That kind of consistency has also been my own experience with LLMs.

For $40, I'd get 20 tok/s * 2.6M seconds per month = 52M tokens of DeepSeek v3.2 per month if I run it 24/7, which is not realistic for most workloads.

On OpenRouter [1], $40 buys 105M tokens from the same model, which is more than 52M tokens, and I can freely choose when to use them.

[1]: https://openrouter.ai/deepseek/deepseek-v3.2

https://i.imgur.com/wFVSpS5.png

and quality vs cost

https://i.imgur.com/fqM4edw.png

But I just noticed that my plot is meaningless because it conflates model quality with provider uptime.

Claude Haiku has a higher average quality than Claude Opus, which does not make sense. The explanation is that network errors were credited with a quality score of 0, and there were _a lot_ of network errors.

https://grep.app/search?regexp=true&q=%5Ba-z%5D%7B8%2C%7D%5C...

For example, license files often contain names and many package managers require a contact person.

When this goes to court, GitHub will probably make the excuse that they somehow did not know that people upload personal data, but the fact that this happens so often that they had to make a secret scanner to stop people from uploading their private keys will prove them as liars.

In addition to the non-empty input, 153 reasoning tokens were produced.

When setting max tokens to 100, the output is empty, and the token limit of 100 has been exhausted with reasoning tokens.

The price hike has been crazy. The Bosgame M5 Mini is $2400 now. I didn't get one last year when they were $1500 because I thought the memory bandwidth was mediocre. However, it doesn't look like we'll get anything better for that price anytime soon.

I've got a 3060 myself, which is nice to play around with the smaller models for free (minus electricity) and with 100% uptime, but I was not able to program anything with them yet that I didn't want to rewrite completely. A heavily quantized Qwen3.5-27B model is getting close though. Maybe in a few months.

It is important to start a new chat so the model is not stuck in its previous mindset, and it is beneficial to have tests to verify that the simplified code still works as it did before.

Telling the model to generate concise code did not work for me, because LLMs do not know beforehand what they are going to write, so they are rarely able to refactor existing code to break out common functionality into reusable functions. We might get there eventually. Thinking models are a bit better at it. But we are not quite there yet.

Anyway, my university did not ban AI, and now most students have degraded to proxies between teaching assistants and ChatGPT.

But I can see why the HN guideline is formulated that way. My students often use the excuse "I did not use AI for writing! I wrote it myself! I only used AI to translate it!" Simply disallowing all kinds of AI usage is much easier than discussing for the thousandth time whether the student actually understands what they have written.

> and while an exact number is hard to compute, let me tell you, it is not 17B or anywhere in that particular OOM :)

I can see ~100B, but that would near the same order of magnitude. I find ~1000B active parameters hard to believe.