DeepMind 员工谈何时用 Gemini Argon、何时用 3.8 Flash

Bogdan Vko @BogdanVko原推发布于 2026.10.03

在 X 上打开原推

I do not always use Gemini Argon, even though I have unlimited & free access to it. Working at Google DeepMind, I choose Argon when I need to perform the initial design phase, as well as critical de-risking exploration/POCs. The goal is to front-load the highly cognitive work and go/no-go decisions, the critical stuff. Later when implementing discrete sections of work (and splitting massive commits into reviewable chunks) I will use Gemini 3.8 Flash...because it is just so darn fast! At this point the project is well-defined and scoped. Another example where I won't use Argon is agentic testing: "Validate that the UI changes work end-to-end by recording me a video demo of the entire feature with local changes" Why would I wait for a big and brainy model to do this easy work if Flash does the job and is much faster at that? Same goes for my agentic review tooling that main agents call. For example, my "minimalist" SKILL scans a given technical doc for complicated sentence structures; the model does not need to be a genius, it just needs to follow a straightforward SKILL I had written for it, which Flash can do just fine (and fast). Of course, some agentic tooling I have needs a big model for that extra reasoning boost. Take my "abstraction-police" SKILL meant to review written code and suggest better abstractions ("borrowed" from @karpathy 😁). I pair this skill with Argon (before I used Opus 5.5/5/4.8/etc), as the ability to navigate ambiguity and make the right calls is paramount. In my opinion, it is important to pick the right tool for the job.

这条是引用着另一条写的：x.com/spicey_lemonade/status/2105775773164638398