Hacker News Reader

  Location: Germany 
  Remote: Yes
  Willing to relocate: No
  Roles: Technical Writer, Software Engineer 
  Homepage: https://kay.is
  LinkedIn: https://www.linkedin.com/in/kay-plößer

k__on LLMs reward expertise

Prompt an image or video generator without knowledge in photography or art skills and your results will look sloppy.

k__on Can an AI Model Obey a Strict Style Guide?

I did a bit of research on that in my last job and got the impression that encoder models might help.

They are well suited to check input for rule violations and are much cheaper to train and run than decoder models.

The downside is, they can't fix issues by themselves.

k__on Qwen3.8-Max: A New Bar for Coding and Cowork

Yeah, I'd assume it's possible to extract all languages as steering vectors from a model and then substract the ones you don't need from its weights.

However, that would just change the weights values and not their dimensions.

k__on More German than many Germans

I think, the main issue with German rules is that we haven't embraced digital technology 100%.

All these rules would be way less cumbersome if they didn't come with a bunch of literal paperwork.

k__on Generative AI floods and dilutes the market for books

The ratio of good books to slop (AI or not) was far to bad for humans to reasonably filter 20 years ago.

k__on DeepSeek V4 Flash 0731 Intelligence, Performance and Price Analysis

Haha, and I almost felt bad after seeing this chart yesterday.

k__on DeepSeek V4 Flash 0731 Intelligence, Performance and Price Analysis

Half OT:

Why do the cache hit rates seem to vary so much between harnesses?

I use pi, which is very minimalist, and I get a hit rate of ~99%. Paying like $1 a day for Flash. Yet, the hit rate mentioned on OpenRouter is only ~79%.

k__on DeepSeek V4 Flash 0731 Intelligence, Performance and Price Analysis

On OpenRouter it's 93 TPS.

k__on DeepSeek-V4-Flash Update

I'm using pi and my caching is ~99%.

k__on DeepSeek-V4-Flash Update

Yeah, it needs quite some hand holding.

I didn't do much agent coding and had a mix experience.

1. It would build something that was in the spirit of what I wanted, but unusable in practice.

2. It would build something quite useful, but only the public APIs were nice, the deeper code layers would get more and more convoluted.

3. It would built what I wanted and it would have okay-ish code.

However, for 3. I also had to add a custom AGENTS.md, many more code example, extra repos as subtrees, and review any code that had new concepts.

Much more work, but still much less than typing it all by hand.

k__on DeepSeek-V4-Flash Update

I'd take more throughput while everything else stays the same.

k__on Kimi K3-256k

While I tend to clear my session after every task, I feel less stressed when my context is as 10% than when it's at 30%

k__on Kimi K3-256k

They remove potentially irrelevant details.

k__on Kimi K3-256k

As I understand it, they would have to train a whole knew model to hard cap it's context to different lengths. That would be cheaper to train and had cheaper inf, but still a huge investment.

So I'd guess it's API level.

k__on Kimi-K3 Releases on HuggingFace 7/27

Right

k__on The front end framework for correctness: built on Effect, architected like Elm

Yeah, I think the Effect team tried building a compiler for once (TS++ or something) but they abandoned it, as it was too much work.

k__on Writing by hand is good for your brain

Do crosswords count?

k__on Qwen 3.8

My 2 weeks with DeepSeek V4:

Pro is ~50% more expensive than Flash.

Both need babysitting.

Plan, split in small tasks, give it docs, types, tests, linter, best practice examples, etc.