An unfortunate confusing title for Mistral's announcement of their first Text-To-Speech model. Apparently includes an open weights model, but also available on their Voxtral API.

Haven't had a chance to dig in yet or see if they offer voice tweaking / cloning, as they only seem to have a limited number of voices. But I'm definitely considering moving my current OpenAI voice workload over to Mistral.

Honestly, I think there's a big divide, and those of us who are using AI intensively might just be increasingly "going dark" & distancing ourselves from those "real world" people. It's becoming detrimental being around people who are so actively negative about AI. It feels like being around people who still insist the sun orbits the earth. Those people are actually happier believing what they believe, so why spend any more time trying to convince them they're wrong?

I spent 2024 on Mastodon and I absorbed their groupthink that AI was useless... I wish I could get that year of my life back. I wish I had that extra year headstart on AI compared to where I am now. So much of my coding frustrations that year that might have been solved from using AI. I am reluctantly back on X - I hate what has been done to Twitter, but that's where so much of the useful information on using AI is being shared.

Well, back to it. Claude has been building another local MCP server tool for me in the background.

> It feels like being around people who still insist the sun orbits the earth.

100% feeling this divide as well.

People that deny the benefit of AI in 2026... I can't even engage with them anymore. I just move on with my life. These people are simply not living in reality, it will catch up to them eventually (unfortunately.)

at which point they will learn to write markdown files too.

Did they make significant improvements in OCR 3? The quality I was getting from Mistral OCR 2 was nowhere near as good as what I could get from just sending the same files to Claude Sonnet via an API call.

I have been finding Voxtral useful though.

Too late to edit / update my comment, but I finally tried Mistral OCR 3 tonight on a PDF file I had. Results were good, and fast... but I actually got better quality output from sending it to Haiku 4.5 instead.

In particular, Haiku 4.5 detected some footers that were on every page and moved them to be the footer at the end of the entire document instead, so that the document read more fluently.

I imagine Mistral OCR 3 might have an edge on speed & pricing, but in my low volume / prioritizing-quality case, seems that Claude is still better than Mistral.

Not sure if this will be considered helpful, but if you include:

<link rel="alternate" type="application/atom+xml" href="https://www.unsungnovelty.org/index.xml" />

in the HEAD of the pages on your website, it makes autodiscovery of the RSS feed a bit easier - not just for crawlers, but also for people with RSS plugins in their browser. It will make the RSS icon appear in their browser's URL field for easy subscription. Took me a while to find the RSS link at the bottom of your pages!

Thanks. Lemme look into that and will take the necessary actions.

There's two criteria, I agree it's hard to skim:

* The blog must have a recent post, no older than 12 months, to meet the recency criteria for inclusion.

* Criteria for posts to show on the website: Blog has recent posts (<7 days old), The website can appear in an iframe

The latter criteria is for the website / post to appear in Kagi's random Small Web feature, where they display the blog post in an iframe. (So I think only posts from the last week are displayed there.) Being on the list should ensure that any new posts could be displayed in Small Web though, and presumably that the website is indexed in Kagi's Teclis index as well. At least, I really hope that the Teclis index is including all of those old blog posts too, and not discarding them.

EDIT: I just realized freediver actually is Vladimir - I'd love to know if Teclis does index all those older blog posts too. I assume it does index everything that is still present in the RSS feeds?

Thank you. I swear I read that three times and missed the other criteria until you pointed it out and then I found it. :/

If that's an issue, and if you don't mind building something out yourself, Marginalia have an excellent API that you can connect to from your own personal non-Javascript meta-search engine. I did that, and I find Marginalia awesome to deal with. They're one of my favorite internet projects.

(Also, thanks for reminding me that it was time I donated something to the Marginalia project: https://buymeacoffee.com/marginalia.nu )

Are there 'public'/'anonymous' API keys I could use to perform a web search with CURL?

(I guess I would get json formatted search result data)

There is! The API Key is literally "public". But apparently it often gets rate limited, because seemingly every Metasearch engine uses that one. I think there might also be a slightly less rate-limited one for Hacker News users if you search around (I no longer remember what it is since I got my own key in the end.)

You can get your own API key for free by emailing, but that would not be anonymous, I guess.

I don't have curl syntax to hand, but hopefully it's easy to figure out from these documents. I may come back and edit later with curl syntax if I get time:

https://about.marginalia-search.com/article/api/

Thx! This is great news.

If their email server does handle self-hosted SMTP server with ip literal email addresses (with the ip from the SMTP, stronger than SPF), indeed, I will probably ask for my mine.

I wish major AI services would do the same or something close.

Yes, I try to avoid the evangelism topics but they're much worse than blockchain and are taking up most of HN now.

Can't help but click sometimes, always see the same arguments, so why not post the same thing as well?

By the way, I never check user names, I just reply to the post content.

Just for anyone else who hadn't seen the announcement yet, this Anthropic 1M context is now the same price as the previous 256K context - not the beta where Anthropic charged extra for the 1M window:

https://x.com/claudeai/status/2032509548297343196

As for retrieval, the post shows Opus 4.6 at 78.3% needle retrieval success in 1M window (compared with 91.9% in 256K), and Sonnet 4.6 at 65.1% needle retrieval in 1M (compared with 90.6% in 256K).

Aren't these numbers really bad? > 80% needle retrieval means every fifth memory is akin to a hallucination.

I don't think it quite means that - happy to be corrected on this, but I think it's more like what percentage it can still pay attention to. If you only remembered "cat sat mat", that's only 50% of the phrase "the cat sat on the mat", but you've still paid attention to enough of the right things to be able to fully understand and reconstruct the original. 100% would be akin to memorizing & being able to recite in order every single word that someone said during their conversation with you.

But even if I've misunderstood how attention works, the numbers are relative. GPT 5.4 at 1M only achieves 36% needle retrieval. Gemini 3.1 & GPT 5.4 are only getting 80% at even the 128K point, but I think people would still say those models are highly useful.

It seems to be the hit rate of a very straightforward (literal matching) retrieval. Just checked the benchmark description (https://huggingface.co/datasets/openai/mrcr), here it is:

"The task is as follows: The model is given a long, multi-turn, synthetically generated conversation between user and model where the user asks for a piece of writing about a topic, e.g. "write a poem about tapirs" or "write a blog post about rocks". Hidden in this conversation are 2, 4, or 8 identical asks, and the model is ultimately prompted to return the i-th instance of one of those asks. For example, "Return the 2nd poem about tapirs".

As a side note, steering away from the literal matching crushes performance already at 8k+ tokens: https://arxiv.org/pdf/2502.05167, although the models in this paper are quite old (gpt-4o ish). Would be interesting to run the same benchmark on the newer models

Also, there is strong evidence that aggregating over long context is much more challenging than the "needle extraction task": https://arxiv.org/pdf/2505.08140

All in all, in my opinion, "context rot" is far from being solved

now that's major news

We never did find out what those drones in New Jersey in 2024 were, did we? One Republican congressman seemed convinced at the time that he'd been informed:

BBC: Mystery New Jersey drones not from Iranian 'mothership' - Pentagon

https://www.bbc.com/news/articles/crrwz91wqd9o

It's certainly a theory / narrative that keeps appearing in the media.

They were flying over military installations, if they were anyone else's drones, they would have been shot down like the weather balloons that spook the government from time to time.

Foreign drones surveilled a military base here and they didn't shoot any down.

Maybe the US reacts differently, but in Europe most military bases have been scouted by Russian drones, and afaik none were shot down.

I've seen the reaction to people flying their toy drones too close to military assets, they send men out with machine guns and megaphones, confiscate the drone and sometimes press charges.

They were Palantir apparently.

It's an 8GB RAM 256GB SSD laptop with a lower spec'd 6-core chip for $599 USD. Seems overhyped to me, PCs have done that for a while, just not as elegantly. Admittedly it probably has far better battery life than a PC, so that's a genuine advantage.