Cross-request KV cache exposure in the GLM-5.3 Workers AI model (155 responses)
A commenter says the issue is not related to GLM-5.3 but is due to lack of user session isolation and memory overlap, and is a SGLANG and Cloudflare bug.
30 stories
Get a notification every evening
You hear about the digest when it is ready. Turn it off any time.
Notifications are on
You will hear about each evening’s digest when it is ready.
To get notifications on iPhone or iPad, add this site to your Home Screen first: tap Share, then Add to Home Screen. Open it from there to turn notifications on.
Notifications are blocked in this browser. Allow them in site settings and reload.
New notification sign-ups are closed for now. You can follow along with RSS.
Notifications could not be turned on. Check your connection and try again.
Notifications could not be turned off and are still on. Check your connection and try again.
The day’s top discussions
A commenter says the issue is not related to GLM-5.3 but is due to lack of user session isolation and memory overlap, and is a SGLANG and Cloudflare bug.
A commenter criticizes the model for not releasing technical details, having excessive safety restrictions, and limited usage quotas.
A commenter says the models Cerebras offers are not good enough for their use, and while fast, they are often wrong.
A commenter says the blog post's title is misleading, and they only needed to configure it correctly.
A commenter criticizes OpenAI's announcement method and suggests the model may have been trained on input from scientists who use OpenAI as a tool.
A commenter says it's unfortunate that the JIT is being removed, as work was progressing quickly but a new PEP was required.
A commenter says they find such open-source projects more important than what big companies do and will always upvote them.
A commenter notes that the model identifies itself sometimes as Kimi and sometimes as ChatGPT, and contains conflicting instructions.
One commenter says Micron's ditching of 4F² is an attempt to reframe its lagging behind in innovation as a choice to stick with old technology.
One commenter observes that good programmers are producing good solutions faster, while bad ones don't even check their mistakes.
One commenter says the protocol should be called Python Type Server Protocol since it only supports Python.
One commenter says they tried Ghostty but can't switch due to Kitty's hints mode.
One commenter quotes a blog post finding that some models write weaker code depending on the target audience.
One commenter describes this as a great example of going from AOT to JIT specializing function calls by types, and says the target is to build faster functions than can be derived at compile time…
A commenter notes that lens size and weight matter more than the body, questioning the point of such a small body that sacrifices features like IBIS and EVF.
A commenter argues that spec-first development doesn't work in real environments because unknowns are not known upfront, and instead suggests using AI for quick prototypes.
A commenter notes that the algorithm uses about half a byte per pixel and saves the path, implying the stack is still present but bounded to the number of pixels.
A commenter says that AI programming has made them care less about which programming language is used, finding it freeing.
A commenter says the three-node setup doesn't make sense and suggests offering user-selectable choices.
A commenter questions if this is another 'vibe coded' OS and whether there are interesting design decisions.
A commenter describes how Partyline was used in the UK during the 90s and 00s to find the freeparty scene.
A commenter says the model's pairing operation (zt,zt+Delta) is cheating and should be (zt,zt-Delta).
A smaller, more technical community
One commenter says what was disclosed was that eMudhra had acted with gross incompetence and that deserves no sympathy.
One commenter praises the content and mentions Daan Leijen's note on division forms.
A commenter praises the article for its explanations and diagrams, and notes the inherent risks in extracting archives to disk.
A commenter says they implemented a Parsec port in Java as a joke at work years ago.