liuliu

10 ms·

151.
by liuliu 1y ago
There are Jupyter kernels for Swift that uses REPL mode of Swift lldb. I used to do it but obvious right now it is not a priority: https://github.com/liuliu/swift-jupyter
152.
by liuliu 1y ago
Marc Brooker said this: https://x.com/MarcJBrooker/status/1960809302333251876 which echos my sentiment. I will just repost it here: > More broadly, I don't think a single definition of 'durable'
153.
by liuliu 1y ago
> So even sqlite doesn't make anywhere near as strong a claim as you make. And? If you write to a disk and later this disk is missing, you don't have durability. SQLite cannot automatically help you to commit your writes to a s
154.
by liuliu 1y ago
Also, let's not forget that Apple / Google is violating PAFACAA right now (the TikTok act, by allowing TikTok in the U.S. AppStore / PlayStore) b/c DoJ is instructed to sue anyone who is following PAFACAA. This will crea
155.
by liuliu 1y ago
> Wow this administration is f*ing batshit insane. No, it is not insane. This creates perfect "everyone violates the law, we can selectively enforce it" scenario. That's how 10% Intel-like condition can be created for othe
156.
by liuliu 1y ago
That's correct. Unless Zuck is OK with using one-generation older TPU, which is a possibility.
157.
by liuliu 1y ago
As I mentioned in the other comment, https://www.sqlite.org/fileformat2.html#the_write_ahead_log WAL uses checksum to make sure there is no holes for its writes. I need to do more analysis, but it goes beyond just rely on f
158.
by liuliu 1y ago
It depends on if there will be holes and whether you communicate "externally". If none of these are concerns (for WAL and using SQLite locally, none of these are), it is OK. To elaborate, for a local app that using WAL, if a trans
159.
by liuliu 1y ago
Does it matter? For all we know, it keeps the serializability. At this point (of computer hardware history), you would care more about serializability than making sure data written to disk after power loss, the latter would now depend on so
160.
by liuliu 1y ago
GCP getting second tier TPU allocation b/c TPU cannot be enough to meet GDM needs. At this point, it would be very stupid for external customers betting on TPUs (I am looking at Apple).
161.
Privacy Pass and Draw Things+ Cloud Compute(releases.drawthings.ai)
1 points by liuliu 1y ago|0 comments
162.
by liuliu 1y ago
Some facts are wrong: - The 27B "MoE" are not the MoE commonly referred to in LLM world. It is not MoE on FFN layers. It simply means two different models used for different denoising timestep ranges (exactly the same as SDXL-Base
163.
Qwen Image Support in Draw Things(releases.drawthings.ai)
4 points by liuliu 1y ago|0 comments
164.
by liuliu 1y ago
You are right. I mis-remembered the sparsity part of K2. The "done wrong" part I was thinking about how the scout -> maverick -> behemoth doesn't scale sparsity according to any formula (less sparse -> sparse -> l
165.
by liuliu 1y ago
In theory, "privacy pass" should help, as you can subpoena content, but cannot know who made these. But that is still thin (and Ollama not doing that too anyway).
166.
by liuliu 1y ago
Any more information on "Privacy first"? It seems pretty thin if just not retaining data. For Draw Things provided "Cloud Compute", we don't retain any data too (everything is done in RAM per request). But that is s
167.
by liuliu 1y ago
Kimi K2 paper said that the model sparsity scales up with parameters pretty well (MoE sparsity scaling law, as they call, basically calling Llama 4 MoE "done wrong"). Hence K2 has 128:1 sparsity.
168.
by liuliu 1y ago
Yeah, not trying to argue either way other than saying contrary to the parent comment, licensing cost is a significant component of the cost equation.
169.
by liuliu 1y ago
Per 100k image. And it is additionally $0.01 per image. Considering H100 is $1.5 per hour and you can get 1 image per 5s, we are talking about bare-metal cost of ~$0.002 per image + $0.01 license cost.
170.
by liuliu 1y ago
16GiB RAM with 8-bit quantization. This is a slightly scaled up SD3 Large model (38 layers -> 60 layers).
171.
by liuliu 1y ago
Thank you! This is what I am looking for in HN comments. Layout engine cannot be that stupid per the article. I guess it is not once you give it enough hint.
172.
by liuliu 1y ago
Only "class" object is reference counted (to some extents, class-like objects too). Int / struct (value-semantics) objects are not reference counted. These are copied eagerly. Swift introduced bunch of ownership keywords to h
173.
by liuliu 1y ago
https://storage.googleapis.com/deepmind-media/gemini/IMO_202...
174.
by liuliu 1y ago
You misunderstood and this is not re-implementing CUDA API. MLX is a PyTorch-like framework.
175.
by liuliu 1y ago
You don't need to guess: https://github.com/cactus-compute/cactus/tree/main/cpp
176.
by liuliu 1y ago
That's the point of experiments right? From my own experience, I found these who use image / video generators are mainly in e-comm and advertisement. Only recently creative works (such as YouTubers) start to use these (they were e
177.
by liuliu 1y ago
There is a simple method to detect this: taking a model "claimed" to be trained scratch, taking the model you suspected is the original, generate a new model = claimed_model * 0.5 + suspected_model * 0.5. If the claimed_model is t
178.
by liuliu 1y ago
HiDream is a separate architecture. OTOH, it might be finetuned on FLUX generated data, we will never know.
179.
by liuliu 1y ago
JPEG XL has the mode that in layman's word, allow bit-by-bit round-trip with JPEG. Original JPEG -> JPEG XL -> Recreated JPEG. Sha256(Original JPEG) == Sha256(Recreated JPEG). That's what people meant by "backward compa
180.
by liuliu 1y ago
It is hard to justify tens-of-millions investment in training to just make it faster without any idea how it scores on benchmarks. It is easier to justify keeping the model intact and spend extra millions to make it faster with exotic means

More ›