Have you heard these exciting AI news? - July 26, 2024 AI Updates Weekly

Hi everybody, welcome to our weekly updates about AI. This week has been particularly exciting with numerous updates to cover.

Top Insight

There's a quote from Derek Austin stating that remote jobs receive thousands of applications each, making it really difficult to apply for jobs. He suggests filtering for recent job postings to have a higher chance of getting through the application process. You can find more insights in his article.

Major Event: Meta’s Release of Llama 31

The major event of this week is the release of Meta Llama 31, which is a set of three models: 8 billion, 70 billion, and 45 billion parameters. This happened on Tuesday. Just to remind you, Llama 2 was released in July 2023, which started a revolution in AI, leading to tens of thousands of derivative models.

Meta aims to be the universal platform for AI, providing free codes and weights for models, enabling fine-tuning and deployment. The largest model, 45B, is on par with GPT-4 and Claude 3.5, making them very competitive.

Model Capabilities

Multilingual: Unlike previous versions that were mostly English-based, these are truly multilingual.
Specialized Training: Specifically trained for coding, reasoning, and tool use.
Context Length: Extended to 128K tokens, allowing up to about 100 pages of text to be processed at once.

Meta has explicitly added to the license that you can create derivatives and deploy them anywhere. The pricing for these models is also very competitive, making them a substantial threat to current market leaders.

Hardware Requirements

Running these models requires significant hardware. For instance, running the largest 45B model with Nvidia H100s would need two nodes, each with eight GPUs. Alternatively, the AMD Instinct MI 300X would need just five GPUs.

Market Impact

Open-source models like the Llama 3.1 are expected to surpass closed-source models in performance and cost-effectiveness, leading to potentially significant market shifts. This is evident from various price comparisons:

Llama 3.1 45B: About $ 6 per million in-out tokens.
GPT-4: $ 15 per million in-out tokens.

The competitive edge in pricing likely means a financial hit for providers like OpenAI and Anthropic.

Additional Releases and Announcements

Nvidia and Mistral Collaboration: Released the Mistral Nemo instruct model with 12 billion parameters.
Stanford's Storm AI: Capable of writing Wikipedia posts.
MZero: A platform for maintaining memory across sessions.
UCLA: AI tool identifying prostate cancer with 84% accuracy.

Vision and Interview with Mark Zuckerberg

Mark Zuckerberg envisions billions of custom models and personalized AI agents powered by Meta's tools.

Noteworthy AI Development

Pandas AI: Improved data analytics using natural language.
Alpha Proof and Alpha Geometry: Achieving top scores in math Olympiads.
AI in Diagnostics: Multiple AI tools showing higher accuracy rates than human counterparts.

Conclusion

This was an exciting week in the AI world with major updates and new releases enriching the landscape. If you have any questions about these developments, feel free to reach out.

Keywords

Meta Llama 31
Open Source AI
Nvidia
Mistral
Multilingual Models
Hardware Requirements
AI Diagnostics
Mark Zuckerberg
AI Market Impact
Personalized AI Agents

FAQ

Q: What is Meta Llama 31? A: Meta Llama 31 is a set of three AI models released by Meta, with parameters of 8 billion, 70 billion, and 45 billion, designed for high scalability and specialized applications.

Q: How do these models compare with GPT-4? A: The largest model, 45B, is on par with GPT-4 and Claude 3.5 in terms of performance and capabilities.

Q: What are the hardware requirements for running these models? A: Running the 45B model requires significant hardware, such as two nodes with eight Nvidia H100 GPUs each, or five AMD Instinct MI 300X GPUs.

Q: What is the vision of Mark Zuckerberg for AI? A: Mark Zuckerberg's vision is to have billions of custom models and personalized AI agents powered by Meta's open-source tools and platforms.

Q: How does the new pricing of Meta Llama 31 compare to other models? A: Meta's models are significantly cheaper, with Llama 3.1 45B costing about $ 6 per million in-out tokens compared to GPT-4's $ 15 per million.

Q: What other significant AI updates were announced this week? A: Other updates include the release of the Mistral Nemo instruct model, Stanford's Storm AI, which writes Wikipedia posts, and AI diagnostic tools demonstrating high accuracy.