francisduvivier

I think mainly that he can move much faster with specific improvements targeting Deepseek on Systems with unified memory (Mac or Strix). It's a lot easier to optimize if you don't need to worry about all the other architectures. So optimize he did and it's just a lot faster than llama cpp for deepseek v4 pro and flash. Also interesting features are more doable, like SSD streaming, which makes it possible to load MOE weights for a model larger than your VRAM, I don't see that landing in llama cpp anytime soon.

Well you have next-i18next and NextAuth.js, those integrations with Next.js help for getting something running quickly.

I recently migrated 2 projects away from Next.js to Vite. It was a lot of work, Next.JS handles quite a lot of stuff: translations, authentication, bundling, css stuff, caching... But it was very much worth it for lowering my daily infuration. I don't think I'll be using NextJS again if I don't have to.

Yes true :D, I kind of just want to blame gpt-5 mini for it and that's one of the bad things about the bad coding, I immediately loose part of the sense of ownership and responsibility. I don't feel like I made it, I just managed it.

Now to be honest I saw this bug, but I decided to just release it anyways because I also already had the v2 in the works which incidentally already had this issue fixed.

I think it's just a tunneling bug that happens when the point that make up the wall get messed up a enough. Almost never happens anymore in the v2[1] that I added. This one also allows you to see these points.

[1] https://francisduvivier.github.io/eternal-struggle-with-spee...

Well that's fixed in the V2 with even more vibe coding:

https://francisduvivier.github.io/eternal-struggle-with-spee...

Some people here were asking for it so I quickly vibe forked a speed control slider for farming some karma here on Hacker News:

https://francisduvivier.github.io/eternal-struggle-with-spee...

Code: https://github.com/francisduvivier/eternal-struggle-with-spe...

Wonder how this compares to open interpreter.

Now that you mention it, the payment system was not working in a ski resort restaurant in Switzerland this noon.

They had to switch to cash.

Lol I can see trolls on reddit using this when people are asking for which npm npm package they should use.

"Just install the everything package, then you will be sure to have the right package"

Pretty cool, but the bigger the display, the more that refresh rate becomes an issue. Say you can refresh the outer pixels on the horizonal axis at 100 fps. Then if you want to have 50 pixels on the outer half circle, then you can change them at 2fps.

Indeed, on the Lex Fridman podcast, Musk said it was a very tough recruiting battle for Ilya. Now it looks like they have a shot at getting him back.

What are the things that you find awful about MacOS?

Correction: looks like it has to do with the quantization rather: 8bit quantization works while less does not not seem to work. Other working model example (no conversion needed): https://huggingface.co/TheBloke//Yarn-Mistral-7B-64k-GGUF/ya...

I also had the same issue, in my case it was because I was trying to use a llama 2 model. When trying with codellama https://huggingface.co/TheBloke/CodeLlama-7B-GGUF/tree/main, which is based on the first llama, it works.

That GPU does pretty well for running an LLM though.

Llama 7B at 2.8 tok/s via Mlc-chat.

I'm setting up an LLM discord bot with it.

Hi, happen to have some reference for this?

Yes, and there are more advantages to the Orange Pi 5 SBCs:

- They have full size HDMI - They have an m.2 connector - They still have an audio jack - They have more powerful performance cores - They have low performance cores - They have a 6 TOPS NPU

On the other hand, that cheap base model has no built-in wifi, but so you can add it in the m.2 slot.

For me, the Raspberry Pi 5 is quite disappointing. But hey, the good thing is that maybe this will push more of the community to the RK3588(S) based boards.

Well I am horrible at playing this, but here you can see how it basically looks: https://github.com/francisduvivier/sedmario_with_gif/blob/ma...

Interesting here how the HN rules seem to be a limitation. I'd say the fact that Amazon is using it's power in a very unexpected way should be highlighted somehow in the title.

But because the original blog writer doesn't do that, current HN rules say you should not do that either. Instead, HN rules say you should post the original article and keep the original title. But in this case, by doing that you are effectively hiding what actually happened. This original title sounds so uninteresting that I'd say it would be warranted here to post the article with the title of https://news.ycombinator.com/item?id=36294748

I'd want to downvote this, but in absence of that option I'll comment here instead to show I disagree. First, although he's indeed also reading the article, it's not like that is the only thing he is doing. He is adding a broader perspective to it. Secondly, the tabloid fodder accusation here is really not any value here, please don't make HN commentary second into something where they just sling random accusations around. Third, I agree that it's not just some podcaster as stated by the other commenters.

Also, the first line of the video description of the video is a link to the medium article.

Since the space/weight for batteries is limited, I bet it will be more like iphone, so only WebKit and tight control to optimize battery usage.

Well that's how they make it sound, but they actually just said not annually.

But it doesn't make a lot of sense to me since influenza will be mutating to escape the immune response.

Haha, I wonder how funny this will be when Canonical eventually really needs to dump it.

Note that they just released the 5B version, which loses the M.2 slot buy gains emmc storage and wifi 6!

I have the original 5 board running for some weeks now with some docker containers and it's running great.

also wondering about this

32GB is not available for sale as far I and gizchina can find.

I have this board, you better add some cooling but I had no issue setting up docker containers with node red, prometheus, grafana, Emby and Home assistant.

Too bad I currently can't succeed in using a raspberry pi camera on it.

There is a crazy amount of SBC's coming out with the RK3588, so I aldo think the chances are pretty good for this one.

But its more than 3 times as dense as PETG still. 1kg of PETG makes is 300+ meter of filament, while 1kg of this stuff is only 100 meter.

So I think the casual observer would notice that. But also means with thar price per KG.