jychang

11 ms·

151.
by jychang 10mo ago
Rybelsus exists already, so there's that
152.
by jychang 10mo ago
Yeah, that's the point of publishing. You get scooped, you lose.
153.
by jychang 10mo ago
Deepseek? It's literally in their research papers. OpenAI? The OpenAI head of research @markchen90 straight up admitted it in a podcast. https://x.com/petergostev/status/1995744289079656834 "In the last
154.
by jychang 10mo ago
That's still behind the times. Even the ancient dinosaur IBM had released a Mamba model [1] before this paper was even put out. > Granite-4.0-Tiny-Base-Preview is a 7B-parameter hybrid mixture-of-experts (MoE) language model featuri
155.
by jychang 10mo ago
Mamba based LLMs aren't even close to novel though. IBM's been doing this since forever [1]. Also, you're off on Deepseek V3.2's param count, the full model's 685B in size with the MTP layer. I don't think ther
156.
by jychang 10mo ago
Kimi K2 is literally a "copy Deepseek's homework" model. Seriously. It's even exactly 61 layers, the same as Deepseek V3/R1.
157.
by jychang 10mo ago
Or like this: https://api-docs.deepseek.com/news/news251201 I don't know what's so special about this paper. - They claim to use MLA to reduce KV cache by 90%. Yeah, Deepseek invented that for Deepseek V2 (an
158.
by jychang 10mo ago
The catch that you're missing is that Deepseek did this ages ago. They're just using MLA, which is well known to reduce KV size by 90%. You know, the MLA that's used in... Deepseek V2, Deepseek V3, Deepseek R1, Deepseek V3.1,
159.
by jychang 10mo ago
What's not to believe? Qwerky-32b has already done something similar as a finetune of QwQ-32b but not using traditional attention architecture. And hybrid models aren't new, MLA based hybrid models is basically just Deepseek V3.2
160.
by jychang 11mo ago
I strongly agree with this position. This is basically the foundation of Control Theory! https://en.wikipedia.org/wiki/Control_theory This is like arguing if "heater on" or "AC on" is better, which
161.
by jychang 11mo ago
That's not how MoE works, you need to train the FFN directly or else the FFN gate would have no clue how to activate the expert.
162.
by jychang 11mo ago
That entirely depends on the number of users. Inference is usually less gpu-compute heavy, but much more gpu-vram heavy pound-for-pound compared to training. General rule of thumb is that you need 20x more vram for training a model with X p
163.
by jychang 11mo ago
Well, the counterargument is that in theory, you can imagine a way to create structural color regardless of substrate. So imagine a technology that shines a laser on a car or a block of concrete and makes it blue; I'd argue that's
164.
by jychang 1y ago
He voted how his electorate would have wanted him to vote. He probably also hoodwinked some rich people to pay him some bucks while he was at it. He's about as "shades of gray" as a politician gets.
165.
by jychang 1y ago
This is over a month old, they released the weights a long time ago.
166.
by jychang 1y ago
Login gated
167.
by jychang 1y ago
> Get everyone to coordinate Lol. Lmao, even. No further comment needed.
168.
by jychang 1y ago
> My feeling is that devs in general are not into the "learning how to use tools" idea Well, the devs learnt how to use Github, didn't they? Seems like people CAN learn things that are useful. I can also make the argument
169.
by jychang 1y ago
So? That seems extremely naive of you. To use an example that a regular human would be familiar with: what if the aliens knew math and physics... and were basically ultra-nazis, and very happy to just subjugate you because "our nazi ph
170.
by jychang 1y ago
Ok, fair enough, I can accept this counterargument at least.
171.
by jychang 1y ago
That's a terrible counterargument for aliens having moral systems incompatible with humans... because it applies to the existence of aliens as well! You might as well as argue "we have no idea if aliens exist, being able to imagin
172.
by jychang 1y ago
Actually, I was explicitly thinking of SCP-3125 there.
173.
by jychang 1y ago
There's also a lot of "universals" that people take for granted as universal when it really isn't universal. Things off the top of my head that humans usually take for granted as "universals": - Separation of m
174.
by jychang 1y ago
That's the kind of thing LLMs would HELP with, though. Comments may go out of date, but LLMs can quickly generate comments that are up to date. LLMs are more likely to prevent the situation that you described. > In many cases AIs wi
175.
by jychang 1y ago
That's only if you're expecting 10 GPUs in use. They're dealing with ~1 GPU in use for a model, just sitting there. Alibaba has a very long tail of old models that barely anyone uses anymore, and yet they still serve. Here&#x
176.
by jychang 1y ago
Tensor parallelism, so you only need to store a fraction of kv cache per gpu.
177.
by jychang 1y ago
They definitely won't idle out- if they idle out, it'll take on the order of up to 60 seconds to load the model back into VRAM, depending on the model. That's an eternity for a request. I highly doubt they will timeout any mo
178.
by jychang 1y ago
You can try running LLMs on your own computer! They have extremely distinct sounds coming from the GPUs. You can hear the difference between GPT-OSS-20b and Qwen3-30b pretty easily just based on the sounds that the gpu is making. The sound
179.
by jychang 1y ago
If you read the article, the SD card was placed there by the camera manufacturer and then the device was sealed so it would withstand pressure, and then sold to divers. Blame the camera manufacturer's engineers. Seems like the SD card
180.
by jychang 1y ago
> have no problems with mechanics. What? Shaq shot free throws at 85% in practice. Players with good mechanics like Curry wouldn't be caught dead shooting 85% during the season, he's always 90%+ at free throws. Curry would prob

More ›