sanxiyn

29 ms·

451.
by sanxiyn 3y ago
> I'm quite sure the entirety of Harry Potter isn't included, but I'm also sure that some of the most popular paragraphs probably are. It's analagous to the kind of stuff people memorize. No, you are wrong about this.
452.
by sanxiyn 3y ago
For detailed investigation of this phenomenon, see Speak, Memory: An Archaeology of Books Known to ChatGPT/GPT-4: https://arxiv.org/abs/2305.00118
453.
by sanxiyn 3y ago
Someone did that calculation and the result is here: https://www.reddit.com/r/slatestarcodex/comments/13u40yf/ 100x GPT-4 to 85%.
454.
by sanxiyn 3y ago
https://arxiv.org/abs/2304.01433 does claim "1.2x-1.7x faster and uses 1.3x-1.9x less power than the NVIDIA A100".
455.
by sanxiyn 3y ago
Amazingly, Firefox retained this performance edge over Chrome for 5 years, and will likely retain for 5 more years. This is really an amazing technological achievement. > Firefox's Stylo engine is fast. Like, really fast. Like, so f
456.
by sanxiyn 3y ago
FWIW, Knuth's chapter on satisfiability mentions MiniSat once for historical interest and discusses Glucose heuristic for five pages. This matches my impression of their technical contributions.
457.
by sanxiyn 3y ago
It's not just GPU. There are not even good multicore SAT solvers. State of the art SAT algorithms are serial and they don't parallelize.
458.
by sanxiyn 3y ago
As I remember, Glucose was a larger advance than MiniSat. MiniSat was well-engineered but not exceptional. Glucose started as a hack, but the heuristic it introduced was so effective that you basically couldn't compete without copying
459.
by sanxiyn 3y ago
You will enjoy "Maieutic Prompting: Logically Consistent Reasoning with Recursive Explanations". It prompts LLM to generate tree of explanations and run MAX-SAT solver over it: https://arxiv.org/abs/2205.11822
460.
by sanxiyn 3y ago
You may want to read "A Time Leap Challenge for SAT Solving". It's just a really fun paper: https://arxiv.org/abs/2008.02215
461.
by sanxiyn 3y ago
Sure, but the point is that Triton is not dependent on CUDA language or frontend. Triton also outputs PTX using LLVM's NVPTX backend. Devils are in the details, but at a very high level, Triton could be ported to AMD by doing s/NV
462.
by sanxiyn 3y ago
They do. (Source: talked to Google adtech engineers.)
463.
by sanxiyn 3y ago
Sue on what grounds? It will be quickly dismissed.
464.
by sanxiyn 3y ago
I don't think that works. LLM-generated contents are not copyrightable.
465.
by sanxiyn 3y ago
They can terminate your account.
466.
by sanxiyn 3y ago
Of course it will hold up in court, it's their service and their terms of service.
467.
by sanxiyn 3y ago
I agree writing something that needs to be compatible with, say, PyTorch is a significant undertaking, but why is that necessary? I also agree some models like MaskRCNN is not static, and people will not change their model code, but I don&#
468.
by sanxiyn 3y ago
geohot wrote tinygrad. This is not about believing his skills to translate to other domains. It is his domain. You definitely shouldn't trust what geohot says about infinitary mathematics or (god forbids) quantum mechanics. On the othe
469.
by sanxiyn 3y ago
There's also OpenAI Triton. People seem to miss that OpenAI is not using CUDA...
470.
by sanxiyn 3y ago
I also know the founder of an AI chip company (ex-AMD), they taped out in 2022, got working chips on the first try. Miraculously, they hired a good software team, some even with previous compiler experience. In their brochure they write thi
471.
by sanxiyn 3y ago
Eh... no? Stable Diffusion works fine on single device. Ditto for smaller LLaMA.
472.
by sanxiyn 3y ago
BTM is very promising, but it is unclear how much it scales, let alone "massively". The paper scaled it to 64 domains and it worked, but you probably want more than 64 nodes. Since domain is actually important to its performance,
473.
by sanxiyn 3y ago
You can, like, just do weighted average. It works. Branch-Train-Merge: Embarrassingly Parallel Training of Expert Language Models (this is Tim Dettmers's previous work!) > These ELMs can be added and removed to update data coverage,
474.
by sanxiyn 3y ago
> It's only a matter of time before people really crack distributed training algorithms that can be run in a less organized swarm configuration. Why do you think this? This continually failed, and seems extremely unlikely to me. Bar
475.
by sanxiyn 3y ago
This work introduces a new quantization scheme, NF4, for 4-bit NormalFloat, based on previous work on quantile quantization, so it's not a simple truncation, but it's also not a GPTQ-like optimization method. Figure 3 of the paper
476.
by sanxiyn 3y ago
For quality, GPTQ-for-LLaMa repository README is already updated for comparison with this work. See under "GPTQ vs bitsandbytes".
477.
by sanxiyn 3y ago
No.
478.
by sanxiyn 3y ago
Sure, but (very unfortunately for clear communication) that's not what "5 nm" written by brancz in "Apple silicon is at 5 nm" means. "Apple silicon is at TSMC N5" would be more accurate.
479.
by sanxiyn 3y ago
If you read the article, it is clear Jensen Huang is commenting only on the price aspect. It used to be the case that as density goes up, price goes down. Huang is saying that price is no longer going down, and he is probably right. But den
480.
by sanxiyn 3y ago
I believe this report by WikiChip from IEDM 2022 is an accurate summary of the situation: > While there were a great number of interesting papers from both academia and industry, it was the one by TSMC that brought frighteningly bad news

More ›