Hieu Lam
lamhieu
AI & ML interests
.-.
Articles
Organizations
lamhieu's activity
reacted to
m-ric's
post with ๐ฅ
about 1 month ago
Post
1160
Emu3: Next-token prediction conquers multimodal tasks ๐ฅ
This is the most important research in months: weโre now very close to having a single architecture to handle all modalities. The folks at Beijing Academy of Artificial Intelligence (BAAI) just released Emu3, a single model that handles text, images, and videos all at once.
๐ช๐ต๐ฎ๐'๐ ๐๐ต๐ฒ ๐ฏ๐ถ๐ด ๐ฑ๐ฒ๐ฎ๐น?
๐ Emu3 is the first model to truly unify all these different types of data (text, images, video) using just one simple trick: predicting the next token.
And itโs only 8B, but really strong:
๐ผ๏ธ For image generation, it's matching the best specialized models out there, like SDXL.
๐๏ธ In vision tasks, it's outperforming top models like LLaVA-1.6-7B, which is a big deal for a model that wasn't specifically designed for this.
๐ฌ It's the first to nail video generation without using complicated diffusion techniques.
๐๐ผ๐ ๐ฑ๐ผ๐ฒ๐ ๐ถ๐ ๐๐ผ๐ฟ๐ธ?
๐งฉ Emu3 uses a special tokenizer (SBER-MoVQGAN) to turn images and video clips into sequences of 4,096 tokens.
๐ Then, it treats everything - text, images, and videos - as one long series of tokens to predict.
๐ฎ During training, it just tries to guess the next token, whether that's a word, part of an image, or a video frame.
๐๐ฎ๐๐ฒ๐ฎ๐๐ ๐ผ๐ป ๐๐ต๐ฒ ๐ฟ๐ฒ๐๐๐น๐๐:
๐ In image generation, Emu3 beats SDXL, but itโs also much bigger (8B vs 3.5B). It would be more difficult to beat the real diffusion GOAT FLUX-dev.
๐ In vision, authors also donโt show a comparison against all the current SOTA models like Qwen-VL or Pixtral.
This approach is exciting because it's simple (next token prediction) and scalable(handles all sorts of data)!
Read the paper ๐ Emu3: Next-Token Prediction is All You Need (2409.18869)
This is the most important research in months: weโre now very close to having a single architecture to handle all modalities. The folks at Beijing Academy of Artificial Intelligence (BAAI) just released Emu3, a single model that handles text, images, and videos all at once.
๐ช๐ต๐ฎ๐'๐ ๐๐ต๐ฒ ๐ฏ๐ถ๐ด ๐ฑ๐ฒ๐ฎ๐น?
๐ Emu3 is the first model to truly unify all these different types of data (text, images, video) using just one simple trick: predicting the next token.
And itโs only 8B, but really strong:
๐ผ๏ธ For image generation, it's matching the best specialized models out there, like SDXL.
๐๏ธ In vision tasks, it's outperforming top models like LLaVA-1.6-7B, which is a big deal for a model that wasn't specifically designed for this.
๐ฌ It's the first to nail video generation without using complicated diffusion techniques.
๐๐ผ๐ ๐ฑ๐ผ๐ฒ๐ ๐ถ๐ ๐๐ผ๐ฟ๐ธ?
๐งฉ Emu3 uses a special tokenizer (SBER-MoVQGAN) to turn images and video clips into sequences of 4,096 tokens.
๐ Then, it treats everything - text, images, and videos - as one long series of tokens to predict.
๐ฎ During training, it just tries to guess the next token, whether that's a word, part of an image, or a video frame.
๐๐ฎ๐๐ฒ๐ฎ๐๐ ๐ผ๐ป ๐๐ต๐ฒ ๐ฟ๐ฒ๐๐๐น๐๐:
๐ In image generation, Emu3 beats SDXL, but itโs also much bigger (8B vs 3.5B). It would be more difficult to beat the real diffusion GOAT FLUX-dev.
๐ In vision, authors also donโt show a comparison against all the current SOTA models like Qwen-VL or Pixtral.
This approach is exciting because it's simple (next token prediction) and scalable(handles all sorts of data)!
Read the paper ๐ Emu3: Next-Token Prediction is All You Need (2409.18869)
reacted to
singhsidhukuldeep's
post with ๐
2 months ago
Post
1627
Just wrapped up a deep dive into the latest lecture on building LLMs, such as ChatGPT, from
@Stanford
CS229 course. Here are my top takeaways:
๐ Understanding the Components: LLMs like ChatGPT, Claude, and others are more than just neural networks; they are a complex blend of architecture, training loss, data evaluation, and systems. Knowing how these components work together is key to improving and scaling these models.
๐ Scaling Matters: Performance improves predictably with more data, bigger models, and greater computational power. However, balancing these factors is crucial to avoid overfitting and resource waste.
๐ Data is King: LLMs are trained on trillions of tokens scraped from the internet, but the quality of this data matters immensely. Rigorous filtering and deduplication processes are essential to maintaining data integrity.
๐๏ธ Pre-Training vs. Post-Training: While pre-training equips the model with general knowledge, post-training (like RLHF) fine-tunes it to follow human-like responses, reducing toxic outputs and improving alignment with human values.
๐ Reinforcement Learning from Human Feedback (RLHF): This technique allows LLMs to maximize outputs that align with human preferences, making models more reliable and accurate.
๐ก Why It Matters: Understanding these processes not only helps us appreciate the complexity behind our everyday AI tools but also highlights the challenges and opportunities in the ever-evolving field of AI.
Whether youโre in tech, data science, or just AI-curious, staying updated on these advancements is crucial. LLMs are not just transforming industries; theyโre redefining the future of human-computer interaction!
I just realized this was almost 2 hours long...
Link: https://www.youtube.com/watch?v=9vM4p9NN0Ts
๐ Understanding the Components: LLMs like ChatGPT, Claude, and others are more than just neural networks; they are a complex blend of architecture, training loss, data evaluation, and systems. Knowing how these components work together is key to improving and scaling these models.
๐ Scaling Matters: Performance improves predictably with more data, bigger models, and greater computational power. However, balancing these factors is crucial to avoid overfitting and resource waste.
๐ Data is King: LLMs are trained on trillions of tokens scraped from the internet, but the quality of this data matters immensely. Rigorous filtering and deduplication processes are essential to maintaining data integrity.
๐๏ธ Pre-Training vs. Post-Training: While pre-training equips the model with general knowledge, post-training (like RLHF) fine-tunes it to follow human-like responses, reducing toxic outputs and improving alignment with human values.
๐ Reinforcement Learning from Human Feedback (RLHF): This technique allows LLMs to maximize outputs that align with human preferences, making models more reliable and accurate.
๐ก Why It Matters: Understanding these processes not only helps us appreciate the complexity behind our everyday AI tools but also highlights the challenges and opportunities in the ever-evolving field of AI.
Whether youโre in tech, data science, or just AI-curious, staying updated on these advancements is crucial. LLMs are not just transforming industries; theyโre redefining the future of human-computer interaction!
I just realized this was almost 2 hours long...
Link: https://www.youtube.com/watch?v=9vM4p9NN0Ts
- 3 replies
Sounds interesting but I think there will be a big breakthrough, a new "architecture/methodology/factor/rethinking" for developing large models. That's what I think, I don't know what it is yet, haha.
reacted to
m-ric's
post with ๐
2 months ago
Post
840
๐ย ๐ช๐ต๐ฒ๐ฟ๐ฒ ๐๐ฐ๐ฎ๐น๐ถ๐ป๐ด ๐น๐ฎ๐๐ ๐ฎ๐ฟ๐ฒ ๐๐ฎ๐ธ๐ถ๐ป๐ด ๐๐ : ๐ฏ๐ ๐ฎ๐ฌ๐ฎ๐ด, ๐๐ ๐๐น๐๐๐๐ฒ๐ฟ๐ ๐๐ถ๐น๐น ๐ฟ๐ฒ๐ฎ๐ฐ๐ต ๐๐ต๐ฒ ๐ฝ๐ผ๐๐ฒ๐ฟ ๐ฐ๐ผ๐ป๐๐๐บ๐ฝ๐๐ถ๐ผ๐ป ๐ผ๐ณ ๐ฒ๐ป๐๐ถ๐ฟ๐ฒ ๐ฐ๐ผ๐๐ป๐๐ฟ๐ถ๐ฒ๐
Reminder : โScaling lawsโ are empirical laws saying that if you keep multiplying your compute by x10, your models will mechanically keep getting better and better.
To give you an idea, GPT-3 can barely write sentences, and GPT-4, which only used x15 its amount of compute, already sounds much smarter than some of my friends (although it's not really - or at least I haven't tested them side-by side). So you can imagine how far a x100 over GPT-4 can take us.
๐๏ธย As a result, tech titans are racing to build the biggest models, and for this they need gigantic training clusters.
The picture below shows the growth of training compute: it is increasing at a steady exponential rate of a x10 every 2 years. So letโs take this progress a bit further:
- 2022: starting training for GPT-4 : 10^26 FLOPs, cost of $100M
- 2024: today, companies start training on much larger clusters like the โsuper AI clusterโ of Elon Muskโs xAI, 10^27 FLOPS, $1B
- 2026 : by then clusters will require 1GW, i.e. around the full power generated by a nuclear reactor
- 2028: we reach cluster prices in the 100 billion dollars, using 10GW, more than the most powerful power stations currently in use in the US. This last size seems crazy, but Microsoft and OpenAI already are planning one.
Will AI clusters effectively reach these crazy sizes where the consume as much as entire countries?
โก๏ธย Three key ingredients of training might be a roadblock to scaling up :
๐ธย Money: but itโs very unlikely, given the potential market size for AGI, that investors lose interest.
โก๏ธ Energy supply at a specific location
๐ย Training data: weโre already using 15 trillion tokens for Llama-3.1 when Internet has something like 60 trillion.
๐คย Iโd be curious to hear your thoughts: do you think weโll race all the way there?
Reminder : โScaling lawsโ are empirical laws saying that if you keep multiplying your compute by x10, your models will mechanically keep getting better and better.
To give you an idea, GPT-3 can barely write sentences, and GPT-4, which only used x15 its amount of compute, already sounds much smarter than some of my friends (although it's not really - or at least I haven't tested them side-by side). So you can imagine how far a x100 over GPT-4 can take us.
๐๏ธย As a result, tech titans are racing to build the biggest models, and for this they need gigantic training clusters.
The picture below shows the growth of training compute: it is increasing at a steady exponential rate of a x10 every 2 years. So letโs take this progress a bit further:
- 2022: starting training for GPT-4 : 10^26 FLOPs, cost of $100M
- 2024: today, companies start training on much larger clusters like the โsuper AI clusterโ of Elon Muskโs xAI, 10^27 FLOPS, $1B
- 2026 : by then clusters will require 1GW, i.e. around the full power generated by a nuclear reactor
- 2028: we reach cluster prices in the 100 billion dollars, using 10GW, more than the most powerful power stations currently in use in the US. This last size seems crazy, but Microsoft and OpenAI already are planning one.
Will AI clusters effectively reach these crazy sizes where the consume as much as entire countries?
โก๏ธย Three key ingredients of training might be a roadblock to scaling up :
๐ธย Money: but itโs very unlikely, given the potential market size for AGI, that investors lose interest.
โก๏ธ Energy supply at a specific location
๐ย Training data: weโre already using 15 trillion tokens for Llama-3.1 when Internet has something like 60 trillion.
๐คย Iโd be curious to hear your thoughts: do you think weโll race all the way there?
-
- 3 replies
posted
an
update
2 months ago
Post
1667
๐ฏ Ghost 8B Beta 1608: Empowering Your AI Assistant
๐ฆ Unlock the Power of Ghost 8B Beta 1608: Build Your Personal AI Companion
Ghost 8B Beta 1608 empowers you to create a safe and multilingual AI assistant tailored to your needs, directly on your personal computer. ๐งโ๐ป Leverage AI's capabilities within your own space! ๐ Ghost 8B Beta 1608 is ready to become your AI companion.
~
๐ฆ ๊ฐ์ธ์ฉ AI ๋ณด์กฐ ๋๊ตฌ๋ก Ghost 8B Beta 1608๋ฅผ ํ์ฉํ์ธ์!
Ghost 8B Beta 1608, AI์ ํ์ ํ์ฉํ์ฌ ์์ ํ๊ณ ๊ฐ์ธํ๋ ์ธ์ด ์ง์์ ์ ๊ณตํ๋ AI ๋ณด์กฐ ๋๊ตฌ๋ฅผ ์ง์ ๊ตฌ์ถํ ์ ์์ต๋๋ค. ๐งโ๐ป ๊ฐ์ธ ์ปดํจํฐ์์ AI์ ํํ์ ๋๋ฆฌ์ธ์! ๐ Ghost 8B Beta 1608๋ ๋น์ ์ AI ํํธ๋๊ฐ ๋ ์ค๋น๊ฐ ๋์ด ์์ต๋๋ค.
lamhieu/ghost-8b-beta-8k
ghost-x/ghost-8b-beta-668ead6179f93be717db4542
๐ฆ Unlock the Power of Ghost 8B Beta 1608: Build Your Personal AI Companion
Ghost 8B Beta 1608 empowers you to create a safe and multilingual AI assistant tailored to your needs, directly on your personal computer. ๐งโ๐ป Leverage AI's capabilities within your own space! ๐ Ghost 8B Beta 1608 is ready to become your AI companion.
~
๐ฆ ๊ฐ์ธ์ฉ AI ๋ณด์กฐ ๋๊ตฌ๋ก Ghost 8B Beta 1608๋ฅผ ํ์ฉํ์ธ์!
Ghost 8B Beta 1608, AI์ ํ์ ํ์ฉํ์ฌ ์์ ํ๊ณ ๊ฐ์ธํ๋ ์ธ์ด ์ง์์ ์ ๊ณตํ๋ AI ๋ณด์กฐ ๋๊ตฌ๋ฅผ ์ง์ ๊ตฌ์ถํ ์ ์์ต๋๋ค. ๐งโ๐ป ๊ฐ์ธ ์ปดํจํฐ์์ AI์ ํํ์ ๋๋ฆฌ์ธ์! ๐ Ghost 8B Beta 1608๋ ๋น์ ์ AI ํํธ๋๊ฐ ๋ ์ค๋น๊ฐ ๋์ด ์์ต๋๋ค.
lamhieu/ghost-8b-beta-8k
ghost-x/ghost-8b-beta-668ead6179f93be717db4542
posted
an
update
3 months ago
Post
3164
๐ Weโre excited to launch Ghost 8B Beta (1608), a top-performing language model with unmatched multilingual support and cost efficiency.
Key Highlights:
- Superior Performance: Outperforms Llama 3.1 8B Instruct, GPT-3.5 Turbo, Claude 3 Opus, GPT-4, and more in winrate scores.
- Expanded Language Support: Now supports 16 languages, including English, Vietnamese, Spanish, Chinese, and more.
- Enhanced Capabilities: Improved math, reasoning, and instruction-following for better task handling.
With two context options (8k and 128k), Ghost 8B Beta is perfect for complex, multilingual applications, balancing power and cost-effectiveness.
๐ Learn More: https://ghost-x.org/docs/models/ghost-8b-beta
ghost-x/ghost-8b-beta-668ead6179f93be717db4542
Key Highlights:
- Superior Performance: Outperforms Llama 3.1 8B Instruct, GPT-3.5 Turbo, Claude 3 Opus, GPT-4, and more in winrate scores.
- Expanded Language Support: Now supports 16 languages, including English, Vietnamese, Spanish, Chinese, and more.
- Enhanced Capabilities: Improved math, reasoning, and instruction-following for better task handling.
With two context options (8k and 128k), Ghost 8B Beta is perfect for complex, multilingual applications, balancing power and cost-effectiveness.
๐ Learn More: https://ghost-x.org/docs/models/ghost-8b-beta
ghost-x/ghost-8b-beta-668ead6179f93be717db4542
reacted to
Xenova's
post with ๐ฅ
3 months ago
Post
13697
I'm excited to announce that Transformers.js V3 is finally available on NPM! ๐ฅ State-of-the-art Machine Learning for the web, now with WebGPU support! ๐คฏโก๏ธ
Install it from NPM with:
๐๐๐ ๐ @๐๐๐๐๐๐๐๐๐๐๐/๐๐๐๐๐๐๐๐๐๐๐๐
or via CDN, for example: https://v2.scrimba.com/s0lmm0qh1q
Segment Anything demo: webml-community/segment-anything-webgpu
Install it from NPM with:
๐๐๐ ๐ @๐๐๐๐๐๐๐๐๐๐๐/๐๐๐๐๐๐๐๐๐๐๐๐
or via CDN, for example: https://v2.scrimba.com/s0lmm0qh1q
Segment Anything demo: webml-community/segment-anything-webgpu
- 5 replies
replied to
their
post
4 months ago
thanks @danielus ๐ค
replied to
their
post
4 months ago
@Dihelson
@llama-anon
@AIWizard76
@danielus
๐ Ghost 8B Beta Released: Game-Changing Language Model
Ghost 8B Beta is a groundbreaking language model developed with a clear vision: to deliver exceptional multilingual support, superior knowledge capabilities, and all while remaining cost-effective. This model comes in two context length variations, 8k and 128k, ensuring flexibility for various tasks. Moreover, it boasts built-in multilingual functionality, making it a powerful tool for global communication and understanding.
- See detailed article: https://huggingface.co/blog/lamhieu/ghost-8b-beta-released-game-changing-language-mode
- Model card: https://huggingface.co/ghost-x/ghost-8b-beta
- Official website: https://ghost-x.org/docs/models/ghost-8b-beta
replied to
their
post
4 months ago
๐ Ghost 8B Beta Released: Game-Changing Language Model
Ghost 8B Beta is a groundbreaking language model developed with a clear vision: to deliver exceptional multilingual support, superior knowledge capabilities, and all while remaining cost-effective. This model comes in two context length variations, 8k and 128k, ensuring flexibility for various tasks. Moreover, it boasts built-in multilingual functionality, making it a powerful tool for global communication and understanding.
- See detailed article: https://huggingface.co/blog/lamhieu/ghost-8b-beta-released-game-changing-language-mode
- Model card: https://huggingface.co/ghost-x/ghost-8b-beta
- Official website: https://ghost-x.org/docs/models/ghost-8b-beta
posted
an
update
4 months ago
Post
2099
๐ Ghost 8B Beta Released: Game-Changing Language Model
--
Ghost 8B Beta is a groundbreaking language model developed with a clear vision: to deliver exceptional multilingual support, superior knowledge capabilities, and all while remaining cost-effective. This model comes in two context length variations, 8k and 128k, ensuring flexibility for various tasks. Moreover, it boasts built-in multilingual functionality, making it a powerful tool for global communication and understanding.
--
* See detailed article: https://huggingface.co/blog/lamhieu/ghost-8b-beta-released-game-changing-language-mode
* Model card: ghost-x/ghost-8b-beta
* Official website: https://ghost-x.org/docs/models/ghost-8b-beta
--
Ghost 8B Beta is a groundbreaking language model developed with a clear vision: to deliver exceptional multilingual support, superior knowledge capabilities, and all while remaining cost-effective. This model comes in two context length variations, 8k and 128k, ensuring flexibility for various tasks. Moreover, it boasts built-in multilingual functionality, making it a powerful tool for global communication and understanding.
--
* See detailed article: https://huggingface.co/blog/lamhieu/ghost-8b-beta-released-game-changing-language-mode
* Model card: ghost-x/ghost-8b-beta
* Official website: https://ghost-x.org/docs/models/ghost-8b-beta
posted
an
update
4 months ago
Post
2112
๐คฏ Ghost 8B Beta emerges as a clear leader, surpassing even proprietary models like xAI Grok 1, OpenAI GPT 3.5, and Mistral Mixtral 8x7B. This dominance extends to its parity with Mistral Medium, further solidifying its position as a top-tier language model. Furthermore, Ghost 8B Beta stands out as one of only three models employing the zero-shot method for evaluation, alongside Claude 2 and Claude 3, showcasing its unique capabilities and potential for groundbreaking applications.
---
๐ฌ Chat with the model here:
- Playground with Ghost 8B Beta (ฮฒ, 8k): lamhieu/ghost-8b-beta-8k
- Playground with Ghost 8B Beta (ฮฒ, 128k): lamhieu/ghost-8b-beta-128k
- Official website: https://ghost-x.org/docs/models/ghost-8b-beta/
---
๐ฌ Chat with the model here:
- Playground with Ghost 8B Beta (ฮฒ, 8k): lamhieu/ghost-8b-beta-8k
- Playground with Ghost 8B Beta (ฮฒ, 128k): lamhieu/ghost-8b-beta-128k
- Official website: https://ghost-x.org/docs/models/ghost-8b-beta/
-
- 2 replies
replied to
their
post
4 months ago
Thank you for your dedication, it sounds great. Here I would like to share some additional information and perspectives so that everyone can better understand the issues we address:
- With language models, when applying in practice we only need it to be understood at 80% or a good overview and combining with RAG will bring better accuracy. So, here we will need a good level of truth telling model and the ability to understand and work with RAG at a very good level to be most effective.
- In Italian, I'm very happy when it speaks well, it proves that my training method and source code for it were correct because it's actually live with the d0x5 version. This is all because Italian was only added later (at the same time as German), responding to the fact that sometimes it can only be described as a translation mays.
- With the ability to reason, I hope you don't misunderstand. It still works well, just when compared to some current superior models like GPT 4o or Claude 3, there will be some songs where it will "lose". It still outperforms a lot of other much larger models. For example, the question "Andrew is free from 11 am to 3 pm, Joanne is free from noon to 2 pm and then 3:30 pm to 5 pm. Hannah is available at noon for half an hour, and then 4 pm to 6 pm. What are some options for start times for a 30 minute meeting for Andrew, Hannah, and Joanne?" taken from OpenAI GPT4 home page.
One note: in reasoning tests, models often set the temperature to 0, with Ghost 8B Beta we always set it to 0.1 as the lowest. The reason is simple because if at this level the model still reasons well, then at level 0.4 (the default level of the current chat) it will still often achieve the same results, and we want to aim for practical efficiency. rather than scores. Let's try to lower the temperature with some reasoning questions to experiment.
After all, you guys are great, thank you so much everyone.
An example of reasoning about time:
An example of a long context with extensive summary capabilities: Paper: Point out the highlights and identify the ideal people to apply it..
replied to
their
post
4 months ago
@Dihelson It's probably because you told the model to do it again. Try telling the model to change each word. Of course, it could still be because the model misunderstood.
replied to
their
post
4 months ago
Try the following conversation: (1) ask to write an article -> (2) ask to translate the article into the languages โโyou want.
replied to
their
post
4 months ago
@AIWizard76 It hasn't gone through any real eval tests to be able to compare, but if we're just talking about ghost 8b beta, it has good translation capabilities for supported languages. It works well for translating long texts and also translating into multiple languages โโsimultaneously.
replied to
their
post
4 months ago
It's simple, currently the base version will not try to lengthen the text and be more "obedient". Maybe tomorrow or the next day I'll put it up for everyone to try.
Note, the current version is running everything from version "disl-0x5", the new version will improve a lot but it may not be ready right now.
replied to
their
post
4 months ago
thank you for your comments and encouragement ๐ค
another question, how do you feel when conversing in Italian?
replied to
their
post
4 months ago
@danielus
let me ask, is this what you want?
replied to
their
post
4 months ago
@danielus I noticed the explanation model because this is what the chat version (ft from ghost 8b beta, base) does for the chat task (base will not try to explain and will respect the system more strictly). The goal of answering with more information is to help users avoid having to learn more or get side answers from just one question. Of course, this can sometimes be a hassle, we'll try to balance it out.