Welcome back!

Open-source AI is not slowing down. A new player has taken the spotlight for coding tasks: DeepSeek Coder V2. It outperforms GPT-4-Turbo in coding at a fraction of the cost, making cutting-edge coding AI more accessible. Let’s unpack this...

Today’s Summary:

• Open-source AI tops GPT-4-Turbo for coding

• Runway latest video generator, Gen-3 Alpha

• DeepMind V2A for soundtracks and dialogue

• Brave browser Leo assistant

• Nvidia Nemotron-4

• TikTok avatars in ads

• Meta pauses AI training in EU

• 2 new tools

TOP STORY

New Open-Source Model Beats GPT-4-Turbo in Coding

The Summary: DeepSeek-Coder-V2, a new open-source language model by a Chinese startup, outperforms GPT-4-Turbo in coding tasks according to several benchmarks. It specializes in generating, completing, and fixing code across many programming languages, and shows strong mathematical reasoning skills. It offers these capabilities at a lower cost compared to the GPT-4-Turbo API.

Key details:

• Supports 338 programming languages and 128K context length

• Released in two versions: 16B and 230B parameters

• The 230B version outperforms GPT-4-Turbo, Claude-3, and Gemini-1.5 Pro in coding and math benchmarks

• Tops leaderboards like Arena-Hard-Auto

• Free model downloads and low-cost API access

Why it matters: The release of DeepSeek Coder V2 is a new milestone for open-source AI. It offers top coding performance at a fraction of the cost of commercial models, democratizing access to advanced AI coding tools for developers and researchers.

TOOLS

Runway Gen-3 Video Generator Offers Improved Controls

The Summary: Runway has announced Gen-3 Alpha, its latest AI model for generating realistic 10-second video clips with high detail and control. Trained on a large multimodal dataset, the model can render human figures with natural movements and expressions, and allows precise control over scene composition, camera motion, and styles.

Key details:

• Gen-3 Alpha can generate 10-second high-definition video clips with photorealistic humans, animals, objects

• Provides precise control over transitions and character movements through descriptive captions

• Designed for cinematic storytelling, with the ability to specify camera angles, lighting, styles

• "Gen-3 Alpha will be available for everyone over the coming days"

Why it matters: Video generation models are advancing rapidly. Runway increased control and photorealism make it a powerful tool for video production pipelines.

RESEARCH

DeepMind V2A Generates Video Soundtracks and Dialogue

The Summary: Google DeepMind has developed a new video-to-audio (V2A) model that uses video input and text prompts to create soundtracks synced with on-screen action, including music, sound effects, and dialogue. V2A allows users to guide the audio output through positive and negative prompts for creative control.

Key details:

• Creates music, sound effects, and dialogue for videos or archival footage

• Trained on video, audio, and annotations

• Uses a diffusion model that refines audio from random noise into an audio waveform matching the visuals

• Includes SynthID watermarking for transparency and safety

• Will be tested for safety before public release

Why it matters: Google DeepMind's V2A technology can bring silent videos to life with realistic soundtracks, opening new creative possibilities for filmmakers, animators, and content creators. The watermarking feature promotes responsible use and transparency.

QUICK NEWS

Quick news

• Brave browser’s private Leo assistant

• Nvidia releases Nemotron-4 340B to generate synthetic data for training other models

• TikTok ads may soon contain AI avatars of creators

• Meta pauses Meta AI in EU over privacy concerns

TOOLS

🥇 New tools

• Mars5 - High quality text-to-speech, open-source

• Olvy - Customer feedback analysis assisted with AI

• Leo - AI assistant built into your browser