There are areas where we seem to be much closer to AGI than most people realize. AGI for software development, in particular, seems incredibly close. For example, Claude Code has bewildering capabilities that feel like magic. Mix it with a team of other capable development-oriented AIs and you might be able to build AI software that builds better AI software, all by itself.
The "G" in AGI stands for "general", so talking about "AGI for software development" makes no sense, and worse than that accepts the AI companies' goalpost-shifting at face value. We shouldn't do that.
But I feel like the point is that, in order to reach AGI, the most important area for AI to be good at first is software development. Because of the feedback loop that could allow.
Claude Code is good, but it is far from being AGI. I use it every day, but it is still very much reliant on a human guiding it. I think it in particular shows when it comes to core abstractions - it really lacks the "mathematical taste" of a good designer, and it doesn't engage in long-term adversarial thinking about what might be wrong with a particular choice in the context of the application and future usage scenarios.
I think this type of thinking is a critical part of human creativity, and I can't see the current incarnation of agentic coding tools get there. They currently are way too reliant on a human carefully crafting the context and being careful of not putting in too many contradictory instructions or overloading the model with irrelevant details. An AGI has to be able to work productively on its own for days or weeks without going off on a tangent or suffering Xerox-like amnesia because it has compacted its context window 100 times.
This is a statistical model, it is as good as the data it averages. So shit from SO in, shit from SO out. Until they have the right dataset that doesn't contain cancerous code from people that can't write code, they can't even create a good agent, let alone AGI.
The real irony is from now on, because people use this magic, it will stay forever. What you can count on in my opinion is that this whole world changes, you don't need to write sw anymore because everything is AI. Hard to imagine, and too far in the future to be relevant for speculations.
You would be surprised at how many prompts in Cursor are required just to adjust a layout and get padding/margins to spec even while providing it the figma link and using a figma MCP, as well as well developed prompts and images/files for context. Still can't figure out why there is 20px padding in a container with no set height.
