What is today’s beat?

RELEASES
🧨 DeepSeek & Silicon Valley’s take on it
🧨 Hugginface to rebuild Deepseek R1

NEWSROOM

🗞️ How Govt Could Reshape IP And Small Business In The US
🗞️ Meet the startup shaking up UK education AI policy: Faculty AI
🗞️ AI Maps Titan’s Clouds with Nvidia Tech
🗞️ Kodiak's Driverless Trucks Deliver for Atlas Energy

BUILDER BYTES
⭐️ Smol Agents Get a Vision Boost
⭐️ VideoLLaMA: Multimodal AI for Video Understanding
⭐️ LLAMA Stack: Meta's AI Dev Toolkit
⭐️ Llasa-3B: Speech-Aware Language Modeling

COMMUNITY
🤩 Google Deepmind PodCast- exploring project Astra
🤩 Big Technology Podcast with Google Deepmind CEO
🤩 StackOverflow: How the internet changed in 2024

FOLLOW US ON SOCIAL FOR MORE

Your FREE newsletter
 share or subscribe 
to show support

🧨 Deepseek R1 🧨

Why is Silicon Valley Excited, Upset, and Motivated?

DeepSeek's R1 model has been launched as a significant competitor to OpenAI's O1, leveraging reinforcement learning to achieve comparable performance at one-third the cost. This development highlights a turning point in the open-source versus proprietary AI landscape, emphasizing affordability and accessibility.

• Who is the company?
  Hong Kong-based quantitative analysis (quant) firm High-Flyer Capital Management started investing in Nvidia GPUs with its extra capital. It turned tech with investment to create their model!

• Resources to build
  Deepseek built Deepseek R1, a direct competitor to OpenAI’s o1 model, at 200x less compute budget for a total of $6M USD for its compute budget. This is what it takes to build a state-of-the-art world class model that is free for everyone to use (with daily cap on usage).

• Benchmarks
  Google Gemini is currently first, and Deepseek R1 and OpenAI o1 are tied for 3rd showing near similar results in all areas. Except Deepseek R1 is 98% cheaper to use than OpenAI o1. So, go ahead and cancel that OpenAI account of yours, and enjoy Deepseek for FREE for daily use (up to 1M tokens per day).

• Open Source
  Code, training data set, and guidelines to train it yourself make this an absolute smash hit across all industries. This truly democratizes AI with efforts like this.
  

DeepSeek R1 represents a critical shift in the AI industry, emphasizing cost-effective innovation and open-source collaboration. We’ll just finish off with a comment that Yann LeCun - the Chief AI Scientist for Meta - posted on on LinkedIn.

“To people who see the performance of DeepSeek and think: ‘China is surpassing the US in AI.’ You are reading this wrong. The correct reading is: ‘Open source models are surpassing proprietary ones.’

⭐️ SPECIAL HIGHLIGHT ⭐️

Huggingface will rebuild Deepseek R1

HuggingFace is redoing the training process to validate that Deepseek R1 is truly open source, and you can follow their progress here on Github.

🗞️ NEWSROOM 🗞️ 

What’s hot in tech right now?

How Govt Could Reshape IP And Small Business In The US
A CrunchBase article that examines how the Trump administration's IP policies shaped challenges and opportunities for small businesses, focusing on patent protections and innovation dynamics.

Meet the startup shaking up UK education AI policy: Faculty AI
It seems the UK is taking a different long-term approach (and slightly under the radar) with education. A look at AI policy innovation in UK education, tackling ethical concerns, equitable access, and AI-driven teaching tools to reshape learning experiences.

AI Maps Titan’s Clouds with Nvidia Tech
Explores how AI and deep learning models unravel the mysteries of Titan's hazy, methane-rich clouds using advanced GPU-powered simulations and data from NASA's Cassini mission.

Kodiak's Driverless Trucks Deliver for Atlas Energy
Covers Kodiak Robotics' milestone of completing its first driverless truck deliveries, marking a significant step in autonomous logistics for the energy sector.

⭐️ BUILDER BYTES ⭐️

What’s hot for builders right now?

Smol Agents Get a Vision Boost
Huggingface Blog explores how Smol Agents integrate vision capabilities, enabling multimodal tasks like image-based question answering and planning. Note, this is part of the world’s smallest large vision models released by HuggingFace.

VideoLLaMA: Multimodal AI for Video Understanding
An Alibaba based research introduces a multimodal AI model combining video and language understanding, enabling tasks like video-based Q&A, summarization, and captioning.

LLAMA Stack: Meta's AI Dev Toolkit
Meta research completes its first complete, comprehensive open source framework releases - Llama stack. View details on training, fine-tuning, and deploying large language models, emphasizing modularity and scalability.

Llasa-3B: Speech-Aware Language Modeling
Top Hong Kong researcher releases trained from scratch TTS model. Showcases a 3B parameter model optimized for integrating speech and text inputs, designed for speech-driven AI applications like transcription and voice assistants with code.

🤩 COMMUNITY 🤩

What’s the latest beat?

THANK YOU

Looking to promote your company, product, service, or event?