Carlos Santana Vega

Just another AI influencer... 🥑🪑

[Carlos Santana Vega @TheDotCSV]

Wow! I wanted to test the new video outpainting feature from Runway, so I recorded a very close-up shot of my face to see if the AI could reconstruct it completely. It did! 🤯

Then I applied outpainting several more times, and the result is impressive!

[Carlos Santana Vega @TheDotCSV]

Here’s another important chart showing something very cool that has happened over the past year: Karpathy published a repository that replicated GPT-2 training and managed to train a similar model in 45 minutes using x8 NVIDIA H100 GPUs.

Since then, the community hasn’t stopped optimizing and improving the training process, bringing the times down to less than 3 MINUTES as of today!!!

Notice that the Y-axis for time is in a logarithmic scale.

[Carlos Santana @DotCSV]

¡FELIZ AÑO 2025 A TODOS!

El año que dejamos atrás ha sido un año espectacular de avances donde se han hecho los deberes y la IA ha madurando como esperábamos, con muy buenos productos e investigaciones, tanto en privado como open source.

¿Lo mejor? Al menos yo me voy tranquilo de haber encontrado las evidencias que necesitaba para creer que el progreso de las capacidades de la IA, por ahora, no tienen techo -I felt the AGI-

Entramos en unos años fascinantes donde la IA cada vez tendrá un impacto más palpable y disruptivo. Y ante la incertidumbre de cómo vivirlos de la mejor forma, lo mejor es continuar como llevamos haciendo desde hace ya 7 años, viviéndolo con pasión y fascinación!

En lo personal, un año donde he primado las cosas importantes de la vida por encima de llenarme de más y más trabajo. Fácil de escribir, difícil de ejecutar en el día a día. Muy orgulloso de ello :)

Nos vemos en DotCSV en 2025 🎇

[Carlos Santana Vega @TheDotCSV]

Here’s the original post, in case you want to know more details 👇https://t.co/Dg9sDyDgWt

[Jason Zada @jasonzada]

Fade Out, a short film made with Google’s Veo 2 (beta). It focuses on an aging rock star’s final years. Focused on film-like details—camera moves, lighting, textures, emotion—to craft a story, not just a montage.

[Carlos Santana Vega @TheDotCSV]

Fade Out is the name of the first audiovisual piece made with AI that I've seen, which is more "audiovisual piece" than "made with AI"... and that's important!

Made with clips from Veo 2 (Google) I think it's a good demonstration of where things are going

[Carlos Santana Vega @TheDotCSV]

When I saw Coca-Cola's AI ad, it seemed to me like it was situated in the uncanny valley, not yet crossing the threshold of being polished enough for a TV commercial—it was very AI-like

This ad is still uncanny, but I do think it crosses that threshold 👍

[Carlos Santana Vega @TheDotCSV]

DeepMind brings us another example of a useful tool for video editing! In this case, using AI to decompose a 2D video into its different layers, decoupling and masking elements from the bg and synthesizing what is hidden.

Very cool and practical 👇
https://t.co/WGm8nnbkns

[Carlos Santana Vega @TheDotCSV]

That said, if I crop the tiny bit visible on the sides from the original video and show the AI my face only through the glass, things get a lot more terrifying 😅

[Carlos Santana Vega @TheDotCSV]

I did another experiment with Runway's outpainting. I wanted to test how much it could reconstruct my face from a video with very little information.

The result is impressive once again! Just with the little that can be seen on the sides it managed to reconstruct my face 😮

[Carlos Santana Vega @TheDotCSV]

@koltregaskes Thank you Kol!

[Carlos Santana Vega @TheDotCSV]

Wow! I wanted to test the new video outpainting feature from Runway, so I recorded a very close-up shot of my face to see if the AI could reconstruct it completely. It did! 🤯

Then I applied outpainting several more times, and the result is impressive!

[Carlos Santana Vega @TheDotCSV]

This is the goal of world building!

[Martin Nebelong @MartinNebelong]

Imagine building worlds like this, fully immersed in the sounds of the jungle with everything responding to the things you do, the words you speak, the brushstrokes you make, the things you sculpt.

I've been dreaming about this for many years, and that's why I'm excited about generative Ai as part of our creative workflows.

[Carlos Santana Vega @TheDotCSV]

Here's another test, this one even more challenging! I orbited the camera around my face to see if the AI could differentiate between the rotation of my head and the rotation of the camera. It nailed it PERFECTLY!

cc @runwayml

[Carlos Santana Vega @TheDotCSV]

I mean, damn, the original video is this one!! And the AI reconstructs my face really well despite only seeing parts of it at different moments in time.

@runwayml check this!

[Carlos Santana Vega @TheDotCSV]

The wall is a lie https://t.co/lJYBIKcWOK

[Carlos Santana Vega @TheDotCSV]

The ChatGPT app adds a new functionality that allows it to interact with other apps to read their context. This means that if you're programming in VSCode, now ChatGPT can read your code from the app.

No control over the app yet, but it's a first step :)

[Carlos Santana Vega @TheDotCSV]

Company behind this is Deep Robotics
https://t.co/B4AktfVoE6

[Carlos Santana Vega @TheDotCSV]

Impressive the evolution that robotic mobility is undergoing!

[Carlos Santana Vega @TheDotCSV]

New technique -FasterCache- that speeds up the inference of video diffusion models, managing to cut times almost by half in models like Mochi 🔥

[Carlos Santana Vega @TheDotCSV]

🔴 ALPHAFOLD 3 RELEASED!

6 months after its announcement, the Alphafold team has released the code and models for academic use.

Great news! 👇

[Carlos Santana Vega @TheDotCSV]

Check all the information below 👇
https://t.co/pTQJb9hYBn

[Qwen @Alibaba_Qwen]

🚀Now it is the time, Nov. 11 10:24! The perfect time for our best coder model ever! Qwen2.5-Coder-32B-Instruct!

Wait wait... it's more than a big coder! It is a family of coder models! Besides the 32B coder, we have coders of 0.5B / 1.5B / 3B / 7B / 14B! As usual, we not only share base and instruct models, we also provide quantized models in the format of GPTQ, AWQ, as well as the popular GGUF! 💖

👉🏻Blog: https://t.co/7FnV3SUHuD

👉🏻Tech Report: https://t.co/Y3JN2Ly7H6

👉🏻Hugging Face: https://t.co/GgfeNq0XML

👉🏻ModelScope: https://t.co/VJwMAvEaHN

👉🏻Kaggle: https://t.co/7GW9GZJYre

👉🏻GitHub: https://t.co/gMGC8b5Hwv

👉🏻Demo [chat]: https://t.co/JxAYwnLM9u

👉🏻 Demo [Artifacts]: https://t.co/cyJEHV30e1

The flagship model, Qwen2.5-Coder-32B-Instruct, reaches top-tier performance, highly competitive (or even surpassing) proprietary models like GPT-4o, in a series of benchmark evaluation, including HumanEval, MBPP, LiveCodeBench, BigCodeBench, McEval, Aider, etc. It reaches 92.7 in HumanEval, 90.2 in MBPP, 31.4 in LiveCodeBench, 73.7 in Aider, 85.1 in Spider, and 68.9 in CodeArena!

[Carlos Santana Vega @TheDotCSV]

🔴 NEW PROGRAMMING MODELS!

Qwen continues to bring joy by releasing new models! Today's case features the Qwen2.5 Coder, which becomes the best model for programming tasks, on par with GPT-4 and Claude.

But it's open source, of course! :)