szopa

Note that capturing a live Russian soldier is worth 10 times more than killing one. Think about how that affects the incentives of Ukrainian soldiers. Avoid committing war crimes, get better gear for your team.

Yeah, that really sucks. I couldn’t quite believe what was the case and spent many hours trying to debug it.

I’m starting to feel that comments about an article being AI-generated are super low value AND super low effort. Who cares? Soon most of the text you’ll be dealing with is gonna be AI generated. But there’s still good and bad AI generated content — start judging it by its merits.

I have been a die hard emacs user for 20 years, and I have a very nice emacs setup (if a little bit idiosyncratic, but all emacs configs are idiosyncratic). However, recently I realized that I read code, but almost never write it. What is more, I spend a lot of time doing it in tmux, over mosh, from my phone. Emacs ergonomy is just not great if all you have is a horrible phone keyboard (and no swiping, because tmux redraws the screen if you swipe). And then I discovered helix. It has all the things I was jealous about vim, BUT it has sane defaults ootb. And truth be told, another thing I use a lot is bat, which is cat with syntax highlighting and an automatic pager.

Thank you for herding all the cats! :)

Some other interesting aspects of the camp:

The event’s currency are Łosie, which you get by taking part in classes and winning tournament games. By the end of each week there’s an auction where you bid for prizes. You can use your Łosie from previous years, but Tasuki implements an inflationary monetary policy to keep old-timers from becoming too rich (every year Łosie rewards get doubled).

Some people have been coming from abroad for many years, and at some point just figured out it makes sense to learn Polish (not the easiest of languages).

Apparently in manga and anime regular characters are often drawn as if they were European-ish (so, some of them are going to have blue eyes or blond hair, not common in Japan). This convention is in part historical (matching American comic books that inspired manga), and in part to make the characters more characteristic and easier to distinguish from each other. But in HnG this applies only to Japanese characters – people from abroad are drawn in a more naturalistic and stereotypical way. Koreans and Chinese will look actually like Asians, and Americans and Europeans will be an even more exaggerated version of themselves. I guess it’s a very different sensitivity than what’s common in the US right now.

I first learned to play Go back in university, but never got very good (it was competing with learning how to program). Many years later, shortly after the war in Ukraine started, I was looking for an activity to share with my 8-year-old son. Life was chaotic then: everyone was anxious, we were hosting a refugee lady, and I could see the stress taking a toll on him. I wanted something where it would be clear we shouldn’t be disturbed – and Go fit perfectly. We started playing, and it was fun. One of the great things about Go is its elegant handicap system, which makes it possible for players at very different levels to still enjoy a fair, challenging game.

Since then, we’ve been going to the local Go club in Warsaw, and it’s become our main hobby. We play each other almost daily, travel to tournaments (sometimes abroad), and even spend our vacations at Go summer camp.

The camp is actually a magical event. It takes place at a campsite in the middle of the Kaszuby Lake District. The conditions are spartan – you either live in a tent or a five-person cabin, and hot water is scarce. But the crowd that gathers there is incredible. Over breakfast you might get an impromptu intro to lambda calculus, in the evening you might end up in a deep philosophical conversation, or hear travel stories from far-off places, or suddenly learn way more about knitting than you thought possible. When we first went, it felt like discovering our long-lost family.

The Go community is much smaller than chess, but also far more tight-knit and welcoming. I’ve heard chess can be more cutthroat, while in Go there’s this unspoken understanding that if you drive people away, you’ll have no one left to play with.

When I travel, I like to drop in on local Go clubs. It’s always been a great experience – I especially enjoyed visiting the San Francisco Go Club in Japantown.

I play almost exclusively over the board. I prefer long, thoughtful games, and I can’t really focus the same way on a screen.

Oh, and the anime about Go, Hikaru no Go, is really good (you should watch it even if you don’t care about the game).

One consideration that that is missing: how familiar are LLMs with this technology? And from this point of view the app has sailed, I’m afraid we are stuck with the frameworks that are available today for eternity, for better or worse. And maybe that is not such a bad thing. I don’t do full stack programming in my day job, but I have this crazy idea that if I ever have a startup idea, I want to be able to code an MVP. So, every two years I do a deep dive and write a toy web app. I’m always learning something new on the frontend (fun!), while on the backend I just use Django, so it just works as it used to, except it usually gets more convenient in many small ways (boring). Sometimes there’s such a thing as too much fun.

One color go is a training method used in go… I wonder if that was her inspiration.

I started at a hybrid company a few months all, after working remotely for a long time (way before Covid). I was initially very skeptical, but then I noticed that coming to the office positively affects my psychological wellbeing. I actually enjoy hanging out with my colleagues.

I really liked the dream of working from wherever, but now I would prefer not to go back to remote work. Granted, this very much depends on your coworkers.

This is simply… incredibly impressive. I wasn’t expecting they would build a computer all the way up from a CPU and memory to an assembler to a compiler, all in the game of life.

Kudos.

I was the founding site lead of a Polish office of U.S. tech company, with an initial target of a hundred developers. After looking at… many resumes I realized I strongly disliked custom resumes. As much as I wanted to appreciate uniqueness and creativity, it got in the way. What I really wanted was something as standard and easy to read as possible. Ideally a LinkedIn profile.

Don’t get me started about cute resumes that were written as code, etc. I hated them and hated myself for how much I hated them.

Why should we be excited about this?

That’s also my favorite song by Louis Armstrong! It was used in Fallout 2, and it started my fascination with Satchmo.

I see your male walking with fluffy dog, and raise you a male with a fluffy dog picking up the poop after the dog. I’m tall, dress in black, and have a black beard (so potentially scary), and after such a maneuver it seems I become the first choice to ask for help.

My favorite Le Carre book is probably The Perfect Spy (especially that it is autobiographical to a great extent).

I find it hilarious how Dogs of War is different as the book and the movie. The movie is a classical '80s thing with big guns and explosions. The book on the other hand gets into great detail about the logistics of organizing a coup, holding accounts, export licenses for arms, etc. Forsyth would've probably preferred to write it as a non-fiction book, but making it a work of fiction vastly extended his life expectancy.

At one point I got so frustrated with langchain that I wrote my own stuff in go: https://github.com/ryszard/agency

Thanks for sharing the exercises! What seems to have helped me was losing weight (not easy at all, but has many additional benefits). There are some additional exercises (look for "hiatal hernia exercises" on YouTube), which are very simple and sometimes can get you instant relief. It's surprising so few people know about them. The comments section under those videos is often full of messages along the lines of "I've been suffering for years, this video changed my life." It's really helpful to know you can do something about it that does not involve taking PPIs every single day.

This sounds very interesting. Can you please elaborate?

Tokenizers seem to be a massive pain in the neck if you are just calling into an API to use your model. The algorithm itself is non-trivial, and they need pretty sizable data to function: the vocabulary and the merges, which just sit there, using memory. I'm writing https://github.com/ryszard/agency in Go, and while there's a good library for the OpenAI tokenization, if you want a tokenizer for the HF models the best I found was a library calling HF's Rust implementation, which makes it horrible for distribution.

However, at some point I realized that I needed not really the tokens, but the token count, as my most important use was implementing a Token Buffer Memory (trim messages from the beginning in such a way that you never exceed a context size number of tokens). And in order to do that I don't need it to be exactly right, just mostly right, if I am ok with slightly suboptimal efficiency (keeping slightly less tokens than the model supports). So, I took files from Project Gutenberg, and compared the ratio of tokens I get using a proper tokenizer and just calling `strings.Split`, and it seems to be remarkably stable for a given model and language (multiply the length of the result of splitting on spaces by 1.55 for OpenAI and 1.7 for Claude, which leaves a tiny safety margin).

I'm not throwing shade at this project – just being able to call the tokenizer would've saved me a lot of time. But I hope that if I'm wrong about the estimates bring good enough some good person will point out the error of my ways :)

Ok, this is indeed a very well kept secret. I have been programming go for a really long time, and was even a readability reviewer at Google, and I had no idea. Thank you, OP.

ChatGPT, GitHub CoPilot, I may need to pay for ngrok soon (interacting with Meta APIs and their webhooks...).

How I wish this post had appeared a few days earlier... I am writing on my own library for some agent experiments (in go, to make my life more interesting I guess), and knowing the number of tokens is important to implement a token buffer memory (as you approach the model's context window size, you prune enough messages from the beginning of the conversation that the whole thing keeps some given size, in tokens). While there's a nice native library in go for OpenAI models (https://github.com/tiktoken-go/tokenizer), the only library I found for Hugging Face models (and Claude, they published their tokenizer spec in the same JSON format) calls into HF's Rust implementation, which makes it challenging as a dependency in Go. What is more, any tokenizer needs to keep some representation of its vocabulary in memory. So, in the end I removed the true tokenizers, and ended up using an approximate version (just split it in on spaces and multiply by a factor I determined experimentally for the models I use using the real tokenizer, with a little extra for safety). If it turns out someone needs the real thing they can always provide their own token counter). I was actually rather happy with this result: I have less dependencies, and use less memory. But to get there I needed to do a deep dive too understand BPE tokenizers :)

(The library, if anyone is interested: https://github.com/ryszard/agency.)

A friend just mentioned to me https://github.com/m1guelpf/plz-cli (whose tagline I unknowingly reproduced). The difference between think and plz is that in think you are able to keep the conversation going, which allows you to easily ask for followup tasks.

Well, the Koreans seem to be aware of this, and they are pushing for changes. They play shorter games, and they've recently moved from byo-yomi to Fischer timing (part of the reasoning was that players are trained to play at the very end of their byo-yomi period, so that introduced a minimum of 30 seconds per move - even if the players had decided in the first 4 seconds).

Those were simply people with not enough imagination. Eg. AlphaGo beating Lee Sedol brought a true renaissance in Go, with papers showing that now humans play objectively better than they did before AI. While we have comfortably superhuman AIs, they are still far from godlike. The doomers were confused about where the ceiling was (and in Go it is much further than we thought).

My father has always been a chess aficionado, but I got turned off by the fact that he was the only available opponent, and he had 37 years of headstart (it didn't occur to me to go to a chess club). When I was in University, I discovered Go, which felt somehow deeper than chess (I do enjoy the occasional game of chess, but compared to Go it feels like a knife fight vs a military campaign). After school I stopped playing (and I had never gotten serious). However, my family needed a distraction when the war in Ukraine started (we live in Poland, so it's right around the corner), so I started playing Go with my son.

Since then we've gotten pretty serious - we play at a local club, went to a tournament, take lessons, etc. However, it's very much a niche hobby - more than A local club it's THE local club. Part of me wishes that I'd have chosen chess, which would've given us an order of magnitude more opponents and materials, both online and offline. On top of that, apparently it would've given my son an additional way to gain social status at school. However, the truth is that we do enjoy Go more than chess, and have already invested quite some energy (I am beginning to be an intermediate with an ELO of around 1150).

There are some perks of course. The Go community is super nice (the joke goes that, unlike with chess, you cannot just go to another club to play IRL if you piss off your club members). It's so small that even as a beginner I got the chance to befriend the reigning Polish champion (and one of the very few European Pros, which is a big deal - they play almost on par with the Asian pros). He turned out to be a fantastical human being, if a bit intimidating at first (he will casually play 5 games simultaneously, maintaining separate teaching conversations over each board, and a bonus conversation about AI; then he will be able to write down every game on paper).

While I feel a hint of jealousy for the popularity of chess, if Go were to disappear, I would just switch to chess and have probably have 85-90% of the fun I'm having with Go (not to mention access to my father's vast chess library). So, the fact that young people enjoy mind games makes me very happy - I feel that playing chess or Go is genuinely good for you.

The word by word translation sounds like it's trying to say that it isn't very competent at lojban, but that it can try to learn lojban if you provide it with parallel examples. All this said in broken lojban, as expected. Quite reasonable, actually.

No, I was explicitly watching for this. In one of the sessions where we asked it to generate Kłeti sentences and the conversation passed the token limit it started inserting characters like ı (the Turkish dotless i). A week earlier I was playing with interpreting go positions, and at some point the model switched to talking about Chess (a bit less subtle than inserting unusual characters).