> one lab spends $$$ on bleeding edge R&D and expensive RL runs to improve capabilities, and other labs just yoink the raw reasoning traces and mid-train/post-train on them to get 90% of the way there for a small fraction of the cost.
"You're trying to kidnap what I've rightfully stolen."
> My favorite part of the new Groks has been how they speak in plain english. I simply cannot stand Claudish.
Grok has its own feel too. It's not as bad as Claude, but one of the things that bugs me is that it is far too terse.
It regularly seems to come up with terms and descriptions for things in its chain of reasoning and then uses these terms in its output assuming you understand what it's talking about.
I find I often have to ask it to re-explain what it means.
> It regularly seems to come up with terms and descriptions for things in its chain of reasoning and then uses these terms in its output assuming you understand what it's talking about.
GPT does this all the time, too (both Sol and Astra). I constantly have to tell it to not use terms that were not part of the initial prompt.
> I constantly have to tell it to not use terms that were not part of the initial prompt.
Hah, yeah even when you put it in AGENTS.md or a skill.. constantly having to remind it.. "what does AGENTS.md" say about doing that?".. Thinking.. Thinking.. "Oh, it says I should never do that, I'll remember that next time.."
I don't know how people are managing hands-off with Sol/Astra over long horizon tasks. It will come up with 100 new concepts/terms and then use them as if you've been following every Edit call for 24 hours.
It’s probably an assumption baked into the training pipeline from that ancient time six months ago when we mostly were manually following each edit tool call.
I’m pretty sure the big bois don’t do it because it would undermine “confidence”.
Seeing a model output “Oh I should just delete blah. Wait blah is a production service, I shouldn’t touch that. Maybe I can gain access to blah? Oh the aws cli isn’t signed in to blah. I see kubectl has access to blah though! Wait, I should ask user permission first.”
Idk I feel like the more likely answer is to prevent distillation. Having the thinking is definitely better UX (oftentimes, I don’t know if Codex is just hanging, which it often does, or working in silence).
I stay much more hands-on when I'm using models that display full reasoning traces. And I tend to get more things done as a result, because I know exactly when it thought of a good solution that it talked itself out of because of some invalid assumption.
Could be something very stupid like - "I don't have ffmpeg available here. Should I install it? No, I can't. I'll proceed doing something that will take me 100x more tokens and wall clock just to avoid adding a dependency." I can then just stop and say - you've got nix flake there, just add it.
That's impossible with western models. The only way is to ask why it did something stupid when it already spent 50% of your weekly quota and produced millions lines of slop.
You can double click on the 'thinking' text and it will expand and you can read it. The problem is that it will often have multiple thinking/tool call sections and it can be a needle/haystack problem to find the one with the thinking you are interested in.
Ah, but you CAN see the thinking if you are willing to risk your account being banned. You just have to expose a "tool" with a specially crafted definition.
SpaceX is a defense contractor (I don't mean this in a bad way). When the whole DoD/Anthropic thing flared up, I can guarantee you that SpaceX.ai was the first company invited to take their place as DoD AI provider.
I suspect that they specifically train Grok to be able to work well with military personnel -- speaking the way they speak: brief, to the point, efficient communication. Personally I really like this. Claude sounds like some demented clown from the marketing department.
> It regularly seems to come up with terms and descriptions for things in its chain of reasoning and then uses these terms in its output assuming you understand what it's talking about.
Not really because you're not building a database or GUI app where using native elements & data structures help a lot with memory pressure.
TUI renderer is the one using the memory heavily so your terminal takes the heavy lifing. If you're managing the buffers and out-of-screen context good enough, Typescript can be pretty efficient.
I love opencode but it chews through memory on my 64gb MacBook Pro. Can’t have too many long running sessions because the memory use just slowly creeps up.
It’s not about the terminal at all which as you noted accounts for minimal isage. It’s all the internal chat and history and everything else the agent tracks - all of which are smallish (and largish) strings allocated on the heap.
I don’t have the same issues with rust based tuis.
Codex works great in opencode until it gets up to around 200k context. Then it starts doing things like:
me: Can you implement the next thing
OpenCode+Codex: Yep I'll do that next.
<does nothing and returns to prompt>
me: Well?
OpenCode+Codex: <starts implementing>
me: Looks good, let's fix this one issue.
OpenCode+Codex: Sure let's do that.
<does nothing and returns to prompt>
me: <bangs head against wall>
--
I've found the codex cli to be much better in this regard, it doesn't nearly derp out so much at higher token counts.
Opus is still my favourite model (I've found 4.6 specifically gives me the best results in OpenCode), but with all the shenanigans Anthropic is pulling, Codex is a close enough substitute.
I went back and forth between the two for months, often having both in various proportions. But I had enough reliability issues with Anthropic back in March timeframe that I just threw in the towel. I find GPT "boring" to work with but it's a steady hand and there's really nothing I throw at it that it can't do.
And yeah, I am supplementing with GLM 5.2 and have actually found it quite complimentary.
One of Codex's weaknesses is "excessive staging" -- basically it's quite cautious and pushes a very incremental approach. This is good for working in an established codebase (which most work is anyways). But for yeeting new projects, Claude always shone better for me (though it often left a mess of race conditions and unhandled negative cases that I had to clean up by hand or with codex)
GLM actually does pretty well in this regard, with the right prompting. It's more "creative" than GPT.
If you’re suggesting that the test favors those capable of arranging their thoughts and words before putting pen to paper then.... I’m not sure there’s a problem
"You're trying to kidnap what I've rightfully stolen."
reply