The bottleneck shifted but didn't disappear. Getting to a working prototype in a weekend is real, but error handling, edge cases, and ops work hasn't gotten much faster. Distribution is completely unchanged too. A lot of these 'where are the AI apps' questions are really asking why there aren't more successful AI businesses, which is a harder and very different problem.
The catalog normalization problem is real and severely underestimated. I ran into this building on top of product data feeds -- even with a single retailer, inconsistency in titles, attributes, and categorization is staggering. LLMs are great at reasoning but they inherit all that messy upstream data, so agentic shopping will keep stumbling until the data layer gets cleaner.
The quality degradation at 2-bit is a real issue. For actual work tasks, a well-tuned 30B at 4-bit usually outperforms a 70B+ at 2-bit in my experience. The expert reduction on top of that compounds things - you're essentially running a fairly different model. Still interesting to see the upper bound of what consumer hardware can attempt, even if the result isn't production-ready.
This matches what I have seen building solo projects -- the activation energy for failure is often lower than for success. Adding friction to the wrong path (like requiring a deliberate step to delete data) works better than willpower alone.
The decision to build this as a TUI rather than a web app is interesting. Terminal-native tools tend to get out of the way and let you stay in flow -- curious how the context management works when you have a large codebase, do you chunk by file or do something smarter?
It’s both! The core is implemented as a server and any UI (the TUI being one) can connect to it.
It’s actually “dumber” than any of your suggestions - they just let the agent explore to build up context on its own. “ls” and “grep” are among the most used discovery tools. This works extraordinarily well and is pretty much the standard nowadays because it lets the agent be pretty smart about what context it pulls in.
The spec-first approach is underrated. Treating the spec as a living artifact the AI can reference across sessions is something I've been experimenting with too. The main challenge is keeping specs short enough to actually stay current.
Rule 2 is the one that keeps biting me. You can spend days micro-optimizing functions only to realize the real bottleneck was storing data in a map when you needed a sorted list. The structure of the data almost always determines the structure of the code.
