OpenAI starts acquiring, Apple continues partnership discussions, while Amazon jumps into the race

Amazon Enters AI Assistant Race with 'Metis' - Leveraging it's Olympus Model to Challenge ChatGPT

Amazon is secretly working on a ChatGPT killer - Amazon is reportedly developing an AI chatbot named "Metis," named after the Greek goddess of wisdom. This project aims to rival OpenAI's ChatGPT and is based on Amazon's proprietary AI model called Olympus, a more potent version of its Titan model. Metis is designed to deliver text and image responses with conversational intelligence and can provide source links, suggest follow-up questions, and generate images. Metis is expected to offer more timely information by incorporating retrieval-augmented generation techniques, potentially including capabilities such as automating complex tasks. The AI market already features significant players like Microsoft, Google, and various startups, with Amazon seeking to enhance its position despite reported hurdles with AI chip demand and performance. CEO Andy Jassy has a direct interest in the project, highlighting its significance to Amazon's AI initiatives. The Metis development team includes members from the Alexa AI team and employs some of Alexa's advanced "Remarkable Alexa" technology. Metis is part of Amazon's AGI team, with tentative plans to launch around an annual Amazon event in September, though it's noted that Amazon might already be playing catch-up in the AI-chatbot field.

• OpenAI Buys a Remote Collaboration Platform - OpenAI has acquired Multi, a startup focused on video-based collaboration for remote teams, marking another strategic move towards enterprise solutions following their recent purchase of Rockset. The acquisition, primarily an acqui-hire, will see most of Multi’s team joining OpenAI, with Multi set to shut down after July 24. This aligns with OpenAI's broader strategy of enhancing its enterprise offerings, evidenced by their corporate partnerships and the success of their business-oriented AI tools, contributing to projected revenues of $3.4 billion this year.

• OpenAI’s first acquisition is an enterprise data startup - OpenAI recently acquired Rockset, an enterprise analytics startup, marking its first purchase involving both technology and team integration. Rockset, having raised $105 million in funding, will support OpenAI's retrieval infrastructure for its products. While OpenAI COO Brad Lightcap highlighted the potential for transforming data into actionable intelligence, Rockset CEO Venkat Venkataramani assured their technology would tackle database challenges at large scale for AI applications. The terms of the deal were undisclosed. Implementation will be gradual, ensuring Rockset's existing customers face no immediate service changes, and only some of Rockset's team will transition to OpenAI.

• Google Develops Challenger to Meta’s Chatbots and Character.AI - Google is developing customizable chatbots that can be modeled on celebrities or created by users, aiming to compete with Meta and Character.AI's similar offerings. These new chatbots, powered by Google's Gemini AI model, will enable users to define their personality and appearance, and may feature partnerships with influencers. The project, led by Ryan Germick and part of Google Labs, highlights Google's effort to leverage generative AI for entertainment while exploring potential integration with YouTube. This initiative reflects Google's push to keep pace with rivals in the rapidly evolving AI landscape.

• OpenAI says GPT-5 will have 'Ph.D.-level' intelligence - OpenAI CTO Mira Murati indicates that the next iteration of their AI, potentially GPT-5, slated for release in approximately a year and a half, could exhibit 'Ph.D.-level' intelligence in specific tasks, marking a significant evolution from GPT-4. However, the timeline has been pushed back from initial rumors suggesting a late 2023 release. Murati clarifies that while these systems may reach human-level proficiency in certain areas, they will not universally achieve such intelligence. This update comes amidst a backdrop of computing and technology recommendations, including discussions of budget-friendly gaming PCs like the HP Omen 40L and Dell laptop selections.

  

• Figure Ships First Fleet of Humanoid Robots to Client - Figure has announced the shipment of its first small fleet of humanoid robots to a client, marking the start of a significant AI data collection effort to train its neural networks. Brett Adcock, co-founder of Figure, emphasized that having an external client increases the pressure to deliver economically viable humanoid robots. The deployment aims to enhance the robots' capabilities and demonstrate their utility in real-world applications.

  

• Apple, Microsoft Shrink AI Models to Improve Them - Tech giants, previously focused on large language models (LLMs) such as Meta's 400-billion-parameter Llama 3, are now exploring small language models (SLMs), including Apple's “Apple Intelligence” models (3 billion parameters) and Microsoft's Phi-3 family (3.8-14 billion parameters). These SLMs are proving competitive with LLMs on benchmarks, surpassing models like OpenAI's GPT-3.5 in understanding various subjects. SLMs benefit from quality-controlled training data, unlike LLMs’ diverse internet-sourced content. OpenAI's CEO suggests a shift away from LLMs to other improvements. SLMs offer advantages in running on local devices, preserving privacy, and being more accessible due to lower hardware costs. They also present opportunities in responsible AI development, language acquisition research, and could inform improvements in generative AI by emulating efficient human language learning.

• Apple and Meta Discuss Potential AI Partnership - Apple and Meta have been in discussions regarding a potential partnership focused on artificial intelligence, as both companies aim to bolster their AI capabilities. This collaboration could leverage Meta's advanced AI models alongside Apple's extensive hardware ecosystem, enhancing user experiences across both platforms. The discussions come amid Apple's recent AI advancements showcased at WWDC 2024 and Meta's ongoing development of AI technologies like its LLaMA models.

  

• Nvidia to Launch in Middle East Amid U.S. Curbs on AI Exports, Ooredoo CEO Says - Nvidia has signed an agreement to deploy its AI technology in Ooredoo's data centers across five Middle Eastern countries, including Qatar, Algeria, Tunisia, Oman, and Kuwait. This move comes despite U.S. export restrictions on advanced AI chips to the region. The partnership will enable Ooredoo to offer its clients advanced AI services, potentially giving them a competitive edge. Ooredoo is also investing $1 billion to expand its data center capacity significantly by the decade's end.

• China's ByteDance Working with Broadcom to Develop Advanced AI Chip - ByteDance is collaborating with Broadcom to develop a 5-nanometer AI chip, navigating U.S.-China tensions over semiconductor technology. The custom chip will be compliant with U.S. export restrictions and manufactured by Taiwan's TSMC. This partnership aims to reduce costs and ensure a steady supply of high-end chips, essential for ByteDance's AI-driven services like TikTok and its generative AI projects. The development highlights ByteDance's strategy to secure advanced AI capabilities amidst stringent U.S. export controls.

  

• Synthesia’s hyperrealistic deepfakes will soon have full bodies - Synthesia has developed hyperrealistic AI avatars that can match expressions and tone of voice to spoken text, with their visuals generated by diffusion models. Initially offering only torso-up avatars, Synthesia is now advancing toward full-body avatars and has improved the creation process. Jack Saunders from the University of Bath noted small imperfections but overall high-quality realism. A new version, launching in late July, will include hands and allow filming from multiple angles, outpacing similar offerings by enhancing hand movements and lip-synching. Creating personalized avatars has become more efficient; a 10-minute recording session using simple equipment suffices, streamlining the previous lengthier and separate recording of face and voice. The updated method simultaneously records facial movements and voice, necessitating user consent and a security protocol.

  

• Google Is Turning Into a Libel Machine - The Atlantic article addresses the dangers of trusting generative AI within search engines, citing an incident where Google Search's AI Overview erroneously claimed that chess player Hans Niemann admitted to cheating by using a chess engine against Magnus Carlsen, which was false. Niemann has defended himself and sued for defamation. The piece explores how AI-generated content, lacking the intent or mental state of humans, poses complex legal challenges in defamation cases. The article highlights recent legal debates on whether tech companies should be held accountable for their AI's false statements, suggesting that product liability frameworks and European risk mitigation strategies could be relevant. It emphasizes the need for existing legal structures to address AI-generated content issues rather than creating entirely new laws, urging the assessment and prevention of potential risks associated with AI technologies.

• Major record labels sue AI company behind “BBL Drizzy” - Major record labels, including Universal Music Group, Sony Music Entertainment, and Warner Records, represented by the Recording Industry Association of America (RIAA), have filed copyright infringement lawsuits against AI music companies Suno and Udio. These companies are accused of using copyrighted music in the training of their AI models to generate songs without proper licensing. The lawsuits, filed in federal courts in Boston and New York, seek damages of up repeat $150,000 per work. The RIAA emphasizes that the AI models produce musical works that too closely imitate the labels' recordings, implying unpermitted use of the originals. The music industry's legal action represents a broader struggle with AI technology firms over allegedly unauthorized use of protected content to train AI tools, impacting musicians' rights and industry revenue streams.

• The Next Phase of Generative AI Presents a Golden Opportunity for Businesses - The evolution of generative AI is poised to revolutionize business operations, offering unprecedented opportunities for innovation and efficiency. Companies can leverage AI to enhance creativity, streamline workflows, and create personalized customer experiences. As AI technology matures, businesses that strategically integrate these advanced tools can gain significant competitive advantages, driving growth and transforming industries. The key to success lies in adopting a forward-thinking approach and investing in AI capabilities that align with long-term business goals.

Awesome Research Papers

• Optimizing AI Inference at Character.AI - Character.AI aims to progress towards Artificial General Intelligence (AGI), pioneering the use of large language models (LLMs) for a variety of applications. The company emphasizes efficient "inference," allowing LLMs to generate replies swiftly and scalably. Their full-stack approach to AI enables tailored optimizations, reducing inference costs and improving efficiency. Innovations include memory-efficient architecture with Multi-Query Attention, Hybrid Attention Horizons, and Cross-Layer KV-sharing, reducing memory demands by over 20x. They also developed a stateful caching system to effectively manage long dialogues and employed int8 quantization to align training and serving precision, enhancing training efficiency. A 33-fold cost reduction in serving costs has been achieved since late 2022, equating to 13.5 times less expense than using leading commercial APIs. Character.AI handles over 20,000 inference queries per second and is dedicated to further advancements in AI to enrich human interactions and productivity.

  

• LongRAG: Enhancing Retrieval-Augmented Generation with Long-context LLMs - The traditional Retrieval-Augmented Generation (RAG) framework relies on retrieving short text units for answer finding, posing efficiency issues. The study introduces LongRAG, which extends the retrieval unit size to 4K tokens, significantly shrinking the searchable corpus while improving retriever performance with remarkable retrieval scores (71% answer recall@1 on NQ). Applied to large language models (LLMs) for zero-shot answer extraction, LongRAG achieves state-of-the-art results without any additional training (62.7% exact match on NQ). The research suggests a promising direction for integrating RAG with long-context LLMs to enhance information retrieval and question answering systems.

• Whiteboard-of-Thought: Thinking Step-by-Step Across Modalities - The paper discusses a novel method known as "whiteboard-of-thought prompting" to enhance the visual reasoning of multimodal large language models. Traditionally, while such models excel in text-based arithmetic and symbolic reasoning, they falter with tasks requiring visual thinking. This method provides the language models with a metaphorical whiteboard to graphically process reasoning steps, significantly improving performance on natural language tasks involving visual and spatial reasoning. Importantly, this is accomplished using existing abilities to write code with graphical libraries, and no need for additional demonstrations or modules. The technique demonstrated state-of-the-art results, raising accuracy up to 92% on tasks where models previously failed completely, including detailed analysis of the technique's effectiveness and error sources.