ryul99 ์œจ ๐Ÿฒ (@ryul99_dev) header
ryul99 ์œจ ๐Ÿฒ (@ryul99_dev) profile picture

Interested in ML(Deep Learning), Linux โญ my tweets are my own

Joined September 2019
Photo by ryul99 ์œจ ๐Ÿฒ (@ryul99_dev)Photo by ryul99 ์œจ ๐Ÿฒ (@ryul99_dev)Photo by ryul99 ์œจ ๐Ÿฒ (@ryul99_dev)Photo by ryul99 ์œจ ๐Ÿฒ (@ryul99_dev)Photo by ryul99 ์œจ ๐Ÿฒ (@ryul99_dev)Photo by ryul99 ์œจ ๐Ÿฒ (@ryul99_dev)Photo by ryul99 ์œจ ๐Ÿฒ (@ryul99_dev)Photo by ryul99 ์œจ ๐Ÿฒ (@ryul99_dev)Photo by ryul99 ์œจ ๐Ÿฒ (@ryul99_dev)Photo by ryul99 ์œจ ๐Ÿฒ (@ryul99_dev)
ryul99 ์œจ ๐Ÿฒ retweeted
DFlash for Gemma 4: Up to 6x Faster. โšกโšก Great to see MTP land natively in Gemma 4 today. If you want to push it further, try DFlash โ€” open source, same quality, more speed!! github.com/z-lab/dflash
Gemma 4: Now up to 3x Faster. โšก Same quality, way more speed. Our new MTP drafters allow Gemma 4 to predict multiple tokens at once, effectively tripling your output speed without compromising intelligence.
74
180
54
1,514
477,609
ryul99 ์œจ ๐Ÿฒ retweeted
Today ggml.ai joins Hugging Face Together we will continue to build ggml, make llama.cpp more accessible and empower the open-source community. Our joint mission is to make local AI easy and efficient to use by everyone on their own hardware.
I've started a company: ggml.ai From a fun side project just a few months ago, ggml has now become a useful library and framework for machine learning with a great open-source community
142
233
68
1,640
313,648
ryul99 ์œจ ๐Ÿฒ retweeted
In today's episode of programming horror... In the Python docs of random.seed() def, we're told "If a is an int, it is used directly." [1] But if you seed with 3 or -3, you actually get the exact same rng object, producing the same streams. (TIL). In nanochat I was using the sign as a (what I thought was) clever way to get different rng sequences for train/test splits. Hence gnarly bug because now train=test. I found the CPython code responsible in cpython/Modules/_randommodule.c [2], where on line 321 we see in a comment: "This algorithm relies on the number being unsigned. So: if the arg is a PyLong, use its absolute value." followed by n = PyNumber_Absolute(arg); which explicitly calls abs() on your seed to make it positive, discarding the sign bit. But this comment is actually wrong/misleading too. Under the hood, Python calls the Mersenne Twister MT19937 algorithm, which in the general case has 19937 (non-zero) bits state. Python takes your int (or other objects) and "spreads out" that information across these bits. In principle, the sign bit could have been used to augment the state bits. There is nothing about the algorithm that "relies on the number being unsigned". A decision was made to not incorporate the sign bit (which imo was a mistake). One trivial example could have been to map n -> 2*abs(n) + int(n < 0). Finally this leads us to the contract of Python's random, which is also not fully spelled out in the docs. The contract that is mentioned is that: same seed => same sequence. But no guarantee is made that different seeds produce different sequences. So in principle, Python makes no promises that e.g. seed(5) and seed(6) are different rng streams. (Though this quite commonly implicitly assumed in many applications.) Indeed, we see that seed(5) and seed(-5) are identical streams. And you should probably not use them to separate your train/test behaviors in machine learning. One of the more amusing programming horror footguns I've encountered recently. We'll see you in the next episode. [1] docs.python.org/3/library/raโ€ฆ [2] github.com/python/cpython/blโ€ฆ
212
475
66
7,699
773,065
ryul99 ์œจ ๐Ÿฒ retweeted
์ดํ•ด ๋ถ€์ฑ„: LLM์ด ๋งŒ๋“  ์ฝ”๋“œ๊ฐ€ ๋‚จ๊ธฐ๋Š” ์‹œํ•œํญํƒ„ - ์ตœ๊ทผ LLM(๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ)์ด ์ƒ์„ฑํ•œ ์ฝ”๋“œ๋ฅผ ์ˆ˜์ •ยท๋ณด์™„ํ•˜๋Š” ๋ฐ ๊ฐœ๋ฐœ์ž๊ฐ€ ๋” ๋งŽ์€ ์‹œ๊ฐ„์ด ๊ฑธ๋ฆฌ๊ณ  ์žˆ์Œ - ๊ธฐ์กด ๋ ˆ๊ฑฐ์‹œ ์ฝ”๋“œ์ฒ˜๋Ÿผ, ์ฝ”๋“œ๋ฅผ ์•ˆ์ „ํ•˜๊ฒŒ ๋ฐ”๊พธ๋ ค๋ฉด ๋จผ์ € ๋ฌด์—‡์„, ์™œ ๊ทธ๋ ‡๊ฒŒ ๊ตฌํ˜„ํ–ˆ๋Š”์ง€ ์ดํ•ดํ•ด์•ผ ํ•˜๋Š”๋ฐ, LLM ์ฝ”โ€ฆ news.hada.io/topic?id=23384
103
30
169
69,782
๐Ÿซข
๋ชฉ์š”์ผ์— ๊ฒฐ๊ณผ๊ฐ€ ๋‚˜์˜ค์ง€๋งŒ ์ œ๊ฐ€ ์ฐธ์—ฌํ•œ ๋…ผ๋ฌธ ํ”„๋กœ์ ํŠธ๊ฐ€ top7์— ํ•ด๋‹นํ•˜๋Š” Best Paper ํ›„๋ณด๊ฐ€ ๋œ ์ƒํ™ฉ์ž…๋‹ˆ๋‹ค. ์ข‹์€ ๊ฒฐ๊ณผ๊ฐ€ ์žˆ๊ธธ...
10
13
52
5,529
Six Dragons Fly Again for the Web: An Interactive Web Application for Everyone @ISMIRConf I developed an interactive web demo for the Six Dragons project that can create an ensemble of Korean court music using our project's model.๐Ÿ‰๐ŸŽต Implementation: crescent.studio/project/six-โ€ฆ
๐Ÿ‰Six Dragons Fly Again: It's time for applying symolic music gen to Korean traditional music๐Ÿ‡ฐ๐Ÿ‡ท Our work on generating Korean court music will be presented @ISMIRConf๐Ÿฅณ Work by @DanbiNaerinHan @MarkGothamMusic DongMin Kim, @crescent_stdio @is_s_yun Web Demo and Concert videoโฌ‡๏ธ
1
5
1
9
3,158
๋ชฉ์š”์ผ์— ๊ฒฐ๊ณผ๊ฐ€ ๋‚˜์˜ค์ง€๋งŒ ์ œ๊ฐ€ ์ฐธ์—ฌํ•œ ๋…ผ๋ฌธ ํ”„๋กœ์ ํŠธ๊ฐ€ top7์— ํ•ด๋‹นํ•˜๋Š” Best Paper ํ›„๋ณด๊ฐ€ ๋œ ์ƒํ™ฉ์ž…๋‹ˆ๋‹ค. ์ข‹์€ ๊ฒฐ๊ณผ๊ฐ€ ์žˆ๊ธธ...
Six Dragons Fly Again for the Web: An Interactive Web Application for Everyone @ISMIRConf I developed an interactive web demo for the Six Dragons project that can create an ensemble of Korean court music using our project's model.๐Ÿ‰๐ŸŽต Implementation: crescent.studio/project/six-โ€ฆ
2
4
3
28
7,830
ryul99 ์œจ ๐Ÿฒ retweeted
Oasis A Universe in a Transformer Oasis is an interactive world model developed by Decart and Etched. Based on diffusion transformers, Oasis takes in user keyboard input and generates gameplay in an autoregressive manner.
3
12
1
83
12,642
ryul99 ์œจ ๐Ÿฒ retweeted
What is the performance limit when scaling LLM inference? Sky's the limit. We have mathematically proven that transformers can solve any problem, provided they are allowed to generate as many intermediate reasoning tokens as needed. Remarkably, constant depth is sufficient. arxiv.org/abs/2402.12875 (ICLR 2024)
106
499
173
2,928
796,080
CPython dict has a huge pitfall: If you have added N keys in dict, no matter how many times you deleted keys, it takes O(N) time to iterate over it.
2
5
2
13
1,730
ryul99 ์œจ ๐Ÿฒ retweeted
Critical flaw found in #Docker Engine allows attackers to bypass authorization plugins (AuthZ) - CVE-2024-41110, CVSS score 10.0. This #vulnerability can lead to severe privilege escalation, affecting numerous Docker versions. Details here: thehackernews.com/2024/07/crโ€ฆ #DevOps
1
78
9
172
72,499
ryul99 ์œจ ๐Ÿฒ retweeted
FuriosaAI's research paper "TCP: A Tensor Contraction Processor for AI Workloads" has been accepted for publication by the International Symposium on Computer Architecture (@ISCAConfOrg), the premier forum for new ideas in silicon design. furiosa.ai/download/FuriosaAโ€ฆ (1/5)
2
9
1
16
933
ryul99 ์œจ ๐Ÿฒ retweeted
๐Ÿ—ž๏ธ Anthropic, AI(LLM)์˜ ๋‚ด๋ถ€ ์ž‘๋™ ํ•ด์„์— ์„ฑ๊ณตํ–ˆ๋‹ค๊ณ ! ๊ทธ๋™์•ˆ AI๊ฐ€ ๋ธ”๋ž™๋ฐ•์Šค๋ผ๋ฉฐ ์–ด๋–ป๊ฒŒ ์ž‘๋™ํ•˜๋Š”์ง€ ๋ชจ๋ฅด๊ธฐ ๋•Œ๋ฌธ์— ๋ฏฟ์„ ์ˆ˜ ์—†๋‹ค, ์œ„ํ—˜ํ•˜๋‹ค๋Š” ์ด์•ผ๊ธฐ๊ฐ€ ๋งŽ์•˜๋Š”๋ฐ์š”. ์ด๋ฒˆ์— Anthropic ํŒ€์—์„œ Claude 3 Sonnet ๋ชจ๋ธ์˜ ๋‚ด๋ถ€ ์ž‘๋™์„ ์ดํ•ดํ•˜๋Š”๋ฐ ์„ฑ๊ณตํ–ˆ๋‹ค๊ณ  ๋ฐœํ‘œํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋™์•ˆ ์ž‘์€ ๋ชจ๋ธ์—์„œ๋Š” ์ผ๋ถ€ ์ง„์ „์ด ์žˆ์—ˆ์ง€๋งŒ, ์ƒ์šฉ์˜ ๋Œ€ํ˜• ๋ชจ๋ธ์˜ ํ•ด์„์ด ๊ตฌ์ฒดํ™” ๋œ ๊ฒƒ์€ ์ฒ˜์Œ์œผ๋กœ ๋งค์šฐ ๋†€๋ผ์šด ๋ฐœํ‘œ์ž…๋‹ˆ๋‹ค. ๊ฐ„๋‹จํ•˜๊ฒŒ ์„ค๋ช…ํ•˜๋ฉด, ๋ชจ๋ธ์˜ ๋‰ด๋Ÿฐ ํ™œ์„ฑํ™” ํŒจํ„ด์„ ์ธ๊ฐ„์ด ํ•ด์„ํ•  ์ˆ˜ ์žˆ๋Š” ์ˆ˜๋ฐฑ๋งŒ๊ฐœ์˜ ๊ฐœ๋…์œผ๋กœ ์‹๋ณ„ํ•˜๊ณ  ์ถ”์ถœํ•˜๋Š” ๋ฐฉ์‹์œผ๋กœ, ๋‚ด๋ถ€ ์ƒํƒœ์— ๋Œ€ํ•œ ๋Œ€๋žต์ ์ธ ๊ฐœ๋… ์ง€๋„๋ฅผ ๋งŒ๋“ค๊ณ  ์‹œ๊ฐํ™”ํ–ˆ๋Š”๋ฐ์š”. ์ด๋ฅผ ํ†ตํ•ด ๋ชจ๋ธ์„ ํˆฌ๋ช…ํ•˜๊ณ  ๊นŠ์ด ์ดํ•ดํ•  ์ˆ˜ ์žˆ๊ฒŒ ๋˜์–ด ๋ชจ๋ธ์˜ ์˜ˆ์ธก๊ณผ ํ–‰๋™์„ ๋” ์ž˜ ์„ค๋ช… ํ•  ์ˆ˜ ์žˆ๊ณ , ๋”๋ถˆ์–ด ์ž ์žฌ์ ์ธ ์œ ํ•ด ํ–‰๋™์„ ๋ฏธ๋ฆฌ ์•Œ ์ˆ˜ ์žˆ๊ฒŒ ๋˜์–ด ์•ˆ์ „์„ฑ์„ ํฌ๊ฒŒ ๋†’์ผ ์ˆ˜ ์žˆ๊ฒŒ ๋˜์—ˆ๋‹ค๊ณ  ํ•ฉ๋‹ˆ๋‹ค. ๋ฌด์—‡๋ณด๋‹ค ์ค‘์š”ํ•œ ๊ฒƒ์€, ๋‰ด๋Ÿฐ(๊ฐœ๋…)๋“ค์„ ์กฐ์ž‘ํ•ด์„œ ์ธ์œ„์ ์œผ๋กœ ์ฆํญํ•˜๊ฑฐ๋‚˜ ์–ต์ œํ•˜์—ฌ ๋ชจ๋ธ์˜ ๋ฐ˜์‘์ด ๋ณ€ํ•˜๋Š” ๊ฒƒ๋„ ํ™•์ธํ–ˆ๋‹ค๋Š” ๊ฒƒ์ธ๋ฐ์š”. ๊ทธ๋™์•ˆ์€ ํ•ด๋กœ์šด ์ถœ๋ ฅ์„ ๋ฐฉ์ง€ํ•˜๊ธฐ ์œ„ํ•ด ํ•™์Šต ๋ฐ์ดํ„ฐ๋ฅผ ํ†ตํ•ด ์‚ฌํ›„์— ์กฐ์ •ํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ์ผ๋Š”๋ฐ, ์ด์ œ๋Š” ๊ทธ๋Ÿด ํ•„์š” ์—†์ด ์‚ฌ์ „์— ๋ชจ๋ธ์ด ๊ฐ€์ง„ ๊ฐœ๋…๋“ค์„ ๋ช…์‹œ์ ์œผ๋กœ ์ดํ•ดํ•œ ๋’ค ์ง์ ‘์ ์œผ๋กœ ๊ฐœ์ž…ํ•  ์ˆ˜ ์žˆ๊ฒŒ ๋˜์—ˆ๋‹ค๋Š” ๊ฒƒ์ด์ฃ . ๋ฐ”๋กœ ๋ฉฐ์น ์ „์— OpenAI์˜ Superalignment ํŒ€์˜ ํ•ด์ฒด๋กœ ์ •๋ ฌ ๋ฌธ์ œ์— ๋Œ€ํ•ด ๋งŽ์€ ์šฐ๋ ค๊ฐ€ ์žˆ์—ˆ๋Š”๋ฐ, ๊ณง๋ฐ”๋กœ Anthropic์ด ๋ฌธ์ œ ํ•ด๊ฒฐ์— ํฐ ์ง„์ „์„ ๋ณด์˜€๋‹ค๋Š” ์†Œ์‹์ด ๋‚˜์˜ค๋„ค์š”. ์ง„์งœ AI ์†Œ์‹์˜ ๋„ํŒŒ๋ฏผ ์ค‘๋…์€ ๋ฉˆ์ถœ ์ˆ˜๊ฐ€ ์—†์Šต๋‹ˆ๋‹ค. ๐Ÿ˜Ž๐Ÿฟ๐Ÿฅค
6
286
37
543
58,506
ryul99 ์œจ ๐Ÿฒ retweeted
Scarlett Johansson is taking legal action against OpenAI after they 'copied' her voice for GPT-4o. According to the actress, Sam Altman tried to hire her voice last September, she said no, and Sam used cloned and used her voice anyway. Insane if true. nitter.cf/BobbyAllyn/status/1792โ€ฆ
Statement from Scarlett Johansson on the OpenAI situation. Wow:
9
37
4
319
78,512
ryul99 ์œจ ๐Ÿฒ retweeted
AI NEWS: Microsoft just announced the world's first AI PCs to rival the MacBook. Plus, major developments from OpenAI, Scarlett Johansson, Elon Musk/xAI, and Geoffrey Hinton. Here's everything going on in AI right now:
56
316
50
2,923
1,982,358
ryul99 ์œจ ๐Ÿฒ retweeted
์›ํŠธ์œ—์—์„œ๋Š” ์žฌ๋ฏธ์žˆ๋Š” ํ˜„์ƒ ์ •๋„๋กœ ํ‘œํ˜„ํ–ˆ์ง€๋งŒ ์ด๊ฑด AI ๊ธฐ์ˆ ์ด ๊ทผ๋ณธ์ ์œผ๋กœ ์‹๋ฏผ์ฃผ์˜์  ์ฐฉ์ทจ์™€ ๋ฐ€์ ‘ํ•จ์„ ๋ณด์—ฌ์ฃผ๋Š” ์‚ฌ๋ก€. (์• ์ดˆ์— ์™œ ๋‚˜์ด์ง€๋ฆฌ์•„์—์„œ ์˜์–ด๋ฅผ ์“ธ๊นŒ?) ๋ฐ”๊ฟ”์„œ ์ƒ๊ฐํ•˜๋ฉด ์˜ˆ์ปจ๋Œ€ ๋‚˜์ด์ง€๋ฆฌ์•„์‹ ์˜์–ด๋ฅผ ๊ตฌ์‚ฌํ•˜๋Š” ํ•™์ƒ์˜ ๊ณผ์ œ๋ฌผ์ด ์ฑ—์ง€ํ”ผํ‹ฐ ์ƒ์„ฑ ํ…์ŠคํŠธ๋กœ ์˜ค์ธ๋ฐ›์•„ ๋‚™์ œ๋ฐ›์„ ์ˆ˜๋„ ์žˆ๋‹ค๋Š” ๊ฒƒ.
ChatGPT๊ฐ€ ์œ ๋‚œํžˆ ์„ ํ˜ธํ•˜๋Š” "delve"๋ผ๋Š” ๋‹จ์–ด์˜ ์ถœ์ฒ˜์— ๋Œ€ํ•œ ๊ธฐ์‚ฌ. RLHF๋ฅผ ์™ธ์ฃผ ์ค€ ๋‚˜์ด์ง€๋ฆฌ์•„๋ฅผ ๋น„๋กฏํ•œ ์•„ํ”„๋ฆฌ์นด์—์„œ ํ”ํžˆ ์“ฐ๋Š” ๋‹จ์–ด๋ผ๊ณ ! ์ธ๊ฑด๋น„๊ฐ€ ์‹ผ ๋‚˜๋ผ์˜ ์˜์–ด ์Šคํƒ€์ผ๋กœ ์ „์„ธ๊ณ„์˜ ์˜์–ด ์Šคํƒ€์ผ์ด ์˜ฎ๊ฒจ๊ฐ€๊ณ  ์žˆ๋Š” ์žฌ๋ฏธ์žˆ๋Š” ํ˜„์ƒ. theguardian.com/technology/2โ€ฆ
1
2,255
62
1,030
328,112
ryul99 ์œจ ๐Ÿฒ retweeted
The latest evals are in, and weโ€™re open sourcing a GitHub repo to run popular evals on various models, including our latest GPT-4 Turbo model. Evaluating models is challengingโ€”weโ€™ve found this useful and want to share with the broader community. github.com/openai/simple-evaโ€ฆ
Our new GPT-4 Turbo is now available to paid ChatGPT users. Weโ€™ve improved capabilities in writing, math, logical reasoning, and coding. Source: github.com/openai/simple-evaโ€ฆ
13
63
8
450
69,058
ryul99 ์œจ ๐Ÿฒ retweeted
For example, when writing with ChatGPT, responses will be more direct, less verbose, and use more conversational language.
48
66
34
1,358
269,410
ryul99 ์œจ ๐Ÿฒ retweeted
The new GPT-4 is #1 On The Human Eval Leaderboard Again...
23
18
1
196
36,128
ryul99 ์œจ ๐Ÿฒ retweeted
github์—์„œ ์ฐธ๊ณ ํ•  ์ฝ”๋“œ ์ „์ฒด๋ฅผ zipํŒŒ์ผ๋กœ ๋‹ค์šด๋กœ๋“œ ๋ฐ›์€ ๋’ค ์ฑ—์ง€ํ”ผํ‹ฐ์—๊ฒŒ ์ฃผ๊ณ  ์–ด๋–ป๊ฒŒ ๋Œ์•„๊ฐ€๋Š” ์ฝ”๋“œ์ธ์ง€ ์„ค๋ช…ํ•ด๋‹ฌ๋ผ๊ณ  ํ•˜๊ธฐ. CI๋กœ ํŒŒ์ผ ์••์ถ•์„ ํ’€์–ด์„œ ํด๋”๊ตฌ์กฐ์™€ ํŒŒ์ผ ์ด๋ฆ„๊นŒ์ง€ ์ฐธ๊ณ ํ•ด์„œ ์„ค๋ช…ํ•ด์ค€๋‹ค. #LLMํ™œ์šฉ์˜ˆ์‹œ
Replying to @mahler83
๊ทธ๋ƒฅ ํŒŒ์ด์ฌ ์Šคํฌ๋ฆฝํŠธ๋กœ๋Š” ๊ตฌํ˜„์ด ์•ˆ ๋˜๋Š” ๋ถ€๋ถ„์ด ์žˆ์–ด์„œ... ์ด๋Ÿฐ ์ง“๊นŒ์ง€ ํ•˜๋„ค
2
88
3
179
31,738