Alex-Programs

That'd be interesting. I'm thinking of writing a guide on how to identify LLM prose, since I seem to be a bit above average at it.

```In a modern conflict zone, jamming L-band means blinding drones, degrading targeting, and cutting off ISR. It’s not accidental. It’s deliberate.

The international treaties that say “don’t transmit here”? Those don’t matter much when you’re trying to survive a drone swarm.```

LLM prose, and it's not the only section that stands out. It's an informative article, so I don't mind it as much, but I think it's a shame people don't write things themselves anymore.

This is neat. I notice that you do similar reverse lemmatisation as my Wiktionary API wrapper: https://dictionary.nuenki.app/get_definition?language=Finnis...

Feel free to use it if you like, it's practically free, and open source if you'd like to host it yourself: https://github.com/Alex-Programs/nuenki-dictionary

Looking at your code, you're processing the actual word in a Finnish-specific way. If you'd like to generalise it, you can also just see if the definitions contains `x of y`, where y is a link, and automatically dereference to `y`.

Is it just me or is cybersecurity... Calming down? I feel like a few years ago there was constant news of ransomware, intrusions, vulnerabilities, etc, but more recently the defensive side seems to have the upper hand.

Thanks for letting me know. The front page items are translated with DeepL, which is used for text that's currently visible. My benchmark claims that it's actually pretty good at French (I haven't got data for Danish) - one of the reasons I'm currently remaking it!

I hadn't heard of Kagi Translate; I'll add it to the benchmarking I do. And I'll see about adding Furigana there.

You can get definitions when you hover, but it depends on the language - it uses Wiktionary.

For some languages it's available but disabled by default. You can change it in the extension popup under "Hover Behaviour".

Thanks! I'm glad people like it; I'm hopeless at marketing it to Language Learners:tm:, but programmers seem to love it and it's nice to get some positive feedback.

DeepL is actually pretty limited in what it supports. Unless I've missed a new language, Nuenki supports all of DeepL's languages.

Some of the additional ones are supported via Claude only and, where permitting, Groq. Groq is far faster than Claude; in languages that DeepL supports, DeepL handles visible text and Claude handles text that you haven't scrolled to yet. Claude-only languages are a bit of a worse experience.

It's pretty easy for me to add a language. It's all stored in a centralised toml file, which happens to be open source - https://github.com/Alex-Programs/nuenki-languages/blob/maste... - and it's about a 20 minute job to add a language, test it, etc. Then it's about half an hour and 5 USD to benchmark whether Llama is any good at it, and if so enable Groq and make the experience a bit more pleasant. I'm currently working on improving the translation quality benchmark (https://nuenki.app/blog/the_best_translator_is_a_hybrid_tran...), because people seem to like it and there's definitely a lot of room for improvement.

That 20 minute number is without updating the big language cloud on the website, which is a bit finicky; iirc I haven't added Vietnamese to it yet.

If anyone here has any requests, I'd gladly add them!

It's also just a hell of a lot more fun.

Language transfer is rather good. I'm not quite sure what it does differently, but there's a reason people recommend it.

I built https://nuenki.app, which follows a fully immersive approach by immersing you while you browse the web. It translates entire sentences at your knowledge level into the target language, and you hover for definitions/the original sentence/etc.

I built a tool[0] that gives you constant input at your level as you browse the web, so you don't need to take time out of your day. You can just learn a little as you browse, and let it compound over time.

It works by estimating the difficulty of English sentences, then translating ones at your level into your target language.

[0] https://nuenki.app

Right, I think I'm extrapolating from pop culture and a distorted lens on the internet.

https://nuenki.app - a browser extension that translates sentences at your level while you browse, so you learn a language through immersion while you go about your day.

I'm also working on and off on a hardware device for blind people.

> One neat thing that the average college professor won't be able to do for you: because the model understands multiple disciplines at once, you can make analogies between what you know well and what you're asking about — and the model knows enough about both subjects to tell you if your analogy is sound: where it holds vs. where it falls apart. This is an incredible accelerator for learning domains that you suspect may contain concepts that are structural isomorphisms to concepts in a domain you know well. And it's not something you'd expect to get from an education in the subject, unless your teacher happened to know exactly those two fields.

I really agree with what you've written in general, but this in particular is something I've really enjoyed. I know physics, and I know computing, and I can have an LLM talk me through electronics with that in mind - I know how electricity works, and I know how computers work, but it's applying it to electronics that I need it to help me with. And it does a great job of that.

As a Brit, the concept of "My lawyer" is slightly unfamiliar. The average Brit doesn't "have a lawyer"; they would only find a lawyer if they had a specific need, eg being accused of a crime or wanting to write a contract etc.

And yet as far as I can tell, most middle class Americans seem to refer to "their lawyer". Do you pay a monthly fee? Are they a criminal defence lawyer, or something broader? How often do you talk to them? How do you find them?

Of sure you find it hard to tell the two away! You lack the gear for tide pull test, and they feel the same here and local.

I hate this.

This is actually pretty cool. I don't have an immediate use for it, but I'm making a mental note, because I'm sure something will come up soon.

> Rust → Memory safety with zero overhead (Mozilla)

I think Rust is interesting because while a lot of its initial appeal came from memory safety, a lot of the usage now is not as a C++ replacement but as a Go/JS/Java replacement. It introduced people to algebraic type systems, its excellent tooling, etc, and now it's moved beyond its initial appeal. When I use other languages I miss Result<>, Option<>, and the fact that everything - from if statements, to matches, to simply {} - is an expression.

There are places where I use it as a C/C++ replacement - STM32 firmware, for example - but I also use it for accelerating my webextension code with WASM and as a CRUD backend.

You might want to obfuscate that email a bit. `hristoapps - gmail` would do the job.

Unobfuscated emails tend to get scraped and put on spam lists.

That's not why I oppose banning advertising.

I've made a product. The people who use it, like it. But I have no online following or presence, and I'm really not the kind of charismatic person who could build one. All the "community" places where I could share it in good faith are incredibly hostile to self promotion, I think because of the wave of people selling vibecoded openai wrappers as language tutors.

I can pay £40 for reddit ads, and while it has negative ROI, it gives me lots of feedback that I can use to iterate. Sure, my project seems to be a commercial dead end - people find it valuable, but most people don't find it quite valuable enough to pay for the high cost of translation - but I still think those ads had a lot of value.

That said, I use an adblocker myself, I wish more intelligent people worked on rockets rather than targeting algorithms, and I do agree that ads have a negative effect in a lot of places - it's just that they do have a real (and IMO moral) utility in some places. If you banned advertising for everything, you'd just encourage bribing moderators to let you self-promote or ensure only people with existing followings can make things.

(it's https://nuenki.app, if anyone's curious)

I used to feel dismayed that I missed that era of the internet and technology (I'm 19). IRC, forums, work-in-progress gifs on personal websites, etc.

I still wish I were there for that, but I'm glad I get to be here for LLMs and the intelligence explosion. I have absolutely no idea what the world will look like in a few years. It certainly isn't the certain high-paying tech job in a largely static world that it looked like a few years ago.

But whatever happens, it's going to be interesting!

I wonder whether I'm spending my time optimally, working on a little SAAS that happens to use LLMs as a downstream commodity, contributing through a niche benchmark.

Brilliant! Incredibly fast.

I've benchmarked its performance at language translation - https://nuenki.app/blog/quasar_alpha_stats

It's very good, top of the leaderboard on almost all languages, and its high score for idiomaticity indicates it's probably quite a large model - and yet it has very fast inference.

That's why you pass it to a big, clever neural network with some reasoning and let it figure it out for you.

Put multiple different sources through multiple different LLMs and take a consensus of them. If there isn't a clear winner, declare it uncertain.

It'd be a bit of a bodge, but it'd work.

If you can get a live feed of the broadcast, you could also try feeding visual information in to be processed by a multimodal LLM.

People will presumably quite quickly start discussing events. Score data can be inferred from that.

Lots of things! I'm in a gap year.

To start with, there's https://nuenki.app. It's a browser extension that selectively translates sentences into the language you're learning, so you're constantly immersing yourself in text at your knowledge level.

I've also been working with a friend on a device to help blind people without light perception. I'm quite new to electronics. It's pretty simple, conceptually - a coin-sized device on the forehead that takes in the light intensity in a ~15 degree cone, then translates it into high resolution haptic feedback to the forehead.

The idea being that it means people without light perception can gain a sixth sense through neuroplasticity, with helps them navigate the room and understand their surroundings. We're planning on open sourcing the files. My mum used to teach blind kids, and there's been quite a lot of interest!

As for Nuenki, I'm pretty bad at marketing, so I'm doing a final lot of work to see if I can make it work financially - seeing if an exceptionally generous affiliate program will do the trick - before putting it on maintenance mode, since I have a small group of users who really like it. I'm burning through my gap year fast, and really want to focus on the electronics project, tutoring, and practicing maths for my physics degree.

Maybe you could feed data from relevant fan subreddits into an LLM to extract results from discussions?

Sure, it's poor quality data, but there's presumably quite a lot of it.

This is cool. The text to speech is OK - it's a bit squeaky, yet super low pitch. Definitely not bad, but it could be improved. The intonation is good. Maybe just let people switch voices? I generally find female voices are better executed with TTS.

I happen to have built a tool with that premise :P https://nuenki.app/ - I presume you're trying to read the news in another language for language learning purposes, not for diversifying your information sources.

Firefox auto translate is great from a technical perspective, and I think it's awesome to see it translate sentence by sentence, but the actual quality isn't great. It tends to keep word order and grammar and mostly just translate the words themselves, as you saw.

Here's the relevant section via DeepL:

In Austria, if a cow or similar livestock dies outside the slaughterhouse due to lightning, a fall, illness or for similar reasons, it is the responsibility of the owner, i.e. usually the farmer, to ensure that the carcass is removed and disposed of in order to protect waterways and maintain an intact landscape.[1] In the Vorarlberg Alps, there are around 20 such deaths per year. If the carcass is located on a mountain pasture or in another place that cannot be reached by truck, it can only be removed by helicopter. This cost 15,000 Austrian schillings in 2001. Adjusted for inflation, this corresponds to approx. 1,500 euros in 2025.[2]

Although at least in Vorarlberg 80 percent of these transportation costs were covered by the federal state, meaning that the farmer only had to pay approx. 3,000 shillings for the helicopter, it was common practice there to remove the animals by blasting on site instead. This only cost 500 shillings in 2001[2], so the farmer was able to save around 2,500 shillings (approx. 250 euros) by blowing up the animal. The explosion tore the animal into smaller pieces, which would then decompose more quickly or be disposed of by scavengers such as birds and foxes. The explosion was either caused by demolition experts, which in turn meant higher costs, or by the farmers themselves.