How we got here

One version of the timeline:

30 Nov 2022: ChatGPT made conversational coding widely accessible.

• It could explain and generate code through dialogue.
• Coding still meant copying snippets between a chat window and an editor.

12 Sep 2024: OpenAI released o1-preview.

• The model spent additional time reasoning before answering, improving results on complex coding, maths, and science tasks.

20 Jan 2025: DeepSeek released R1.

• DeepSeek published an MIT-licensed reasoning model, its training report, and smaller distilled models.
• Strong reasoning is now accessible: developers could inspect, adapt, and self-host the weights.

02 Feb 2025: Karpathy named “vibe coding”.

• “I ‘accept all’ always, I don’t read the diffs anymore.”
• The phrase captured a cultural shift toward describing an outcome and letting an AI produce the implementation.
• It did not remove the need for engineering judgment, evidence, or review.

24 Feb 2025: Anthropic released Claude Code as a research preview.

• The model moved into the terminal, where it could search and read a repository, edit files, run tests, use command-line tools, and commit changes.
• Now you can delegate bounded engineering tasks to an agent operating inside the development environment.

24 Nov 2025: Claude Opus 4.5 pushed coding agents further.

• Anthropic reported stronger software-engineering, planning, tool-use, and long-horizon agent performance, with fewer tokens needed for comparable or better results.
• The practical shift was from agents that could attempt a change to agents that could plan, execute, and verify longer tasks with less intervention.

2026: Powerful local models became a practical choice.

• Open-weight releases such as Qwen3.6, Gemma 4 12B, and Nemotron 3 Nano brought strong coding, reasoning, tool use, and multimodal capabilities to developer-owned hardware.
• Support for runtimes such as llama.cpp, MLX, Ollama, and LM Studio made local deployment increasingly routine. Hardware needs still varied from an ordinary laptop to a high-end workstation.

2026: Frontier-grade open models became a category.

• Models from several labs, including GLM-5.2, Kimi K3, MiniMax M3, and DeepSeek V4, competed near the proprietary frontier on reasoning, coding, tool use, multimodality, and long-horizon agent work.
• Efficiency-focused releases such as MiMo-V2-Flash showed that the open ecosystem was broadening across both capability and deployment cost.
• These models could be inspected and self-hosted, but the largest required datacentre-scale hardware. Open and local had become separate dimensions.

Beliefs to let go

Habits from before Nov 2025 to unlearn:

• “The code must be written by hand”.
  • Who types doesn’t matter; maintainable code that solves the right problem does
• “Every line must be manually reviewed”.
  • If you don’t trust the results, devise a workflow to figure out what you should manually review.
• “First attempts should be good”. Expect a high garbage rate; iterate fast instead
  • Conversely, with good models, first attempts can be good, if you plan. This was one-shotted.
• “More typing equals more learning”. Reviewing agent builds teaches more than hand-typing!

Letting go isn’t ceding control. You’re encoding it: conventions, constraints, review, and that’s the thesis of this course.