[Ferret7446]

I feel like this kind of tool has been completely obsoleted by LLMs (even local models). I've used similar tools such as tiny for Emacs, which is a DSL for generating text based on numeric ranges. Now, it's simply more efficient to ask AI.

[burntsushi]

Quite plausible. I built most of it in May 2025 (1 year ago) before I even looked at LLMs for coding, and announced it in Oct 2025[1]. I started it principally as a mechanism to dog-food Jiff[2]. Jiff has not been obsoleted by LLMs and is exactly the sort of thing an LLM would add as a dependency to a Rust project. This ended up being an extremely useful exercise because it did lead to some Jiff improvements.

Fun fact: ripgrep started as something to dog-food the regex crate (with a focus on performance). I didn't originally build it to release to end users funnily enough. To be clear, I'm not implying bttf will follow the same path. I honestly probably agree with you at this point. I wouldn't have 1 year ago though.

[1]: https://news.ycombinator.com/item?id=45608547

[2]: https://github.com/BurntSushi/jiff

[illiac786]

I believe AI-for-everything will become unsustainable financially for many and I’m genuinely curious to see how people deal with it. When to use it? When is it wasteful?

My big hypothesis is that tokens are going to get much more expensive. Either that or OpenAI/anthropic are going bankrupt. I’m almost excited to find out, I have to admit.

Your remark just reminded me of this, I went a bit off topic, I admit.

[gandreani]

Have you tried DeepSeek V4 Flash? It's very competent and extremely cheap.

I think Gemma 4 is also a good example of a capable small model.

I mention these not only because they're cheap but because they can run on consumer devices. The "every year bigger and more capable SOTA model" trend is mirrored by "the every year smaller and more capable open source model" trend.

[illiac786]

256GB is what deepseek v4 flash with Q4 requires I believe. It is really still very far from “running locally on your device”. And it’s getting further away every day, looking at how the electronic market prices are surging.

I need to find stats on average RAM of personal devices, but I expect it will be so low, we are light years away from running a frontier model (from today) locally on a smartphone, let’s stop dreaming (and I really would love having it).

I do agree local models are progressing and I am to this day in awe at what a 50GB file can do – it still feels like black magic to me.

Also granted, something like Gemma 2 2B seems to have similar performance to ChatGP 3.5 and only require 2GB of RAM. But I think the RAM/performance ratio curve over time is logarithmic and not linear, it’s moving slower and slower.

[pchristensen]

Won't the most efficient LLMs just learn about and use tools like this, instead of crunching all the tokens to do it themselves?

[Ferret7446]

Not really. Similarly to humans, it's easier to write code than read it; LLMs aren't particularly less efficient at doing off the cuff tasks vs reading in all of the tokens of documentation for a specific tool (assuming you did the legwork of manually feeding that tool to the LLM, which makes it less efficient for you) and then coming up with a solution. And that argument would only apply if there were custom tools for each of the infinite possible tasks, which there aren't.