echo hamletBOT
Hey @everyone, excited to announce our journal club / presentation titled Practical Applications of ML in Biology and Drug Discovery!
In this journal club we're going to go over various models and AI algorithms that have practical applications in molecular biology and drug discovery. The presentation is going to include a number of SOTA models that can currently be used in your existing workflows, as well as some of their strengths and limitations.
Hope to see you all there.
echo hamletBOT
@here Apologies in advance for the third consecutive ping ð . We changed the event start time to 3 PM EST instead of 2 PM EST to accommodate those who are attending the weekly DNA diffusion meetup.
Once again my apologies for any confusion.
sterile cargoBOT
ð¢ Harmonai Announcement - Office Hours, Beat Synced AI Animation Workshop, and Production Challenge ð
Attention, @everyone! We have some exciting news and events to share with you. Please read below for important updates:
1ïžâ£ Harmonai Office Hours is starting now! ð
Come by and hear about the latest from Harmonai and ask the team questions
2ïžâ£ ** Beat Synced AI Animation Workshop by Purz next Tuesday, May 31** ðšðµ
We are thrilled to announce a special workshop by the talented artist, Purz! Join us for an immersive session on Beat Synced AI Animation. Purz will guide you through the creative process of combining music and animation. This workshop promises to be an incredible opportunity to expand your artistic horizons. This workshop is scheduled for 31st May.
https://twitter.com/PurzBeats
3ïžâ£ **Production challenge submissions are due tomorrow! ** â°
For all the ambitious creators out there, the Harmonai Production Challenge is well underway! Just a friendly reminder that the challenge submissions are due this Friday. Showcase your skills and innovation by crafting amazing compositions with Harmonai. We can't wait to see your entries!
4ïžâ£ ** Harmonai Challenge Showcase next Monday, May 29th** ð
Mark your calendars! Next Monday, May 29th, we will be hosting the Harmonai Challenge Showcase. This is your chance to witness the incredible talent within our community. Join us as we celebrate the outstanding compositions created during the production challenge. Be prepared to be inspired, amazed, and motivated by the diverse range of masterpieces made by our community!
old ventureBOT
@everyone pickle is an egregiously unsafe file format to use to distribute ML models. A malicious pickle file can execute arbitrary code on your computer when you call unpickle, meaning that a bad actor can upload what looks like a model but is actually a Trojan giving them full access to your computer.
While this exploit is rate in the wild and there are no known cases of it happening with ML algorithms, itâs clearly not suitable to be the default format for distributing ML models.
Weâve been working with Hugging Face and Stability AI on this, and are now ready to announce several improvements to the security of the OS AI ecosystem:
0. SafeTensors, a library that is safe and is designed to support the needs of large scale AI researchers.
Weâre going to have a waiting period before making it the default to let users iron out any bugs we missed, but we will be making it the default format across our sectors of the ecosystem and we anticipate much if the rest of the OS LLM ecosystem following suite.
Note that it is absolutely safe to use your own or otherwise trusted pickle files, the issue is focused on when you download and unpickle files you find on the internet. When LLaMA was leaked, this was a serious concern I had. Fortunately, the leaked files were quickly confirmed to be identical to the officially released ones.
I want to give a huge shout out to the AI Village, a DEF CON group that focuses on AI and Security for being the reason I personally know about this issue and for making the introduction to Trail of Bits. The AIV has a discord server you can join here https://discord.gg/uz3hnZgnJN to discuss ML Security (itâs also found in #732688974337933322)
@everyone Our first work on mechanistic interpretability just came out! âCan Transformers Learn to Solve Problems Recursively?â by @dylanzhang @Curt Tigges @Stella Biderman (she/her) Maxim Raginsky, and @taliaringer trains small transformers on recursive tasks like computing the binary successor or transversing a tree.
In contrast to most ME work we find that the model fails to fully learn the tasks we study. This gives us a new way to validate our results though: looking at our reverse engineered algorithm we are able to identify a type of inputs that the model should fail in. Lo and behold, when we looked at the accuracy breakdown we found that it failed on those inputs 100% of the time and that they compressed 91% of all model failures!
The paper has a lot more, including documenting how the LR influences generalization strategies, an Abstract State Machine formalization, and looking at how different presentations of the same problem changes model behavior!
Thereâs a lot of directions for future work, some of which we are already discussing in our new channel #1110611369574793306 If youâre interested in using LLMs to study formal reasoning, definitely stop by and say hi.
Paper: https://arxiv.org/abs/2305.14699
Twitter thread: https://twitter.com/TaliaRinger/status/1661786081249964050?s=20
sterile cargoBOT
@everyone
ð¢ Harmonai Challenge Showcase - Starting Now! ð
Join us for the Harmonai Challenge Showcase, starting now in the Center Stage!
We'll be listening to challenge submissions from @Taera, @Greg White, @crlandsc, @lyra âš, @NeuralNotWork, @jmoso13, @Jimney, @Silvio and our very own @ODDS
Come on through! ðµâš
ancient spearBOT
@everyone We're excited to share our first preprint since our public launch:
ð§ ðïž MindEye!
Our state-of-the-art fMRI-to-image approach that retrieves and reconstructs images from brain activity
Project page: https://medarc-ai.github.io/mindeye/
arXiv: https://arxiv.org/abs/2305.18274
sterile cargoBOT
Hey @everyone, Harmonai Office Hours is starting now!
Come join the Harmonai team in the Center Stage to hear about the latest AI Music developments.
sterile cargoBOT
@everyone Starting now in the centre stage!
https://discord.gg/Sn9gmgyv?event=1113884419003007037
hazy otterBOT
@everyone
ãæ°äŒç»ç¬¬äžåŒŸãStable Diffusion Life - ããªãã®æŽ»çšæ³å€§åéããéå¬ïŒ
ä»äºã趣å³ãæãã¬å Žæã§äœ¿ã£ãŠãã #StableDiffusion ã®æŽ»çšæ³ããšããœãŒããæããŠãã ããïŒ
å
¥éžè
ã«ã¯DreamStudioã¯ã¬ãžãã20ãã«åâïž3人ããã¬ãŒã³ãðïŒå
¬åŒnoteãtwitterã§çŽ¹ä»ãããŸãïŒ â¶[https://note.com/stabilityai/n/n357cd4bbc960]
ãŸã ãŸã å¿åãå°ãªãã®ã§ããã²ãåå ãã ããïŒ
ã¡ãªã¿ã«20ãã«åã¯5000ååãStable Diffusionãéã¹ãŸãïŒ
hazy nacelleBOT
https://twitter.com/jsuarez5341/status/1672306455092027392?s=20 @Neural-MMO ðŠŸ
hazy nacelleBOT
@model release Some exciting news from our awesome @code.ai team!! We finally are live with what we have been working on the past few months: StableCode our first step towards helping the world get to 1 billion developers! https://twitter.com/StabilityAI/status/1688931312122675200
Twitter
hazy otterBOT
@everyone
ãæ°å
¬éã Stability AI Japan ã¯æé«æ§èœã®æ¥æ¬èªèšèªã¢ãã«ãJapanese StableLM Base Alpha 7BããšãJapanese StableLM Instruct Alpha 7Bããå
¬éããŸããïŒ
https://ja.stability.ai/blog/japanese-stablelm-alpha
ããã§ãä»å€18æããå
¬åŒDiscordã§LLMéçºè
ãšã®ããŒã¯ã€ãã³ããéå¬ããŸããæè¡çãªå
容ã«ã¯ãªããŸããåããŠã®æ¹ã倧æè¿ã§ãïŒè³ªåã»ãã·ã§ã³ããããŸãã®ã§ãçæ§ãã²ãåå ãã ããð ð
https://discord.com/invite/KGTF3m2U?event=1139018615358767145
hazy nacelleBOT
Big congrats to @Shahbuland @Tanishq on the initial release of DRLX! Featuring an accelerated DDPO trainer and various reward functions. Stay tuned for RLHF-enhanced image gen!
https://twitter.com/carperai/status/1692696352583557190?s=20
sterile cargoBOT
@everyone
ð¢ We're super excited to share the release of Stability AI's new text-to-audio platform: Stable Audio! ð¢
ðµ Try it out now at https://www.stableaudio.com and share what you make ðµ
ð¥ Training and inference code, as well as open-source models will be coming soon! ð¥ïž
If you want to join the discussion, head on over to the #stable-audio channel in the Stable Foundation server: https://discord.gg/k3vkc5vE
We'll be hosting our Harmonai Office Hours in a little over half an hour to discuss Stable Audio, and our upcoming open-source releases. Come on through! https://discord.gg/92kynGB4?event=1151585333012611104
sterile cargoBOT
Hey @everyone, Harmonai Office hours are starting! https://discord.gg/ETrJTYNk?event=1162087709309931643
Also, we've made our training repo for our audio generation models public!
Check out https://github.com/Stability-AI/stable-audio-tools to see what we've been working on. No pre-trained models there yet, those will be coming soon
old ventureBOT
@everyone
Continuing EleutherAIâs mission of pushing forward open research and broadening access to the tools that make this possible, we are releasing Llemma, a family of powerful base models for mathematics trained via continued pretraining of CodeLlama on a general mathematics dataset for up to 200B tokens.
Why is this important?
A year ago, Google published Minerva, a LLM with impressive mathematical reasoning abilities. Minerva isnât publicly accessible, preventing research from building on these advances. This has hindered outside progress in the Math+AI subfield greatly!
Just like CodeLlama has helped spur advances in open AI for code research, we hope that others will build on Llemma to be a strong platform for furthering the study of AI for mathematics! We release our models, datasets, and training, evaluation, and analysis code.
This is the beginning of our research on this topic, not the end. Our current plans include furthering few-shot theorem proving and finetuning for full-proof generation, and much more. Come join us in https://discord.com/channels/729741769192767510/1112407059359600662â to get involved. We also meet on Thursdays at 2pm US Eastern Time.
Work done through collaboration between EleutherAI and several academic labs, by @zhangir_azerbayev @Hailey Schoelkopf @keirp @dsantosmarco @mcaleste @Albert Jiang Jia Deng @Stella Biderman (she/her) @wellecks !
Blog post: https://blog.eleuther.ai/llemma
ArXiv paper: https://arxiv.org/abs/2310.10631
Project page: https://github.com/EleutherAI/math-lm
Sample explorer: https://llemma-demo.github.io/
old ventureBOT
Hi all! @NLP
We just released a new paper on Quality-Diversity through AI Feedback (QDAIF), a way for LLMs to automatically generate meaningfully diverse, high-quality text responses in creative domains (like generating stories and poems).
For many tasks we want a diverse range of high-quality outputs from models to choose from. QD algorithms aim towards this, but it's challenging to define measures for quality and diversity by hand in subjective domains like creative writing. Inspired by RLAIF, what if LLMs assessed qualitative features of diversity, too? That way LLMs could generate, diversify, and improve their own responses.
QDAIF enables this search for diverse, high-quality solutions, overcoming the limitations of hand-crafted measures in creative writing domains (opinions, stories, poetry). We found QDAIF to be better suited in creative writing domains at covering the search space with diverse, high-quality stories, poems, etc., compared to baselines and verified the grounding of QDAIF through human evaluation.
This work was part of a research collaboration between EleutherAI, CarperAI, StabilityAI, Aleph Alpha, UBC, and others. Shoutout to @andmany ,@joellehman , and@lactoseintol (and any others I've missed), not to mention @gooseluvr for the support!
Project page: https://qdaif.github.io/
ArXiv: https://arxiv.org/abs/2310.13032
Tweet: https://x.com/andrewdai99/status/1716913881816383805?s=20
hazy nacelleBOT
Hi all!
We just released a new paper on Quality-Diversity through AI Feedback (QDAIF), a way for LLMs to automatically generate meaningfully diverse, high-quality text responses in creative domains (like generating stories and poems).
For many tasks we want a diverse range of high-quality outputs from models to choose from. QD algorithms aim towards this, but it's challenging to define measures for quality and diversity by hand in subjective domains like creative writing. Inspired by RLAIF, what if LLMs assessed qualitative features of diversity, too? That way LLMs could generate, diversify, and improve their own responses.
QDAIF enables this search for diverse, high-quality solutions, overcoming the limitations of hand-crafted measures in creative writing domains (opinions, stories, poetry). We found QDAIF to be better suited in creative writing domains at covering the search space with diverse, high-quality stories, poems, etc., compared to baselines and verified the grounding of QDAIF through human evaluation.
This work was part of a research collaboration between EleutherAI, CarperAI, StabilityAI, Aleph Alpha, UBC, and others. Shoutout to @andmany @joellehman , and @lactoseintol (and any others I've missed), not to mention @canadagoose1 for the support!
Project page: https://qdaif.github.io/
ArXiv: https://arxiv.org/abs/2310.13032
Tweet: https://x.com/andrewdai99/status/1716913881816383805?s=20
old ventureBOT
@everyone We've had a number of papers accepted for publication recently, including both previously released ones and new ones. In lieu of many posts today, I figured batching them probably makes more sense ð I've provided some annotations to help guide which might be new to you.
Apologies if I'm missing your tag or paper! Please DM me with corrections.
EMNLP
RWKV: Reinventing RNNs for the Transformer Era (Findings) by @blinkdl @hypnopump @Quentin Anthony, et al.
trlX: A Framework for Large Scale Reinforcement Learning from Human Feedback by @alexhavrilla @maxreciprocate @duyphung.ai @1.69 @Ryan Gosling @Stella Biderman (she/her) @Quentin Anthony Ethan Kim and @ihateihatelouis paper forthcoming
NeurIPS
The Goldilocks of Pragmatic Understanding: Fine-Tuning Strategy Matters for Implicature Resolution by LLMs (spotlight) by @Laura Ruis Akbir Khan, @Stella Biderman (she/her), @sara_hooker, Tim RocktÀschel, and Edward Grefenstette major paper update
LEACE: Perfect linear concept erasure in closed form by @norabelrose @dsj @shauli_ @rcotterell @edwardraff @Stella Biderman (she/her)
Reconstructing the Mind's Eye: fMRI-to-Image with Contrastive Learning and Diffusion Priors by Scotti et al. incl @ilovescience (spotlight) previously unannounced
Emergent and Predictable Memorization in Large Language Models by @Stella Biderman (she/her) @Orz @lintangsutawika @Hailey Schoelkopf @Quentin Anthony @Ryan Gosling and @edwardraff
Math-AI Workshop
OpenWebMath: An Open Dataset of High-Quality Mathematical Web Text by @keirp @dsantosmarco @zhangir_azerbayev and Jimmy Ba
Llemma: An Open Language Model For Mathematics by @zhangir_azerbayev @Hailey Schoelkopf @keirp@dsantosmarco @mcaleste @Albert Jiang @jianga2718@Stella Biderman (she/her) @wellecks
Socially Responsible Language Modelling Research Workshop
Eliciting Language Model Behaviors using Reverse Language Models (spotlight) by Pfau et al. incl. @alexinfanger and @ai_waifunew paper
@Stella Biderman (she/her) has an invited panel.
Workshop on Backdoors in Deep Learning
Detecting Backdoors with Meta-Models by Langosco et al. incl. @Hyperion new paper
__Workshop on Attributing Model Behavior at Scale (ATTRIB) __
Sparse Autoencoders Find Highly Interpretable Features in Language Models by @hoagy @aidan ewart @loganriggs @Robert_AIZI @leesharkey
Apache Cassandra Conference
@picocreator has an invited talk on RWKV
Nature
Roleplay with Large Language Models by Murray Shanahan and @repligate previously unannounced
Preprints
The OpenELM Library by @Hyperion @Honglu Ryan Zhou, Daniel Scott, and @joellehman new paper
Meet-ups
If you want to meet up with EleutherAI check out #1171809525280550932 #1171291697561477170 #1182032181921587200 respectively.
knotty pagodaBOT
Happy holidays to you all! ð
echo hamletBOT
@everyone Wishing you a Joyous New Year! As we step into this exciting chapter, we extend our heartfelt gratitude to each of you who joined our community. Whether you enriched our open research initiatives or engaged in spirited discussions across our channels, your contributions have been invaluable. A sincere thank you to everyone, and we look forward to sharing more meaningful moments with you in the upcoming year! Cheers to a fantastic year ahead! ð
https://tenor.com/view/funny-animals-kittens-smiling-cute-cats-happy-gif-12099363
ancient spearBOT
@everyone We're excited to share some of our first work from our LLM efforts!
Our goal is to eventually train open LLMs that have SOTA medical capabilities, but first we must understand how current LLMs perform.
That's what we do in our new blog post!
Highlights:
⢠We've implemented Google's MultiMedQA suite of tasks in
EleutherAI's lm-eval-harness for easy eval of open LLMs
⢠We've discovered that SOTA generalist open LLMs like Qwen-72b outperform Med-PaLM (SOTA in Nov. 2022) and even openly released medical LLMs like Meditron-70b, all without any special prompting
⢠We perform a dataset contamination analysis and don't observe any strong signs of test set contamination
⢠Lots of future directions to explore for medical LM evals, this blog post is just part 1!
Read here â https://www.medarc.ai/blog/medarc-llms-eval-part-1
MedARC tweet â https://twitter.com/MedARC_AI/status/1750506121545359862
Twitter
old ventureBOT
@everyone Do neural nets learn features in a predictable order?
Our results suggest the answer is âyesââ networks learn statistics of increasing complexity. Early-training networks only use low-order moments (mean & covariance) of the input distribution.
Specifically, we show that networks automatically learn to perform well on maximum-entropy distributions whose low-order statistics match those of the training set early in training, then lose this ability later.
We also extend our theory to language models by proving an equivalence between token n-gram frequencies and the moments of embedding vectors. Empirically, we find a fascinating double descent phenomenon: Pythia does well on unigram & bigram sequences in the first ~256 steps, then gets worse as it learns higher-order n-gram statistics, then gets better again by using in-context learning to adjust to the new distribution.
Finally we use optimal transport methods to surgically edit the low-order statistics of one class to match those of another, and show that early-training networks treat the edited samples as if they were drawn from the target class.
Paper: https://arxiv.org/abs/2402.04362
Code: https://github.com/EleutherAI/features-across-time
Twitter thread: https://x.com/norabelrose/status/1755680678736547910?s=20
Thanks to @quintinpope @luciaquirke @Alex Mallen @.xfern
old ventureBOT
@everyone Most large language models trained last year were multilingual, but our tooling for evaluating models trained in languages other than English and Chinese are quite limited. Often times, organizations will even evaluate their models by translating evaluation benchmarks. However the kinds of questions of interest to people in different countries and cultures differs, and sometimes the correct answer to an allegedly objective question differs by language!
To improve evaluation practices for Korean language we've been working with Korean NLP researchers and industry practitioners to build two new evaluation datasets:
Hae-Rae Bench: This evaluation benchmark contains six tasks across four domains: vocabulary, history, general knowledge, and reading comprehension. HAE-RAE Bench emphasizes a model's aptitude for recalling Korean-specific knowledge and cultural contexts, presenting a greater challenge to non-native models, by disturbing abilities and knowledge learned from crosslingual transfer. This work was done by @GSON @albert_h_lee .
K-MMLU: This benchmark replicates the methodology that produced MMLU, but using examinations common in Korea. We manually annotate a subset of the questions as to whether they require Korea-specific knowledge and also designate a KMMLU-Hard subset that current models find especially challenging. This work was done by @GSON @albert_h_lee @lliy8786 @muennighoff @Stella Biderman (she/her) .
Hae-Rae Bench has been accepted to LREC-COLING 2024, and KMMLU is under review at ACL. Both of them can be run today via the Language Model Evaluation Harness.
If you speak a language other than English or a non-mainstream culture in an English-speaking country and would like to talk about designing a benchmark to measure language model competencies that matter to you, come join us in the new #1208111628051152969 channel. If you're interested in evaluation more generally or want help using our evaluation framework, come by #755950983669874798
hazy otterBOT
@everyone
çæ§ããã€ãããããšãããããŸããStability AI JapanãããïŒç¹ã®ãã¥ãŒã¹ããå±ãããŸãã
Stability AI å ¬åŒããŒãããŒäŒæ¥ãAICU Inc.ãããããç»åçæAI Stable Diffusion ã¹ã¿ãŒãã¬ã€ãããçºå£²ãããŸãã
åœç€Ÿä»£è¡šJerry Chiã¯ãããStability AI Japanã¡ã³ããŒãã¬ãã¥ãŒã«å
¬åŒåå ããŠãããŸãã
æè¡è
ã®æ¹ããAIã€ã©ã¹ãã¬ãŒã¿ãŒã®æ¹ãŸã§ãStable Diffusionã®åºç€ããå¿çšãŸã§åŠãã§é ããŸãã
æ¡åŒµæ©èœãçæãã¯ããã¯ã«ã€ããŠã解説ãããŠãããåå¿è
ã®æ¹ã§ãæ¬æžã«ãã£ãŠæãéãã®ã€ã©ã¹ããçæã§ããŸãã
3æ29æ¥(éææ¥) çºå£²äºå®ã§ãããäºçŽã¯ä»¥äžã®ãªã³ã¯ããããã²ãé¡ãããããŸãã
https://j.aicu.ai/SBXL
NVIDIAäž»å¬ãGPU Technology Conference 2024ïŒGTC2024ïŒãã3æ19æ¥(ç«)ãã22æ¥(é)ãŸã§éå¬äžã§ãã
ãããèšå¿µããŠããNVIDIA CEO ãžã§ã³ã»ã¹ã³ã»ãã¢ã³ ãµã€ã³å
¥ãGeForce RTX 4090ããåœãããã£ã³ããŒã³ãå®æœããŸãã
GTC2024 AIã«ã³ãã¡ã¬ã³ã¹ã®ã奜ããªã»ãã·ã§ã³ãèŠèŽããã ãã§ãæœéžã«å¿åã§ããŸãã
å¿åæ¹æ³
å
¬åŒX @StabilityAI_JP ããã©ã㌠& ãªãã¹ãããŠå¿å
https://twitter.com/StabilityAI_JP/status/1767737620736659863
以äžã®ãªã³ã¯ãã#GTC24 ã«ç¡æåå ç»é²ãã»ãã·ã§ã³ã1ã€ä»¥äžèŠèŽ
â»ç»é²æãLocation ã JapanïŒæ¥æ¬ïŒã«ããŠããæ¹ã察象ã§ãã
https://nvidia.com/ja-jp/gtc/?ncid=ref-inpa-805225
å¿å㯠3æ27æ¥(ç«ææ¥) ååŸ4æ59å ãŸã§ã
NVIDIA CEOãµã€ã³å
¥ãGPUãåœéžãã貎éãªãã£ã³ã¹ã§ãããã²å¥®ã£ãŠãç³èŸŒã¿ãã ããã
hazy otterBOT
@everyone
Stable Code 3BãããŒã¹ã«ããæ°ããæ瀺åŠç¿æžã¿LLMããStable Code Instruct 3BãããªãªãŒã¹ããŸããã
ãã®ã¢ãã«ãå©çšããããšã§ãèªç¶èšèªããã³ããã«ãã£ãŠã³ãŒãçæïŒæ°åŠïŒãã®ä»ã®ãœãããŠã§ã¢ãšã³ãžãã¢ãªã³ã°é¢é£ã®åºåãªã©ã
æ§ã
ãªã¿ã¹ã¯ãåŠçããããšãã§ããŸãã
Codellama 7B InstructãDeepSeek-Coder Instruct 1.3Bãªã©ãåç以äžã®ãµã€ãºã®ã¢ãã«ã«å¹æµããæ§èœãæã¡ãŸãã
ãŸããStability AIã¡ã³ããŒã·ãããååŸããããšã§ãåçšå©çšããããšãã§ããŸãã
詳现ã¯ä»¥äžã®èšäºãã芧ãã ããã
https://ja.stability.ai/blog/stable-code-instruct-3b
hazy otterBOT
@everyone
Stable Audio 1.0ãããŒã¹ã«æ§ç¯ãããæ°ã¢ãã«ãStable Audio 2.0ãã®ãã¢ãµã€ããå
¬éããŸããã
ãã®ã¢ãã«ã¯ã44.1KHzã¹ãã¬ãªã§æ倧3åéã®é«å質ãªãã«ãã©ãã¯ãçæã§ããŸãã
ãªãŒãã£ãªãããªãŒãã£ãªãžã®å€ææ©èœãåããŠããŸãã
ãªãŒãã£ãªãµã³ãã«ãšèªç¶èšèªã«ããããã³ãããçšããŠãããŸããŸãªãµãŠã³ããçæã§ããŸãã
ãŸãããµãŠã³ããšãã§ã¯ãã®çæãšã¹ã¿ã€ã«ã®è»¢éãæ¡åŒµããã
ã¢ãŒãã£ã¹ãããã¥ãŒãžã·ã£ã³ã«æè»æ§ãšã³ã³ãããŒã«æ§ãæäŸããã¯ãªãšã€ãã£ããªããã»ã¹ãåäžãããŸãã
ãã®ã¢ãã«ã¯ãã¢ãµã€ãã§ç¡æã§äœ¿çšããããšãå¯èœã§ããå¶äœãå§ããŠã¿ãŠãã ããã
https://stableaudio.com/
詳现ã¯ãã¡ãã®èšäºãã芧ãã ããã
https://ja.stability.ai/blog/stable-audio-20
sterile cargoBOT
@everyone
ð£ ðµ We're thrilled to announce the launch of Stable Audio 2.0! ðµ ð£
This new model enables higher quality outputs up to three minutes with improved prompt fidelity, and is now available to all users at https://www.stableaudio.com.
We've also enabled the ability to upload your own audio for style transfer!
Check out our full release blog post here, with a few details on the model implementation: https://stability.ai/news/stable-audio-2-0
We're also super excited to bring back our 24/7 Stable Radio with music made entirely by Stable Audio! Check it out here: https://www.youtube.com/watch?v=yvOXZ6SV2Rk
hazy otterBOT
@everyone
æšæ¥ãªãªãŒã¹ãããŠé ããStable Diffusion 3 APIã«ã€ããŠããã詳现ãªæ
å ±ãæäŸãããŠé ããŸãã
ãã²ãæ¥æ¬èªããã°ããäžèªãã ãããŸãã
Stable Diffusion 3ã®ãªãµãŒãããŒããŒã§æããã«ãããŠããããã«ããã®ã¢ãã«ã¯ã人éã®å奜è©äŸ¡ã«åºã¥ããŠãDALL-E 3 ã Midjourney v6 ãªã©ã®ããã¹ãç»åçæã·ã¹ãã ãã¿ã€ãã°ã©ãã£ãšããã³ããã®å¿ å®æ§ã«ãããŠäžåã£ãŠããŸãã
æ°ããMultimodal Diffusion Transformer (MMDiT)ã¢ãŒããã¯ãã£ã¯ãç»åè¡šçŸãšèšèªè¡šçŸã«å¥ã ã®ãŠã§ã€ãã»ããã䜿çšããããšã§ãStable Diffusionã®æ§ããŒãžã§ã³ãšæ¯èŒããŠãããã¹ãç解ãšã¹ããªã³ã°æ©èœãåäžããŠããŸãã
ã¢ãã«ã¯æ¬æ¥ããAPIãéããŠå©çšå¯èœã§ãããç§ãã¡ã¯ãªãŒãã³ãªãªãªãŒã¹ã«å ç«ã¡ãã¢ãã«ã®æ¹åã«ç¶ç¶çã«åãçµãã§ããŸããç§ãã¡ã®ãªãŒãã³ãªçæAIãžã®åãçµã¿ã«åºã¥ããè¿ãå°æ¥ã«ãStability AIã¡ã³ããŒã·ããã§ã¢ãã«ã®ãŠãšã€ããå©çšã§ããããã«ããããšãç®æããŠããŸãã
ãã詳ããæ
å ±ã¯ã以äžã®æ¥æ¬èªããã°ðŸ ãã芧ãã ããïŒ
https://ja.stability.ai/blog/stable-diffusion-3-api
ãŸããStable Diffusion 3 APIãç¡æã§ããè©ŠããColabãäœæããŸããããå©çšãã ããã
https://x.com/xqdior/status/1780618334607942054
Twitter
hazy otterBOT
@everyone
æ¥æ¬èªå€§èŠæš¡èšèªã¢ãã«ãJapanese Stable LM 2 1.6BãããªãªãŒã¹ããŸããã
ãã¡ãã®ã¢ãã«ã¯Stability AI ã¡ã³ããŒã·ããã«ãå å ¥ããã ãããšã§åçšå©çšãå¯èœã§ãã
詳现ã¯ä»¥äžã®ããã°ãã芧ãã ããïŒ
https://ja.stability.ai/blog/japanese-stable-lm-2-16b
hazy otterBOT
@everyone
Discord äžã§Stability AIã®ã¢ãã«ãçŽæ¥äœ¿çšã§ããããStable ArtisanãããªãªãŒã¹ããŸãã
DiscordãµãŒããŒïŒhttps://discord.gg/stablediffusion
詳现ã¯ä»¥äžã®ããã°ãã芧ãã ããïŒïŒè±èªïŒ
https://ja.stability.ai/blog/stable-artisan
echo hamletBOT
Hey @everyone, some extremely cool people prepared an AlphaFold3 letter to the Nature editor criticizing the journal for not upholding their policies about making code available to reviewers and alongside publications and also expressing concern about the precedent this sets. The letter and a form to collect endorsements are at https://docs.google.com/forms/d/e/1FAIpQLSf6ioZPbxiDZy5h4qxo-bHa0XOTOxEYHObht0SX8EgwfPHY_g/viewform
sterile cargoBOT
@everyone
ð¢ Stable Audio Open 1.0 is available to the public! ð¢
Model weights are available at https://huggingface.co/stabilityai/stable-audio-open-1.0
This model is trained to generate sound effects, samples, and field recordings. Great for making samples for your music!
You can use the stable-audio-tools repo to fine-tune this model on your own sample libraries and create your own custom Stable Audio models.
We're so excited to get this out, can't wait to see what y'all make with it!
hazy otterBOT
@everyone
æ¬æ¥ãStable Diffusion 3ã·ãªãŒãºã®ææ°ãã€æãå
é²çãªããã¹ãããç»åãžã®AIã¢ãã«ã§ããStable Diffusion 3 Mediumã®ãªãŒãã³ãŠãšã€ããçºè¡šã§ããããšãå¬ããæããŸãïŒð
ãã®æ°ãããªãªãŒã¹ã¯ããžã§ãã¬ãŒãã£ãAIã®é²åã«ãããéèŠãªãã€ã«ã¹ããŒã³ã§ããããã®ãã¯ãã«ãªãã¯ãããžãŒãæ°äž»åãããšããç§ãã¡ã®ã³ãããã¡ã³ããç¶ç¶ãããã®ã§ãã
SD3 Medium ã¯ãSD3ã®20åãã©ã¡ãŒã¿ãŒã¢ãã«ã§ãããã€ãã®ç¹çãã¹ãç¹åŸŽãåããŠããŸãã
詳ããã¯ãã¡ãð
https://ja.stability.ai/blog/stable-diffusion-3-medium
hazy otterBOT
@everyone
Stable Video 4Dã¯ããŠãŒã¶ãŒã1ã€ã®ãããªãã¢ããããŒãããã ãã§ã8ã€ã®æ°ããã¢ã³ã°ã«ã®ãã€ãããã¯ãªããã«ã»ãã¥ãŒã»ãããªãåãåãããšãã§ãã
æ°ããªã¬ãã«ã®å€æ§æ§ãšåµé æ§ãæäŸãããStability AI åã®video-to-video çæã¢ãã«ã§ãã
詳现ã¯ä»¥äžã®ããã°ãã芧ãã ããïŒ
https://ja.stability.ai/blog/stable-video-4d
old ventureBOT
Hey @everyone! The HPC team lead by @Quentin Anthony has been hard at work keeping our GPT-NeoX library at the forefront of large scale AI training. The most recent major feature, with @dmayhem93 (Super Saiyan Aligned) @Not Not Louis e/ð and @nathanthinks at SynthLabs and @ai_waifu, is the introduction of post-training to GPT-NeoX. Now you can do SFT, DPO, and KTO finetuning native to the GPT-NeoX library itself and we have other algorithms including REINFORCE and PPO on the way.
Our testing shows a 30% performance improvement over HuggingFace's trl library at the 13B scale, with the added bonus of being scalable to massive computing systems that trl doesn't support.
This is part of a broader push to improve the GPT-NeoX library and continue to power open research at scale on frontier HPC systems. Preference learning joins other new features such as:
main if you don't want to wait for the stable release though! GPT-NeoX 3.0 is currently in pre-release bug testing so if you give it a try stop by #730090096287547444 and let us know what your experience is like.Check out our blog post (and SynthLabs' here) to learn more or head over to the GPT-NeoX library to give it a try.
hazy otterBOT
@everyone
ãã€ã³ã
ãªãªãŒã¹ã¢ãã«
Stable Diffusion 3.5 Large ããã³ Stable Diffusion 3.5 Large Turbo ã¯ãçŸåš Hugging Face ããããŠã³ããŒãã§ããGitHub ã§ã¯æšè«ã³ãŒããå
¥æå¯èœã§ãã
詳现ã¯ä»¥äžã®ããã°ãã芧ãã ããïŒ
https://ja.stability.ai/blog/introducing-stable-diffusion-3-5
echo hamletBOT
@everyone Wanted to give a massive shout-out to @de_muedi, @utterly_butterly, and the entire @DNA-LLM team on their recent preprint Life as a Function: Why Transformer Architectures Struggle to Gain Genome-Level Foundational Capabilities.
It's still an early version and they plan to extend the work presented here as well as other downstream projects/applications.
sterile cargoBOT
@everyone
ð¢We're excited to announce the release of Stable Audio Open Small, now available for download on Hugging Face!ð¢
This is a smaller (341M parameters), more efficient version of our Stable Audio Open 1.0 model, optimized for quick inference.
To read about the new ARC post-training method we used to accelerate this model, check out our new research paper on arXiv!
We also partnered with Arm on this release to enable further optimization of the model for deployment on CPUs. You can check out their new learning path to see how you can enable fast edge deployment of this new model.
ð»Weights: https://huggingface.co/stabilityai/stable-audio-open-small
ðPaper: https://arxiv.org/abs/2505.08175
ðArm learning path: https://learn.arm.com/learning-paths/mobile-graphics-and-gaming/run-stable-audio-open-small-with-lite-rt
old ventureBOT
@everyone Let's say you're working with an AI as a co-scientist and you ask it to proof read the paper and report back on any mistakes. How likely is it to find real errors? We built SPOT, a benchmark of papers that were retracted or errata'd to find out! SPOT comprises 83 research papers with 91 author-validated error annotations across 10 STEM fields. We categorize errors into 6 error types â equation/proof, figure duplication, data inconsistency, statistical reporting, reagent identity, and experiment setup - and include information on where in the paper the error is, whether it lead to an erata or a retraction, and an expert-human authored description of the error.
We benchmarked 10 top models, both closed and open, and the results are sobering â the best results are from o3 which has a precision of 6% and a recall of 21%. All other models score below 4% precision and 10% recall. We also look at what happens when you run a model multiple times: across 8 trials models rarely discover the same errors and generally assign a near-zero confidence to their claims. The appendix contains breakdowns by field, error type, ablations for when figures are omitted, and more.
This work was lead by @GSON with contributions by @Honglu @mrgonao myself, and several others who aren't on Discord.
