mzl

Why? I'm not being glib here, but I don't see what it really changes fundamentally that makes it noteworthy except as a curiosity.

Such releases often show weaknesses if the intelligence index, because the models that are ranked jointly do not have the same percieved level of quality.

Conversely, LSP is quite limited in comparison to what can be done with a real IDE integration in JetBrains product suite, so it is not so clear cut. See this post for some details: https://matklad.github.io/2023/10/12/lsp-could-have-been-bet...

Skills for creating good and repeatable benchmarking scripts.

A knowledge base for my research area, with tools for paper ingestion and search.

An md file to html presentation tool, there are several but this one helps me.

A review tool that splits a PR or branch intelligently into modules, and does per module reviews and global reviews for different aspects, and then summarizing that into a report. Can be used with multiple different harnesses. Written as a Python project, but build-time assembled into a single-file Python script with uv run --script shbang line.

If you modify the work, that creates a derived work from whatever copyright the original works has, not a new work that is fully copyrightable.

As the article says in the Tl;DR at the top the code may be contaminated by open source licenses

> Agentic coding tools like Claude Code, Cursor, and Codex generate code that may be uncopyrightable, owned by your employer, or contaminated by open source licenses you cannot see

Kimi K2.5 and K2.6 are both >1T

It is tricky to build good infrastructure for prompt caching.

Which version of Kimi and served from where?

Composer-2 is based on Kimi K2.5, but with extensive RL. Cursor estimated 3x more compute on their RL than the original K2.5 training run (some details in https://cursor.com/blog/composer-2-technical-report).

Composer-2 seems very useful in Cursor, while K2.6 according to AA seems to be a really useful general model: https://artificialanalysis.ai/articles/kimi-k2-6-the-new-lea...

I've been prescribed slightly more than 5g per day (2 x 650mg tablets every 6 hours) for pain after an operation jointly with ibuprofen, which is scarily close to the limits.

I've heard people saying the study is bad, but whenever I've asked about why the answers have been pretty bad. Do you have a good source for why we should disregard it?

Dan Luu had some interesting analysis about car safety, comparing how different auto-makers fared on newly introduced crash tests: https://danluu.com/car-safety/

The main take-away for me from that page is that very few manufacturers seem to design for actual safety (only Volvo had good results), and Tesla was angry that a new test had been introduced which feels indicative of a bad safety culture.

There was an interesting scandal in Sweden where Oracle managed to sell the Millenium system to a regions hospitals even though they did not fulfill the requirements, and then when it inevitably crashed and burned they had to do an emergency rollback to the previous system after just a few days.

Here is an article in English: https://www.heise.de/en/news/Scrapping-the-millennium-introd...

No, but a lot of AI-adjsuted wordings have the very idiosyncratic AI-style that is prevalent in the AI-slop that is everywhere, and that style has quickly become associated with writing that is generally void of content and insight. So it is natural to get gut-reactions to the typical phrasings that have become associated with AI.

As others have said, this is more of a constraint programming system than Wave Function Collapse. Whatever one wants to call it, I liked it.

For guiding the search, you might want to consider search steps that select only one feature, for example that a pair of adjacent tiles should be connected by a road, and just propagate that information. That could be used as a way to guide the search on high-level features first, and then later realize the plans by doing the normal search.

I have a (very slight) beef with the name Algorithm X, as it is more of a data-structure to manage undo-information for the backtracking than an algorithm. It is a very fun, useful, and interesting data-structure, but it doesn't really change what steps are performed in the backtracking search.

In my view, Scrum is a way to force dysfunctional teams to have some process, it is not useful for a team that is already delivering and working in a samll-a agile manner.

Are you using the Model GPU memory snapshotting for this?

I like the intelligence per watt and intelligence per joule framing in https://arxiv.org/abs/2511.07885 It feels like a very useful measure for thinking about long-term sustainable variants of AI build-outs.

The cost of running things like prompt caching is defined by the implementation as that gives the infrastructure costs.

Saying that it is just in index from string prefixes into KV Cache misses all the fun, interesting, and complicated parts of it. While technically the size of the prompt-pointers is tiny compared with the data it points into, the massive scale of managing this over all users and requests and routing inside the compute cluster makes it an expensive thing to implement and tune. Also, keeping the prompt cache sufficiently responsive and storing the large KV Caches somewhere costs a lot as well in resources.

I think that the OpenAI docs are pretty useful for the API level understanding of how it can work (https://developers.openai.com/api/docs/guides/prompt-caching...). The vLLM docs (https://docs.vllm.ai/en/stable/design/prefix_caching/) and SGLang radix hashing (https://lmsys.org/blog/2024-01-17-sglang/) are useful for insights into how to implement it locally for one computer ode.

The prompt cache caches KV Cache states based on prefixes of previous prompts and conversations. Now, for a particular coding agent conversation, it might be more involved in how caching works (with cache handles and so on), I'm talking about the general case here. This is a way to avoid repeating the same quadratic cost computing over the prompt. Typically, LLM providers have much lower pricing for reading from this cache than computing again.

Since the prompt cache is (by necessity, this is how LLMs work) prefix of a prompt, if you have repeated API calls in some service, there is a lot of savings possible by organizing queries to have less commonly varying things first, and more varying things later. For example, if you included the current date and time as the first data point in your call, then that would force a recomputation every time.

Depends on which cache you mean. The KV Cache gets read on every token generated, but the prompt cache (which is what incurs the cache read cost) is read on conversation starts.

Technically, Cerebras solution is really cool. However, I am skeptical that it will be economically useful for models that are larger in size, as the requirements on the number of racks scales with the the size of the model to fit the weights in SRAM.

I find it interesting that the spark version seems worse than the gpt-oss version (https://simonwillison.net/2025/Aug/5/gpt-oss/)

An LLM model only outputs tokens, so this could be seen as an extension of tool calling where it has trained on the knowledge and use-cases for "tool-calling" itself as a sub-agent.

I said exactly what I wanted to say, in as simple terms as I am capable. The fact that some people insist on reality being simpler than it is does not make it true.

https://zayenz.se - personal site with blog and research publications

Oh, I'm also a fan. It is really cool to see what they've done. However, in the current systems they have available, they would (as far as I've understood it) just need way to many racks to be able to serve the full Deepseek model for it to have any kind of economics. The main limiting factor is the amount of sram available per wafer.

I don't think it would be economically viable to serve the full DeepSeek models on Cerebras hardware.