AI Weekly: 07/31/23

Good morning and welcome to this week’s edition of AI Weekly! In a relatively quieter week of AI news, Stability has released its latest version of Stable Diffusion, a 3.5 billion parameter model that is capable of producing full 1-megapixel resolution images in seconds.

Also, Amazon has expanded its generative AI offerings in healthcare with the release of AWS HealthScribe, a platform offering AI tools for clinicians to transcribe and analyze conversations with patients.

In AI media and entertainment news, a startup called Channel 1 is building a personalized news outlet that features AI news anchors delivering users’ news, while another TV channel in India has officially employed its first AI news anchor named Lisa.

Enjoy more about this past week’s AI happenings below!

- ZG

Here are the most important stories of the week:

IMAGE/VIDEO

Stability AI launches Stable Diffusion XL 1.0, its most advanced text-to-image model yet. Link.

• The model contains 3.5 billion parameters and can produce full 1-megapixel resolution images in seconds with multiple aspect ratios.

• Stable Diffusion XL 1.0 is customizable and improved in text generation, including legible logos, calligraphy, and fonts.

• The model supports inpainting, outpainting, and image-to-image prompts, with multi-part instructions in short prompts.

• The open source version raises ethical concerns as it could be misused to generate harmful content, deepfakes, and porn.

• Stability AI is addressing the challenges with filtering unsafe imagery and respecting artists' requests to remove their work from training data sets.

AI-generated news anchors have made their debut in India, with one named Lisa presenting news on Odisha TV. Link.

• Lisa's appearance resembles a human presenter, but her robotic monotone and lack of natural tone shifts are noticeable.

• She is part of a growing trend of AI news presenters employed in newsrooms across various countries.

• Lisa's introduction raises questions about the impact on other jobs in her newsroom and whether it was driven by cost-saving measures.

• Odisha TV's managing director sees AI anchors like Lisa as partners, handling repetitive and data analytical tasks so newspeople can focus on creative work.

• The era of AI news readers is already here, with AI-generated presenters becoming a common presence in the media industry.

Channel 1 News, a startup, plans to use generative AI to create a new video news channel with fine-tuned LLMs serving as news presenters, offering an immersive experience and news stories in multiple languages. Link.

• The company's founders are looking ahead to future advancements in AI technology and believe it will eventually become indistinguishable from human presenters.

• Channel 1 News intends to source stories from credible places and use a team of editors to verify accuracy in its news reports.

• Social media reactions to AI news channels vary, with some expressing concerns and others seeing potential in the technology. Additionally, Google is reportedly testing AI-powered news story generation with various news organizations.

CODE/DEVTOOLS

Stack Overflow introduces OverflowAI with generative AI capabilities for improved technical question answering. Link.

• OverflowAI includes updated AI search, a Visual Studio Code extension, and a Slack integration for enterprise users.

• The goal is to make it easier for developers and enterprises to find and use information efficiently.

• OverflowAI doesn't replace the Stack Overflow community but complements it with natural language processing-based queries.

• It benefits users of all experience levels by providing accurate generated results without potential hurdles from community feedback.

• OverflowAI also enhances Stack Overflow for Teams, allowing direct code querying and generation within Visual Studio Code.

Amazon launched Agents for Bedrock, enabling companies to build AI apps that can automatically perform tasks for users, like booking flights or making restaurant reservations. Link.

• AI agents combine generative AI models with customized data to automate tasks without relying on extra help like plug-ins.

• Amazon positioned itself as an infrastructure provider in the AI space via AWS, offering tools like Agents for Bedrock to startups and enterprises.

• Other tech companies, including Meta and OpenAI, are also working on AI agents for broader user adoption.

• AWS also announced AWS HealthScribe, an AI-powered service for automating clinical documentation in healthcare, aiming to save time for healthcare professionals.

• AWS made its Entity Resolution and CodeWhisperer tools generally available, enhancing machine learning-powered analytics and real-time code suggestions for data analysts.

HARDWARE/ROBOTICS

Nvidia is making its cloud-based AI supercomputing service, DGX Cloud, widely accessible on Oracle Cloud Infrastructure (OCI). Link.

• DGX Cloud provides infrastructure and software for training advanced models in generative AI and other AI fields.

• It offers purpose-built infrastructure to meet the demands of generative AI for training large, complex models.

• DGX Cloud simplifies complex infrastructure management, allowing developers to focus on experiments and prototypes without worrying about infrastructure.

• Nvidia claims that DGX Cloud enables faster training with speedups of two to three times compared to traditional cloud computing.

• Biotechnology firm Amgen is using DGX Cloud to expedite drug discovery and reported three times faster training of protein LLMs and up to 100 times faster post-training analysis with Nvidia RAPIDS.

Google DeepMind has introduced RT-2, which allows robots to transfer concepts learned from small datasets to different scenarios, improving generalization capabilities and semantic understanding. Link.

• RT-2 can interpret new commands and perform rudimentary reasoning, such as reasoning about object categories or high-level descriptions, without explicit training.

• The system can determine the best tool for a specific novel task based on existing contextual information, making it more scalable for handling diverse tasks.

• The efficacy rate of executing new tasks has improved from 32% to 62% in the transition from RT-1 to RT-2.

• DeepMind's robotics team developed Robotics Transformer (RT-1) last year, using a database of 130,000 demonstrations to achieve a 97% success rate in over 700 tasks like picking and placing objects.

• The robotics industry faces challenges in handling complex tasks due to the infinite variables involved, but recent breakthroughs in robotic learning are improving adaptability.

HEALTHCARE

Amazon introduces AWS HealthScribe, a platform offering AI tools for clinicians to transcribe and analyze conversations with patients. Link.

• HealthScribe's generative AI helps create transcripts, extract details, and generate summaries for electronic health record (EHR) systems.

• The AI-powered platform categorizes transcripts based on clinical relevance and extracts structured medical terms like medications and medical conditions.

• HealthScribe's AI capabilities are powered by Bedrock, Amazon's platform for generative AI-powered apps using pretrained models.

• The platform allows clinicians to review AI-generated notes before finalizing records and emphasizes security and privacy in processing customer data.

• Amazon also announces AWS HealthImaging, a service for storing, transforming, and analyzing medical imaging data at a petabyte scale with dynamic pricing options.

POLICY/LAW/ETHICS

Google, Microsoft, OpenAI, and Anthropic are forming the Frontier Model Forum to ensure safe and responsible progress in AI technology. Link.

• The forum aims to improve AI safety research, establish best practices for responsible development and deployment, and share knowledge about AI safety risks with policymakers and academics.

• The forum will also support the development of AI solutions for global challenges like climate change and cancer detection.

• Membership is open to companies that develop and deploy frontier models, demonstrate a strong commitment to frontier model safety, and are willing to collaborate on joint initiatives.

• The forum defines "frontier models" as large-scale machine-learning models with capabilities beyond current standards.

• The forum emphasizes the need for appropriate safeguards to mitigate AI risks and stresses the importance of keeping AI technology under human control.

A coalition of open-source AI stakeholders, including Hugging Face, GitHub, EleutherAI, Creative Commons, LAION, and Open Future, released a policy paper calling on EU policymakers to protect open-source innovation in the EU AI Act, the world's first comprehensive AI law. Link.

• The coalition seeks to ensure that the AI Act supports open AI development practices without imposing overbroad obligations that could favor closed and proprietary AI development, disadvantaging the open AI ecosystem.

• They emphasize that open-source AI innovation is essential and should not be hindered by regulations.

• The EU AI Act classifies AI systems based on the risk they pose to users, with higher-risk systems facing more regulation.

• The coalition aims to ensure that the AI Act's regulations do not inadvertently burden open-source developers, who are often hobbyists, nonprofits, or students, as they have different development models than companies using APIs.

• The "Brussels Effect" of the EU's regulatory influence in tech makes the coalition's efforts potentially influential in shaping global regulatory conversations, including in the US.

OTHER

Alphabet's Q2 2023 earnings call highlighted the growing adoption of the company's cloud and product offerings, as over 70% of generative AI startups rely on Google's cloud infrastructure and AI capabilities. Link.

• Google Cloud is providing startups access to resources like its Vertex AI platform, advanced TPU chips, and a library of over 80 pre-trained AI models to accelerate development.

• The cloud division saw strong growth, with revenue up 28% to $8 billion in Q2, turning profitable in Q1 with an operating profit of $395 million.

• Google is incorporating generative AI in its own products, making services like Search and Workspace more helpful and enhancing ads creatives, optimization, and campaign management.

• Google products like Bard conversational AI and Duet in Workspace are seeing strong adoption, and the company is prioritizing the development of its most promising AI applications.

Netflix is investing in AI for entertainment with job postings for machine learning platform product manager and technical director for generative AI. Link.

• The machine learning product manager role offers an annual salary of $300,000 to $900,000, far exceeding many actors' pay.

• Netflix plans to use AI not just for content recommendations but also to create original content and optimize production.

• The streaming giant already airs a Spanish reality dating series using AI-generated "deepfakes" and employs generative AI in its gaming studio for narrative composition and dialogue.

• Striking actors rejected a proposal offering a one-time $200 day rate for AI-enhanced CGI scans, fearing loss of ownership and compensation rights for eternity.

Protect AI, a platform that aims to address security weaknesses in AI systems and focuses on practical threats in the AI and machine learning supply chain, has raised $35 million in a Series A round led by Evolution Equity Partners, bringing the total raised to $48.5 million. Link.

• The funds will be used to enhance Protect AI's platform, expand research efforts, and launch new open source projects.

• Its flagship tool, AI Radar, provides visibility into AI model components, generates a "machine learning bill of materials" (MLBOM), and scans for vulnerable software in pipelines.

• Protect AI offers tools to mitigate AI attacks, such as prompt injection attacks, and evaluates Jupyter notebooks for potential vulnerabilities.

• The startup faces competition in the AI-defending security tools space, but has attracted high-profile customers in various industries.

Effectiv, a real-time fraud and risk management platform, raised $4.5 million in seed funding, led by Better Tomorrow Ventures, bringing its total raise to over $9 million. Link.

• Effectiv is combating bank transfer and payment fraud, which amounted to $1.6 billion in 2022.

• The platform provides AI-driven tools, including payment fraud detection, customer and business onboarding verification, and compliance management.

• It trains models to detect simulated behavior and applies anomaly scores to identify abnormal transactions for further scrutiny.

• Since launching in 2022, Effectiv has gained 15 financial institutions and fintech clients, including BHG Money, Cardless, and USALLIANCE Financial.

• The new funding will be used for product development, including the release of DeviceAI, a device biometrics technology that flags suspicious user behavior on a device.

Scientists are utilizing human brain cells in AI chips to make them faster and stronger. Link.

• The "DishBrain," a semi-biological AI chip, integrated human and mouse brain cells and quickly learned to play Pong in five minutes.

• The project received a grant of $407,000 from Australia's National Intelligence and Security Discovery Research Grants Program.

• These human brain AI chips blend AI with biological computing, potentially surpassing traditional silicon-based hardware.

• The research could lead to state-of-the-art computer chips with advanced AI capabilities, impacting fields like robotics, drug discovery, and brain-machine interfaces.

• The grant will support the development of more advanced AI machines that replicate the learning capacity of biological neural networks.

LinkedIn is reportedly developing an AI tool called "LinkedIn Coach" to support users in job application processes, skill development, and networking on the platform. Link.

• The leak comes from app researcher Nima Owji, who discovered the upcoming features.

• The AI assistant may answer user queries about how it works and provide insights into company cultures, including Microsoft's, as LinkedIn is owned by Microsoft.

• LinkedIn's spokesperson confirmed that the company is continuously exploring ways to enhance user experience on the platform, hinting at upcoming developments.

• Microsoft has been integrating chatbots into various products, including Bing, Copilot, and GitHub, showcasing its commitment to AI-powered tools.

• The specific release date for LinkedIn Coach and its full range of features are yet to be officially announced.

London-based startup AutogenAI has raised $22.3 million in funding from Blossom Capital for their generative AI tool to help businesses write stronger pitches, improving their strike rate in winning deals. Link.

• The recent funding round values AutogenAI in the "hundreds of millions" range.

• AutogenAI has attracted 28 clients in less than a year, but it doesn't disclose customer names due to the perception that using AI might detract from work quality.

• The platform combines large language models from OpenAI and others with a client's proprietary data, helping users create pitches based on successful past work.

• AutogenAI's software could speed up the pitch-writing process by up to 800%, significantly reducing procurement costs and allowing more focus on strategic additions.

• The startup's focus on practical applications, like pitch-writing, appeals to businesses seeking clear ROI on AI tools that save time and resources.

Stay Ai, a platform that helps Shopify brands build up their subscription programs and increase customer lifetime value, has raised $15.1 million in Series A funding, with $8.6 million led by Telescope Partners and an additional $7.5 million led by Watchfire Ventures and individuals, including Nik Sharma and Adam Turner. Link.

• The platform includes a recurring commerce enablement system for subscriptions, clubs, and memberships with loyalty tools like SMS chatbot, surveying, and rewards programs.

• Stay Ai's ExperienceEngine is an AI-powered system that allows merchants to perform marketing procedures like A/B testing and leverage predictive analytics and machine-learning models for better decision-making.

• Its tools have been successful in reducing cancellations and increasing subscriber revenue for clients like Olipop.

• Stay Ai is used by around 300 Shopify merchants, including Vita Coco, Momofuku, and Snow, and its revenue is pacing 10x from 2022.

• The funding will be used to expand the team, accelerate the innovation roadmap, and make acquisitions, like the recent acquisition of churn reduction app RetentionEngine.