The context window is limited to 64k or 128k. If you’re using this with a coding agent, there’s going to be a lot of compactions. I found that I had a subagent whose compacted context plus prompts and such was over the window and it errored out in opencode.

As we have more powerful AI plus agents that can run on your phone, keeping everything as a pile of markdown plus some compute on top makes a lot of sense

Exactly

This is glitch art for text, I love it

If this much JSON could fit in a microcontroller’s memory (7.5MB in text), and it’s a performance issue, maybe it’s worth upgrading the JSON parser?

The page has comparisons with Qwen 3.6 27b and I’d love to see comparisons with Qwen 3.8 27b, the newer one is much more capable!

It's somewhat close, but a lot worse at code it seems. Qwen 3.8 is a wild improvement over 3.6.

                               Qwen 3.8-27B     Ornith-1.5-35B
    
    Terminal-Bench 2.1         73.0             67.8
    SWE-bench Pro              61.7             59.6
    DeepSWE (1.1)              42.2             22.0
    NL2Repo                    42.3             46.2
    GPQA Diamond               89.2             89.2
    Humanity's Last Exam       30.8             25.6

Google is making private AI practical with Gemma4 something that you can run without an Internet connection.

All of the proofs of privacy rely on us getting the math right. All of the privacy from unplugging your internet cable is there by default.

Neat! Similar to my graduate thesis, NoDictionaries: https://nodictionaries.com/cato/de-agri-cultura/156

Nice.

May I point out that in "Crudam si edes, in acetum intinguito", that "edes" is more likely to be the future of edere/esse "to eat"? (Just guessing by context.)

Fixed!

I remember encountering NoDictionaries years ago, and love it. Thank you for building it!

Homer is the Homer of America, unless you believe in fencing off universal properties of mankind into exclusive tribal ownership.

(cf Saul Bellow asking “who is the Tolstoy of the Zulus”, and Ralph Wiley in the Atlantic saying that Tolstoy is the Tolstoy of the Zulus.)

Happy to answer any questions :)

Self Hosted object storage looks neat!

For this project, where you have 120GB of customer data, and thirty requests a second for ~8k objects (0.25MB/s object reads), you’d seem to be able to 100x the throughput vertically scaling on one machine with a file system and an SSD and never thinking about object storage. Would love to see why the complexity

(Author here) that's more or less what I have right now – one machine with a file system and an SSD. S3 API on top is there to give multiple web servers shared access to the same storage. I could have used something else instead of S3 – say, NFS – but there was a feature request for S3 [1] and S3 has a big ecosystem around it already.

[1] https://github.com/healthchecks/healthchecks/issues/609

The complexity for that is almost always for redundancy and for ease of deploys.