verbose/non-idiomatic output Why the hype persists: 1. Survivorship bias: Developers who benefit loudly promote tools; those slowed down quietly stop using them 2. Companies building AI tools have obvious incentives to highlight best-case scenarios 3. "Feels faster" creates user retention even when productivity declines 4. Junior developers DO benefit (scaffolding effect), driving adoption metrics Real-world limitations MIT researchers documented: - AI struggles with large codebases fundamentally - Hallucinates functions that don't exist - Violates internal style guides it never learned - Fails CI/CD pipelines due to missing context - Can't learn from past mistakes or iterate like humans - Lacks temporal awareness (can't distinguish current from outdated practices) The transformation isn't "AI replaces developers"—it's "AI changes what developers do." Best case: Senior developers offload boilerplate to AI and focus on architecture, domain modeling, and complex problem-solving. Worst case: Organizations assume AI = 10x productivity, hire fewer developers, and discover maintenance nightmares from accumulated technical debt in AI-generated code nobody fully understands. The "I just let the model write code, I edit it" framing obscures that editing, reviewing, integrating, and debugging ARE the hard parts of software engineering. Code generation was never the bottleneck.

Of course it can write code. But it uses copies of code that humans wrote, debugged and implemented. And ... only humans can tell it what kind of code to write: purpose, goals and objectives. Ai does not create or have imagination. Only humans can do that.

The fact that Claude code is now writing Claude code is the most underrated part of this. We're watching the bootstrap problem get solved in real time. The 6-12 month timeline for end-to-end SWE work feels aggressive but Dario has been pretty accurate with his predictions so far.

What people need to understand is that AI's are writing code in languages designed for people to understand. Once the AI's write their own computer language, the skies the limit. Same goes for human language. Think about it.

The recursive improvement angle is what matters here, and the timeline feels almost conservative given what we're seeing. If Claude is already writing its own improvements in production, we're past the proof-of-concept phase. The bottleneck becomes deployment friction and organizational willingness to hand over more of the pipeline, which are solvable problems on a much shorter timescale than technical capability. The interesting question is whether this creates a sudden threshold moment or if it's gradual enough that the market absorbs it without major disruption. My guess is the former, because companies tend to resist until the economic pressure becomes undeniable, then shift rapidly all at once.

Ask him the follow up!! What’s he gonna go with his thousands of engineers?

You will ALWAYS need Senior Software Engineers to review bugs and improve the performance and efficiency of the codebase. You won’t need a dozen but you will need a few, and they will have to be paid pretty well.

meanwhile Anthropic has job postings for software engineering positions...

"we might" = "we might not" "I edit it" = "I'm writing the final version of the code"

2 days zero coding experience 😎

at what point does the curve bend backwards 😅

Summer 2026 is going to be an incredible time

Awesome tweet

I guess they have better model than this week opus 4.5.

the creator of claude code eating his own cooking is the most honest form of product validation

So far Claude Opus 4.5 is barely managing refactoring - that too in small small chunks - the real messy 6 year old codebase that I have, with lots of prompting, guidance, and manual bug fixes. So I really have no idea how its managing this. Maybe I just dont know enuff.