HNHacker News
TopNewBestAskShowJobs

dajonker

967 karma · joined August 19, 2015

submissionscomments
dajonker··on Launch HN: RunAnywhere (YC W26) – Faster AI Inference on Apple Silicon
I use voxtype on my Linux machine with parakeet. Super fast and regularly even gets the tech lingo correct. You can configure prompts and keywords to help with that as well.
dajonker··on Apple Studio Display and Studio Display XDR
Look at how many people only use their 14 inch laptop screen, it's ridiculous and terribly unergonomic.
dajonker··on Qwen3.5 122B and 35B models offer Sonnet 4.5 performance on local computers
Radeon R9700 with 32 GB VRAM is relatively affordable for the amount of RAM and with llama.cpp it runs fast enough for most things. These are workstation cards with blower fans and they are LOUD. Otherwise if you have the money to burn get a 5090 for speeeed and relatively low noise, especially if you limit power usage.
dajonker··on Museum of Plugs and Sockets
I like the UK sockets because they have a switch.
dajonker··on GPT-5.3-Codex
There was never a $100 billion deal. Only a letter of intent which doesn't mean anything contractually.
dajonker··on Claude Code daily benchmarks for degradation tracking
Wouldn't be surprised if they slowly start quantizing their models over time. Makes it easier to scale and reduce operational cost. Also makes a new release have more impact as it will be more notably "better" than what you've been using the past couple of days/weeks.
dajonker··on Trinity large: An open 400B sparse MoE model
You are equally understating past performance as you are overstating current performance.

One year ago I already ran qwen2.5-coder 7B locally for pretty decent autocomplete. And I still use it today as I haven't found anything better, having tried plenty of alternatives.

Today I let LLM agents write probably 60-80% of the code, but I frequently have to steer and correct it and that final 20% still takes 80% of the time.

dajonker··on Qwen3-Max-Thinking
These LLM benchmarks are like interviews for software engineers. They get drilled on advanced algorithms for distributed computing and they ace the questions. But then it turns out that the job is to add a button the user interface and it uses new tailwind classes instead of reusing the existing ones so it is just not quite right.
dajonker··on GLM-4.7-Flash
Yes! This update works great. Seems to be pretty good at first glance. I'll have to setup an interesting task and see how different models approach the problem.
dajonker··on Show HN: Sweep, Open-weights 1.5B model for next-edit autocomplete
I use llama.vim with llama.cpp and the qwen2.5-coder 7B model. Easily fits on a 16 GB GPU and is fast even on a tiny RTX 2000 card with 70 watts of power. Quality of completions is good enough for me, if I want something more sophisticated I use something like Codex
dajonker··on GLM-4.7-Flash
Update: I'm experiencing issues with OpenCode and this model. I have built the latest llama.cpp and followed the Unsloth guide, but it's not usable at the moment because of:

- Tool calling doesn't work properly with OpenCode

- It repeats itself very quickly. This is addressed in the Unsloth guide and can be "fixed" by setting --dry-multiplier to 1.1 or higher

- It makes a lot of spelling errors such as replacing class/file name characters with "1". Or when I asked it to check AGENTS.md it tried to open AGANTS.md

I tried both the Q4_K_XL and Q5_K_XL quantizations and they both suffer from these issues.

dajonker··on GLM-4.7-Flash
I don't know whether it just doesn't work well in GGUF / llama.cpp + OpenCode but I can't get anything useful out of Devstal 2 24B running locally. Probably a skill issue on my end, but I'm not very impressed. Benchmarks are nice but they don't always translate to real life usefulness.
dajonker··on GLM-4.7-Flash
Yes I usually run Unsloth models, however you are linking to the big model now (355B-A32B), which I can't run on my consumer hardware.

The flash model in this thread is more than 10x smaller (30B).

dajonker··on GLM-4.7-Flash
Great, I've been experimenting with OpenCode and running local 30B-A3B models on llama.cpp (4 bit) on a 32 GB GPU so there's plenty of VRAM left for 128k context. So far Qwen3-coder gives the me best results. Nemotron 3 Nano is supposed to benchmark better but it doesn't really show for the kind of work I throw at it, mostly "write tests for this and that method which are not covered yet". Will give this a try once someone has quantized it in ~4 bit GGUF.

Codex is notably higher quality but also has me waiting forever. Hopefully these small models get better and better, not just at benchmarks.

dajonker··on IPv6 just turned 30 and still hasn't taken over the world
Of course you can subnet ipv6, in fact I run several ipv6 subnets at home. You have to delegate a different prefix to each subnet.
dajonker··on T-Ruby is Ruby with syntax for types
At least CPython and CRuby (MRI), the most common implementations of each language, ignore all type hints and they are not able to use them for anything during compile or runtime. So the performance argument is complete nonsense for at least these two languages.

Both Python and Ruby (the languages themselves) only specify the type hint syntax, but neither specifies anything about checking the actual types. That exercise is left for the implementations of third party type checkers.

dajonker··on T-Ruby is Ruby with syntax for types
Well said. There are many problems you have to deal with when writing code and type annotations only solve one particular kind. And even type annotations can be wrong: when you're dealing with data from external sources, dynamic languages like Python, JavaScript and Ruby will happily parse any valid JSON into a native data structure, even if it might not be what you specified in your type hints. Worse yet, you may not even notice unless you also have runtime type checks.

The kind of messy code base that results from (large) numbers of (mediocre) developers hastily implementing hacky bug fixes and (incomplete) specifications under time pressure isn't necessarily solved by any technical solution such as type hints.

dajonker··on Ruby 4.0.0
I don't have a great code example at hand unfortunately, but I found that people often tend to write more "nominally" typed code (expecting explicitly named classes) rather than taking advantage of duck typing (interfaces, structural types), meaning the code becomes more rigid, harder to change and more time wasted on resolving all the type checks, even if the code otherwise is perfectly reasonable and free of bugs.

In other words, I found that the resulting code often looked more like Java but with weaker guarantees about types and much worse performance.

dajonker··on Ruby 4.0.0
In my honest opinion, if you can't live without static typing, Ruby just isn't for you.

Adding static typing to a dynamic language mostly gives you the disadvantages of both, without a lot of benefits. It's better to stick to languages that were designed with static types from the start.

I love programming in Ruby, having to worry about type annotations and the additional constraints that come with them would take a lot of the fun out of that.

dajonker··on NextSilicon reveals new processor chip in challenge to Intel, AMD
Even if the hardware is really good, the software should be even better if they want to succeed.

Support for operating systems, compilers, programming languages, etc.

This is why a Raspberry Pi is still so popular even though there are a lot of cheaper alternatives with theoretically better performance. The software support is often just not as good.

dajonker··on DeepSeek OCR
Try MinerU 2.5 with two-step parsing. It gives good results with bounding boxes per block. Not sure if you can get it to do more detailed such as word or character level.
dajonker··on Garbage collection for Rust: The finalizer frontier
Implementing a doubly linked list without either unsafe or some very confusing code that could arguably win an obfuscation contest.
dajonker··on Garbage collection for Rust: The finalizer frontier
If I understand the article correctly it's for those cases where you want memory safety (i.e. not using "unsafe") but where the borrow checker is really hard to work with such as a doubly linked list, where nodes can point to each other.

For the rest you'd still use non-GC rust.

dajonker··on AV2 video codec delivers 30% lower bitrate than AV1, final spec due in late 2025
I believe patents play a big role here as well. Anything new must be careful to not (accidentally) violate any active patent, so there might be some tricks that can't currently be used for AV1/AV2
dajonker··on The product of the railways is the timetable
That graphic from the Swiss Railways is absolutely wonderful

https://sma-partner.com/storage/app/media/Dokumente/Netzgraf...

dajonker··on Python developers are embracing type hints
A dynamic language with type hints has all of the disadvantages of a dynamic language with all the disadvantages of a static language on top of it.
dajonker··on Cloudflare Email Service: private beta
I haven't experienced any price increase on the cheapest Postmark tier over the past 3 years or so? In any case they deliver excellent service and as a business earning money and sending emails per transaction it's almost free.
dajonker··on Zed's Pricing Has Changed: LLM Usage Is Now Token-Based
The edit prediction model is open source https://huggingface.co/zed-industries/zeta

It's based on qwen 2.5 coder 7B and you should be able to run it locally quite easily since it would only require about 8 GB of VRAM for the 8-bit version. Not sure if Zed supports this though, I'm not a Zed user myself.

dajonker··on A staff engineer's journey with Claude Code
Exactly, makes it feel almost like an advertorial for Anthropic, who likely need most customers to pay 1000 bucks a month to break even.
dajonker··on The Framework Desktop is a beast
I guess that depends on your definition of "decent". For the smaller models that can run on a 16/24/32 GB nvidia card, the chip is anywhere between 3x and 10x slower compared to say a 4080 super or a 3090 which are relatively cheap used.

Biggest limitations are the memory bandwidth which limits token generation and the fact it's not a CUDA chip, meaning longer time until first token for theoretically similar hardware specifications.

Any model bigger than what fits in 32 GB VRAM is - in my opinion - currently unusable on "consumer" hardware. Perhaps a tinybox with 144 GB of VRAM and close to 6 TB/s memory bandwidth will get you a nice experience for consumer grade hardware but it's quite the investment (and power draw)

← PreviousPage 2 of 9Next →