I do find value in LLMs as a tool for specific uses, but no way I’m paying through the nose for an expensive subscription to support greedy, shithead tech bros plundering the world for profit and government agencies using the tech to spy on everyone.
Open weight models running on consumer hardware are getting pretty good, though. After learning some tricks from the Codacus YouTube channel, I’m running the Qwen 3.6 MoE model on my 4060 gaming laptop with 200k context and usable performance for about the same electricity cost as playing a game. Open source agents like OpenCode are also getting really good. I’m using the beta desktop version of OpenCode and it’s pretty close in capability and UX to the Claude agent I am forced to use at work.
There are still some rough edges and performance issues, but the local LLM scene is already quite decent and is rapidly evolving. Qwen 3.7 open weight models are supposed to be released soon and I’m also looking for things like TurboQuant, speculative prefill (PFlash), etc. to mature and smooth the rough edges.
Anyone outsourcing critical thinking to LLMs is using them wrong. With software development, for example, quality is speed. Yes, a codebase riddled with technical debt will always collapse under its own weight, regardless of whether humans or LLMs wrote the shitty code, so using LLMs to generate slop that senior engineers aren’t carefully reviewing, or letting LLMs make architectural decisions is the wrong approach. However, skillful use of LLMs can actually yield better quality code than humans alone. Here are some examples:
LLMs often find issues when reviewing PRs that the human reviewers missed in my experience. Different, focused agents can be used to review code from different perspectives, like security, performance, etc.
Test coverage can be higher with LLMs automating the test writing. They can often also come up with useful test cases that humans didn’t consider and also make it possible to maintain more varied types of tests that the team otherwise might not take on.
LLMs are great at fixing lint issues. I’ve seen codebases where lint warnings accumulate because only the errors are getting fixed, but LLMs can fix everything quickly.
Once solid patterns have been established in a codebase by senior engineers and documented in agent MD files, LLMs can implement features and follow the existing patterns for routine features and bug fixes without getting “creative” and introducing new tools and patterns that complicate things out of boredom.
LLMs can be used to look at codebases for open source dependencies to better understand how they work in order to utilize them better
LLMs can suggest alternative libraries, approaches, patterns, etc. with the pros and cons of each. While humans still need to do their own research, this can often be a useful starting point that helps make more informed decisions
You can use Claude code with your own LLM too. It seems to be better at automatically creating memories and stuff. Downside is of course that it’s closed source.
Codex is also open source and usable with your own LLMS. Same for Mistral Vibe but that’s been much less useful for me, opencode with their free or Go plan models is much better. Opencode Go with Deepseek Flash gets you so much usage it’s hard to run ouy at 10 bucks a month
I do find value in LLMs as a tool for specific uses, but no way I’m paying through the nose for an expensive subscription to support greedy, shithead tech bros plundering the world for profit and government agencies using the tech to spy on everyone.
Open weight models running on consumer hardware are getting pretty good, though. After learning some tricks from the Codacus YouTube channel, I’m running the Qwen 3.6 MoE model on my 4060 gaming laptop with 200k context and usable performance for about the same electricity cost as playing a game. Open source agents like OpenCode are also getting really good. I’m using the beta desktop version of OpenCode and it’s pretty close in capability and UX to the Claude agent I am forced to use at work.
There are still some rough edges and performance issues, but the local LLM scene is already quite decent and is rapidly evolving. Qwen 3.7 open weight models are supposed to be released soon and I’m also looking for things like TurboQuant, speculative prefill (PFlash), etc. to mature and smooth the rough edges.
Have you considered not outsourcing critical thought to a glorified autocorrect?
I dunno, just a thought.
You’re part of the problem.
Anyone outsourcing critical thinking to LLMs is using them wrong. With software development, for example, quality is speed. Yes, a codebase riddled with technical debt will always collapse under its own weight, regardless of whether humans or LLMs wrote the shitty code, so using LLMs to generate slop that senior engineers aren’t carefully reviewing, or letting LLMs make architectural decisions is the wrong approach. However, skillful use of LLMs can actually yield better quality code than humans alone. Here are some examples:
You can use Claude code with your own LLM too. It seems to be better at automatically creating memories and stuff. Downside is of course that it’s closed source.
Codex is also open source and usable with your own LLMS. Same for Mistral Vibe but that’s been much less useful for me, opencode with their free or Go plan models is much better. Opencode Go with Deepseek Flash gets you so much usage it’s hard to run ouy at 10 bucks a month