2 ms·

Fine, I’ll try GLM 5.2 since it has arrived in my OpenCode subscription. I haven’t used it much, but GLM 5.1 hasn’t blown me away. If it really is that good, a

Fine, I’ll try GLM 5.2 since it has arrived in my OpenCode subscription. I haven’t used it much, but GLM 5.1 hasn’t blown me away. If it really is that good, a minor release seems to undersell it. I found Anthropic’s .5 jump a bit silly, but it does its job in communicating the advancement.

aerodexis 4mo ago
How do you go about comparing models? Personally, I've found that the quality of output is overwhelmingly dictated by the quality of the input I give it - and that the quality of input differs massively depending on the task at hand, how much documentation exists, what kind of mood I'm in, etc. Hence, I'm beginning to believe that unless you're doing very specialized work - the model itself has become a commodity, incremental quality is irrelevant and that economics and privacy are what actually matter now.
- dsign 4mo ago
  > the model itself has become a commodity Indeed. So the winner is going to be whoever can also commoditize the rest, i.e. hardware and energy costs. In the long run, whoever has the factories to make chips, windmills and solar panels is going to be ahead. I think this is well understood by the American elites, and it can very well explain their exacerbated hawkishness.
- ignoramous 4mo ago
  > How do you go about comparing models? I do this often: I'd have a prompt in mind before I manually find the fix for subtle bugs, or plan out how I'd implement a feature. Then, I prompt the models I want to test (with the prompt I originally had in mind) and see if any of them get it right. To my surprise, the Claudes always perform better, and so, it seems like my prompting behaviour is either attuned to them, or Claudes are good at "figuring it out". That said, in my experiments, I've always found that given a "good" prompt (with just the right detail), the top coding models (including from Chinese labs, which I now increasingly prefer) have no discernable difference; in which case, I usually defer to the cheaper & fastest of those models (presently, that's either DeepSeek v4 Flash & Pro, MiMo v2.5 & Pro, MiniMax M3; these are a coding tier below GLM5.2, Kimi K2.7, and Qwen 3.7 Max, imo) as ranked by OpenRouter: https://openrouter.ai/rankings?view=month https://openrouter.ai/rankings?view=month
  - aerodexis 4mo ago
    wow - so you pmuch baked this kind of testing into your day to day work. Might not be a bad startup idea to create a harness that can do this automagically.

aerodexis 4mo ago

How do you go about comparing models? Personally, I've found that the quality of output is overwhelmingly dictated by the quality of the input I give it - and that the quality of input differs massively depending on the task at hand, how much documentation exists, what kind of mood I'm in, etc. Hence, I'm beginning to believe that unless you're doing very specialized work - the model itself has become a commodity, incremental quality is irrelevant and that economics and privacy are what actually matter now.

dsign 4mo ago
> the model itself has become a commodity Indeed. So the winner is going to be whoever can also commoditize the rest, i.e. hardware and energy costs. In the long run, whoever has the factories to make chips, windmills and solar panels is going to be ahead. I think this is well understood by the American elites, and it can very well explain their exacerbated hawkishness.
ignoramous 4mo ago
> How do you go about comparing models? I do this often: I'd have a prompt in mind before I manually find the fix for subtle bugs, or plan out how I'd implement a feature. Then, I prompt the models I want to test (with the prompt I originally had in mind) and see if any of them get it right. To my surprise, the Claudes always perform better, and so, it seems like my prompting behaviour is either attuned to them, or Claudes are good at "figuring it out". That said, in my experiments, I've always found that given a "good" prompt (with just the right detail), the top coding models (including from Chinese labs, which I now increasingly prefer) have no discernable difference; in which case, I usually defer to the cheaper & fastest of those models (presently, that's either DeepSeek v4 Flash & Pro, MiMo v2.5 & Pro, MiniMax M3; these are a coding tier below GLM5.2, Kimi K2.7, and Qwen 3.7 Max, imo) as ranked by OpenRouter: https://openrouter.ai/rankings?view=month https://openrouter.ai/rankings?view=month
- aerodexis 4mo ago
  wow - so you pmuch baked this kind of testing into your day to day work. Might not be a bad startup idea to create a harness that can do this automagically.