“I understand your desire to write C but I cannot afford to write C when Rust is an option”
23.06.2025 14:38
👍 3
🔁 0
💬 0
📌 0
“I understand your desire to write C but I cannot afford to write C when Rust is an option”
As far as I remember Anthropic recommends XML (and even those hacky parsers) for quite some time. So it might be Claude thing
One can also probably feed billions of people with insect protein from mosquitoes that live in Siberia
Any American mind sharing their thinking process on **this**?
Great summary! Uv indeed feels very magical and as there was a team that polished that for like 15 years.
The time has come. The prophecy is accomplished.
We are going to review one year of uv usage to ponder the pros, the cons, and whether you should migrate.
It's a long article, but I have a 10 lines TL;DR at the top, you can pretend you read the whole thing :)
open.substack.com/pub/bitecode...
Extremely weird to see supposedly rust-first project being a thin wrapper around Java based Apache Tika and C++ Tesseract. github.com/yobix-ai/ext...
We've done full circle
This is great news huggingface.co/blog/modernb...
Not everything needs to be fed through LLM and for that we now have much better foundation for tons of apps that work with texts but not necessarily need to generate any, so classification, NER, similarity scores.
Old quant types (some base model types require these): - Q4_0: small, very high quality loss - legacy, prefer using Q3_K_M - Q4_1: small, substantial quality loss - legacy, prefer using Q3_K_L - Q5_0: medium, balanced quality - legacy, prefer using Q4_K_M - Q5_1: medium, low quality loss - legacy, prefer using Q5_K_M New quant types (recommended): - Q2_K: smallest, extreme quality loss - not recommended - Q3_K: alias for Q3_K_M - Q3_K_S: very small, very high quality loss - Q3_K_M: very small, very high quality loss - Q3_K_L: small, substantial quality loss - Q4_K: alias for Q4_K_M - Q4_K_S: small, significant quality loss - Q4_K_M: medium, balanced quality - recommended - Q5_K: alias for Q5_K_M - Q5_K_S: large, low quality loss - recommended - Q5_K_M: large, very low quality loss - recommended - Q6_K: very large, extremely low quality loss - Q8_0: very large, extremely low quality loss - not recommended - F16: extremely large, virtually no quality loss - not recommended - F32: absolutely huge, lossless - not recommended
Learning about quantization suffixes while `ollama pull llama3.3` download completes (fyi, quantization for the default 70b is q4_K_M)
• make-ggml .py: github.com/ggerganov/ll...
• pull request: github.com/ggerganov/ll...
It's a bit sad though AWS have probably the most challenging "try it out" out of all big LLM providers.
docs.aws.amazon.com/nova/latest/...
* There are typos in the examples (s/client/bedrock)
* Non-OpenAI like API (fixed by LiteLLM)
* Bruh
aws.amazon.com/ai/generativ... Looks quite impressive. I really appreciate the direction towards cost and speed optimization rather than accuracy for most cases I care about.
So called "static graph" corresponds to the process of embedding images into the html code. That is also the reason people use the word "embedding" so often now.
3. Structured outputs are awesome and make tremendous value in real business processes which involve "go / no go" constraints. You don't want to wiggle your way to include those constraints in prompt in English. Most software is deterministic, struct outputs help to blend LLMs with it.
2. It's not completely obvious how providers implement their "advanced" features like JSON mode. One needs to read the docs (obv) and play around with it to get a good grasp of how is it REALLY implemented.
Structured outputs make no negative impact on reasoning abilities of LLMs (I've also observed that in practice).
Key takeaways:
1. Proper design of prompts is important
Even the most senior researchers can do wrong. Many builtin prompts in packages are also bad.
blog.dottxt.co/say-what-you...
Not with that attitude
Also: you can also use variables (or expressions?!) for the formatting information! #Python is cool...
More details and explanation at fstring.help
Absolutely/Based
My outsider opinion is that EU have chosen societal stability (in all aspects, including significant freeze in social lifts) in exchange for dynamism. I do think it is deeply human to seek comfort and lack of surprises, so I do not think EU is dumb. But they will need to readjust in times of change.
Maybe one can also interact with aging data as well.
Remind that there was something really interesting back then, but forgotten for a while in Read-It-Later style of apps for example. Or liberate the users like Arc browser does with their auto-discarded tabs.
Rustified apps are such a nice improvement of quality of life.
Also signifies that people *do* in fact care about performance despite all the "good enough performance is enough" affirmations.
uv makes python management really painless to a degree of it becoming non-issue at all.
It's so strange to feel immediate ick on Twitter/X given their update that
1) made videos autoplay
2) pushed too many "meme"/tiktok/youtube shorts like entertainment videos
3) pushed me to Followed tab
to only find extremely low activity of people whom I actually following there.