Some LLM APIs let you give a schema or regex for the answer. I think it works because LLMs give a probability for every possible next token, and you can filter that list by what the schema/regex allows next.

I'm sure the reason is the plethora of markdown data is was trained on. I personally use ``` stuff.txt ``` extremely frequently, in a variety of places.

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.

Very common struggle, but a great way to prevent that is prefilling the assistant response with "{" or as much JSON output as you're going to know ahead of time like '{"response": ['

I think this is the first time I stumped upon someone who actually mentions LSTM in a practical way instead of just theory. Cool!

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?

I had similar issues with local models, ended up actually requesting the backticks because it was easier this way, and parsed the output accordingly. I cached a prompt with explicit examples how to structure data, and reused this over and over. I have found that without examples in the prompts some llms are very unreliable, but with caching some example prompts this becomes a non-issue.

Funny, I do the same. Additionally, one can define a json schema for the output and try to load the response as json or retry for a number of times. If it is not valid json or the schema is not followed we discard it and retry.

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 most mainstream APIs by now have a way for you to conform the generated answer to a schema.

I do use backticks a lot when sharing examples in different format when using LLMs and I have instructed them to do likewise, I also upvote whenever they respond in that matter.

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

Yeah, that’s infuriating. They’re getting better now with structured data, but it’s going to be a never ending battle getting reliable data structures from an LLM.

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

Nooo you are definitely not alone. gpt-5-nano even is slowest model I’ve used since like 2023, second only to gpt-5-mini

https://phrasing.app

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=...

Here is an example of the formatting I desired: https://x.com/barrelltech/status/1963684443006066772?s=46&t=...

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)

I provide it with data and ask it to convert it to prose in specific formats.

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

It’s user facing, so will just have an option for users to regenerate the explanation. It happens rarely enough that it’s not a huge issue, and normally doesn’t effect content (I think once I saw it go a little wonky and end the sentence with a few random words). Just sometimes switches to mono space font in the middle of a paragraph, or it “spells” a word wrong (spell is in quotes because it will spell `chien` as `chi§en`).

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

I did not! It’s not on azure yet and I’ve still got some credits to burn. That’s exciting though, hopefully it will iron out this weird ghost character issue.

What's their unique value? How are they differentiated vs. OAI/Anthropic/etc. who have way more money/distribution/etc.?

Lovable has the worst moat I have ever seen for a company.

Our engineer used lovable for about a day, then just cloned the repo and used cursor since it it was much more productive.

Aren't there a billion Lovable clones now that do the exact same thing?

I could never get anything useful out of Lovable and was frustrated with the long editing and waiting process.

I'd prefer a site builder template with dropdowns. Lovable feels like that type of product, just with an LLM facade.

I don't hate AI, I just wasn't getting into the groove with Lovable.

I just couldn't love it, and frankly I don't get the hype around it. I recently found that all my use cases can be served by either:

1. A general purpose LLM chat interface with high reasoning capacity (GPT-5 thinking on web is my go to for now)

2. An agent that has unrestricted token consumption running on my machine (Claude Code with Opus and Amp are my go to for now).

3. A fine-tuned, single purpose LLM like v0 that is really good at one thing, in this case at generating a specific UI component with good design aesthetics from a wireframe in a sandbox.

Everything else seems like getting the worst of all worlds.

>I live in the US, but if I were a European, I would be all in on supporting Mistral. Strengthen your own country and region

That's a bit of a double edged sword. My support goes as far as giving local offerings a try when I might not have done otherwise. But at that point they need to be able to compete on merit.

https://proton.me/support/lumo-privacy

> Lumo is powered by open-source large language models (LLMs) which have been optimized by Proton to give you the best answer based on the model most capable of dealing with your request. The models we’re using currently are Nemo, OpenHands 32B, OLMO 2 32B, and Mistral Small 3. These run exclusively on servers Proton controls so your data is never stored on a third-party platform.

Mistral small and large are open weight, so they are likely self hosting?

I don't think France would allow that to happen - they would block it on national interest grounds.

I find that American AI companies are being incredibly wasteful - they've famously been shown up by frugal Deepseek, but generally smarter architectures outweigh more raw resources.

Maybe also transaction search (helpful for customer support), just the currency conversion ratio or balances, helpful for accounting etc. lots of read only usecases around

I got Perplexity Pro one year for free, what makes Opus 4.1 search so good? Is it worth switching?

By research you mean web search/deep research or only use knowledge embedded in llm? I use ChatGPT most of time, didn't find Claude work better for me, maybe I should switch if there is a big gap in performance.

After testing it for myself, I find it's pretty solid for a lot of general use cases but basically breaks as soon as it needs to make multiple tool calls.

A bit of a shame if they're going to launch these connectors that the model isn't suitable for them.

You're right!

Yes on the single market being a huge issue, but the other "cultural" differences are total BS, and just modern media narratives.

Western Europe and the US had essentially the same level of government safety nets (and government spending and economic growth) from the 1950s to the 1990s.

Who do you think started all of the European industrial giants that are still globally competitive today? Europe had no problem competing in the industrial revolution, if these were actually European values there would be no competitive European industry like there is no competitive European tech. It's only the digital revolution that Europe has struggled with.

China is communist in name only. In fact they are the most capitalist major country on earth right now.

Government spending makes up a smaller % of Chinese GDP than in the US, so definitionally their economy is more privatized than the US economy. China is at 33%, US at 36%, Europe is at 50%.

For every Jack Ma, there's a million other Chinese businesses flourishing in every niche imaginable with very little oversight from the CCP.

Not a big deal, but FYI there's an 'h' in the company name "Anthropic".

Inspired by the Greek word for human: Anthropos / ἄνθρωπος, the same etymology as English words like anthropology, the study of humans.

(I'd hazard a guess that your first language is something like a Romance language such as French, where people would pronounce that "anthro..." as if there is no h? So a particularly reasonable letter to forget when typing!)

I'm curious how relevant this actually is as a USP with the proliferation of open weight models and a glut of technical consultants.

So are Gemini Flash (Lite) and GPT mini/nano.

True, but we should also remember that some services like the fast responses and the image generations (may?) run in US data centres also for Mistral. So that part of the data, in principle, may end up in the ends of other extra European countries.

This said, I am really supportive of Mistral, like their work, and hope that they will get more recognition and more EU-centric institutional support.

how does the 'memory' feature in mistral work wrt GDPR if i type in my personal information ?

Is that due to usage (or lack thereof)?

"Eschew flamebait. Avoid generic tangents."

https://news.ycombinator.com/newsguidelines.html

I'm glad to see a European company succeeding here, especially since Mistral has released open weights models. But you're deluding yourself if you think Mistral is any more moral than its American counterparts.

how you come you are helping ycombinator by posting here then?

[flagged]