- Evolving AI Insights
- Posts
- 🚀 Meta releases Llama 3
🚀 Meta releases Llama 3
Also: Microsoft's crazy deepfake tool
Welcome, AI enthusiasts
As usual on Thursdays and Fridays AI companies compete for the biggest headlines and this week is no different. Meta has just released and open-sourced Llama 3, essentially giving away a $100 million product for free. Microsoft has revealed a tool that produces realistic deepfakes from just one photo and Boston Dynamics showcased their new all-electric Atlas robot which has taken the internet by storm. Let's dive in!
In today’s insights:
Meta Releases the most capable openly available LLM to date, Llama 3.
Microsoft's VASA-1 Produces a Deepfake from just one photo
Boston Dynamics' Atlas Robot Goes Electric
Read time: 4 minutes
🗞️ LATEST DEVELOPMENTS
Evolving AI: Meta just open-sourced Llama 3, the most capable openly available LLM to date.
Key Points:
Improved response diversity and reduced false refusals.
Better reasoning and programming capabilities
Outperforms competitors in key benchmarks
Details:
Meta today announced the first two models in the next generation of Llama. The series, called Meta Llama 3, initially includes pre-trained and instruction-tuned LLMs with 8 and 70 billion parameters, respectively. It excels in the MMLU (Massive Multitask Language Understanding) general knowledge benchmark, notably surpassing Google’s models Gemma and Gemini, among others. The improvement in answering variety and the capacity to write better code mark significant progress. Llama 3 has been pre-trained on over 15 trillion tokens from publicly available sources. The dataset is seven times larger than Llama 2, contains four times more code, and covers over 30 languages.
Why This Matters:
We're entering a new world where GPT-4-level models are open-source and freely accessible. Additional Llama 3 models with up to 400 billion parameters and new features such as multilingualism are under development. The models will soon be available on various cloud platforms and can be downloaded from Meta. Absolutely massive.
85% of all AI Projects Fail, but AE Studio Delivers
If you have a big idea and think AI should be part of it, meet AE.
We’re a development, data science and design studio working with founders and execs on custom software solutions. We turn AI/ML ideas into realities–from chatbots to NLP and more.
Tell us about your visionary concept or work challenge and we’ll make it real. The secret to our success is treating your project as if it were our own startup.
Evolving AI: Microsoft reveals VASA-1, a framework turning photos into talking, singing visuals.
Key Points:
VASA-1 animates still headshots with just an audio file.
Uses include ads and digital replicas but risks deepfake misuse.
Demonstrations include a talking Mona Lisa.
Details:
Microsoft's VASA-1 can animate a single photo into a video that mimics real human expressions and movements, using just an audio track. This model integrates complex facial dynamics and head movements to create videos that look and feel alive, pushing boundaries in AI content creation. However, Microsoft stresses this technology's demonstration status, highlighting its potential uses and the inherent risks of deepfake capabilities. The technology could revolutionize advertising, entertainment, and digital interactions, but it also necessitates careful ethical considerations.
Why It Matters:
Microsoft researchers consider VASA-1 a crucial development towards realistic digital AI avatars for numerous applications, but also caution about the potential for misuse. Thus, Microsoft will not release VASA-1, but intends to continue its improvement. Could this be the future of digital media, or a Pandora's box for misinformation?
Here’s a real-time demo of the tool in use:
VASA-1: Lifelike Audio-Driven Talking Faces Generated in Real Time from Microsoft. Start with any image, add any speech clip. That's all it takes. Reminds me of the Emote Portrait Alive paper from February. 2nd attempt at uploading some of these.
— Andrew Curran (@AndrewCurran_)
1:23 AM • Apr 18, 2024
BOSTON DYNAMICS
🤖 Boston Dynamics' Atlas Robot Goes Electric
Evolving AI: Boston Dynamics introduces a new electric version of its Atlas humanoid robot, marking a significant shift from its previous hydraulic models.
Key Points:
The new Atlas model features a quieter, more fluid motion compared to its predecessors.
It retains the "Atlas" name as it moves towards commercial use, breaking from Boston Dynamics' usual rebranding strategy.
Pilot testing is set to begin at Hyundai facilities next year, with broader production planned for the future.
Details:
Boston Dynamics has unveiled an all-electric version of its Atlas humanoid robot. This new version has smoother movements, a stark contrast to the hydraulic jerks of earlier versions. This design also sports a more friendly, cartoonish look aimed at commercial markets, which aligns it with competitors like Agility’s Digit and Apptronik’s Apollo. While the mechanical skeleton has been slimmed down, removing exposed cables and bulky hardware, it maintains the iconic traffic-light head of its predecessors.
Why This Matters:
In collaboration with Hyundai, Boston Dynamics aims to trial and refine the new Atlas through practical implementations and industrial challenges in the coming years. Atlas will be enhanced with AI features like reinforcement learning and computer vision, enabling it to better navigate and respond to intricate real-world scenarios. The objective is to surpass human abilities, achieving maximum task efficiency.
🔎 AI in research
Researchers at the University of Cambridge are using artificial intelligence to speed up the search for new treatments for Parkinson's disease, showing promising results in the journal Nature Chemical Biology.
Key Takeaways:
Machine learning quickly found five strong inhibitors of alpha-synuclein aggregation, important for tackling Parkinson's disease.
AI made the screening process 1000 times cheaper and 10 times faster, helping speed up the development of treatments.
With over six million people affected by Parkinson's worldwide and numbers expected to triple by 2040, faster treatment discovery is critical.
The AI method focuses on key protein aggregates, leading to more effective and efficient treatment approaches and quicker moves to clinical trials.
💡 Tip of the Day
With Boston Dynamics' new Atlas, the robotics company is moving closer to creating robots that can assist us in our everyday lives. The company has been on this quest for decades.
Here's an incredible video showcasing the 40-year development of their Atlas robot, from 1983 to 2023. Our personal favorite is 1999.
🎯 SNAPSHOTS
Direct links to relevant AI articles.
⚡️ Energy consumption: AI computing is on pace to consume more energy than India.
🤖 GPT-4: GPT-4 performed close to the level of expert doctors in eye assessments.
📲 Google: Google is combining its Android software and Pixel hardware divisions to more broadly integrate AI.
📈 Trending AI Tools
🎙️ Podsemble - AI powered tool for podcast production with research, transcription, content generation, and management (link)
📧 Varolio - AI-powered automation tool for email prioritization and centralized communication (link)
🖼️ Zocket - A generative AI platform to create and deploy social media ads (link)
🦉 Trademark Owl - AI trademark registration tool to minimize complications (link)
📸 Caption My Photos - AI-powered tool that automatically generates creative and personalized captions for photos (link)
🔎 Glimpse - AI powered tool designed to streamline online research (link)
🤝 Venturefy - A tool to verify corporate proof to increase trust with customers (link)
What'd you think of today's edition? |
Reply