Rendered at 12:26:38 GMT+0000 (Coordinated Universal Time) with Cloudflare Workers.
dchftcs 58 minutes ago [-]
What are "extensible strategy shapes" for those who don't speak LLM?
adrian_b 7 hours ago [-]
The comparison results seem very plausible.
From the conclusion, I agree with:
> I wouldn't make either one the top-level coordinator by default.
But I do not agree with the follow-up sentence:
> The best shape is still a frontier coordinator or judge above them: GPT-5.5 or Claude Opus deciding what to delegate, checking the finished work, and rerunning narrow pieces when the answer looks wrong. These models make the worker layer much more serious, not the coordinator layer unnecessary.
For the coordinator or judge above them I would put myself, not a too expensive LLM under the control of an external entity, achieving thus simultaneously higher quality, lower cost and greater security.
incrudible 2 minutes ago [-]
> For the coordinator or judge above them I would put myself
You will not be able to keep up with the sheer volume, or alternatively you're never gonna ingest as much information as the LLM, so you're gonna miss out. Input tokens are relatively cheap.
Think of yourself as the CTO, they can't possibly make a judgement call on every detail, but an LLM can, and if you're gonna let an LLM do that, might as well go with frontier, and if you're not gonna let an LLM do that, you're stuck with whatever the lower-tier LLMs provided you with.
That doesn't mean you shouldn't read or judge the code at all, but you're still gonna want to use the LLM as the lever.
throwa356262 6 hours ago [-]
A lot of LLM discussions is driven by people who cannot code themselves.
There are multiple AI influencers on youtube who can't code 5 lines of python to save their lives. But they do own 3 DGX spark and a stack of maxed out mac minis...
(Not complaining, AI is supposed to be democratic)
throwa356262 5 hours ago [-]
Since I quit my Claude subscription, every month I spend $20 (the cost of CC pro plan) playing around with new models and new providers.
Currently testing M3 for agentic tasks. It works OK and their token plan is very cheap. Highly recommend for claw / hermes type of work.
Tested GLM 5.1 for coding last month and it burned through my tokens a bit too quickly, but it worked well enough.
scottchiefbaker 14 hours ago [-]
FWIW Opencode Go is giving 3x MiniMax M3 access right now. According to their chart you get almost 10x as much access to MM3 vs GLM 5.2.
Considering how close the models are, the extra free queries may be worth it.
oceanwaves 9 hours ago [-]
Yes, that's what I'm finding too. There seems to be a concerted promotional pricing campaign tied to M3's release across providers. Since their differences are subtle, it makes a lot of sense to fan-out to M3.
mt42or 1 hours ago [-]
All software benchmark are bullshit currently because none mesure capacity of doing same tasks after 1000 first warmed commit of random stuff. It's always easier to build something from scratch but nobody rebuild their feature from 0 every day.
oceanwaves 20 hours ago [-]
GLM 5.2 edges as the safer pick when tasks are more challenging from-scratch builds and the result needs to arrive as a complete, runnable project. MiniMax M3 is the value pick for a lot of worker traffic.
ashenke 8 hours ago [-]
I'd love to see a comparison with both Deepseek v4 models as well
killingtime74 7 hours ago [-]
I've used both and they are great. Would be better to have a GPT or Opus benchmark
From the conclusion, I agree with:
> I wouldn't make either one the top-level coordinator by default.
But I do not agree with the follow-up sentence:
> The best shape is still a frontier coordinator or judge above them: GPT-5.5 or Claude Opus deciding what to delegate, checking the finished work, and rerunning narrow pieces when the answer looks wrong. These models make the worker layer much more serious, not the coordinator layer unnecessary.
For the coordinator or judge above them I would put myself, not a too expensive LLM under the control of an external entity, achieving thus simultaneously higher quality, lower cost and greater security.
You will not be able to keep up with the sheer volume, or alternatively you're never gonna ingest as much information as the LLM, so you're gonna miss out. Input tokens are relatively cheap.
Think of yourself as the CTO, they can't possibly make a judgement call on every detail, but an LLM can, and if you're gonna let an LLM do that, might as well go with frontier, and if you're not gonna let an LLM do that, you're stuck with whatever the lower-tier LLMs provided you with.
That doesn't mean you shouldn't read or judge the code at all, but you're still gonna want to use the LLM as the lever.
There are multiple AI influencers on youtube who can't code 5 lines of python to save their lives. But they do own 3 DGX spark and a stack of maxed out mac minis...
(Not complaining, AI is supposed to be democratic)
Currently testing M3 for agentic tasks. It works OK and their token plan is very cheap. Highly recommend for claw / hermes type of work.
Tested GLM 5.1 for coding last month and it burned through my tokens a bit too quickly, but it worked well enough.
Considering how close the models are, the extra free queries may be worth it.