AI Weekly: 07/24/23

Good morning and welcome to this week’s edition of AI Weekly! In this week’s news, Apple made headlines with their ChatGPT competitor, “Apple GPT,” reportedly brewing behind the scenes, while Wix launched its new AI-driven website generator called AI Site Generator, raising questions about smaller entrants’ long term defensibility.

In AI entertainment news, an AI-created South Park episodes has made waves, showcasing the company The Simulation’s generative product called Showrunner.

In regulation news, Microsoft, Google, Amazon, Meta, OpenAI, Anthropic, and Inflection have committed to external testing of new AI systems before public release and clear labeling of AI-generated content, as part of voluntary commitments agreed upon with the White House. Enjoy reading more about last week’s AI happenings below!

- ZG

Here are the most important stories of the week:

TEXT

Apple is reportedly developing its own AI-powered chatbot, referred to as "Apple GPT," using its LLM framework called "Ajax," running on Google Cloud and built with Google JAX. Link.

• The chatbot is not yet planned for public release, and Apple has multiple teams working on the project, considering potential privacy implications.

• While other tech giants like Meta, Microsoft, and Google have released generative AI products to the public, Apple has been relatively quiet in this space, though it has been using the Ajax-powered chatbot internally.

• Apple's most famous AI system is Siri, and the company has been gradually incorporating AI into its software.

• Tim Cook, Apple's CEO, mentioned in a recent interview that AI is something the company is closely monitoring, and Apple has been working on its AI initiatives under the leadership of John Giannandrea and Craig Federighi.

• Apple is expected to make a significant AI-related announcement sometime next year, though specific plans in the AI space remain unclear.

Google is developing an AI tool for news publishers that can generate article text and headlines, potentially transforming the journalism industry. Link.

• The tech giant aims to partner with news outlets to use the AI tool in newsrooms, giving journalists the choice to enhance their work and productivity.

• The project, internally known as "Genesis," has been pitched to media companies like The New York Times, The Washington Post, and News Corp (owner of The Wall Street Journal).

• Google emphasizes that the tool is not intended to replace journalists but to assist them in reporting, creating, and fact-checking articles.

• While generative AI features are being rapidly developed and deployed in workplace applications to streamline tasks and boost productivity, concerns still exist about potential inaccuracies and "hallucination" of responses.

• News outlets like CNET and Gizmodo have experienced issues with AI-generated stories but still plan to use the technology.

IMAGE/VIDEO

The AI photo editing app Remini has gone viral on TikTok, allowing users to create professional headshots for sites like LinkedIn using their own selfies. Link.

• The app's generative AI feature, which was added last year, has led to a surge in downloads and made it the No. 1 Overall app on the U.S. App Store, overtaking Instagram Threads.

• To create AI headshots, users select 8-12 photos with different backgrounds and angles, and the AI generates the shots based on a chosen "model" image and curriculum style.

• The app is free to download, but a subscription is required for full access to its features, priced at $9.99/week or $79.99 per year.

• Some users have expressed dissatisfaction with the app's AI results, as it often made unwelcome adjustments to their bodies in the photos, leading to concerns about body image and mental health impacts.

• Remini has seen around 40 million downloads in a short period and stated that it is working to refine its AI technology to minimize inaccuracies in body representation. The app also offers other photo enhancement features, such as restoring old photos and using the Official Barbie filter on selfies.

The Simulation (formerly Fable Studio) has introduced Showrunner, an AI application that can create generative AI TV episodes of popular shows, demonstrated with a 22-minute South Park episode called "Westland Chronicles." Link.

• Showrunner uses generative AI to create complete episodes, including story, script, voice recording, animation, editing, and scoring, with minimal human involvement.

• The episode is a demo and has no commercial purpose, but it highlights the power and potential threat of AI in TV content creation.

• The technology is currently only available to researchers and journalists and not intended for public release due to safety concerns.

• The Simulation aims to create a simulation where AI characters have rich stories and lives, generating 22-minute episodes of their activities every week.

• Showrunner could potentially enable writers to own the means of production, allowing them to pitch movies as animated mock-ups created by AI, giving them more power and control over their work.

CODE/DEVTOOLS

GitHub is expanding its generative AI offerings, aiming to integrate a context-aware conversational assistant, Copilot Chat, into integrated development environments (IDE) like Microsoft Visual Studio and VS Code. Link.

• Copilot Chat is specifically designed for developers and can assist with complex tasks through simple prompts, empowering all team members to build applications and debug codebases quickly.

• The AI acts as a pair programmer, helping with tasks like starting a file in an unfamiliar coding language, autocompleting boilerplate code, conducting debugging, and writing unit tests.

• Copilot Chat demonstrates contextual awareness, understanding and interpreting the code developers write and the error messages that appear, offering tailored guidance.

• GitHub's controlled study shows that developers achieved tasks 55% faster using GitHub Copilot, and on average, 46% of code across all programming languages is constructed with it, with Java developers using it even more (61%).

• The AI can also review code within the IDE, identifying potential security issues and suggesting remediation, leading to fewer vulnerabilities in security scans.

Wix, a popular web builder, is embracing generative AI with its new AI Site Generator tool to streamline the website creation process for customers. Link.

• The AI Site Generator allows users to describe their intent, and the AI system generates a complete website with homepage, inner pages, text, images, and even business-specific sections like events and bookings.

• Wix believes AI can simplify back-end functionality, infrastructure, and payments for websites, especially benefiting small businesses struggling to launch and maintain sites.

• The AI Site Generator uses a combination of in-house and third-party AI systems, including ChatGPT, to create custom web templates based on prompts and user specifications.

• The AI Site Generator is unique in its ability to automatically incorporate e-commerce, scheduling, food ordering, and event ticketing components based on customer requirements.

• Wix acknowledges that AI still cannot fully replace human designers, and users have access to Wix's full suite of editing tools to customize their generated websites.

Microsoft announced new AI features for Azure, including Vector Search, which uses machine learning to enhance search capabilities for unstructured data like images and text. Link.

• Vector Search uses vectorization to encode the meaning of words and images, enabling faster and more accurate search results, and can be used in chat-based apps, personalized responses, and data pattern identification.

• The Document Generative AI solution integrates AI-powered document processing services with OpenAI's language models, allowing apps to read documents and generate responses based on the content.

• Microsoft will introduce OpenAI's Whisper model, an automatic speech recognition model, to Azure's AI speech services, providing transcription and translation capabilities.

• Microsoft also launched Real-time Diarization, an AI-driven speech service to identify speakers in real-time, and expanded the availability of Custom Neural Voice, allowing customers to closely reproduce or create synthetic voices.

• To prevent misuse, Custom Neural Voice includes controls like voice actor consent, speaker verification, and a code of conduct for customers, with watermarking and detection tools to identify audio clips created with the service.

Cleanlab, a startup offering a data curation solution for LLMs used in enterprise AI, has raised $5 million in seed funding led by Bain Capital Ventures. Link.

• The company's open-source product uses a method called "confident learning" to identify, understand, and clean incorrect labels in data, improving the effectiveness of machine learning models by addressing the "dirty data problem."

• Cleanlab offers two products: Cleanlab Open Source, a free Python library for applying confident learning, and Cleanlab Studio, a cloud-based SaaS product with a user-friendly interface and advanced features for data curation, integrating with popular LLM frameworks and platforms.

• The startup has already attracted over 10,000 users for its open-source project and more than 100 customers for its cloud product, including Fortune 500 companies, government agencies, research institutions, and startups across various industries.

• Cleanlab plans to use the funding to expand its team, scale product development, and grow its customer base.

• Investors at Bain Capital Ventures believe Cleanlab is addressing a significant and underserved problem in the enterprise AI space, and the company is part of a broader emphasis on artificial intelligence at Bain, which has invested in several AI startups this year.

Splunk has announced Splunk AI, a suite of AI-driven solutions to enhance its unified security and observability platform, combining automation with human-in-the-loop experiences for better detection and response capabilities. Link.

• The new Splunk AI Assistant uses generative AI to provide an interactive chat experience, allowing users to create SPL queries through natural language, increasing accessibility to valuable data insights.

• The AI empowers SecOps, ITOps, and engineering teams to automate data mining, anomaly detection, and risk assessment, allowing them to focus on strategic tasks and reducing errors.

• The model is integrated with domain-specific LLMs and ML algorithms, leveraging security and observability data for productivity and cost efficiency.

• The new capabilities aim to optimize alerting speed and accuracy, enhance digital resilience, and streamline anomaly detection workflows.

• Splunk also introduced ML-powered foundational offerings, such as the Splunk Machine Learning Toolkit (MLTK) 5.4, providing guided access to ML technology for all skill levels and enabling users to upload pre-trained models to apply them seamlessly to their data.

Unstructured.io, a startup aiming to help enterprises adopt LLMs by extracting and staging enterprise data in a way that LLMs can understand and leverage, has raised $25 million in a Series A funding led by Madrona and Bain Capital Ventures, among others. Link.

• The company was co-founded in 2022 by Brian Raymond, Matt Robinson, and Crag Wolfe, who previously worked together at Primer AI.

• Data processing and preparation for AI development workflows can be time-consuming, and Unstructured.io provides tools to clean up and transform enterprise data for LLM ingestion.

• Unstructured offers various tools to remove unwanted objects, perform optical character recognition, and create processing pipelines for specific types of documents.

• The startup has close ties to defense agencies, and its board includes former members of the U.S. intelligence community and the Pentagon's Joint Artificial Intelligence Center.

AI21 Labs has launched Contextual Answers, a plug-and-play generative AI engine as a dedicated API to help enterprises leverage LLM technology on their data. Link.

• The API can be embedded into digital assets, allowing business employees or customers to interact with accurate information through a conversational experience without needing to engage with different teams or software systems.

• It provides a solution for enterprises to implement LLMs into their data stack without significant engineering efforts, making it accessible to developers without NLP or AI expertise.

• The AI engine is optimized to adapt to internal jargon, acronyms, and project names within the uploaded knowledge base, while separating organizational knowledge from external sources and maintaining data security and confidentiality.

• AI21 Labs plans to integrate the feature into its writing platform Wordtune to allow users to retrieve select information from uploaded documents.

• Other players in the data ecosystem, such as Databricks and Snowflake, have been working on similar projects using LLM technology to answer specific queries on data and get insights from unstructured documents.

POLICY/LAW/ETHICS

OpenAI is withholding key features of its powerful AI tool, GPT-4, due to ethical and legal concerns. Link.

• GPT-4 is designed to be multimodal, capable of processing both images and text, enabling new applications like identifying car engine issues from uploaded photos.

• The AI can also help visually impaired users by vocalizing descriptions of images, including solving captcha puzzles.

• Blind users have already tested the visual analysis function, successfully identifying objects and gathering information.

• OpenAI has restricted certain features like facial recognition over privacy concerns, fearing user backlash and potential legal challenges.

• The company wants to address safety concerns and seeks public input before releasing the image analysis tool to the public.

Microsoft, Google, Amazon, Meta, OpenAI, Anthropic, and Inflection have committed to external testing of new AI systems before public release and clear labeling of AI-generated content, as part of voluntary commitments agreed upon with the White House. Link.

• The commitments aim to make AI systems and products safer and more trustworthy while comprehensive regulations are developed for the AI industry.

• President Joe Biden met with top executives from the seven companies at the White House to discuss the commitments.

• The companies will also invest in cybersecurity, develop mechanisms for third parties to report system vulnerabilities, prioritize research on societal risks of AI, and create AI systems to address societal challenges.

• External red-teaming exercises will be introduced, allowing outside experts to test AI systems for potential issues before public release.

• White House officials are working on executive actions and legislation to regulate AI more comprehensively in the future. The commitments are voluntary, and enforcement mechanisms are not in place.

Over 8,000 published authors, including prominent writers like Margaret Atwood and Dan Brown, are demanding payment from tech companies for the use of their copyrighted works in training AI tools. Link.

• The authors accuse AI companies of unfairly profiting from their works without compensation.

• They argue that the AI systems rely on copyrighted books, articles, essays, and poetry for their training data, and it's only fair that the authors are compensated for their use.

• The open letter was addressed to the CEOs of OpenAI, Meta, Google, Stability AI, IBM, and Microsoft.

• The authors are also calling on AI companies to seek permission before using copyrighted material and to pay writers when their work is featured in the results of generative AI.

• There has been increasing pressure on tech companies regarding intellectual property violations in the training of AI models.

OTHER

Some experts, including Stability AI CEO Emad Mostaque, believe that AI is heading towards a bubble similar to the dot-com bubble of the late 1990s. Link.

• Investors are eagerly pouring money into AI companies, with around $1.7 billion invested in generative AI companies in the first quarter of the year, doubling the previous year's total.

• Mostaque's Stability AI has already received over $100 million in investments, even before the public release of its product, ChatGPT.

• Other AI and investment experts have drawn parallels between the current AI boom and the dot-com era, cautioning that the AI hype may be overhyped in the short term.

• Some experts believe AI is currently experiencing a surge similar to the investment boom in internet companies in 1995, not necessarily at the point of the dot-com bubble burst in 1999.

• Opinions on the long-term outlook for AI companies vary, with some suggesting potential overvaluation while acknowledging the unpredictable nature of short-term market momentum.

Qualtrics plans to invest $500 million in AI over the next four years, focusing on generative AI solutions tailored to enterprise experience management use cases. Link.

• The new AI-integrated platform, XM/os2, will bring the power of generative AI to every part of Qualtrics' platform, according to CEO Zig Serafin.

• Details about how the investment will be divided among business divisions and specific internal efforts remain unclear.

• This investment trend in generative AI is not unique to Qualtrics, with other tech giants and VC firms also pouring significant amounts of capital into the exploding generative AI category.

• McKinsey estimates that generative AI could add $4.4 trillion annually to the global economy, but some strategists warn that the AI boom may not lead to massive profits and could be reminiscent of the tech bubble of the 1990s.

Cohere is partnering with global consulting firm McKinsey's AI division, QuantumBlack, to offer secure, enterprise-grade generative AI solutions tailored to McKinsey clients' needs. Link.

• The collaboration will define generative AI use cases, design IT architecture, develop and train AI models, build employee capabilities, and implement organizational changes to meet clients' needs.

• Cohere has seen significant growth and funding in recent months, providing AI services on-premises or on a client's preferred cloud platform, distinguishing itself from consumer-facing models like OpenAI's ChatGPT.

• McKinsey's clientele includes major firms like GM, Ford, Exxon, Pepsi Co., and American Express, and the partnership aims to create value on the ground for clients.

• Initial case studies show promising results, with generative AI used to manage customer feedback in multiple languages and aid product development, leading to savings and faster time-to-market.

• As generative AI becomes a practical tool for value creation, the alliance represents a significant move in enterprise-grade AI.

OpenAI's head of trust and safety, Dave Willner, announced his plans to step down from his position and transition into an advisory role to spend more time with his family. Link.

• The exit comes amid growing scrutiny from lawmakers, regulators, and the public over the safety of OpenAI's products, especially after the viral success of AI chatbot ChatGPT.

• OpenAI's CEO, Sam Altman, has called for AI regulation and expressed concerns about AI being used to manipulate voters and spread disinformation.

• Willner's role at OpenAI has grown significantly since he joined, and his work has been foundational in operationalizing the company's commitment to safe and responsible technology use.

• OpenAI's Chief Technology Officer, Mira Murati, will serve as the interim manager of the trust and safety team, and Willner will advise the team until the end of the year.

• OpenAI, along with six other leading AI companies, has made voluntary commitments to make AI systems and products safer and more trustworthy, including outside testing and clear labeling of AI-generated content.

OpenAI has signed a two-year deal with the American Journalism Project (AJP) to fund local news outlets' experiments with AI technology. Link.

• The deal is part of OpenAI's effort to collaborate with journalism companies on news and tech-sharing agreements.

• The funding, totaling $5 million, will be distributed via grants to ten of AJP's portfolio organizations, allowing them to experiment with AI in their newsrooms, products, and revenue teams.

• A new product studio within AJP will be established to support local news outlets in their AI experiments and act as a feedback loop between OpenAI and the local journalism industry.

• OpenAI will provide up to $5 million worth of credits to AJP's portfolio companies to access its tech products for experimentation.

• The news industry is eager to experiment with AI, but there are concerns that technology will not fully replace the work of journalists and that AI companies may withdraw funding in the future.

A research team has received a $600,000 grant from Australia's Office of National Intelligence to explore merging human brain cells with artificial intelligence. Link.

• The team, in collaboration with Cortical Labs, has already demonstrated the ability of brain cells in a Petri dish to play a game of "Pong."

• The goal is to merge biology with AI to advance machine learning technology for self-driving cars, drones, and delivery robots.

• The researchers claim the technology may eventually surpass existing silicon-based hardware in performance.

• The project aims to allow machine intelligence to learn throughout its lifetime, like human brain cells, retaining old skills while learning new ones and applying existing knowledge to new tasks.

• The team plans to grow brain cells in a lab dish using the DishBrain system to investigate "continual lifelong learning," but the project is expected to be highly ambitious and time-consuming.