dust42

12 ms·

121.
by dust42 11mo ago
It just depends on what you are doing. A green field react app in typescript with a CRUD API behind? The LLMs are a mind blowing assistant and 1000t/s is crazy. You are doing embedded development or anything else not as mainstream as w
122.
by dust42 1y ago
I usually use unsloth quants, in this case https://huggingface.co/unsloth/Qwen3-Coder-30B-A3B-Instruct-... - the Q4_K_M variant. On 128GB I would definitely run a larger model, probably with ~10B active parameters. All
123.
by dust42 1y ago
On a Macbook pro 64GB I use Qwen3-Coder-30B-A3B Q4 quant with llama.cpp. For VSCode I use continue.dev as it allows to set my own (short) system prompt. I get around 50token/sec generation and prompt processing 550t/s. When giving
124.
by dust42 1y ago
Interesting project but I wonder if this is mostly academic. Wheatstone bridges are very sensitive to tampering. While the reflectometry is pretty cool I wonder if for practical purposes the extra sensitivity is needed for physical intrusio
125.
by dust42 1y ago
The graph also only starts in 2001. I worked as a student at the EMBL (European Molecular Biology Laboratory) in the bio-physics instrumentation group in the mid 90ies. The lab group was developing prototypes of thin-film electrophoresis DN
126.
by dust42 1y ago
Thinking about it, I am sure it is only a matter of time until a self driving car or a robot will be used to kill a human. Or on a lower level for a DDoS attack - all cars/robots going to the white house.
127.
by dust42 1y ago
> How does this square with the video where they showed it running continuously for an hour doing an actual Amazon package sorting job? https://www.youtube.com/watch?v=lkc2y0yb89U The video shows several of glitches. Fro
128.
by dust42 1y ago
Probably they were next to sports stadiums. Most of the time there is ample unused cell capacity. Also: the New York marathon hosts 50k+ runners. I don't easily believe that 100k phones distributed over several sites on a 35 mile radiu
129.
by dust42 1y ago
Is the trained model available?
130.
by dust42 1y ago
It is available behind an experimental flag: --experimental-transform-types
131.
by dust42 1y ago
Exactly this. And that is why I like this question because the amount of correct details and the amount of nonsense give a good idea about the quality of the model.
132.
by dust42 1y ago
The 120B model badly hallucinates facts on the level of a 0.6B model. My go to test for checking hallucinations is 'Tell me about Mercantour park' (a national park in south eastern France). Easily half of the facts are invented. N
133.
by dust42 1y ago
Looking in their github, it says that it uses openCV and Tensorflow. The motion detection is done with openCV and will be immune against any attack unless you move so slow that you are under the detection threshold. Tensorflow for the objec
134.
by dust42 1y ago
I tried with Claude Sonnet 4 and it does *not* work. So looks like GLM-4.5 Air in 3bit quant is ahead. Chat is here: https://claude.ai/share/dc9eccbf-b34a-4e2b-af86-ec2dd83687ea Claude Opus 4 does work but is far behin
135.
by dust42 1y ago
I am really afraid of this as well. When using one of the plugins for vscode, I would easily use a few million tokens a day. Thus with Claude Code I assume it isn't much different under the hood. The prices on openrouter are roughly $5
136.
by dust42 1y ago
Minor correction, BERT is an encoder (not encoder-decoder), ChatGPT is a decoder. Encoders like BERT produce better results for embeddings because they look at the whole sentence, while GPTs look from left to right: Imagine you're tryi
137.
by dust42 1y ago
Look, it takes two prompts and two minutes to create this page: https://claude.ai/share/20b36bd3-d817-4228-bc33-aa7c4910bc2b (the preview seems to only work in Chrome, for Firefox you have to download the html). Plus m
138.
by dust42 1y ago
As user murshudoff mentioned elsewhere in the discussion, openrouter has an endpoint to get the prices. Takes 1 minutes to get them.
139.
by dust42 1y ago
Well, openrouter lists currently 473 models of dozens of providers. Three providers and 26 models is 2 minutes of Google search. I stand by my word to call that low effort.
140.
by dust42 1y ago
tldr; low effort website that only contains 26 Google, OpenAI and Anthropic models and only input and output prices but no info about prompt cache and prompt cache prices. For a list of 473 models of 60+ providers with input, output, contex
141.
by dust42 1y ago
Same happened in Germany: in the region of Mittelbaden, harmful per- and polyfluorinated chemicals (PFAS or PFC) have been contaminating groundwater and arable land for years. The district of Rastatt, which is particularly affected, holds a
142.
by dust42 1y ago
Actually Le Chat is french for the (male) cat. Also 'Le Chat' is a well known laundry detergent (of german origin - Henkel company). The headline 'Le Chat takes a deep dive' means 'the cat takes a deep dive'.
143.
by dust42 1y ago
I have one for you: implement gemma 3n multimodel support in llama.cpp
144.
by dust42 1y ago
Apart from the AI getting stuck for at least 20 seconds, an impressive experience. More professional than 80% of screening calls I have received in the past. Did correctly understand and summarise my input, not much noticable delay - except
145.
by dust42 1y ago
Two weeks ago the same website (heise.de) reported that the german military will use Google cloud services [1]. If there is any place where local cloud and open standards would make sense, then it seems to me the military would be it. I can
146.
by dust42 1y ago
If your local timezone is GMT>=0, then you wont see it. A special form of a Heisenbug that you only see easily when located in the Americas, while when in Europe/Africa/Asia it is invisible unless you switch your local time zon
147.
by dust42 1y ago
You do yourself. This service exists in Germany and likely in many other countries since a quarter of a century. Cost: 15€/month. The paper letters are collected and once per month forwarded to you. [1] https://www.deutschep
148.
by dust42 1y ago
It is a 4-bit quant gemma-3-4b-it-Q4_K_M.gguf. I just use "describe" as prompt or "short description" if I want less verbose output. As you are a photographer, using a picture from your website gemma 4b produces the foll
149.
by dust42 1y ago
To add some numbers, on MBP M1 64GB with ggml-org/gemma-3-4b-it-GGUF I get 25t/s prompt processing 63t/s token generation Overall processing time per image is ~15secs, no matter what size the image is. The small 4B
150.
by dust42 1y ago
I have a MBP M1 Max 64GB and I get 40t/s with llama.cpp and unsloth q4_k_m on the 30B A3B model. I always use /nothink and Temperature=0.7, TopP=0.8, TopK=20, and MinP=0 - these are the settings recommended for Qwen3 and they make

More ›