Vision + Language

I wanted to share a very good practical on vision + language. I worked with Yuki Asano to develop this notebook. If you got any interest in this, it could help a lot.

Basically, in this vision+lang models you will have a joint image and text embedding model which means that it maps both text and images to the same embedding space.

Notebook🚀

Nice to see such things. I genuinely hope they make it open source.


Post Link

Analysis for Llama 3.1.

1. 15.6T tokens, Tools & Multilingual
2. Llama arch + new RoPE
3. fp16 & static fp8 quant for 405b
4. Dedicated pad token
5. <|python_tag|><|eom_id|> for tools?
6. Roberta to classify good quality data
7. 6 staged 800B tokens long context expansion


Data mixture
- 50% general knowledge
- 25% maths & reasoning
- 17% code data and tasks
- 8% multilingual data


Source

Question of the day.


3 different image generators with all different algorithms etc but almost the same output,
any reason/insight why?

It's been an incredibly exciting week in the world of AI:

- OpenAI launched a new search tool called SearchGPT
- Meta updated its Llama language model to version 3.1
- Mistral AI released a new and improved Mistral Large 2 model
- DeepMind's AI achieved a silver medal at the International Math Olympiad
- Elon Musk announced plans to develop Grok 2 and 3

Btw it's depressing that almost all👆 of them are almost inaccessible, either beta version or needs high gpus.

I expect
1. bank rates will rapidly converge to the current parallel market rate, maybe a bit lower.
2. market rate will not go down but it will rise more slowly, over the next 6 months, % change of ETB/USD will be less than the last 6 months.
3. longer term, ETB will strengthen


Source from Nemo Semret

🔋Ranking Programming Languages by Energy Efficiency


Compiled languages “tend to be” the most energy-efficient and fastest-running.

...the five slowest languages were all interpreted: Lua, Python, Perl, Ruby and Typescript. And the five languages which consumed the most energy were also interpreted ones.


Paper

Who do you think are some of the smartest or coolest or awesome people in AI/ML?

I'll start Alec Radford (the real guy behind almost every OpenAI projects and amazing things)

NGL my only paid AI service is ChatGPT and it's been so worth it!

Sometimes all it takes is going to a conference and you will get it for free. Who hates free stuff, right.

🚨 Exciting News! 🚨

Join us in just ONE HOUR for an incredible episode of the Techኢት Podcast S02E07 🎙
🎉 We're featuring BEKA, a top developer turning unique ideas into reality! 🚀

Spread the word and share this with anyone you think will love it! 🌐✨

👉👉👉Join

@Techinethio

Join this podcast to see how boring @beka_cru is 😂😂😂


Jk, you will learn a lot

If you usually use generated images and you want to enhance their resolution use AuraSR-v2: a GAN-based Super-Resolution for upscaling generated images,

Try the model on hf spaces with your images for free: https://huggingface.co/spaces/gokaygokay/AuraSR-v2


Btw a little about GANs, they have of two neural networks, the generator(create data that is similar to the real data) and the discriminator(distinguish between real data (from the training set) and fake data (produced by the generator), which are trained together through a process of adversarial learning where the generator and discriminator are in sort of a competition.

Linear Algebra for Data Science

Covers a vector of topics from matrices, vector spaces, orthogonality and projections, singular value decomposition, determinants, eigen values and vectors, big theorems, to name a few!

https://kyunghyuncho.me/linear-algebra-for-data-science/

Let's talk about Sam😁

Introducing Meta Segment Anything Model 2 (SAM 2) — the first unified model for real-time, promptable object segmentation in images & videos.

SAM 2 is available today under Apache 2.0 so that anyone can use it to build their own experience.

I will try to test it on something very soon and will post the results here.

Source