Llama 3 Goes for the Gold, OpenAI Tests SearchGPT Prototype, Apple Commits to AI Safety ... and more

Welcome to AI Weekly Breakthroughs, a roundup of the news, technologies, and companies changing the way we work and live.

Llama 3.1 Goes for the Gold

Meta has introduced Llama 3.1, their most advanced AI model series to date, emphasizing open access to artificial intelligence. The flagship model, Llama 3.1 405B, is a groundbreaking open-source AI model with unmatched capabilities in general knowledge, multilingual translation, and tool use, designed to rival the best closed-source models. This release includes upgraded versions of the 8B and 70B models, all supporting a context length of 128K and multiple languages. Meta has enhanced its training infrastructure, utilized extensive datasets, and adopted innovative training techniques to achieve high performance and scalability. The Llama ecosystem is supported by major tech partners, and new tools like Llama Guard 3 and Prompt Guard aim to ensure responsible AI development. By making these models and their components openly available, Meta seeks to foster innovation and broaden the accessibility of AI technology globally.

OpenAI Tests SearchGPT Prototype

OpenAI is testing SearchGPT, a prototype of new search features designed to combine the strength of its AI models with information from the web to give users fast and timely answers with clear and relevant sources. It will be launched to a small group of users and publishers to get feedback. While this prototype is temporary, OpenAI plans to integrate the best of these features directly into ChatGPT in the future.

Microsoft Intros Bing Generative Search 

Microsoft has introduced Bing generative search, its counterpart to Google's AI-powered search, to a limited user base. This new feature leverages generative AI models to compile and present summaries of web-sourced information in response to search queries. Users can opt to view traditional search results if they prefer. Microsoft aims to enhance the search experience through gradual feedback and testing. However, concerns persist about AI-generated content accuracy and its potential to reduce traffic to source websites, an issue observed with similar AI features from other companies. Microsoft claims to be monitoring the impact on web traffic but has not provided detailed data.

OpenAI In Discussions with $700 Billion Chip Giant

OpenAI is in talks with Broadcom, a major chip manufacturer, to develop a new AI chip capable of challenging Nvidia's dominance in the AI hardware market. Broadcom, a $700 billion company, is larger than Intel, AMD, Arm, and Qualcomm combined, and is known for its wide range of semiconductor and infrastructure software solutions. The discussions are part of OpenAI’s broader strategy to reduce dependency on Nvidia, which has become the leading player in the AI revolution. This initiative might be connected to a data center project named "Stargate," potentially powered by AMD. Although it's uncertain if Broadcom's involvement is directly related to Stargate, OpenAI's commitment to developing its own hardware is evident, underscored by hiring former Google engineers with experience in Tensor processors. Despite this, any new chip developed is unlikely to rival Nvidia's offerings immediately due to the extensive R&D required.

US Lawmakers Request Government Access to OpenAI

Senate Democrats and an independent lawmaker have sent a letter to OpenAI CEO Sam Altman, questioning the company's safety standards and treatment of whistleblowers. Key inquiries include whether OpenAI will allow U.S. government agencies to test and review its next foundation model before deployment and whether the company will commit 20% of its computing resources to AI safety research. This scrutiny follows whistleblower allegations of inadequate safety measures for GPT-4 Omni and retaliatory actions against those raising concerns. The letter highlights growing regulatory pressures on OpenAI, evidenced by Microsoft and Apple stepping down from OpenAI’s board amid increased scrutiny. Former OpenAI employee William Saunders has also voiced fears that future iterations of OpenAI’s models could pose existential threats to humanity.

NIST Releases AI Model Risk Testing Tool

The National Institute of Standards and Technology (NIST) has re-released Dioptra, an open-source tool designed to test and measure the risks associated with AI models, particularly those subjected to malicious attacks like data poisoning. Initially launched in 2022, Dioptra allows companies and researchers to assess, analyze, and track AI risks by exposing models to simulated threats in a controlled environment. This tool aims to help evaluate the performance claims of AI developers and improve model safety. Dioptra's release coincides with documents from NIST's AI Safety Institute on mitigating AI dangers and follows the U.K. AI Safety Institute's similar toolset, Inspect. Although Dioptra offers valuable insights into potential attacks on AI systems, it currently only supports locally downloadable models and does not work with API-gated models like OpenAI's GPT-4o.

Apple Commits to AI Safety 

Apple has signed the White House’s voluntary commitment to developing safe, secure, and trustworthy AI, joining 15 other tech companies such as Amazon, Google, and Microsoft. The commitment involves rigorous testing of AI models before public release, maintaining the confidentiality of AI model weights, and developing systems to label AI-generated content. Apple plans to integrate its generative AI, Apple Intelligence, into its core products, impacting 2 billion users. This move, starting with embedding ChatGPT into the iPhone, signals Apple's intent to comply with regulatory expectations as AI regulations evolve. The Department of Commerce will soon release a report on open-source AI models, highlighting the growing regulatory focus on AI safety and accessibility.

You Can Now Fine-tune GPT4o-mini 

You can now customize GPT-4o mini for your application with fine-tuning. Available today to Tier 4 and 5 users, OpenAI plans to gradually expand access to all tiers. The first two million training tokens a day are free, through Sept 23.

Mistral AI Announces Mistral Large 2

Mistral AI has announced the release of Mistral Large 2, a significant upgrade to their flagship model, which excels in code generation, mathematics, reasoning, multilingual support, and function calling. With a 128k context window and 123 billion parameters, it supports numerous languages and over 80 coding languages. Mistral Large 2 is designed for single-node inference and offers enhanced accuracy, reliability, and cost efficiency. The model is available for research and non-commercial use under the Mistral Research License, with commercial use requiring a separate license. It outperforms previous models and competitors in various benchmarks and is now accessible via cloud platforms like Google Cloud, Azure AI, Amazon Bedrock, and IBM watsonx.ai.

Gemini Introduces Upgrades with Launch of 1.5 Flash

Gemini has introduced several upgrades with the launch of 1.5 Flash, enhancing its performance with faster and higher-quality responses. This update, available in over 40 languages and more than 230 countries, includes a quadrupled context window of 32K tokens, improving conversation length and complexity. New features include the ability to upload files for analysis and integrated related content links to provide deeper information. Gemini is also expanding its availability through Google Messages in the EEA, UK, and Switzerland, and the mobile app is reaching more countries. Additionally, Gemini for Teens is being rolled out globally with new safeguards and AI literacy resources to ensure safe and productive use.

Kling AI Video Generator Now Available Worldwide 

Kuaishou’s Kling AI has officially launched globally, expanding its availability beyond China and positioning itself as a strong competitor to OpenAI's Sora. Kling AI offers users 66 free daily credits for video generation and supports both text-to-video and image-plus-text-to-video formats. It generates videos at speeds of about 1 minute per second of footage and supports multiple resolutions and camera controls. The platform excels in producing realistic videos and has strong multilingual capabilities, though it struggles with certain styles and fine patterns. Future updates will include enhanced features like high-definition video and extended generation times. Kling AI also enforces content restrictions to prevent violent or inappropriate outputs.

Stability AI Introduces Stable Video 4D

Stable Video 4D is a groundbreaking AI model that transforms a single video into multiple dynamic viewpoints from eight different angles. This model, available on Hugging Face, can generate five frames per angle in about 40 seconds and provides a comprehensive 3D representation by allowing users to specify camera angles. It represents an advancement from image-to-video generation, moving into full 3D video synthesis. With potential applications in game development, video editing, and virtual reality, Stable Video 4D offers improved consistency and detail compared to previous methods. Currently in its research phase, it promises further enhancements and real-world applicability as development continues.

Cohere Launches Rerank 3 Nimble for Faster Reranking 

Cohere has launched Rerank 3 Nimble, an upgraded foundation model designed to enhance enterprise search and Retrieval-Augmented Generation (RAG) systems. Rerank 3 Nimble offers approximately three times the throughput and reduced latency compared to its predecessor, Rerank 3, making it ideal for high-volume workloads. It supports both English and over 100 other languages, handles long documents efficiently, and excels with multi-aspect and semistructured data. This model improves search relevancy and reduces processing costs in RAG systems by reranking documents before they are used for grounded generation.

Llama 3.1 405B Closes the Gap between Closed- and Open-sourced Models

This figure shows the performance disparity between closed-source and open-source models. The gap has been consistently closing since the start of 2023 and the latest open-source model Llama 3.1 can rival top performing closed-source models such as GPT4o and Sonnet 3.5.

AI Gets Silver Medal in International Mathematical Olympiad Problems

AI has made significant strides in mathematical reasoning with the introduction of AlphaProof and AlphaGeometry 2. These systems demonstrated impressive performance by solving four out of six problems from the International Mathematical Olympiad (IMO), achieving a silver-medal equivalent score of 28 out of 42 points. This marks a notable achievement in AI's ability to tackle advanced mathematical problems. Both systems represent a leap forward in AI’s ability to assist with mathematical proofs and problem-solving, with implications for future research and collaboration between mathematicians and AI tools.

AI Bot Calls Out Distracted Belgian Politicians 

Belgian artist Dries Depoorter has created "The Flemish Scrollers," a facial recognition system that monitors Flemish politicians' phone usage during parliamentary sessions. This system, developed in response to an incident where Flemish minister-president Jan Jambon was caught playing Angry Birds during a policy discussion, scans YouTube live streams to detect when politicians are distracted by their phones. Using Python and machine learning, it identifies the politicians and tags them on social media with reminders to stay focused. While the system aims to highlight the potential misuses of technology, some politicians argue that they use their devices for work-related tasks during sessions.

Bee AI raises $7M for its wearable AI assistant that learns from your conversations  

ZoomInfo alum raises $15M for startup that builds AI sales engineers

AI4 - Las Vegas - August 12 - 14 

The AI Conference 2024 - San Francisco - September 10 - 11 

World Summit AI - Amsterdam - October 9 - 10 

Gitex Global - Dubai - October 14 - 18 

Big Data Conference Europe - Vilnius - November 19 - 22