?? Summer of AI: ??1 Million for the Best AI! ?? GraphRAG, Mamba ?? vs Transformers ??, Moshi, strawberry ....??

July 22, 2024

Spanish

Welcome to this exciting edition of our featured news! Buckle up for a thrilling ride through the latest and greatest in the AI world, where groundbreaking advancements and fierce competitions await. ??

Subscribe to our newsletter for More Updates!

Don't miss our latest post in the AI for People of Letters (AI-for-All) block. It's currently free.

-?? ??1 Million Prize for the ARC-AGI Challenge!

Until mid-November, with $500,000 going to the first team that achieves an 85% resolution. If no winner emerges in three months, the prize increases. Get all the details at arcprize.org.

Key Insights for Programmers

For those of us who program, Keras (by Fran?ois Chollet) remains an essential Python library for creating deep learning models. Fran?ois Chollet, a Google engineer and researcher, has garnered attention with the ARC-Prize initiative, offering a significant reward for the AI that can master the ARC benchmark.

Understanding the ARC-Prize

The ARC-Prize consists of a series of logic tests designed to evaluate AI models' ability to avoid reliance on memory. Fran?ois Chollet, who launched this initiative, believes that current large language models (LLMs) rely too heavily on memorized templates, preventing them from generalizing beyond their training data. Unlike human intelligence, which combines reasoning and memory, LLMs are primarily interpolative. Chollet has also criticized OpenAI for not being transparent about its achievements and for promoting excessive enthusiasm around LLMs.

Ver charla completa de Fran?ois Chollet

-?? Our Medals of Honor for the French Mistral!

Our neighboring country is standing out with two impressive AI models:

• Codestral Mamba: Based on the Mamba2 architecture, this architecture is a strong alternative to Transformers. This model is available under the Apache 2.0 license, it excels in coding and advanced reasoning tasks. As an interesting anecdote, Mistral dedicates Codestral Mamba to Cleopatra: “whose glorious destiny ended in tragic circumstances with a snake...” ??
• Mistral NeMo: This state-of-the-art language model, developed in collaboration with NVIDIA, features 12 billion parameters and a context window of 128,000 tokens. It is multilingual and optimized with Tekken, and is now available under the free Apache 2.0 license.

1 . ???? COOKING UP NEW TECHNIQUES WITH OLD CONCEPTS:??

Innovative advancements in AI are transforming how we approach traditional concepts, leading to more efficient and accurate models. Here are some of the latest techniques:

• ?? Agent Mixing and Route LLM: While agent mixing enhances local responses, Route LLM goes a step further by classifying queries and directing them to the most suitable model. This approach not only improves the quality of answers but also reduces costs. Check out the implementation here: Route LLM
• ?? Microsoft Graph RAG: Microsoft Research has introduced Graph RAG, a technique that utilizes semantic search through graph architecture for text retrieval. It offers greater accuracy than traditional RAG but demands more computational resources and time. Learn more about it here.
• ??♀? Graph Reader: This improvement node exploration by incorporating an agent with a reasoning plan and a memory system to track relevant nodes. This method advances depth-first search efficiency. For further details, see the research paper.
• ?? LangGraph Cloud: This managed hosting service simplifies the deployment and scaling of applications built with the LangGraph framework. It enables developers to launch production-ready APIs with a single click. The beta version is available now, and you can learn more from Harrison Chase's discussion here.

While new techniques are emerging, others like Fine Tuning are being challenged. According to a paper, Fine Tuning would increase hallucinations linearly if the model lacks prior domain knowledge paper

2. Multimedia Models

??? Moshi: A Revolutionary Voice Model

French AI sensation Moshi has outpaced OpenAI with a groundbreaking model that generates realistic voices, expressing over 70 different emotions and speech styles. This tool is perfect for creating surreal and entertaining stories. Try it out and explore its capabilities here.

??? Web Sim Tool: Creating Websites from Descriptions

Experience the revolutionary concept of a generative and “dreamed” internet with Web Sim Tool. This innovative tool allows you to create complete websites and applications from a simple description or URL, using powerful models like Claude 3.5 Sonet, currently free to use. It simulates a browser interface, turning your prompt into a fully functional web or application with its own URL. Explore this new technology here.

??? Eleven Labs: Celebrity Voices for Digital Texts

Eleven Labs, pioneers in voice technology, have launched an application that allows users to listen to digital texts narrated by famous actors like Judy Garland, Burt Reynolds, and Laurence Olivier. This has been made possible through partnerships with the actors' heirs, though it has sparked some controversy. Check out the application here.

3. ???? BIG-TECH:

?? OPENAI

???? OpenAI's Path to AGI: 5 Strategic Levels

OpenAI has outlined a classification system to guide its progress toward Artificial General Intelligence (AGI), broken down into five strategic levels:

• ?? Level 1: Chatbots
• Level 2: Reasoners
• Level 3: Agents
• Level 4: Innovators
• Level 5: Organizations

Currently, OpenAI is at Level 1 with its chatbot technology.

?? Introducing GPT-4 Mini

OpenAI has launched GPT-4 Mini, a cost-effective AI model that surpasses GPT-3.5 Turbo in various tasks. This model aims to provide advanced AI capabilities at a lower cost. Learn more about it here.

?? Strategic Partnerships and Health Initiatives

OpenAI has announced a significant partnership with Alamos National Laboratory, known for its role in the development of the atomic bomb, to implement AI in bioscience. Additionally, OpenAI is entering the health market with initiatives to combat chronic diseases through personalized advice based on individual data. They are also collaborating with Color Genomics for early cancer detection.

?? Project Strawberry

OpenAI is working on a new technology called Strawberry, which aims to replace Q*. This technology focuses on post-workout scenarios and adapts base models to automate search, planning, reinforcement learning, and self-training. This development could be related to Stanford's Quiet-STaR paper, which you can read here.

There is speculation that models tested in the LMSYS Chatbot Arena, such as “upcoming-gpt-mini,” “column-u,” and “column-r,” may indicate the impending launch of a GPT-mini for mobile devices, similar to the pre-launch phase of GPT-4o. Check out the chat comparison models here.

OPENAI CONTROVERSIES

?? Cybersecurity Incident

Last year, OpenAI was targeted by a cybercriminal who managed to access an internal server. While no data was compromised, the company chose to handle the incident internally without notifying the FBI or relevant authorities. This decision has raised security concerns, highlighted by former employee Leopold Aschenbrenner.

?? Employee Complaints and SEC Investigation

Complaints from former employees have continued to emerge, leading to a request for an investigation by the US Securities and Exchange Commission (SEC). The complaints center around complex confidentiality agreements that allegedly prohibit communication with the SEC and require the waiver of incentives. OpenAI denies these allegations, asserting that its policies are designed to protect employees' rights.

?? Tensions with the European Union

OpenAI is also facing scrutiny from the European Union over its partnership with Microsoft. This comes amid a period of transition, with Microsoft's observers leaving OpenAI's board. Additionally, OpenAI's new partner, Apple, will not participate as an observer on the board.

?? META: Innovation's Papers and Controversies

3D Gen Meta: Revolutionizing 3D Image Generation

Meta has unveiled a new AI paper, 3D Gen, which can generate high-fidelity 3D images from text descriptions in less than a minute. This model utilizes Physically Based Rendering (PBR) for realistic re-illumination in real-world applications. Although it is not yet available to the public, its release is anticipated soon. For more details, check out the Meta 3D Gen publication.

Mobile LLM: AI on the Go

Meta is also releasing Mobile LLM, an architecture optimized for mobile devices. This advancement aims to bring sophisticated AI capabilities to smartphones and other mobile technology. Learn more about Mobile LLM in their research paper.

Potential Buyout and Regulatory Challenges

Meta is reportedly considering the acquisition of EssilorLuxottica, the parent company of Ray-Ban and Oakley eyewear. However, both the Wall Street Journal and the Financial Times have reported that Meta will not implement its generative models in Europe, citing the region's unpredictable regulatory environment. This decision affects the potential market for mobile-ready multimodal frames and smart glasses. Meta may face a significant fine, up to 10% of its annual turnover, for not adhering to user data protection regulations with their new consent model for free users.

Other News in Tech:

• Gemma2: Google has expanded its context window to 2 million tokens with their latest model, Gemma2. Privacy Issues: Google is under scrutiny for scanning Google Drive without users' permission, raising significant privacy concerns.
• Rabbit Code: A major data breach has been uncovered.
• Soju: A Breakthrough in AI Hardware. Soju, the first chip designed specifically for Transformer models, promises to be 20 times faster and 15% more efficient than current Nvidia GPUs. This innovation could significantly enhance the performance and efficiency of AI models.

Subscribe to our newsletter for More Updates!