28 February 2025
7 min read

The world debated AI governance while developers stopped writing code

February 2025 packed three genuinely significant AI moments into less than two weeks. Andrej Karpathy, a founding member of OpenAI and former head of AI at Tesla, posted a short message on February 2nd coining the term "vibe coding." OpenAI launched ChatGPT Deep Research on February 3rd. And on February 10th, sixty-one countries gathered in Paris for the AI Action Summit, where the US and UK declined to sign the collective statement on inclusive and sustainable AI. It was a lot.

Vibe coding: programming without knowing how to program

Karpathy's description was intentionally casual. Vibe coding, as he framed it, means giving an AI a rough description of what you want built, running what it produces, telling it when something breaks, and repeating until it works. You are not writing code. You are directing code generation through natural language.

This was happening already, but naming it gave the practice permission to be taken seriously. Hobbyists started building apps that would have required a professional developer six months earlier. Small founders shipped internal tools without hiring engineers. The concept spread quickly because it described something many people were already doing but felt slightly embarrassed to admit.

The obvious tension is quality. Vibe-coded software often works well enough for personal use or quick prototypes, but tends to accumulate technical debt at a rate that becomes painful at scale. The code is frequently correct but not maintainable. Still, for a broad class of use cases, "correct enough" is sufficient.

ItemValue
Cursor
Best for: Codebase-aware editing, refactoring
Limitation: Still requires dev judgment
GitHub Copilot Agent
Best for: IDE integration, file-level changes
Limitation: Multi-file reasoning gaps
Claude in Projects
Best for: Long-context code review
Limitation: No direct execution
Replit Agent
Best for: Full app generation from scratch
Limitation: Infrastructure decisions are opaque
v0 by Vercel
Best for: React UI from a description
Limitation: Limited to frontend/component scope
Fig. 1. Vibe coding tools — best use case and primary limitation per platform.

ChatGPT Deep Research: a 30-day task in a few hours

OpenAI launched Deep Research on February 3rd, initially to ChatGPT Pro subscribers. The system can autonomously browse the web, read documents, and produce a structured research report on complex topics. The reported benchmark was tasks that would take a human researcher up to 30 days to complete.

In practice, Deep Research works well for literature surveys, competitive analysis, and pulling together information that exists but is scattered across many sources. It struggles with tasks requiring primary research, access to paywalled content, or nuanced judgment about source credibility. But for a large class of knowledge-work tasks, the tool genuinely compresses research time from days to hours.

30 days
Human equivalent task time
5-30 min
Deep Research typical run time
Fig. 2. Research task duration — human equivalent vs Deep Research run time (hours).

The Paris AI Summit: 61 countries, one division

The AI Action Summit ran on February 10th and 11th in Paris. It was co-chaired by France and India and attended by representatives from 61 countries. The stated goals were to discuss innovation, regulation, and ensuring AI benefits are shared globally.

The headline outcome was a declaration calling for "inclusive and sustainable" AI development. Fifty-nine countries signed it. The United States and the United Kingdom did not.

ItemValue
Total attendees
61 countries
Declaration signatories
59 countries
Countries that refused to sign
Key themes: Global access, safety, governance frameworks
USA, UK
French AI investment pledged
Over coming years
€109 billion
Fig. 3. Paris AI Summit outcomes — attendees, signatories, and investment pledged.

The US position, broadly, was that multilateral agreements on AI governance risk constraining innovation without meaningfully improving safety. The UK's reasoning was similar. Both governments had just watched DeepSeek demonstrate that even resource-constrained actors can produce frontier-level models, which complicates the argument that agreements between wealthy nations can effectively govern the technology.

Elon Musk and the $97 billion OpenAI bid

On February 10th, Elon Musk led a group of investors in a $97.4 billion offer to acquire OpenAI. This was not a routine acquisition attempt. Musk was a co-founder of OpenAI who left the board in 2018 over disagreements about the lab's direction. He has since started his own AI company, xAI. The offer was widely interpreted as either a genuine acquisition attempt or a disruptive move designed to complicate OpenAI's planned conversion from a nonprofit to a for-profit entity.

OpenAI declined the bid relatively quickly. But the episode highlighted an ongoing tension in the AI industry: the question of whether AI development should be controlled by a small number of well-capitalised private entities, and who gets to make that argument convincingly.

What February 2025 actually told us

The month illustrated a gap that has been widening for some time. The technical capability of AI tools was moving faster than either policy frameworks or workforce adaptation. Developers were using AI to skip traditional coding entirely. Researchers were using it to compress knowledge work. Governments were meeting in Paris arguing about values and governance while the tools themselves kept shipping new capabilities.

Vibe coding became a concept because it named something real. Deep Research shipped because the underlying capability had crossed a usefulness threshold. The Paris Summit revealed that countries still disagree fundamentally about whether AI should be governed as a global commons or a national competitive asset. All three of those things are still true.