I'll also note that the output isn't quite right --- the top number should be 13 rather than 1!

I mean, the specification for the hour marks (angle_i) starts with a mark at angle 0. It just followed that spec. ;)

I've pursued a similar strategy via CDs found in resale shops. I've gotten a good collection going, but there's a very specific selection of CDs popular enough for people to buy but not so good that they kept it.

I'm assuming it isn't sensitive for your purposes, but note that Google will train on these interactions, but not if you pay.

I think it'll be hard to find a LLM that actually respects your privacy regardless whether or not you pay. Even with the "privacy" enterprise Co-Pilot from Microsoft with all their promises of respecting your data, it's still not deemed safe enough by leglislation to be used in part of the European energy sector. The way we view LLM's on any subscription is similar to how I imagine companies in the USA views Deepseek. Don't put anything into them you can't afford to share with the world. Of course with the agents, you've probably given them access to everything on your disk.

Though to be fair, it's kind of silly how much effort we go through to protect our mostly open source software from AI agents, while at the same time, half our OT has build in hardware backdoors.

I agree, Google is definitely the champion of respecting your privacy. Will definitely not train their model on your data if you pay them. I mean you should definitely just film yourself and give them everything, access to your files, phone records, even bank accounts. Just make sure to pay them those measly $200 and absolutely they will not share that data with anybody.

You're thinking of Facebook. A lot of companies run on Gmail and Google Docs (easy to verify with `dig MX [bigco].com`), and they would not if Google shared that data with anybody.

It’s not really in either Meta or Google’s interests to share that data. What they do is to build super detailed profiles of you and what you’re likely to click on, so they can charge more money for ad impressions.

Meta certainly shares the data internally. https://www.techradar.com/computing/cyber-security/facebooks...

LLMs add a new thread model. If trained on your data, they might very well leak some of its information in some future chat.

Meta, Alphabet might not want that, but it is impossible to completely avoid with current architectures.

Honestly, there are plenty of more profitable things to do with such information. I think ad impressions being the sole motivator for anybody, is sorta two decades ago.

Big companies can negotiate their own terms and enforce them with meaningful legal action.

I don't care. From what I understand of LLM training, there's basically 0 chance a key or password I might send it will ever be regurgitated. Do you have any examples of an LLM actually doing anything like this?

> For centuries, communities gathered around the fire to hear the same saga recited for the hundredth time. The listeners didn't grow impatient with the familiar opening formulas, the elaborate genealogies, the detailed descriptions of weapons and weather.

I generally agree with the broad strokes of this post, but this description gives me pause. How do we know that listeners didn't grow impatient? Though I suppose it would take a good deal of compression to answer this :)

He's retired, so I'm guessing more about the clout. Or even just "love of the game"? He had a fairly popular tweet thread a couple years back where he wrote out 80 tips for competitive programming -- that feels less likely to be clout based

He did use a little autocomplete apparently, but used [Vscode](https://x.com/jacob_posel/status/1945585787690738051).

And it's not against the rules to use LLMs apparently in the competition. (https://atcoder.jp/posts/1495). I'd be curious what other competitors used.

Interesting, thanks for the links! I had read this part of the article:

> All competitors, including OpenAI, were limited to identical hardware provided by AtCoder, ensuring a level playing field between human and AI contestants.

And assumed that meant a pretty restricted (and LLM-free) environment. I think their policy is pretty pragmatic.

Ugh

Cool project! I have a couple of questions that would be nice in the writeup: * How did you generate your example problems? Did you take an existing benchmark? Or did you have LLMs generate the problems? * Do you have any thought to adding a second "base programming language" to alter? I'm not sure that there's enough variation as there is. (Another thought would be to generate 4 or 5 different new languages, each quite different, and then run the benchmark on each of those languages? I'm not sure how much the fact that it is randomly generated each time matters that much?)

But overall, a clever idea!

Generating the problems: I just thought up a few simple things that the computer might be able to do. In the future, I hope to expand to more complex problems, based upon common business situations: reading CSVs, parsing data, etc. I'll probably add new tests once I get multi-shot and reliability working correctly.

New base programming languages would be great, but what would be even better is some sort of meta-language where many features can be turned on or off, rather than just scrambling the keywords like I do now.

I did some vibe testing with a current frontier model, and it gets quite confused and keeps insisting that there's a control structure that definitely doesn't exist in the TiānshūBench language with seed=1.