Connor Neblett

Connor Neblett

@connor.neblett.com

Coder and person who wants to leave the world better than he found it. Building @supernotes.app to make note-taking better than he found it.

46 followers34 following21 posts

Some unasked for thoughts on the current state of AI/LLM developer tooling: SOTA thinking models (e.g. Gemini 2.5 Pro and OpenAI o3) are *really* good at working through complex problems. Bugs that took me hours of reading docs/StackOverflow and trial-and-error now take a 10-20 minute convo... 1/

View full thread

Zed has some interesting ideas but their pace is a little slow compared to the rest of the space. "The entire LLM chat is a text buffer" is powerful in theory, but I haven't found use for it in practice. Maybe a skill issue. 9/

Some unasked for thoughts on the current state of AI/LLM developer tooling: SOTA thinking models (e.g. Gemini 2.5 Pro and OpenAI o3) are *really* good at working through complex problems. Bugs that took me hours of reading docs/StackOverflow and trial-and-error now take a 10-20 minute convo... 1/

View full thread

Aider is decent and obviously being able to use other models is a key feature, but atm the claude models are still strongest for agentic coding and claude code is so deeply integrated it's hard to compete. 7/

Speaking of Claude, gonna echo the sentiments of others and say that I'm not sure Claude 3.7 was much of an upgrade. It can be a bit cleverer, but also is prone to go off on a tangent and stick a bunch of dumb stuff in my codebase if there are issues with the one-shot solution. 8/

Some unasked for thoughts on the current state of AI/LLM developer tooling: SOTA thinking models (e.g. Gemini 2.5 Pro and OpenAI o3) are *really* good at working through complex problems. Bugs that took me hours of reading docs/StackOverflow and trial-and-error now take a 10-20 minute convo... 1/

View full thread

The chat sidebars in IDEs (Cursor/Cline/Roo/Copilot/etc) work well for single-file work, but they struggle when it comes to tasks that need agenting all over the codebase (especially an existing complex codebase). For that I prefer the agentic CLIs like claude code. 5/

Claude code is really great at running through an entire code-base and making changes, as long as they are not too complex. e.g. you can just say "add a new API route" and it will not only do that but also make sure all the plumbing is there as well, like auto-writing tests and such... 6/

Some unasked for thoughts on the current state of AI/LLM developer tooling: SOTA thinking models (e.g. Gemini 2.5 Pro and OpenAI o3) are *really* good at working through complex problems. Bugs that took me hours of reading docs/StackOverflow and trial-and-error now take a 10-20 minute convo... 1/

View full thread

Generalizing for all of these tools, they are a much better coding companion for greenfield projects that they helped create (i.e. vibe coding) than they are at interacting with projects that have been built by messy humans over years. 3/

Cursor still has the best/smartest tab-complete in the game. Copilot has been making progress but Cursor's autocomplete knows what I want to do next 95% of the time, esp when it comes to any sort of re-factoring. 4/

Some unasked for thoughts on the current state of AI/LLM developer tooling: SOTA thinking models (e.g. Gemini 2.5 Pro and OpenAI o3) are *really* good at working through complex problems. Bugs that took me hours of reading docs/StackOverflow and trial-and-error now take a 10-20 minute convo... 1/

I didn't feel this way about previous models because there were still too many hallucinations when dealing with hairy problems. CoT seems to do a great job of cutting down on that. DeepSeek R1 was pretty good at this too and I used it for a while, but the above two have eclipsed it imo. 2/

You are up to date.