AI News May 2024: In-Depth and Concise

Welcome to The AI Track's comprehensive monthly roundup of the latest AI news!

Each month, we compile significant news, trends, and happenings in AI, providing detailed summaries with key points in bullet form for concise yet complete understanding.

This page features AI News for May 2024. At the end, you will find links to our archives for previous months

AI NEWS May 2024

[30 May] OpenAI introduces ChatGPT-EDU, a tailored version of ChatGPT for educational institutions.

OpenAI has launched ChatGPT Edu, an advanced AI tool designed to enhance higher education by providing powerful text and vision reasoning capabilities, data analysis, and enterprise-level security.

Key Points:

Capabilities: ChatGPT Edu, powered by GPT-4o, excels in text and vision reasoning, coding, and mathematics. It also offers advanced tools for data analysis, web browsing, and document summarization.
Implementations: Successful applications at universities like Columbia and Wharton, aiding in tasks from personalized tutoring to large dataset analysis and reflective learning.
Security and Accessibility: Provides enterprise-level security, supports over 50 languages, and includes robust administrative controls and higher message limits compared to the free version of ChatGPT.
Customization and Sharing: Allows universities to build and share custom GPT versions within their workspaces.
Scalability: Designed to make AI accessible and beneficial across various educational settings, promoting transformation in academic and operational frameworks.

[30 May] Google explains that the viral mistakes made by AI Overviews were due to misinterpreting queries, nuances in language, and data voids.

Google explains that the viral mistakes made by AI Overviews were due to misinterpreting queries, nuances in language, and data voids. The company emphasizes that AI Overviews are integrated with core web ranking systems to ensure accuracy and include relevant links. Google has implemented improvements to better detect nonsensical queries, limit satire and user-generated content, and enhance quality protections for news and health topics.

Key Points:

AI Overviews are designed to assist with complex queries and include relevant links.
Errors stem from query misinterpretations and data voids.
Google has implemented technical improvements and quality protections.

[30 May] AMD, Broadcom, Cisco, Google, Hewlett Packard Enterprise, Intel, Meta, and Microsoft (but not Nvidia …!!) have formed the Ultra Accelerator Link (UALink) Promoter Group to enhance AI connectivity in data centers.

AMD, Broadcom, Cisco, Google, Hewlett Packard Enterprise, Intel, Meta, and Microsoft have formed the Ultra Accelerator Link (UALink) Promoter Group to create an industry standard for high-speed, low-latency AI communication in data centers, enhancing scalability and performance for AI workloads.

Key Points:

UALink aims to establish open standards for AI accelerator interconnects.
The 1.0 specification will connect up to 1,024 accelerators in an AI computing pod.
This initiative supports AI/ML, HPC, and cloud applications.

[29 May] OpenAI has partnered with Vox Media and The Atlantic

OpenAI has entered licensing agreements with The Atlantic and Vox Media, allowing their content to train its AI models and be featured in ChatGPT, aiming to avoid copyright lawsuits and enhance attribution practices.

Key Points:

OpenAI is rapidly securing partnerships to license training data, preventing copyright issues.
The Atlantic and Vox Media announced their deals with OpenAI on May 29, 2024.
These deals follow similar agreements with major publishers like News Corp, Axel Springer, and The Associated Press.
Terms for The Atlantic and Vox Media were undisclosed.
Content from these publishers will receive attribution links in ChatGPT.
Vox Media will utilize OpenAI’s technology to enhance its affiliate commerce and ad data platforms.
The Atlantic is creating a microsite, Atlantic Labs, to develop AI tools for journalism.

[28 May] OpenAI has formed a new Safety & Security Committee

OpenAI has formed a Safety and Security Committee to enhance its safety protocols and oversee critical security decisions as it advances its AI capabilities.

Key Points:

The committee includes Bret Taylor (Chair), Adam D’Angelo, Nicole Seligman, and CEO Sam Altman.
The committee will evaluate OpenAI’s safety processes over the next 90 days and provide recommendations.
The initiative comes amidst the development of OpenAI’s next frontier model aimed at reaching AGI.
The committee also includes technical and policy experts like Aleksander Madry and Lilian Weng.
Recent departures of safety researchers highlighted concerns over the company’s safety culture.

[28 May] OpenAI has begun training GPT-5, the successor to GPT-4

OpenAI has announced the commencement of training for GPT-4’s successor, highlighting ongoing advancements in AI capabilities. This new model is expected to further enhance the performance and applications of AI, pushing the boundaries of what is achievable with language models.

Key Points:

OpenAI is training a successor to GPT-4.
Expectations are high for significant improvements and new capabilities.
The new model will likely address current limitations and expand AI applications.
This development marks another step in the rapid evolution of AI technology.

[26 May] Elon Musk's xAI Raises $6 Billion in Series B Funding Round

xAI, founded by Elon Musk, secured $6 billion in a Series B funding round to enhance its AI product offerings, infrastructure, and research, positioning itself competitively within the AI industry.

Key Points:

xAI announced a $6 billion Series B funding round on May 26, 2024.
Investors include Valor Equity Partners, Vy Capital, Andreessen Horowitz, Sequoia Capital, Fidelity, and Prince Alwaleed Bin Talal.
Funds will be used to bring xAI’s first products to market, build advanced infrastructure, and accelerate R&D.
xAI’s recent advancements include Grok-1, Grok-1.5 with long context capability, and Grok-1.5V with image understanding.
The company focuses on developing truthful and competent AI systems beneficial to humanity.
xAI continues to hire for various roles to support its mission.
The investment raised xAI’s valuation to $18 billion, with a pre-money valuation of $18 billion.
xAI’s chatbot, Grok, a competitor to OpenAI’s ChatGPT, has undergone several updates, including the introduction of Grok 1.5 with multimodal capabilities.
The funding positions xAI to compete with AI leaders like Microsoft and OpenAI.

[22 May] Microsoft Build 2024 emphasized Microsoft's strategic expansion of AI integration across its product portfolio

Microsoft Build 2024 demonstrated Microsoft’s extensive push to embed AI into its ecosystem, unveiling advanced AI features across various products, including new AI agents, enhanced developer tools, and innovative device integrations, all aimed at boosting productivity and user experience.

Key Points

AI Integration: Introduction of Copilot AI agents for automating tasks like email monitoring and data entry.
Phi-3-Vision: A multimodal AI model for mobile devices, capable of text and image analysis.
Edge Browser: AI-powered real-time video translation for platforms like YouTube.
Custom Emoji in Teams: Personalized emoji feature launching in July.
Snapdragon Dev Kit: High-spec development kit for Windows.
File Explorer Integration: Git support for tracking coding projects.
Advanced Paste in PowerToys: AI-powered clipboard features requiring an OpenAI API key.
Build Event Highlights: Announcements on AI-powered “Copilot Plus PCs,” Arm-powered devices with Prism for app compatibility, and new AI tools for developers.
Windows Volumetric Apps: Extending Windows apps into 3D space on Meta Quest headsets.
Partnerships: Collaboration with Khan Academy for AI-powered educational tools.
New Models: Availability of OpenAI’s GPT-4o on Azure and hints about upcoming GPT-5.
Developer Tools: Enhanced support for AI app development on Windows, new Copilot extensions, and improved Teams features for developers.

[22 May] News Corp has signed a licensing deal with OpenAI to allow the use of its content for AI training, marking a significant move in the media industry's adaptation to AI technologies.

News Corp has signed a licensing deal with OpenAI to allow the use of its content for AI training, marking a significant move in the media industry’s adaptation to AI technologies.

Key Points:

The deal includes content from The Wall Street Journal, New York Post, and The Times.
This partnership aims to protect against copyright issues while benefiting from AI advancements.
The agreement reflects a broader trend of media companies collaborating with AI firms.

[21 May] At the 2nd AI Safety Summit, 16 major AI companies, agreed to the "Frontier AI Safety Commitments"

At the second international AI Safety summit in Seoul, 16 major AI companies, including Amazon, Google, Microsoft, Meta, and OpenAI, agreed to the “Frontier AI Safety Commitments.” These commitments aim to ensure the safe development and deployment of AI models by setting and adhering to safety thresholds and being transparent about managing risks. This agreement marks a significant step towards consistent accountability and transparency in AI development globally.

[21 May] Anthropic has made significant progress in understanding the inner workings of large language models (LLMs)

Anthropic has made significant progress in understanding the inner workings of large language models (LLMs) by using dictionary learning to identify millions of features within their AI model, Claude Sonnet, enhancing AI safety and interpretability.

Key Points:

Core Discovery: Anthropic identified millions of features within Claude Sonnet, representing various concepts like cities, people, and abstract ideas such as bugs in code and gender bias.
Technique Used: The team applied dictionary learning to decode these features and understand their impact on the model’s behavior.
Manipulation and Implications: By amplifying or suppressing specific features, researchers could alter Claude’s responses, demonstrating causal relationships between neuron activations and AI behavior.
AI Safety: This research has significant implications for AI safety, providing tools to monitor and steer AI models toward safer outcomes by understanding and manipulating internal representations.
Potential for Ethical AI: Insights gained can help develop more reliable and secure AI systems, addressing concerns about bias, deception, and misuse.
Collaborative Effort: The research invites further collaboration to enhance AI interpretability and safety.

This work represents a major step towards transparent and trustworthy AI systems, providing a conceptual map of how LLMs process and represent information internally.

[20 May] Microsoft has announced the AI-infused Copilot+ PCs the most powerful Laptops ever built

AI Integration: Copilot+ PCs include a Neural Processing Unit (NPU) capable of 40 trillion operations per second, enabling advanced AI functionalities such as local AI assistant operations.
Hardware Specifications: These laptops must have at least 16GB of RAM, 256GB of storage, and the Snapdragon X Elite or X Plus processors.
Exclusive Features: Copilot+ PCs feature unique AI capabilities like Recall, Cocreate, and Live Captions, enhancing productivity and creativity. Recall offers a searchable “photographic memory” of past activities.
Manufacturer Collaboration: Available from multiple manufacturers including Acer, Asus, Dell, HP, Lenovo, Samsung, and Microsoft, starting at $999.
Battery Life: Devices offer up to 15 hours of web browsing and 20 hours of video playback.
Availability: Pre-orders are available with shipping starting on June 18th, 2024.
Performance Claims: Microsoft claims Copilot+ PCs outperform Apple’s MacBook Air by over 50% in sustained performance.
Future Prospects: Microsoft plans to include Intel and AMD processors in future Copilot+ PCs once they meet the required NPU performance benchmarks.

Around 700 to 800 staff, mainly Chinese engineers, have been offered transfers to countries like the U.S., Ireland, Australia, and New Zealand.

Microsoft confirmed the offer but did not specify the number of employees affected. This move aligns with U.S. efforts to limit China’s development of advanced AI technology for potential military use.

Despite this, Microsoft maintains its commitment to the region, employing approximately 7,000 engineers in its Asia-Pacific R&D group, primarily in China.

[16 May] Microsoft asks China-based AI and cloud employees to consider relocating due to U.S. tech restrictions

Microsoft has reportedly asked its China-based employees involved in cloud computing and artificial intelligence to consider relocating due to U.S. restrictions on China’s access to advanced technology.

Around 700 to 800 staff, mainly Chinese engineers, have been offered transfers to countries like the U.S., Ireland, Australia, and New Zealand.

Despite this, Microsoft maintains its commitment to the region, employing approximately 7,000 engineers in its Asia-Pacific R&D group, primarily in China.

[16 May] OpenAI partners with Reddit to enhance AI models with real-time content

OpenAI has entered into a partnership with Reddit to utilize its real-time, structured content for training its AI models, particularly ChatGPT. This deal, approved by OpenAI’s independent board, will also see the development of new AI-powered features for Reddit users and moderators.

Sam Altman, OpenAI’s CEO, has an 8.7% stake in Reddit but recused himself from the approval process. Reddit’s focus on data licensing as a growth strategy, following its public listing, includes agreements with multiple companies worth over $200 million.

This deal boosted Reddit’s stock by 11%. However, concerns from users about data monetization persist, similar to past protests on Stack

Overflow and Reddit’s opposition to Vana’s data DAO.

[16 May] OpenAI disbands its safety research team, raising concerns about the prioritization of product development over AI safety

OpenAI has disbanded its safety research team, a move that has sparked significant concern within the AI community. This decision comes just after the launch of GPT-4o, OpenAI’s most advanced generative model. The disbanding of the team, which was focused on developing controls to prevent “superintelligent” AI from going rogue, has led to the resignation of key figures including Jan Leike and co-founder Ilya Sutskever. This shift indicates a prioritization of product development over safety measures at OpenAI, raising questions about the company’s commitment to AI safety.

[14 May] Google's I/O 2024 emphasized AI's integration across its services, advancements in Gemini AI and introduced groundbreaking technologies like Project Astra for AR glasses.

Google’s I/O 2024 emphasized AI’s integration across its services, particularly with advancements in Gemini AI and introducing groundbreaking technologies like Project Astra for AR glasses.

Key Points:

Gemini 1.5 Flash and Pro: Introduced new Gemini models enhancing speed, efficiency, and capabilities, including a large context window for extensive document handling.
Project Astra: Introduced an AI that integrates with AR glasses to offer real-time, contextual interactions, aiming to transform everyday experiences with minimal response lag, positioning Google at the forefront of wearable AI technology.
AI Integration in Google Search: Expanded the Search Generative Experience to all U.S. users, incorporating AI-generated insights directly into search results.
Advancements in Gmail: Gemini AI is set to revolutionize how users interact with email through features like email summarization, enhanced reply suggestions, and in-depth search within email content, making email management more efficient.
Veo and Imagen 3: Announced new generative models for creating video and images, pushing the limits of AI in media creation.
Google Photos Upgrades: The introduction of ‘Ask Photos’ allows users to search for specific images within Google Photos using natural language queries, enhancing the usability of photo libraries through advanced AI search capabilities.
TalkBack Features: Google is enhancing its accessibility features with Gemini Nano, providing richer, clearer descriptions for images to aid blind and low-vision users, demonstrating the company’s focus on accessibility.
Music AI Sandbox: A new tool for musicians, the Music AI Sandbox allows the creation of music loops from AI prompts, reflecting Google’s ongoing innovation in creative AI fields.
Project Gameface on Android: Originally a hands-free gaming mouse, Project Gameface has been extended to Android, enabling users to control devices through facial gestures and head movements, with open-source code available for developers.

[14 May] U.S.-China AI Safety Talks Kick Off in Geneva

U.S. and China to Discuss AI Safety and Risk Management in Geneva amid their tight competition to dominate the AI landscape.

Key Points:

First Diplomatic AI Talks: The meeting marks the first under an intergovernmental dialogue agreed upon by President Joe Biden and President Xi Jinping in November.
Focus Areas: Discussions will center on AI risks, safety, and the development of trustworthy AI, with both nations emphasizing national security and economic growth.
Delegation Leads: The U.S. team is led by Tarun Chhabra and Seth Center, while China’s delegation is co-led by the Ministry of Foreign Affairs and National Development and Reform Commission.
Non-Negotiable Policies: Washington’s technology protection policies are not up for negotiation, focusing on mitigating AI risks without promoting technical collaboration.
Nuclear Weapon Control: The U.S. has urged China and Russia to commit to human-only decision-making for nuclear weapons deployment.
AI in Security: China’s rapid AI deployment across various sectors is a concern for U.S. security and its allies.
Guardrails for AI Models: The Biden administration plans to implement safeguards on U.S.-developed AI models to protect them from adversarial nations.
Global AI Standards: The U.S. and China are vying to shape global AI rules while exploring universally acceptable standards.
Legislative Actions: U.S. Senate Majority Leader Chuck Schumer plans to propose AI risk mitigation recommendations, influenced by competition with China.

[13 May] OpenAI has introduced GPT-4o, a new and improved AI model with enhanced voice and multimedia analysis features

OpenAI has introduced GPT-4o, a new and improved AI model offering faster and more accurate interactions, along with enhanced voice and multimedia analysis features, aimed at advancing human-machine interaction in a competitive AI landscape.

Key Points:

Introduction of GPT-4o: OpenAI launched its latest AI model, GPT-4o, which will be available to free users, enhancing speed and accuracy. The “o” stands for omni, integrating voice, text, and vision capabilities into a single model.
New Features: Updates include improved international language support, the ability to analyze images, audio, and text documents, and advanced voice capabilities. GPT-4o can handle real-time text, audio, and video inputs, responding to audio inputs in as little as 232 milliseconds, similar to human conversation speed.
Demonstrations: Live demos showcased the AI’s ability to create emotive voice responses, solve math problems via camera input, analyze facial expressions, and provide real-time translation between languages. Some glitches occurred, including audio cutouts and unexpected responses.
Voice and Interaction: GPT-4o offers real-time translations and the ability to interrupt the bot with new queries for dynamic interactions. It can respond to verbal questions with audio replies almost instantly.
Access and Availability: The new features will be gradually rolled out to free users, with certain capabilities already available to paying ChatGPT Plus and Team users. The GPT Store will be accessible to all users, offering custom chatbots previously limited to paid customers.
Safety Measures: OpenAI is rolling out features gradually to prevent misuse, particularly in voice and facial recognition. Extensive safety measures, including filtering training data and refining model behavior, are in place to ensure responsible use.
Competitive Landscape: OpenAI’s announcement coincides with growing competition from other AI firms like Google, Anthropic, and Cohere, who are also releasing advanced models. Google is expected to unveil more AI updates at their upcoming I/O conference.
Speculations and Future Plans: Speculation about OpenAI’s next moves includes the potential development of a search product to rival Google. CEO Sam Altman hinted at more forthcoming innovations but quelled rumors about an imminent launch of GPT-5.
Partnerships and Legal Issues: OpenAI is negotiating to integrate its AI with Apple’s iPhone OS and faces lawsuits over alleged copyright violations. The company continues to expand its product offerings and develop new AI technologies.
Model Capabilities: GPT-4o matches GPT-4 Turbo in text and coding performance, with significant improvements in non-English languages, vision, and audio understanding. It offers a 50% reduction in API costs and is twice as fast, making it a more efficient and cost-effective option.

[13 May] IBM Makes Its Mark in AI Competition by releasing its open-source Granite LLMs

IBM has made significant strides in the field of AI by open-sourcing its Granite AI models, aiming to empower developers and businesses by providing accessible, high-performance tools for software creation and other programming tasks.

Key Points:

IBM has open-sourced its Granite AI models, making them available under the Apache 2.0 license for both research and commercial use, thereby encouraging broader innovation in AI applications.

The Granite models are specialized for programming, trained on extensive code data, supporting application modernization and development across various programming languages.

IBM’s strategy focuses on specific business use cases rather than general applications, highlighting the models’ suitability for enterprise environments.

IBM’s specific business use cases for their AI models primarily target software development within enterprise environments. This includes automating tasks like generating unit tests, writing documentation, running vulnerability tests, and modernizing applications, particularly for complex programming languages such as COBOL. These AI models are designed to improve efficiency and accuracy in these specific, technical aspects of enterprise software development.

Available tools like InstructLab and Ollma facilitate the deployment of these models, lowering the entry barriers for developers and enhancing the usability of AI in software development.

Tools like InstructLab and Ollma make it easier for people to use IBM’s AI models. These tools help developers set up and use AI without needing a lot of tech skills or resources, making it simpler for them to create software with AI.

IBM ensures the models are responsibly created, with data cleaned of inappropriate content, making them safe and reliable for professional use.

[8 May] AlphaFold 3 by Google’s DeepMind: A New Era in Biomolecular Modeling and Drug Discovery

AlphaFold 3, developed by Google DeepMind, introduces groundbreaking advancements in biomolecular structure prediction, poised to transform drug discovery and biological research, potentially creating a multi-billion-dollar business landscape.

Key Points:

Revolutionary Model Advancements: AlphaFold 3, launched by Google DeepMind, significantly enhances the predictive modeling of complex biological structures, such as proteins, DNA, RNA, and their interactions, offering unparalleled accuracy and broadening its application scope beyond its previous iterations.
Potential Commercial Impact: Demis Hassabis, CEO of DeepMind, projects the AI-driven drug discovery market to exceed $100 billion, with AlphaFold 3 positioned as a central technology. Isomorphic Labs, an Alphabet subsidiary, aims to capitalize on this by applying AlphaFold to real-world drug design challenges.
Strategic Partnerships: Isomorphic Labs has forged partnerships with pharmaceutical giants like Eli Lilly & Co. and Novartis, potentially valued near $3 billion, highlighting the commercial viability and high stakes of integrating AI in drug development.
Innovative Platform Deployment: The AlphaFold Server, launched alongside AlphaFold 3, offers an accessible platform for scientists globally to utilize the advanced capabilities of AlphaFold 3 for non-commercial purposes, promoting widespread adoption and innovation in the scientific community.
Enhanced Accessibility and Utility: The server simplifies the interaction with complex AI models, allowing biologists and researchers easy access without needing deep technical knowledge, thus democratizing high-level scientific inquiry.
Future Prospects and Challenges: While the technology heralds a new era in biotechnology and AI, the industry faces challenges like proving clinical success, navigating regulatory landscapes, and addressing public and market expectations.

This synthesis incorporates insights from the recent Bloomberg interview and article on DeepMind’s strategic advancements with AlphaFold 3, underscoring its potential to reshape the future of drug discovery and biological research significantly.

[7 May] Microsoft has developed a groundbreaking generative AI model, specifically for US intelligence agencies

Microsoft has developed a groundbreaking generative AI model, based on GPT4, specifically for US intelligence agencies. This AI model operates in an isolated, air-gapped cloud environment, ensuring secure processing and analysis of highly sensitive information without the risk of data leaks or cyber-attacks. The system is designed to handle classified data without learning from it, limiting access to approximately 10,000 authorized personnel and providing functionalities such as answering questions and writing code for intelligence operations

[6 May] Apple working on AI chips for data centers

Apple has been actively developing its own AI chips for data centers, known as Project ACDC, with no set timeline for completion. This initiative is part of Apple’s broader strategy to catch up in the AI sector, especially in AI inference for server chips, an area currently dominated by Nvidia. Despite the collaboration with Taiwan Semiconductor Manufacturing Co., the outcomes of these efforts remain uncertain. Apple’s CEO, Tim Cook, has emphasized significant investments in AI, with more details expected at the upcoming Worldwide Developers Conference in June

[6 May] OpenAI has formed a partnership with Stack Overflow

OpenAI and Stack Overflow have partnered to provide developers with access to Stack Overflow’s OverflowAPI, enhancing AI models and developer experiences on both platforms.

Summary

OpenAI and Stack Overflow announced a new API partnership aimed at empowering developers with access to highly technical content and AI development models.
The partnership involves OpenAI utilizing Stack Overflow’s OverflowAPI product to enhance AI model performance, with a focus on improving content and receiving feedback from the Stack Overflow community.
Stack Overflow will incorporate OpenAI models into their OverflowAI development, utilizing insights from internal testing to maximize model performance.
The collaboration aims to deepen engagement with content and foster efficiency and collaboration within the developer community.
Both companies emphasize the importance of serving diverse communities and striving for socially responsible AI.
The first set of integrations and capabilities resulting from the partnership will be available in the first half of 2024.
The partnership enables Stack Overflow to reinvest in community-driven features and sets new standards for vetted, trusted, and accurate data in technology solutions.

[6 May] Microsoft is reportedly developing a language model named MAI-1, with aprox. 500 billion parameters, positioning it between GPT-3 and ChatGPT-4

Microsoft is reportedly developing a language model named MAI-1, led by Mustafa Suleyman, with aprox. 500 billion parameters, positioning it between GPT-3 and ChatGPT-4, expected to be debuted at the Build developer conference.

Key Points

MAI-1 is being developed in-house by Microsoft, marking the first time the company has ventured into such a project since investing in OpenAI.
MAI-1 is reported to have around 500 billion parameters, making it significantly larger than Microsoft’s previous models like Phi-3, but still smaller than OpenAI’s GPT-4.
Mustafa Suleyman, former Google AI leader and CEO of Inflection AI, is leading the development of MAI-1 after Microsoft acquired a majority of Inflection’s staff and IP for $650 million.
Microsoft’s approach to AI involves both small locally run language models for mobile devices and larger state-of-the-art models powered by the cloud, similar to Apple’s strategy.
The model’s training process involves a “large cluster of servers” equipped with Nvidia Corp. graphics cards and compiling training data from various sources.
Microsoft might utilize training data and assets from Inflection AI to power MAI-1, including text generated by GPT-4 and web content.
The exact purpose of MAI-1 within Microsoft is not yet determined, and its ideal use will depend on its performance.
MAI-1’s complexity suggests it’s unlikely to run on consumer devices, indicating Microsoft will deploy it in its data centers for integration into services like Bing and Azure.
MAI-1 is expected to make its debut as early as this month, possibly during the Build developer conference starting on May 16.