liuliu

20 ms·

301.
by liuliu 3y ago
My number of ops == your number of expression nodes. Trying to say that for certain models, the number of expression nodes is not trivial (in the range of a few thousands). LLM have smaller number of expression nodes / parameter count
302.
by liuliu 3y ago
And 10k ops is not a lot! SDXL has around 4k ops (from my memory) on the forward pass, from this test, that means you spend 40ms each iteration on autodiff!
303.
by liuliu 3y ago
One of these days I will find time to write more about model inference optimizations without going through distillation / quantization. Case in point: switching llama.cpp from custom kernel to cublas's GEMM implementation reduces
304.
by liuliu 3y ago
Everyone uses gguf / safetensors? How model management is done for this kind of tool?
305.
by liuliu 3y ago
Like others said. Waymo One in San Francisco is great. Smooth / confident drive. Good situation awareness (several times when it made unexpected action, only later I realized there is a person or a car it tried to avoid). Looking forwa
306.
by liuliu 3y ago
> Nearly every Apple product has become objectively better since his departure. The complaints actually got addressed. Customers are being treated as valid sources of input to the product. Hard to see an issue here. This is a subjective
307.
by liuliu 3y ago
I've briefly touched it in another blog post. It is not just trusting the institutions to not use the data for training. Optimizations you apply to making LLM inference fast might present new security landscape that we don't know
308.
by liuliu 3y ago
Both of your SDXL and M1 Max number should be faster (of course, it depends on how many steps). But the point stands, for SDXL, 3090 should be 5x to 6x faster than M1 Max and should be 2x to 2.5x faster than M2 Ultra.
309.
by liuliu 3y ago
> In Marmot every row is uniquely mapped to a JetStream. This guarantees that for any node to publish changes for a row it has to go through same JetStream as everyone else. and > This means there is NO serializability guarantee of a
310.
by liuliu 3y ago
The author "Aras P" worked for Unity close to 2 decades (retired beginning of this year). Doesn't invalidate what you said, just some background on why Unity :)
311.
by liuliu 3y ago
Should be in a few days. I asked Stability to clarify whether I can deliver weights through my Cloudflare bucket and whether qualifying as non-commercial is who runs the model, not who delivers the model.
312.
by liuliu 3y ago
I am sorry if you feel that way. I joined HN when it was a small tight-knit community without much of marketing presence. The "obvious" comment is more like "people know other people" kind of thing. I didn't try to
313.
by liuliu 3y ago
Agreed on UX feedback. It accumulated a lot of crufts from the old technologies to the new. This just echos my early feedback that co-iterating UI and the technology is difficult, you'd better pick the side you want to be on and there
314.
by liuliu 3y ago
I am. I thought this is obvious. My statement is objective. I would go as far as: it is the only app works at 8GiB macOS devices with SDXL-family models.
315.
by liuliu 3y ago
You have to make trade-off in software development. Fooocus trades on the best picture rather than the most beautiful interface, and also simplicity in its use. I think it is a good trade-off given the technology is improving at breaking-ne
316.
by liuliu 3y ago
Again, try Draw Things, it runs well for SDXL on 8GiB macOS devices.
317.
by liuliu 3y ago
Yeah, Fooocus is much better if you are going for the best local generated result. Lvmin puts all his energy into making beautiful pictures. Also it is GPL licensed, which is a + in my book.
318.
by liuliu 3y ago
You should check out Draw Things on macOS. It works well enough for SDXL on 8GiB macOS devices.
319.
by liuliu 3y ago
> PS, will you be adding Turbo directly to the app in the near future? I need to get some clarity from Stability given the new Membership thing. I personally see no reason why I cannot as long as I tell people these weights are non-comme
320.
by liuliu 3y ago
Instruction here for how to use it on iPhone / iPad / Mac today: https://twitter.com/drawthingsapp/status/1729633231526404400 (note that you need to convert 8-bit version yourself if you want to use it o
321.
by liuliu 3y ago
Yes. But my post (and the one I replied to) are talking about Jessica's early donation to OpenAI non-profit. Did that somehow "transferred" to be shares in the for-profit or did Jessica purchased some during the transition (h
322.
by liuliu 3y ago
These are donations. How that becomes investment / financial stakes? (It is a question, since how the transition to capped-profit left a lot of questions unanswered).
323.
by liuliu 3y ago
Why? The only check is to fire the CEO. He is un-firable. May as well have a board of one, at least someone cannot point to the non-profit and claim "it is a non-profit and can fire me if I am diviated from the mission".
324.
by liuliu 3y ago
The company is not destroyed. Board is not shutting down the company, they fired the CEO. The other ~700 people chose to quit. Not sure why it is "life-ruined" other than probably some tender offers withdrawn (and even this bit is
325.
by liuliu 3y ago
Even if you have limited PTO, nothing limits you to set it in such a way can go negative. It is pretty obvious what's "unlimited" means in this case.
326.
by liuliu 3y ago
Several low profile OAI employees retweeted with ♥ (same as gdb) indicating the returning is done and through from my interpretation.
327.
by liuliu 3y ago
Maybe auto-correction is off because you need to type too much acroynms and jargons and auto-correction is annoying?
328.
by liuliu 3y ago
Any examples? I felt the other way.
329.
by liuliu 3y ago
Yeah. It is still useful for them to share these. My take away: 1. data is all you need to generate these amazing videos with the right gait (gait is something I focused on). 2. nobody doing new network structure, it is animatediff beefed u
330.
by liuliu 3y ago
Emu image is not significantly slower than SDXL or similar. So you would expect to have similar performance as Hotshot. The upscaler (8 frame to 37 frame) version probably would take significantly longer.

More ›