AI Revolution: Meta's Llama 3.1, Open Source Models, and Video Breakthroughs

Meta Releases Llama 3.1

Meta has unveiled Llama 3.1, an upgraded version of their open-source language model. This new release comes in three sizes:

• 8 billion parameters
• 70 billion parameters
• 405 billion parameters

The larger models generally offer improved reasoning and math capabilities. Some key features of Llama 3.1 include:

• Tool use
• Multilingual agents
• Enhanced complex reasoning
• Improved coding abilities

Benchmark tests show Llama 3.1 outperforming other state-of-the-art models like GPT-4 and Claude 3.5 in many areas. The 8 billion parameter version surpasses other models of similar size across most benchmarks.

What makes Llama 3.1 particularly powerful is its open-source nature. Developers can download, customize, and fine-tune the models for specific applications. This gives the open-source AI community a model comparable to closed-source options like GPT-4, but with more flexibility.

There is one caveat to the open-source license - organizations with over 700 million monthly active users must request a special license from Meta.

Accessing Llama 3.1

While the full 405 billion parameter model requires significant computing resources, there are several ways to access and use Llama 3.1:

• Meta AI chatbot
• WhatsApp
• Instagram Messenger
• Facebook Messenger
• Meta Ray-Ban smart glasses (coming soon)

The Meta AI chatbot confirms it is running the 70 billion parameter version of Llama 3.1.

Groq has also integrated Llama 3.1 into their platform, offering fast inference for the 8B, 70B, and 405B versions. The 405B model is currently limited to enterprise users on Groq.

Perplexity AI has added the 405B Llama 3.1 model as an option for Pro subscribers, allowing direct comparison with other top models like Claude 3.5.

Mistral AI Releases Mistral Large

Competing with Meta in the open-source model space, Mistral AI launched Mistral Large, a 123 billion parameter model. Mistral claims this model outperforms Llama 3.1 70B in math performance.

Benchmark comparisons show Mistral Large 2:

• Slightly underperforming GPT-4 on coding tasks
• Outperforming Claude 3.5 and Llama 3.1 405B on coding
• Matching or exceeding other top models on various benchmarks

Mistral Large 2 showed particular strength in code generation across multiple programming languages.

The emergence of powerful open-source models like Llama 3.1 and Mistral Large represents a significant moment for AI development. These models can be freely modified and optimized, potentially leading to rapid advancements.

Apple Enters the Open-Source AI Arena

Apple has also joined the open-source AI movement, releasing smaller 7 billion and 1.4 billion parameter models. Their new model outperformed Mistral 7B and claims to be approaching the capabilities of Llama 3 and Google's Gemma (though this was before Llama 3.1's release).

Google Upgrades Gemini

Google announced upgrades to their Gemini model:

• Free tier users now have access to Gemini 1.5 Flash
• Improved quality and latency, especially for reasoning and image understanding
• Expanded token limit to 32,000 for free users
• Upcoming ability to upload files via Google Drive for context
• Display of source links for fact-checking
• Integration with Google Messages on select Android devices

OpenAI Announcements

OpenAI made several notable announcements:

1. GPT-4 fine-tuning is now available, with up to 2 million free training tokens per day through September 23rd.

2. Introduction of Search GPT, a new AI-powered search prototype. Features include:

   • AI-generated answers with sources
   • Image integration
   • Traditional search results alongside AI responses

3. Voice feature alpha rollout for ChatGPT Plus subscribers starting next week.

Anthropic Faces Scraping Controversy

Anthropic, the company behind Claude AI, has come under fire for aggressive web scraping practices. Reports indicate their bots are scraping websites at extremely high rates, even when explicitly asked not to in robots.txt files and terms of service.

This highlights the ongoing debate about data collection practices for AI training, which is likely to intensify in the coming months and years.

Elon Musk's AI Predictions

Elon Musk made several bold claims about upcoming AI developments:

1. Grok 2.0 is coming soon, supposedly on par with GPT-4 or Claude 3.5.
2. Grok 3.0, claimed to be "the most powerful AI in the world," is expected by December.
3. xAI (Musk's AI company) has started training on what he calls "the most powerful AI training cluster in the world" - 100,000 liquid-cooled H100 GPUs.
4. Tesla plans to have "genuinely useful humanoid robots" in low production for internal use next year, with high production for other companies by 2026.

Bing AI Redesign

Microsoft's Bing search engine is testing a new AI-integrated design:

• AI-generated answers appear on the left side of the results page
• Traditional search results are moved to a right sidebar
• Sources for AI responses are clearly displayed

The rollout appears to be gradual, with some users still seeing the old design.

AI Video Generation Advancements

Luma AI's "Loop" Feature

Luma AI has introduced a new "Loop" feature in their Dream Machine tool. This allows users to create infinitely looping animations from still images or short clips. Examples include:

• Flaming loops
• Spaceships perpetually flying
• Animals in continuous motion
• Abstract, colorful swirling backgrounds

The feature works particularly well when starting with an existing image rather than pure text-to-video generation.

Cing AI Improvements

Cing, considered by many to be the best text-to-video generator currently available, has made some updates:

• Easier access: Users can now register with just an email, removing the need for phone verification
• Free daily credits: Users receive 66 credits every 24 hours, enough for about 6 video generations
• Advanced options: Camera movement controls, negative prompts, quality settings, and aspect ratio choices

Cing's output quality is approaching what we've seen from more advanced models like Sora, especially for text-to-video generation.

Runway's Training Data Controversy

A report from 404 Media suggests that Runway, another popular AI video tool, may have trained on thousands of YouTube videos without permission. While not officially confirmed, an anonymous source provided a spreadsheet of YouTube channels allegedly used for training data.

This raises complex questions about fair use, copyright, and the ethics of AI training data collection. The situation is further complicated by issues like:

• Commentary channels using copyrighted footage
• The transformative nature of AI-generated content
• Potential harm to content creators

A Twitter poll on the subject showed mixed opinions, with many people conflicted about the ethics of using publicly available content for AI training.

Stability AI's Stable Video 4D

Stability AI has released Stable Video 4D, a model that can transform a single object video into multiple novel view videos from 8 different angles. This allows for the creation of 3D-like representations from 2D video input.

The model is available on Hugging Face, though easy-to-use online interfaces are not yet widely available.

Adobe's AI Integration

Adobe has introduced new AI features in Illustrator and Photoshop:

• AI-powered pattern and texture generation based on text prompts
• The ability to fill shapes with AI-generated designs
• Automatic pattern extension and repetition

These tools aim to speed up the creative process for designers and illustrators.

Leonardo AI Teams Feature

Leonardo, an AI image generation platform, has introduced a new Teams feature allowing for collaborative work. Key aspects include:

• Shared team collections
• Consistent outputs across team members
• Fine-tuned models for specific project needs
• Shared team feed

This feature is particularly useful for game development and other collaborative creative projects.

Sakana AI's Ukiyo-e Generator

Sakana AI, founded by ex-Google employees, has released an AI model specifically designed to generate traditional Japanese ukiyo-e artwork. The model, based on a fine-tuned version of Stable Diffusion XL, is available on Hugging Face.

Suno AI Music Generation Updates

Suno, an AI music generation company, has introduced a new "Stems Pro" feature. This allows users to separate vocals and instrumentals from generated songs, providing greater control and creative flexibility.

AI in Game Development

The upcoming NCAA college football video game is using AI to rapidly scan and integrate player likenesses. This demonstrates how AI can significantly speed up game development processes, especially for games with large rosters of real-world athletes.

Conclusion

The past week has seen a flurry of AI advancements across multiple domains. From powerful new language models to innovative video and image generation techniques, the AI landscape continues to evolve at a rapid pace. As these tools become more accessible and powerful, we can expect to see even more creative applications and potential disruptions across various industries.

However, these developments also bring important ethical and legal questions to the forefront, particularly regarding data collection and copyright issues. As AI capabilities grow, so too does the need for thoughtful consideration of its implications and responsible development practices.

Article created from: https://youtu.be/ULO-jshyOeY?feature=shared