Amazon Shows its AI Ascent with Collaborations

Amazon spends $2.75 billion on AI startup Anthropic in its largest venture investment yet

Amazon is solidifying its position in the competitive artificial intelligence (AI) landscape with a $2.75 billion investment in Anthropic, a San Francisco-based AI firm considered a key player in generative AI. This investment, part of a potential $4 billion commitment announced previously, gives Amazon a minority stake without board influence. Anthropic, valued at $18.4 billion, has amassed over $10 billion in funding to date, including this round. The firm's AI model, Claude, rivals OpenAI's ChatGPT and has recently been upgraded to Claude 3, touted to surpass OpenAI's GPT-4 and Google’s Gemini Ultra in benchmark tests. Anthropic will harness Amazon Web Services (AWS) for cloud computing and adapt Amazon's custom chips for AI development, strengthening a symbiotic relationship between the two. Amidst a venture capital boost for AI, with $29.1 billion invested in 2023, Amazon leads Big Tech's strategic shift from acquisitions to venture-style investments in AI, paralleled by Microsoft and Google's AI initiatives. This trend invites scrutiny, however, as regulators investigate the implications of Big Tech's interlinked partnerships in the AI domain.

Amazon also announced a collaboration with Anthropic, AWS and Accenture to bring trusted AI solutions targeted at enterprise clients.

• Signs of Apple's AI App Store could surface in June - Apple is reportedly planning to revolutionize its approach to artificial intelligence (AI) software by introducing an enhanced AI App Store, potentially announced at the Worldwide Developers Conference (WWDC). Analyst Ben Reitzes from Melius Research speculates that instead of solely focusing on developing its own AI apps, Apple aims to create a platform where consumers can acquire AI apps from various vendors, drawing parallels to how Apple previously built the iTunes and iPhone App Stores. Details on purchasing AI apps and possible upgrades to Siri and other services could be unveiled at WWDC 2024. Discussions with other tech giants, including Google, suggest Apple's interest in licensing AI technologies and possibly integrating them into the new AI ecosystem. The official announcement and specifics regarding WWDC 2024 remain to be disclosed.

• Claude Opus > GPT-4 - using Arena votes, Claude unseats GPT-4, impressing on speed, context length and capabilities.

• Here’s why AI search engines really can’t kill Google - AI-driven search tools are challenging the traditional role of search engines like Google in providing information online. While they seem more intuitive, offering direct answers to complex questions, their effectiveness is mixed when handling the diverse purposes Google serves, such as navigating to websites, answering factual queries, or exploring broad topics. AI tools struggle with navigational prompts and real-time information, frequently lagging behind Google's speed and accuracy. They excel, however, at extracting concise answers from a sea of internet data for simple, evergreen questions and buried information queries. The future of search may leverage AI's potential, but replacing Google's comprehensive role as a quasi-operating system for the internet is not straightforward, requiring a blend of speed, accuracy, and multifunctionality that goes beyond current AI chatbot capabilities.

• Behind the plot to break Nvidia’s grip on AI by targeting software - A coalition of tech companies, including Qualcomm, Google, and Intel, has formed the UXL Foundation to challenge Nvidia's dominance in the AI chip market by developing an open-source project called OneAPI, which aims to create a suite of software and tools that can power various types of AI accelerator chips. The UXL Foundation's goal is to establish a standard programming model for AI computing and attract a critical mass of developers to its platform, eventually supporting Nvidia hardware and code. Despite facing competition from the UXL Foundation and numerous well-funded startups, Nvidia maintains a strong position in the AI chip market, largely due to its widely adopted CUDA software platform and its extensive developer ecosystem built over 15 years.

• Google AI could soon use a person’s cough to diagnose disease - Google scientists have developed a machine-learning tool, named Health Acoustic Representations (HeAR), designed to diagnose and monitor health conditions by analyzing human sounds such as coughs and breathing. HeAR is notable for its use of a vast dataset from YouTube, enabling it to outperform existing models in detecting COVID-19 and tuberculosis. The system employs self-supervised learning, requiring less labeled data for refinement. Although still in the research stage and not commercially available, HeAR’s foundation model promises a non-invasive, efficient approach for potential future diagnosis and disease monitoring, pending further validation and regulatory approval.

• Microsoft Bing Chief Exiting Role After Suleyman Named AI Leader - Microsoft Corp.'s Mikhail Parakhin, head of the company's Bing search engine and advertising businesses, will exit those roles and search for a new position within the company, following the appointment of Mustafa Suleyman to oversee consumer artificial intelligence work. Parakhin's departure from his core role in Microsoft's consumer AI products marks the first major shuffle resulting from CEO Satya Nadella's decision to install Suleyman, a co-founder of AI researcher DeepMind, as head of a unified push in the AI space. The move reflects Nadella's dissatisfaction with his team's efforts, as Microsoft has integrated AI into various products over the past year, yet Bing has made limited gains against Google, and other AI-powered products remain works in progress.

• Why the AI Hyperrealists at Databricks Spent $10 Million to Beat Meta’s LLM - Databricks, valued at $43 billion, has joined the race to develop the best open-source large language model (LLM) by releasing DBRX, which slightly outperforms other open-source models and OpenAI's GPT-3.5 in various benchmarks. The company spent $10 million and used 3,100 Nvidia H100 chips over two months to develop DBRX, showcasing the efficiency and cost-effectiveness of their approach compared to competitors like Meta's Llama 2. By open-sourcing DBRX, Databricks aims to attract AI talent and corporate buyers who want to use LLMs while staying within their budgets, highlighting the importance of price in the real-world adoption of AI technologies.

• AI and data infrastructure drives demand for open source startups - The Runa Open Source Startup (ROSS) Index, compiled by Luxembourg-based Runa Capital, highlights the growth of commercial open source software (COSS) startups, particularly those focused on AI and data infrastructure, with LangChain, an open source framework for large language models, leading the 2023 index. The report also reveals the global origins of many COSS startups, a notable increase in European startups, and a shift in developers' preferences towards TypeScript, Python, and Rust. Funding for these startups reached $513 million in 2023, demonstrating significant growth, and the ROSS Index's methodology, which combines GitHub stars growth rate and manual selection, specifically targets open source startups to provide insights into this sector of the tech industry.

Awesome Research Papers

MacGyver: Are Large Language Models Creative Problem Solvers? - The paper investigates the ability of LLMs to solve creative problems using the "MacGyver" task, which involves using everyday objects to solve unexpected problems. The authors found that while LLMs can generate creative solutions, they struggle with tasks requiring physical reasoning, understanding object affordances, and sequencing multiple steps. However, they propose a "step-by-step verify" strategy, where the LLM iterates on the solution with user verification, leading to significant improvements in the quality and creativity of the solutions.

MathVerse: Does your Multi-modal LLM Truly See the Diagrams in Visual Math Problems? - The development of MathVerse, a comprehensive visual math benchmark, aspires to thoroughly evaluate the interpretive abilities of Multi-modal Large Language Models (MLLMs) in math problem-solving with visual elements. Researchers collated 2,612 diverse math problems with diagrams and generated 15,000 test instances in six variants, each with different information levels. Additionally, a new evaluation method, the Chain-of-Thought (CoT), has been introduced. It intricately analyzes the reasoning steps generated by models like GPT-4(V), allowing for a nuanced examination of MLLMs' reasoning processes. MathVerse aims to provide actionable insights to refine MLLMs, pushing forward their development.

ObjectDrop: Bootstrapping Counterfactuals for Photorealistic Object Removal and Insertion - Diffusion models have improved image editing but often generate images that violate physical laws, particularly the effects of objects on the scene. The authors propose a practical solution by creating a counterfactual dataset and fine-tuning a diffusion model on it to remove objects and their effects on the scene. To address the challenge of photorealistic object insertion, they propose bootstrap supervision, which leverages their object removal model trained on a small counterfactual dataset to synthetically expand the dataset, outperforming prior methods in modeling the effects of objects on the scene.

AI generates high-quality images 30 times faster in a single step - MIT CSAIL researchers have presented a groundbreaking framework that revolutionizes the generation of AI-created art through diffusion models. Traditional models require a multi-step iterative process, but the new method, distribution matching distillation (DMD), distills this into a single step. The DMD employs a teacher-student dynamic, where a simpler model learns from a more complex one, maintaining the quality of image generation while achieving a 30x speed increase. Their approach overcomes the stability challenges faced by GANs and could benefit various applications, including faster content creation and drug discovery. DMD achieves near-parity with multi-step methods in quality benchmarks and offers potential for real-time visual editing. The research, advancing the capabilities of AI in visual content generation, will be presented at an upcoming conference.

DBRX: A New State-of-the-Art Open LLM - Databricks introduces DBRX, a new benchmark-setting open-source Large Language Model (LLM) that demonstrates superior performance on various standard benchmarks compared to existing models like GPT-3.5 and CodeLLaMA-70B. DBRX utilizes a fine-grained mixture-of-experts (MoE) architecture for greater training and inference efficiency, with a 40% reduction in parameter count and double the FLOP-efficiency compared to dense models. DBRX achieves high performance on language understanding, programming, and mathematics tasks, and offers enhanced efficiency with competitive edge over some closed model APIs. It's available through Databricks' API and via Hugging Face, enabling customers to train or extend the model using Databricks' tools and infrastructure. The model's development utilized a variety of Databricks' data and AI tools, and it's positioned as a key element in Databricks' next-generation GenAI products, aiming to empower enterprises in managing their data and AI destiny.

Awesome Updates

Monetizing your GPT - OpenAI is partnering with a select group of builders to test GPT earnings based on usage.

GPT Vision Free - ChatGPT is apparently testing a new GPT-3.5 model called "text-davinci-002-render-paid" with vision capabilities.

Adobe and Microsoft partner to bring new generative AI capabilities to marketers as they work in Microsoft 365 applications - At the Adobe Summit, Adobe and Microsoft announced a new collaboration to integrate Adobe Experience Cloud with Microsoft Copilot in Microsoft 365, aiming to improve efficiency for marketers by streamlining workflows and breaking data silos. This synergy will embed relevant marketing insights and tools within commonly used Microsoft applications like Outlook, Teams, and Word, enhancing marketers' ability to create content, manage campaigns, and track performance without constant application switching, which has been identified as a creativity impediment for 43% of marketing professionals. Initial features will offer easy access to campaign insights, automated content creation, and in-context notifications to keep projects on track.