You can't run Plex directly off the device like a DS224+ would.
3,007 karma · joined October 23, 2012
You can't run Plex directly off the device like a DS224+ would.
He’s not trying to defend his hypothesis at a PhD thesis defense in front of a panel.
It’s more like he’s on the stage at a small TEDx talk. He’s not trying to offer proof for each statement, he’s trying to almost play stand up comedian keeping the crowd engaged.
Is this just extremely overfitted?
Is there a way for us to test this? Or even if the model isn't open source, I'd pay $1 to upload a capture from my wifi card on my linux box and upload it to the researchers and have them generate a picture and see if it's accurate
Which means it has ~3b parameters active per token.
Qwen3-32b has 32b params active per token
It's such a simple question: "For someone who does not want to run the model locally, what is the difference between these 2 models on the API?" and yet nobody can answer that question.
They also released today Qwen3-VL Plus [1] today alongside Qwen3-VL 235B [2] and they don't tell us which one is better. Note that Qwen3-VL-Plus is a very different model compared to Qwen-VL-Plus.
Also, qwen-plus-2025-09-11 [3] vs qwen3-235b-a22b-instruct-2507 [4]. What's the difference? Which one is better? Who knows.
You know it's bad when OpenAI has a more clear naming scheme.
[1] https://modelstudio.console.alibabacloud.com/?tab=doc#/doc/?...
[2] https://modelstudio.console.alibabacloud.com/?tab=doc#/doc/?...
[3] https://modelstudio.console.alibabacloud.com/?tab=doc#/doc/?...
[4] https://modelstudio.console.alibabacloud.com/?tab=doc#/doc/?...
A 2x 3090 build is okay for inference, but even with nvlink you're a bit handicapped for training. You're much better off with getting a 4090 48GB from China for $2.5k and just using that. Example: https://www.alibaba.com/trade/search?keywords=4090+48gb&pric...
Also, this phrasing is concerning:
> WARNING - these components don't fit if you try to copy this build. The bottom GPU is resting on the Arctic p12 slim fans at the bottom of the case and pushing up on the GPU. Also the top arctic p14 Max fans don't have mounting points for half of their screw holes, and are in place by being very tightly wedged against the motherboard, case, and PSU. Also, there's probably way too much pressure on the pcie cables coming off the gpus when you close the glass.
If nothing else, it’ll inspire the next generation of mathematicians.
At least have it be split into some files and structured in some way.
Even if you can capture every single bit of extra energy from a tshirt, you'd end up with a tiny fraction of 100W. Certainly not enough to power a mobile device like this.
I'm sure a lot of engineering hours were spent on getting the door handle on your car to the exact safety/cost/functionality requirements, and at the end of the day, it's a door handle. Replace "door handle" with 99% of hardware and software that you ever see, and the same thing still applies. And yet, imagine using a car without a door handle.
Most important work isn't sexy, it's banal stuff that's boring until you remove it and realize how important it is.
Well, OpenAI and other companies training AI models have shown that the architecture of the model matters less than the quality of data fed into it. Same applies for humans.
I understand that the Great Firewall is mostly about censoring dissent, but it's also to keep Chinese citizens away from junk food media sources. The type of videos you see on Douyin vs Tiktok is a great example of the difference.
Yes, the videos on Douyin are politically censored, but they're also a lot less brainrot than Tiktok videos. The Tiktok algo is optimized for ad impressions and profit, whereas the Douyin algo is more tuned to some nebulous concept of Confucian social harmony, for better or worse.
A more nuanced take is that I don't think it's useful to measure Chinese govt behavior just mapped to "amount of suppressing political dissent". I actually think the level of censorship is above the level required for that. It's more useful to recognize that "suppressing political dissent" is actually a subset of Confucian "promote social harmony"- which is not strongly valued in the USA but is at least important enough to be paid lip service in China- and I suspect a big chunk of educated members of government may truly believe in that ideal. It explains behaviors like "why the Douyin algo is so different from Tiktok" and other overreaches of the Chinese govt, because it's not solely about suppressing dissent.
Llama 4's release in 2025 is (deservedly) panned, but Llama 3.1 405b does not deserve that slander.
https://artificialanalysis.ai/#frontier-language-model-intel...
Do not compare 2024 models to the current cutting edge. At the time, Llama 3.1 405b was the very first open source (open weights) model to come close to the closed source cutting edge. It was very very close in performance to GPT-4o and Claude 3.5 Sonnet.
In essence, it was Deepseek R1 before Deepseek R1.
Let's say you have a model that generates the string "The 44th president of the United States is ___ ___". Your model will generate "Barack" as the n+1 token, and the MTP layer probably does a good enough job to generate "Obama" as the n+2 token (even though that MTP layer is a mere <2bil parameters in size). Then you just check if "Obama" is correct via the same speculative decoding process, which is a lot faster than if you had to start over from layer 1-48 and generate "Obama" the regular way.
Instead of generating tokens one at a time, you generate the second one as well, and then use speculative decoding on that second token (instead of having it be produced by a draft model like Qwen 0.6b). If the token is checked and is correct, then the 2nd token gets generated MUCH faster.
If it's wrong, you have to generate it again the normal way (a lot slower than just checking it). Usually, it's correct, so inference is a lot faster.
Deepseek R1 also has a MTP layer (layer 61) https://huggingface.co/deepseek-ai/DeepSeek-R1/blob/main/mod...
But Deepseek R1 adds embed_tokens and shared_head.head tensors, which are [129280, 7168] or about 2GB in size at FP8.
Qwen3-Next doesn't have that: https://huggingface.co/Qwen/Qwen3-Next-80B-A3B-Instruct/blob...
So it saves a few GB in active parameters for MTP, which is a Big Deal. This is one of the changes that helps significantly speeds up inference.
I'm pretty sure that's just a frontend for Uptime Kuma https://github.com/louislam/uptime-kuma
Seriously, this is extremely mild as far as issues go. There's basically no incentive to fix this problem, because I bet even the people who lost a few pennies of electricity would still prefer the convenience of ollama not having auth.
Plus, that's the worst case scenario, in real life even if some black hat found an exposed ollama service, they have no interest in generating tokens for <insert random LLM here at 4 bit quant> at a slow speed of <50tok/sec.
Plenty of cops have shot/executed innocent people, and were declared not guilty due to qualified immunity.
Plus, a phone in China can't execute you either. It'd be the police goons who do that.
Everyone should remember, limitations of technology is not meant to define society. Instead, we build edge cases into technology to better match society’s general expectations.
A website owner saying “yes normal humans, no bad bots, EXCEPT good bots” is totally fine.