Not sure how accurate this is, but found contextarena benchmarks today when I had the same question.
It appears only gemini has actual context == effective context from these. Although, I wasn't able to test this neither in gemini cli, nor antigravity with my pro subscription because, well, it appears nobody actually uses these tools at Google.
Claude is Anthropic's property which they rent to the government. Is there any other place where rental agreements don't come with clauses on how the property can and can't be used?
Enforce this with deterministic guardrails. Use strictest linting config you possibly can, and even have it write custom, domain specific linters of things that can't happen. Then you won't have to hand hold it that much
Yeah, people seem to forget one of the L's in LLM stands for Language, and human language is likely the largest chunk in training data.
A cli that is well designed for humans is well designed for agents too. The only difference is that you shouldn't dump pages of content that can pollute context needlessly. But then again, you probably shouldn't be dumping pages of content for humans either.
It's not obvious that human language is or should be the largest amount of training data. It's much easier to generate training data from computers than from humans, and having more training data is very valuable. In paticular, for example, one could imagine creating a vast number of debugging problems, with logs and associated command outputs, and training on them.
I learned TS after a few years with JS. I thought having strict types was cool. Many of my colleagues with much more (JS) experience than me thought it was a hassle. Not sure if they meant the setup or TS or what but I always thought it was weird.
Isn't that NVME also very expensive to replace because it's tied to hardware identifiers? If you keep swapping all the time, surely NVME would be the first part to fail
> But you are forgetting your skills (seen it first hand), and you're not learning anything new.
This is just false. I may forget how to write code by hand, but I'm playing with things I never imagined I would have time and ability to, and getting engineering experience that 15 years of hands on engineering couldn't give me.
> Your next interview won't be testing your AI skills.
Which will be a very good signal to me that it's not a good match. If my next interview is leetcode-style, I will fail catastrophically, but then again, I no longer have any desire to be a code writer - AI does it better than me. I want to be a problem solver.
> getting engineering experience that 15 years of hands on engineering couldn't give me.
This is the equivalent of how watching someone climb mountain everest in a tv show or youtube makes you feel like you did it too. You never did, your brain got the feeling that you did and it'll never motivate you to do it yourself.
This is only true for fully unsupervised "vibe coding". But you'll find this will not work for anything beyond a basic todo list app.
You'll free up your time from actually writing code, but on the other hand, you'll have to do way more reading, planning, making architectural decisions etc. This is what engineering feels like should be.
If you want an answer to the OP question, just ask AI to analyze the session jsonl files in your user directory and give you statistics of what's in there.
You'll find that at least half of it is noise.
If you put that in commits, you lose the ability to add "study git commits to ground yourself" in your agents.md or prompts. Because now you'll have 50%+ noise in your active session's context window.
Context window is precious. Guard it however you can.
It appears only gemini has actual context == effective context from these. Although, I wasn't able to test this neither in gemini cli, nor antigravity with my pro subscription because, well, it appears nobody actually uses these tools at Google.
https://contextarena.ai/?showLabels=false
reply