Thanks for the read. It seems to confirm that resource limits are an important factor for terminal benchmarks:
> The extra resources enable the agent to try approaches that only work with generous allocations, such as pulling in large dependencies, spawning expensive subprocesses, and running memory-intensive test suites.
> An agent that writes lean, efficient code very fast will do well under tight constraints. An agent that brute-forces solutions with heavyweight tools will do well under generous ones. Both are legitimate things to test, but collapsing them into a single score without specifying the resource configuration makes the differences—and real-world generalizability—hard to interpret.
So changing the resource limits changes the benchmark. Yet their score table claims their score to be for Terminal-Bench 2.1, not Terminal-Bench 2.1 with raised limits.
I also love exploring ideas like this. The problem that made me stop using Gemini is that it would always try really hard against all my prompting to ask a stupid question at the end of its response that would completely derail my train of thought.
I love the UX of voice mode. I’ve always hated voice models.
Gemini’s being particularly egregious (always ending in some deranged question, can not reliably be prompted away) prompted me to build my own client for my real harness that simply does STT -> model -> TTS (both being independently useful).
I guess I see some value in a model responding quickly and with more nuance, but it’s not much. I can wait for it to finish. I’d much rather have it be actually useful. I’m not looking for a digital friend.
The delegation feature lets me see some value in a voice model for orchestration type features. But in either case, I don’t really like (or understand why others would like) talking to a model with different features and quirks just because I’m using a different medium to communicate over.
Instacart sells data to Papa Johns for better pizza ad timing.
Papa Johns isn’t somehow able to predict this based on their own data (which is the impression I initially got from the headline), they are buying it from Instacart.
Deepseek regularly does this for me, at least when actively exploring something (multi-turn). It massively varies its reasoning budget based on the task.
This is such a classic video. I love showing it to people. It even made its way into non-software circles due to its satisfying sounds, here is a big streamer (who has nothing to do with software) reacting to it: https://youtu.be/4ItTg2ibo7c
Finally I get to tell one of these personal stories: I met Timo at a GPN in Karlsruhe because he had a cool LED strip project on him. He showed me some videos of other projects he had done and they had that distinct sound ...
The ones playing political games dressing them up as technical ones are the worst:
“My approach is technically correct and I won’t change it even though it causes issues down the line”. I’ve seen this a lot in the Gnome/Wayland world.
> I don't think people get into kernel maintenance to play politics.
I’m not a kernel developer and the projects I’ve worked on haven’t even been that big, but even there it was necessary to cater to multiple stakeholders and consider multiple viewpoints. I’d go as far as saying that software development in general gets pretty political pretty quickly, as soon as you depend on somebody else’s work or somebody else depends on yours. Every decision will impact somebody and different options will do so differently - these are political considerations.
I can’t imagine this being less of an issue in the kernel.
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.
Yeah. The output is magic either way, with or without reasoning.
For daily use I actually like the reasoning summary to be brief/quick to scan.
That said, I understand the author’s desire for the real thing. It just feels better to have that access, especially when Anthropic will give it to you, but encrypted.
Meh. I‘d be happy if my AI built my program just the way that I want it, but it turns out to be a lot of work go get it to do that (Deepseek, Claude might be a bit less involved).
I still have to make a bunch of decisions based on its research, so it still turns out to be hours of work that can go into it.
Granted, I can probably optimize some things by encoding my style preferences. But so far it has been a lot more involved than „make me an app, make no mistakes“.
That‘s why I‘d call it in a large part my work. It‘s my vision, in the same way an architect that has never personally laid a brick can call a building „their work“.
Why not, really? People have different (though likely conflicting) views on the company and it’s an interesting (and potentially profitable) space to follow.
I‘d argue it’s more meaningful than sports, which makes it more interesting to follow.
Of course the sports aspects are silly, but maybe they‘re just a fun way to keep track. Happens to other popular companies in the tech sector as well.
> The extra resources enable the agent to try approaches that only work with generous allocations, such as pulling in large dependencies, spawning expensive subprocesses, and running memory-intensive test suites.
> An agent that writes lean, efficient code very fast will do well under tight constraints. An agent that brute-forces solutions with heavyweight tools will do well under generous ones. Both are legitimate things to test, but collapsing them into a single score without specifying the resource configuration makes the differences—and real-world generalizability—hard to interpret.
So changing the resource limits changes the benchmark. Yet their score table claims their score to be for Terminal-Bench 2.1, not Terminal-Bench 2.1 with raised limits.