tarruda

8 ms·

91.
by tarruda 8mo ago
They seem to be the same company that released ACEStep music generation model: https://acestep.io/ Though the only mention I found was in ComfyUI docs: https://docs.comfy.org/tutorials/audio/ace-st
92.
by tarruda 8mo ago
This is probably one of the most underrated LLMs releases in the past few months. In my local testing with a 4-bit quant ( https://huggingface.co/ubergarm/Step-3.5-Flash-GGUF/tree/mai... ), it surpasses every o
93.
by tarruda 8mo ago
Considered getting a 512G mac studio, but I don't like Apple devices due to the closed software stack. I would never have gotten this Mac Studio if Strix Halo existed mid 2024. For now I will just wait for AMD or Intel to release a x86
94.
by tarruda 8mo ago
Mac Studios or Strix Halo. GPT-OSS 120b, Qwen3-Next, Step 3.5-Flash all work great on a M1 Ultra.
95.
by tarruda 8mo ago
At this point I wouldn't be surprised if your pelican example has leaked into most training datasets. I suggest to start using a new SVG challenge, hopefully one that makes even Gemini 3 Deep Think fail ;D
96.
by tarruda 8mo ago
Would love to see a Qwen 3.5 release in the range of 80-110B which would be perfect for 128GB devices. While Qwen3-Next is 80b, it unfortunately doesn't have a vision encoder.
97.
by tarruda 8mo ago
Love the idea of keeping the agent filesystem in a single file!
98.
by tarruda 8mo ago
I'm only interested in the local, single user use case. Plus I use a Mac studio for inference, so vLLM is not an option for me.
99.
by tarruda 8mo ago
These days I don't feel the need to use anything other than llama.cpp server as it has a pretty good web UI and router mode for switching models.
100.
by tarruda 10mo ago
AFAIK MPS cannot be used on Asahi, so it has to be done using Vulkan which will definitely be much slower.
101.
by tarruda 10mo ago
A VM is displayed as a window on the host OS and Emacs is the window manager within that VM window. What's the difference from running emacs directly as an application on the host?
102.
by tarruda 10mo ago
> but it’s light years behind on compute. Is that the only factor though? I wonder if pytorch is lacking optimization for the MPS backend.
103.
by tarruda 10mo ago
> It's fast (~3 seconds on my RTX 4090) It is amazing how far behind Apple Silicon is when it comes to use non- language models. Using the reference code from Z-image on my M1 ultra, it takes 8 seconds per step. Over a minute for th
104.
by tarruda 10mo ago
> Do you disagree with that? I think that Qwen3 8B and 4B are SOTA for their size. The GPQA Diamond accuracy chart is weird: Both Qwen3 8B and 4B have higher scores, so they used this weid chart where "x" axis shows the number
105.
by tarruda 10mo ago
Here's what I understood from the blog post: - Mistral Large 3 is comparable with the previous Deepseek release. - Ministral 3 LLMs are comparable with older open LLMs of similar sizes.
106.
by tarruda 10mo ago
Deepseek is already a MoE
107.
by tarruda 10mo ago
You can run at ~20 tokens/second on a 512GB Mac Studio M3 Ultra: https://youtu.be/ufXZI6aqOU8?si=YGowQ3cSzHDpgv4z&t=197 IIRC the 512GB mac studio is about $10k
108.
by tarruda 10mo ago
Only a matter of time before using coding agents with local LLMs is a viable alternative.
109.
by tarruda 11mo ago
I had a terrible first impression with Gemini CLI a few months ago when it was released because of the constant 409 errors. With Gemini 3 release I decided to give it another go, and now the error changed to: "You've reached the d
110.
by tarruda 11mo ago
Interesting that the 8B of the Qwen3-VL family 9th place, above a few proprietary models. This thing can run locally with llama.cpp on modest hardware.
111.
by tarruda 11mo ago
> Why not just reject papers authored by LLMs and ban accounts that are caught? Are you saying that there's an automated method for reliably verifying that something was created by an LLM?
112.
by tarruda 1y ago
> LLMs do not have a mechanism for sampling from given probability distributions Would a LLM with tool calls be able to do this?
113.
by tarruda 1y ago
> dunno if the life vest bit comment of yours was sarcastic, but it is a funny remark for sure :-) It was a quote of the linked article: "Holtec International, which owns the closed nuclear facility, reported the worker was a contra
114.
by tarruda 1y ago
> Here is how I do it on my Hetzner bare-metal servers using Ansible: https://gist.github.com/fungiboletus/794a265cc186e79cd5eb2fe ... It also works on VMs. As someone with zero ansible experience, can you elaborate
115.
by tarruda 1y ago
I'm wondering what would be the use case for a laptop with 24TB storage.
116.
by tarruda 1y ago
I have implemented a simple asyncio compatible micro event loop library in python. The goal was to understand the underlying mechanisms behind python's async/await and to help coworkers understand how event loops work under the ho
117.
by tarruda 1y ago
> More stable than what? IIRC the whole drama began because Kent was constantly pushing new features along with critical bug fixes after the proper merge window. I meant stable in the sense where most changes are bug fixes, reducing the
118.
by tarruda 1y ago
I hope it eventually comes back once it is more stable. Would be great to have an in kernel alternative to ZFS for parity RAID.
119.
by tarruda 1y ago
It is not the same as COSMIC, but Krohnkite is very customizable and has like 9 tiling modes.
120.
by tarruda 1y ago
I have recently replaced COSMIC by KDE + Krohnkite on the newly released Debian 13. After some tweaking of the key bindings, I managed to make it behave very similarly to COSMIC.

More ›