Hacker News: fumeux_fume

New comment by fumeux_fume in "Mistral OCR 4.1"

fumeux_fume — Fri, 14 Aug 2026 04:53:03 +0000

I think people misunderstand the utility of Mistral's OCR. It's not going to beat SOTA models for extraction on edge-case docs, but it's MUCH cheaper and faster and does an excellent job on simple ones. I've been working on converting PDFs to EPUBs and Mistral has been making steady improvements. On a chapter of Bleak House it was able to extract and tag the header, titles, and references at the bottom every time. The only thing it struggled on was line numbers in the right margin which it correctly tagged as "aside text" 3/5 times, but always separated from the core text each time. The important thing to keep in mind is that there's no prompting needed, just upload the PDF and voila!. There's even a batch mode with a 50% discount.

New comment by fumeux_fume in "Gemini 3.7 Flash"

fumeux_fume — Fri, 14 Aug 2026 04:27:02 +0000

Nice! I test if models know which tubes I can use for an amp given the power and number of pins.

New comment by fumeux_fume in "BMW Spider-Man in-car advertising"

fumeux_fume — Sat, 01 Aug 2026 03:45:44 +0000

> BMW's own PressClub release described the advertisement as a special surprise for drivers...

That one got a chuckle out of me

New comment by fumeux_fume in "KOReader"

fumeux_fume — Wed, 29 Jul 2026 13:01:48 +0000

Lots of nice features but the UX and UI are so painful that it's not worth jumping from the stock kobo reader for me.

It's nice to have another option though and if the UX improves one day maybe I'll use it as my main reader.

New comment by fumeux_fume in "Hannah Fry Wins the Leelavati Prize in 2026 for Mathematics Outreach"

fumeux_fume — Sat, 25 Jul 2026 16:50:43 +0000

> As Fry says, the limit to someone's understanding is never their ability to understand. The limit is always in their motivation to care. And she sees her role as being to provide that motivation.

> "I think what you need to do," she says about the key to connecting with a wide audience, "is to create a hole in people’s imagination. If I can create just a moment where you want to know the next bit – you know exactly the size and shape it needs to be to fulfill your craving for that exact piece of information – then everybody's in."

I wish we had more teachers that thought like this.

New comment by fumeux_fume in "OpenAI’s accidental attack against Hugging Face is science fiction that happened"

fumeux_fume — Fri, 24 Jul 2026 06:02:28 +0000

Hardly science fiction. I respect a lot of Simon's work, but his credibility gets chiseled away a bit each time he participates in these kinds of echo chamber posts that are basically secondhand marketing.

New comment by fumeux_fume in "Kimi K3: Open Frontier Intelligence"

fumeux_fume — Fri, 17 Jul 2026 03:59:53 +0000

Runs great on mobile!

New comment by fumeux_fume in "NotebookLM is now Gemini Notebook"

fumeux_fume — Fri, 17 Jul 2026 00:45:53 +0000

I wonder how much of the annoying aspects of the podcasters can be controlled by the custom user instructions. I hate sitting through the patter of how "we're not political, just reporting the views of the sources."

New comment by fumeux_fume in "Book prizes don't work how you think"

fumeux_fume — Wed, 15 Jul 2026 23:56:26 +0000

Except for the prize committees outsourcing roughly the same sets of judges, they work a lot like how you’d expect. The judges pickup a bunch of books and choose the ones they like the most. Since the author makes it clear how subjective the process truly is, you can assume that personal biases play a huge part in how winners are chosen.

New comment by fumeux_fume in "The kids with phones are alright"

fumeux_fume — Tue, 14 Jul 2026 16:19:11 +0000

> Many other people, observing our current policy context, have also called out how smartphone and social media bans for young adults (and we are talking about that particular group here, not toddlers and primary schoolers) risk swaddling them in cotton wool and then releasing them into the world, without critical adulting skills, on the day they hit a magic birthday.

New comment by fumeux_fume in "The kids with phones are alright"

fumeux_fume — Tue, 14 Jul 2026 14:26:00 +0000

So this is the new hot take? That kids need unfettered access to smart phones and the internet to toughen them up? Absolutely cringe.

New comment by fumeux_fume in "Delta flight hit by firework while landing at Midway Airport on Fourth of July"

fumeux_fume — Mon, 06 Jul 2026 06:52:11 +0000

My opinion about how people use fireworks changed when I adopted a dog from a shelter and discovered how deeply traumatizing booms are to her. From chatting with other dog owners this seems to be common. If my neighbors could see the extent fireworks affect the same friendly, silly pup their children love greeting everyday, they would probably think twice about how they use them. We can manage the night of the 4th by traveling out of town, but the 1-2 weeks of random booms that follow the holiday are really tough for us.

New comment by fumeux_fume in "Delta flight hit by firework while landing at Midway Airport on Fourth of July"

fumeux_fume — Mon, 06 Jul 2026 06:39:55 +0000

In rural places where wildfire risks are very high, firework bans are usually enforced and respected.

New comment by fumeux_fume in "Qwen 3.6 27B is the sweet spot for local development"

fumeux_fume — Mon, 29 Jun 2026 19:41:39 +0000

M series macs are usually used for running these LLMs locally because the GPU and CPU share the same pool of RAM at very low latency. If you upgrade your RAM on a different kind of chipset without the Unified Memory Architecture, then it'll be much slower to produce all the tokens you need. Just another data point to add to your upgrade equation.

New comment by fumeux_fume in "Qwen 3.6 27B is the sweet spot for local development"

fumeux_fume — Mon, 29 Jun 2026 18:26:34 +0000

I suspect with those specs, you're not in the game right now for reliably using local models for code generation. The easiest way in is a MacBook with at least 32GB of RAM. This should be able to run a 4bit quantization of qwen 3.6 using the MLX format really well.

New comment by fumeux_fume in "Alan Greenspan has died"

fumeux_fume — Mon, 22 Jun 2026 19:10:23 +0000

A very Economist answer to Ryan's question. So basically, no.

New comment by fumeux_fume in "The Token Compression Illusion: Why I'm Skeptical of RTK"

fumeux_fume — Thu, 18 Jun 2026 23:32:53 +0000

https://en.wikipedia.org/wiki/Brandolini%27s_law

New comment by fumeux_fume in "Banned book library in a wi-fi smart light bulb"

fumeux_fume — Tue, 16 Jun 2026 13:29:04 +0000

So many objects have their own networks now. With a clever SSID and placement in a room full of potential targets, it could be pretty tough for someone to narrow it down to the bulb.

New comment by fumeux_fume in "Travel locally, where you are"

fumeux_fume — Thu, 11 Jun 2026 23:46:12 +0000

I couldn't help reading this article in the voice of Ted Flanders. It's such an insipid point and the writing is so bad. Not the writers fault, but when you find out he lives in Switzerland it's kinda hilarious.

New comment by fumeux_fume in "Disagreement Among Frontier LLMs on Real-World Fact-Checks"

fumeux_fume — Thu, 28 May 2026 15:26:58 +0000

This is a good pattern because it would allow all the models to "think" a bit before giving an answer even if they don't have reasoning or thinking turn on. Just make sure you have the reasoning output before the final answer. A mistake I see all the time is having the answer outputted first then the explanation after which leaves more room for models to rationalize bad answers.

Good pattern: {"explanation": , "answer": }

Bad pattern: {"answer": , "explanation": }