• AI Emergence
  • Posts
  • New, Open-source models from Mistral, Nvidia

New, Open-source models from Mistral, Nvidia

Along with: Google’s experiments & struggles and the chance to win our GenAI Course

Hey there, 

I love this boxing arena between Open Source and Proprietary models. After the SORA punch from OpenAI a couple of weeks ago, the last 2 weeks have been filled with new-exciting open source models. If you are new here, we are building Human Intelligence about Artificial Intelligence while AI is Emerging!

Open-source models made headlines this week, with Mistral Large rivaling GPT-4 in reasoning, Phind 70b, and Starcoder2 excelling in code generation.

Meanwhile, Gemini's struggles highlight that generative AI remains in its early stages, with much progress still needed.

What would be the format? Every week, we will break the newsletter into the following sections:

  • The Input - All about recent developments in AI

  • The Tools - Interesting finds and launches

  • The Algorithm - Resources for learning

  • The Output - Our reflection 

  • Question to ponder before we meet next!

Table of Contents

This week, Mistral announced its new flagship model, Mistral Large. This model could now be the second-ranked model after OpenAI's GPT-4, achieving impressive results on industry benchmarks like MMLU (common sense and reasoning) with an 81.2% score. 

What makes this significant is Mistral's partnership with Microsoft. Through this deal, Mistral's open and commercial language models will become available on Microsoft's Azure AI platform along with Mistral's own "la Plateforme."

Now - Microsoft is in partnership with the top 2 LLM providers! (source)

Phind 70B, another open-source LLM, is making a dent in the LLM landscape. It’s one of the best-performing code generation models, even outperforming GPT-4 Turbo with an 82.3% score on the HumanEval benchmark. Best of all, Phind 70B runs 4x faster than GPT-4!

Building upon the Code Llama 70B model, Phind 70B's impressive gains come from fine-tuning with an additional 50 billion tokens. It also supports a massive 32K token context window. (source)

ServiceNow, Hugging Face, and NVIDIA have announced StarCoder2, a new family of open-access large language models (LLMs) designed for code generation. 

This state-of-the-art model, trained in 619 programming languages, represents a significant leap forward in generative AI performance.

This collaborative effort enables the release of robust base models, providing the community with the tools to develop a diverse array of applications more efficiently. With a focus on data and training transparency, StarCoder2 facilitates innovation while maintaining ethical standards in AI development. (source)

Apple Inc. has ended its decade-long electric car initiative, Project Titan.

Team members from the discontinued Project Titan are being transitioned to Apple’s generative AI initiatives, highlighting the company's commitment to leveraging its strengths in the technology sector. (source)

Stability AI has unveiled an early preview of Stable Diffusion 3 - its most advanced text-to-image model to date. This new version offers major improvements in handling complex prompts, generating higher-quality images, and even improving spelling accuracy.

 With a flexible range of parameters (from 800M to 8B), Stable Diffusion 3 provides users with a variety of options for scalability and quality. The model blends a diffusion transformer architecture with flow matching for optimal results.

Stability AI has opened a waitlist ahead of the full launch. Joining the waitlist grants early access and lets users help shape the model's performance and safety. (source)

Figure AI, a company pioneering human-like robots, is attracting major investment from tech giants like Nvidia, Amazon, and Microsoft,

Already backed by OpenAI and other prominent names, Figure AI is now securing significant funding. Jeff Bezos, through Explore Investments LLC, leads the charge with a $100 million commitment. Microsoft follows closely with $95 million, while Nvidia and an Amazon-related fund each contribute $50 million. (source)

Google recently launched a private program offering select independent publishers beta access to an unreleased generative AI platform. This program helps under-resourced publishers expand their content output.

Participating publishers must use the tools to generate a set amount of content over 12 months. In exchange, they receive a monthly stipend (totaling a five-figure sum annually) and gain the ability to create audience-relevant content cost-effectively.

The beta tools allow publishers to efficiently index reports from sources like government agencies and other news outlets. They can then summarize and republish this information as new articles. (source)

Google's Genie: Turning prompts into playable games!

Google DeepMind recently unveiled Genie, AI model that can generate playable 2D platformer games based on various prompts.

Genie learned game dynamics through unsupervised training on a massive dataset of 2D platformer videos (over 200,000 hours). It can generate assets and predict how the game will respond to player input. Remarkably, Genie adapts to text, sketches, or images to create unique, playable experiences.

Though not yet publicly accessible, Genie showcases the immense potential of AI for content generation and interactive experiences. (source)

Last week, we touched upon the update that Reddit may be partnering with an unknown company. Turns out, It's Google! This partnership, reportedly valued at $60 million per year, grants Google access to Reddit’s data API, which delivers real-time content from Reddit’s platform.

This arrangement not only provides Google with an efficient and structured method to tap into Reddit's vast content repository for AI training but also opens up new avenues for showcasing Reddit content across Google’s products. (source)

AI companies crave the treasure trove of data held by content platforms.  A recent report suggests even Tumblr and WordPress could strike deals to share data with OpenAI and Midjourney for training purposes.

Security researchers have uncovered a critical vulnerability in Hugging Face's Safetensors conversion service. This flaw could allow attackers to hijack AI models, leading to system compromise and the potential spread of malicious code.

The vulnerability allowed attackers to embed malicious code within model conversion requests. If a user unknowingly processed this request, the attacker could gain control of the system. This access could enable the theft of sensitive data, access to private AI models, and even further the spread of malicious code.

Hugging Face has patched the vulnerability, emphasizing the importance of vigilance in the rapidly evolving AI landscape. (source)

Now, you can not only give feedback but also rate the answers of ChatGPT – and that too, directly to the builder.

The "About" section of GPT will include the builder's social profiles, ratings, categories, number of conversations, conversation starters, and other GPTs created by the builder. This is essentially to help train GPT to improve. (source)

Singapore is taking a commendable step by encouraging its citizens aged 40 and older to hit the books again, with a focus on staying abreast of AI advancements.

The government is footing the bill for this retraining initiative, demonstrating a savvy approach to governance in these rapidly changing times. (source)

We're shaking things up in the Tools section! This week, you'll put generative AI image tools to the test in a design contest.

This Week's Tool Challenge:

Design a New Podcast Logo for Leading With Data: This exclusive interview series, hosted by Kunal Jain, features top industry leaders and experts in Data Science, Machine Learning, and AI.

The Mission: Create an eye-catching new logo for the "Leading with Data" podcast. Use your favorite image generation tool, experiment with prompts, and submit:

  • The tool you used

  • Your crafting prompt

  • The awesome logo you generated

The Prize: The best logo wins a shoutout in next week's newsletter and free access to our Generative AI course!

Let's see your design skills shine!

  • DeepLearning.AI introduces "Prompt Engineering with Llama 2," a course on mastering Llama 2 AI models, developed in partnership with Meta. Learn the best practices for prompting and building applications with Llama 2.

  • A recent Berkeley Artificial Intelligence blog highlights the power of compound AI systems, which use multiple components instead of relying on a single model.

  • The article 'RAG is Dead, Long Live RAG!' on Vectorize delves into the intricacies of Retrieval-Augmented Generation (RAG) and the Gemini 1.5 AI model. It offers valuable insights into the advancements of new AI models while highlighting the enduring significance of RAG.

  • In the recent episode of Leading with Data, I engaged in a fascinating conversation with Harshad Khadilkar. We delved into his journey and discussed the intriguing experiments he's leading in reinforcement learning methodologies at Franklin Templeton, providing a unique perspective on the cutting-edge trends in AI.

But can open-source models truly challenge commercial giants? I expect a few more rounds of jabs and punches before we see a winner!

What do you think?

In 12 months, Which models would lead the LLM leaderboard?

Login or Subscribe to participate in polls.

What about 24 months? Which Models do you think would be leading the LLM leaderboard in 24 months?

Login or Subscribe to participate in polls.

If you have followed AI emergence and the way I have set the survey - it is not too difficult to guess what I think 🙂

How do you rate this issue of AI Emergence?

Would love to hear your thoughts

Login or Subscribe to participate in polls.

Reply

or to participate.