Brillant breakdown of the model selection economics. The GLM-4.7 comparison really nails how specialized models outperfrom general ones at specific tasks while costing way less. I've seen similar patterns where a cheaper model with tight context produces more reliable outputs than larger models with creative freedom. The mode:all config solving the duplicate agent problem is such a clean solution.
thanks! as the quality of SLMs increase for specific tasks, i'd be curious to see how much of dev workflows slowly evolve into using a mix of multiple models instead of a few big ones. I'm testing some models like qwen-coder-30b with 128k window and its doing ok for most low-creativity-repetitive work.
Great detail on the tradeoffs as well as the comparison between models! I’m yet to try OpenCode (still working on getting the best out of Claude Code) but this is a great read
It's interesting how you frame the discussion around developer setups, placing such a clear emphasis on personal journey and optimization over direct recommendations. What if this level of tailored complexity, while undoubtedly maximising individual efficiency, inadvertently introduces unique challenges regarding onboarding or cross-functional support in a collaborative team environment?
Valid point. I believe strongly that engineers will be left with no choice but to do this to be effecient and innovative at their work. Since agents are writing the actual code, engineers are expected to communicate their thoughts clearly to get high quality output. Even the best context engineering recommendations/hacks/templates dont work for everyone because each engineer thinks differently. I'm sure that the chief of staff agent I've created for myself will not be effective for anyone else and might actually be more harmful. What's fungible, though, is the how (process, tools, workflow , best practices, learnings).
Now, addressing team challenges is another huge challenge. I wrote about it in my earlier post (https://www.devashish.me/p/why-5x-engineers-dont-make-5x-teams), but we're trying to address this by codifying the knowledge and streamlining a lot of repetivie processes. wip
Brillant breakdown of the model selection economics. The GLM-4.7 comparison really nails how specialized models outperfrom general ones at specific tasks while costing way less. I've seen similar patterns where a cheaper model with tight context produces more reliable outputs than larger models with creative freedom. The mode:all config solving the duplicate agent problem is such a clean solution.
thanks! as the quality of SLMs increase for specific tasks, i'd be curious to see how much of dev workflows slowly evolve into using a mix of multiple models instead of a few big ones. I'm testing some models like qwen-coder-30b with 128k window and its doing ok for most low-creativity-repetitive work.
thank you for making us understand 'whats the difference in opencode - under the hood version' details.
Great detail on the tradeoffs as well as the comparison between models! I’m yet to try OpenCode (still working on getting the best out of Claude Code) but this is a great read
It's interesting how you frame the discussion around developer setups, placing such a clear emphasis on personal journey and optimization over direct recommendations. What if this level of tailored complexity, while undoubtedly maximising individual efficiency, inadvertently introduces unique challenges regarding onboarding or cross-functional support in a collaborative team environment?
Valid point. I believe strongly that engineers will be left with no choice but to do this to be effecient and innovative at their work. Since agents are writing the actual code, engineers are expected to communicate their thoughts clearly to get high quality output. Even the best context engineering recommendations/hacks/templates dont work for everyone because each engineer thinks differently. I'm sure that the chief of staff agent I've created for myself will not be effective for anyone else and might actually be more harmful. What's fungible, though, is the how (process, tools, workflow , best practices, learnings).
Now, addressing team challenges is another huge challenge. I wrote about it in my earlier post (https://www.devashish.me/p/why-5x-engineers-dont-make-5x-teams), but we're trying to address this by codifying the knowledge and streamlining a lot of repetivie processes. wip