umajho

I actually did something similar last month. I just asked the LLMs:

> Do Simon Willison's classic pelican test. Recall the specification first and then draw it.

Qwen3.8-27b didn't know that the test is about riding a bicycle, while all other larger models I tested (DeepSeek V4 Pro, Qwen3.8 Max, GLM 5.3, Kimi K3) drew the pelican riding a bicycle.

logs: https://gist.github.com/umajho/c0e20d245d721d7c472a32d317640...

rendered: https://gist.github.com/umajho/b1fdf01d31c741bb11bdb5a49c275...

It seems Litestream will soon support arbitrary s3-compatible destinations.[^1] Neat.

[^1]: https://github.com/benbjohnson/litestream/pull/731

If you want to get the grapheme length in JavaScript, JavaScript now has Intl.Segmenter[^1][^2].

  > [...(new Intl.Segmenter()).segment(THAT_FACEPALM_EMOJI)].length
  1

[^1]: https://developer.mozilla.org/en-US/docs/Web/JavaScript/Refe...

[^2]: https://caniuse.com/mdn-javascript_builtins_intl_segmenter_s...

That makes sense.

> …. For instance, “a trait is a bit like an interface” is wrong, …

Wait. What's wrong with this statement?

MinIO also switched to AGPLv3 a while back, and they stated that “the AGPL license requires that all software connecting with MinIO be 100% open source for you/your users not to be in violation of the license.”[^1] Since Redis and MinIO are somewhat similar, (Both can be used to store and retrieve data. Redis uses a custom protocol, and MinIO uses an S3-compatible API.) Should I assume that this statement also applies to Redis?

[^1]: https://github.com/minio/minio/issues/13308#issuecomment-929...

I am currently using the Q4_K_M quantized version of gemma-3-27b-it locally. I previously assumed that a 27B model with image input support wouldn't be very high quality, but after actually using it, the generated responses feel better than those from my previously used DeepSeek-R1-Distill-Qwen-32B (Q4_K_M), and its recognition of images is also stronger than I expected. (I thought the model could only roughly understand the concepts in the image, but I didn't expect it to be able to recognize text within the image.)

Since this article publishes the optimized Q4 quantized version, it would be great if it included more comparisons between the new version and my currently used unoptimized Q4 version (such as benchmark scores).

(I deliberately wrote this reply in Chinese and had gemma-3-27b-it Q4_K_M translate it into English.)