Would you like to elaborate further on how the experience was with it? What was your approach for using it? How did you generate synthetic data? How did it perform?

This is maybe more maybe less insidious. It will literally just insert a random character into the middle of a word.

I work with an app that supports 120+ languages though. I give the LLM translations, transliterations, grammar features etc and ask it to explain it in plain English. So it’s constantly switching between multiple real, and sometimes fake (transliterations) languages. I don’t think most users would experience this

I got this format from writing markdown files, it’s a nice way to share examples and also specify which format it is.

In slack/teams I do it with anything someone might copy and paste to ensure that the chat client doesn't do something horrendous like replace my ascii double quotes with the fancy unicode ones that cause syntax errors.

In readme files any example path, code, yaml, or json is wrapped in code quotes.

In my personal (text file) notes I also use ``` {} ``` to denote a code block I'd like to remember, just out of habit from the other two above.

It also helps with having a field of the json be the confidence or a similar pattern to act as a cut for what response is accepted.

I think they are saying that if highest probability phrase fails the regex, the LLM is able to substitute the next most likely candidate.

e.g. DOMINO https://arxiv.org/html/2403.06988v1

It’s pretty rare though. Really solid model, just a few quirks

Would you be willing to share an example prompt? I'm curious to see what it'sesponding well to.

Mistral medium is ranked #8 on lmsys arena IIRC, so it’s probably just not your style?

I’m also comparing this to gpt-5-mini, not the big boy

For example, I ask, what are the most common targets of removal in magic: the gathering? Mistral's answer is so-so, including a slew of cards you would prioritize removing, but also several you typically wouldn't, including things like mox amber, a 0 cost mana rock. Gemini flash gave far fewer examples, one for each major card type type, but all of them are definitely priority targets that often defined an entire metagame, like Tarmogoyf.

Was looking to both decrease costs and experiment out of OpenAI offering and ended up using Mistral Small on summarization and Large for the final analysis step and I'm super happy.

They have also a very generous free tier which helps in creating PoCs and demos.

I have no affiliation with mistral, I just have recent experience with them and wanted to share

Also @mistral hmu if you want to arrange something!

I don’t write much about elixir (or clojure) as I’m terrible at technical writing, but I am a die hard fan of both languages

I’m making an app to learn multiple languages. This portion of the pipeline is about explaining everything I can determine about a work in a sentence in specifically formatted prose.

Example: https://x.com/barrelltech/status/1963684443006066772?s=46&t=...

I'd love to try Phrasing, but there is no way that I'm going to give my credit card information if I've never seen how it works. I'm not asking for special treatment. I think the 4 Euro 14 day trial should be free. Not because I don't want to pay, but because I don't want my credit card details shared with every service I ever try.

Yes, I understand that your credit card provider is safe and that is not my worry. It's just that handing over that information is a very large hurdle, and one that users get over only after you built trust with them. If we can never see what we're getting, we have no reason to get over that hurdle and develop that trust.

My main task this week (besides QAing the new setup flow) is updating examples on the home page and in the blog. It’s all a one person show, so things move a little slowly, but I’m working as fast as I can :)

The paid trial gets you credits at cost — Phrasing is bootstrapped by me, and there’s no way I can afford a $4 loss leader for everyone who signs up. I would have to 100x the subscription prices (or more) in order to afford that.

There is a way to use the new setup process to “try” the application without spending money, but that’s not an avenue I want to intentionally funnel users through, so you won’t see any copy around it. Feel free to get in touch via the app though if you need more details.

I have plans for a free trial eventually, but its blocked by a few key product developments, so it will be a year at least before that’s an option!

Is there an example you can show that tended to fail?

I’m curious how token constraint could have strayed so far from your desired format.

Yes I use(d) structured output. I gave it very specific instructions and data for every paragraph, and asked it to generate paragraphs for each one using this specific format. For the formatting, I have a large portion of the system prompt detailing it exactly, with dozens of examples.

gpt-5-mini would normally use this formatting maybe once, and then just kinda do whatever it wanted for the rest of the time. It also would freestyle and put all sorts of things in the various bold and italic sections (using the language name instead of the translation was one of its favorites) that I’ve never seen mistral do in the thousands of paragraphs I’ve read. It also would fail in some other truly spectacular ways, but to go into all of them would just be bashing on gpt-5-mini.

Switched it over to mistral, and with a bit of tweaking, it’s nearly perfect (as perfect as I would expect from an LLM, which is only really 90% sufficient XD)