AI Weekly: 09/11/23

Good morning and welcome to this week’s edition of AI Weekly! In this week’s news, Anthropic launched their ChatGPT Plus competitor, Claude Pro, that includes higher usage capabilities and early access to new features.

More news has come out surrounding Apple’s AI plans, as the long-time tech juggernaut is reportedly building foundation models for text and image/video modalities. While Apple has been a bit slower to the party than players like OpenAI, Google, and others, their market share over mobile devices may give them a competitive edge as they integrate LLMs natively into their products.

In healthcare news, Microsoft is partnering with digital pathology provider Paige to create the world's largest image-based AI model for cancer identification, training on billions of images to diagnose both common and rare cancers.

Continue reading below to learn more about the exciting AI developments from last week!

- ZG

Here are the most important stories of the week:

TEXT

Anthropic has launched Claude Pro, a paid version of its generative AI chatbot Claude, priced at $20 per month—identical to ChatGPT Plus. Link.

• Claude Pro offers users 5x more usage compared to the free version, the ability to send more messages, priority access during high-traffic periods, and early access to new features.

• Claude is powered by Anthropic's large language model, Claude 2, and is positioned as a competitor to ChatGPT.

• Anthropic was founded by former OpenAI senior employees and positions itself as a more ethical and responsible alternative to OpenAI.

• The company has conducted extensive research on safely training generative AI models and has articulated principles for responsible AI development.

• With heavy investment from Google and a focus on AI ethics, Anthropic is seen as a strong contender that could potentially displace OpenAI as a leader in the AI space.

Salesforce announced it is preparing to pilot new AI features for Slack this winter, collectively called Slack AI, which are built natively into the platform for easy access. Link.

• One key feature of Slack AI is the ability to generate channel highlights, summarizing the most important parts of conversations for users to catch up or use in reports.

• Slack AI can also produce one-click summaries for specific conversation threads, providing not just highlights but also suggested next steps based on the conversation.

• In addition to conversation summaries, Slack AI enhances the platform's search functionality by wrapping up messages, files, and channels with keywords into a summarized package for easier parsing.

• Salesforce revealed these updates nearly a month after introducing a redesigned interface for Slack, aimed at reducing clutter and improving focus, featuring a new sidebar and dedicated tabs for various functionalities like DMs and Activity.

• More details about Slack AI and other new products are expected to be unveiled at Salesforce's Dreamforce event, taking place from September 12 to 14 and available for streaming on Salesforce+.

Apple is reportedly investing millions of dollars per day in artificial intelligence, with multiple teams working on various AI models. Link.

• The company's conversational AI unit, called "Foundational Models," is led by John Giannandrea and includes several former Google engineers. Giannandrea has expressed skepticism about the utility of AI-powered chatbots.

• In addition to conversational AI, Apple has a Visual Intelligence unit focused on image generation and another team researching "multimodal AI," which can handle both text and visual data.

• Apple is developing a chatbot for AppleCare customer interactions and is also working on automating multistep tasks with Siri.

• Apple's most advanced language model, known internally as Ajax GPT, has been trained on "more than 200 billion parameters," making it reportedly more powerful than OpenAI's GPT-3.5.

• Despite its advanced capabilities, Ajax GPT is primarily for internal use and remains restricted within the company, according to previous reports.

Intuit has launched Intuit Assist, an AI assistant integrated into its financial software products like TurboTax, QuickBooks, and Mailchimp, designed to provide personalized answers and insights to users. Link.

• The assistant uses generative AI to understand natural language queries and provide clear, accurate responses, and can perform tasks like creating invoices, sending reminders, and optimizing marketing campaigns.

• Intuit Assist appears as a sidebar in the user interface and can also connect users to human experts for additional support.

• The launch is the culmination of years of R&D and represents Intuit's significant investment in generative AI technology to enhance value for its 100 million small business and consumer customers.

• Intuit has developed an orchestration layer called GenOS, which acts as an operating system for AI, and allows for the injection of real-time, customer-specific data into AI models, making the assistant context-aware and consistent across all Intuit products.

• The assistant aims to simplify tasks for users, as demonstrated by an example where a user could ask questions like "How are my tenants paying their rent?" or "How much profit did I make last month?" and receive instant, accurate answers.

ChatGPT's worldwide monthly website visits have been declining for three months, and average time spent on the platform has decreased from 8.7 minutes in March to 7 minutes in August, according to Similarweb data. Link.

• The rate of decline in total visits leveled off in August, dropping by only 3%, and in the US, visits even increased slightly by 0.4%.

• The number of unique visitors also saw a small increase in August, from 180 million to 180.5 million.

• A theory suggests that the decline over the summer could be attributed to school being out, and that traffic stabilized in August as students returned to school.

• ChatGPT was initially the fastest-growing app ever, reaching 100 million users in two months, before being surpassed by Meta's Threads.

• Users and academic papers have reported a decline in GPT-4's performance, with significant drops in task accuracy, such as identifying prime numbers.

IMAGE/VIDEO

Zoom is integrating a new AI assistant, called AI Companion, into its video conferencing software to maintain its industry leadership and enhance meeting productivity. Link.

• The AI Companion will offer features like meeting highlights, summaries, and a chatbot for questions. These features will be available for free to all paid Zoom users.

• Zoom is using its own proprietary AI models for this feature, in addition to models from Meta's Llama 2, OpenAI, and Anthropic.

• Amid previous concerns about using customer conversations to train AI, Zoom clarified that the new AI features will be off by default, and account owners and IT administrators will have the option to enable them.

• Competitors like Google and Microsoft are also integrating AI into their products. Google's Duet AI bot and Otter's OtterPilot bot can attend and record meetings, while Microsoft Teams offers similar functionalities.

• Starting next year, Zoom plans to introduce additional AI features that allow users to prepare for meetings by searching through content across the platform, including past meetings, chats, and select third-party apps.

Artisse is the newest AI photo creation app to gain steam, rivaling competitors like Remini by offering more input and output flexibility and greater realism in generated photos. Link.

• Users upload 15 selfies to train the AI, which can then create new photos based on text or image prompts.

• The app allows users to modify generated images by changing styles or adding additional prompts, and aims to be more inclusive by working on flexibility in terms of body shape and skin tones.

• Founded by William Wu, who has a background in investment and strategy, Artisse aims to democratize the creation of "perfect" personal photos, making them accessible to people regardless of their expertise or resources.

• The app takes longer to process images (30-40 minutes for training the AI and additional minutes for generating photos), but claims to offer superior realism compared to competitors.

• Artisse offers the first 25 photos for free and plans to charge around 20 cents per photo afterward.

• A subscription model offering HD and sizing features is planned for the next release. The company also offers end-to-end consulting services for B2B clients.

• Currently bootstrapped and based in Hong Kong, Artisse is in negotiations for venture capital funding and plans to release a web version next month.

SPEECH/AUDIO

Otter.ai has expanded its offerings beyond audio transcription to introduce OtterPilot for Sales, an AI assistant aimed at helping sales teams close deals and provide coaching at scale. Link.

• The new tool can be used during sales calls to automatically generate transcriptions and extract various Sales Insights, such as next steps, action items, and key metrics like BANT and MEDDPIC.

• These insights are automatically integrated into CRM and productivity tools like Salesforce and Hubspot, providing real-time data and insights to the sales team.

• Greg Holmes, Otter.ai advisor and former Zoom CRO, highlighted that sales reps spend less than a third of their time actually selling, and OtterPilot aims to make that time more productive by automating data capture and analysis.

• OtterPilot for Sales will be available exclusively for enterprise customers, expanding Otter.ai's market reach.

• In addition to OtterPilot, sales teams can also utilize Otter AI Chat, which acts as a virtual meeting participant, providing immediate feedback and generating meeting-specific content.

CODE/DEVTOOLS/INFRA

OpenAI is hosting its first developer conference, OpenAI DevDay, in San Francisco on November 6, offering a platform for developers to discuss ideas and preview new tools and technologies. Link.

• The event will include breakout sessions led by OpenAI's technical staff and will feature a keynote address, although many details about the itinerary and location are still undisclosed.

• DevDay is primarily an in-person event, but the keynote address will be available for online streaming. Registration for in-person attendance is available on OpenAI's website with limited spots.

• Sam Altman, CEO of OpenAI, expressed excitement about showcasing the company's latest work to enable developers to build new things.

• OpenAI currently has a user base of more than two million developers who are utilizing its generative AI technologies for app development and enhancements.

• Ethical considerations surrounding the integration and innovation of generative AI are expected to be a topic of discussion during the event.

HEALTHCARE

Microsoft is partnering with digital pathology provider Paige to create the world's largest image-based AI model for cancer identification, training on billions of images to diagnose both common and rare cancers. Link.

• Paige specializes in AI solutions for pathologists, who play a crucial role in diagnosing cancer. The company aims to modernize the traditional workflow of pathologists to improve accuracy and efficiency.

• Paige has FDA approval for its FullFocus viewing tool, which allows pathologists to examine digital slides. It also has an AI model for identifying breast, colon, and prostate cancer, and is the only company with FDA approval for AI as a secondary tool in prostate cancer diagnosis.

• The new AI model, built in collaboration with Microsoft, is "orders of magnitude larger" than existing models, training on 4 million slides.

• Microsoft's cloud storage and supercomputing infrastructure are key to the partnership, but the tech giant is also contributing to the development of new algorithms, detection methods, and diagnostics.

• While the technology is still under development and peer review, it aims to solve storage issues for health systems and expedite the diagnostic process, potentially reducing patient wait times for diagnoses from weeks to days.

Glass Health was founded by Dereck Paul and Graham Ramsey to initially serve as a "personal knowledge management system" for medical professionals but later pivoted to use generative AI for diagnosis and treatment recommendations. Link.

• The AI tool takes physician-inputted patient summaries and suggests 5-10 possible diagnoses along with a clinical plan, also providing editable case assessment paragraphs for medical records.

• The company aims to alleviate healthcare provider burnout and improve medical practice, securing $1.5 million in pre-seed funding and acceptance into Y Combinator’s Winter 2023 batch.

• Despite its promise, generative AI in healthcare has faced scrutiny for inaccuracies and ethical concerns, including potential biases and misleading or outdated information.

• Glass Health employs a team of academic physicians to create and review clinical guidelines, positioning its AI tool as an assistant to clinicians rather than a replacement, partly to avoid regulatory issues.

• With over 59,000 users, the company plans to pilot an HIPAA-compliant, electronic health record-integrated offering and has secured a total of $6.5 million in funding for further development and research.

POLICY/LAW/ETHICS

Microsoft announced the 'Copilot Copyright Commitment,' pledging to cover the legal costs of customers who face copyright infringement claims while using Microsoft's Copilot services, provided they use built-in guardrails and content filters. Link.

• Brad Smith, Microsoft's vice chair and president, justified the commitment by stating that since Microsoft charges for Copilot, it should assume responsibility for any legal issues.

• Artificial intelligence models like ChatGPT and Meta's LLaMA have faced lawsuits for training on copyrighted material without the consent or compensation of the original creators, such as authors and artists.

• A report by The Atlantic revealed that LLaMA was trained on over 170,000 books without the authors' knowledge, permission, or compensation.

• Nell Watson, a prominent AI researcher, views Microsoft's move as a step towards professionalizing the Large Language Model space and addressing intellectual property concerns, potentially increasing the technology's accessibility.

OTHER

Imbue, formerly Generally Intelligent, has raised $200 million in Series B funding from investors that include the Astera Institute, Nvidia, and executives from Cruise and Notion. Link.

• The company aims to develop AI systems capable of robust reasoning and coding, with the goal of creating practical AI agents that can work safely in real-world scenarios.

• Initially focused on researching the fundamentals of human intelligence through tasks in complex 3D worlds, Imbue has shifted its approach to developing models that are "internally useful," including those that can code.

• Imbue sets itself apart by emphasizing the importance of "robust reasoning" in AI, which involves dealing with uncertainty, making decisions, and adapting to real-world complexities.

• The company is training very large models with over 100 billion parameters on a compute cluster co-designed by Nvidia, focusing on optimizing these models for reasoning tasks.

• While not aiming to productionize its current work, Imbue is laying the groundwork for future, more general-purpose AI systems. It is also developing its own AI tooling, including debugging prototypes and visual interfaces, to facilitate the creation of custom models by users in the future.

Ello, a startup focused on eradicating childhood illiteracy using artificial intelligence and child speech recognition technology, recently secured $15 million in Series A financing to support product development and expand access to consumers. Link.

• Ello offers a subscription-based service for children from kindergarten to Grade 3, delivering five books monthly for $24.99.

• The Ello app determines a child's reading level and interests, and reading experts hand-select books for them.

• Ello's proprietary technology listens to children read aloud, analyzes their speech for corrections, and employs phonics-based strategies to teach reading skills.

• The company plans to partner with schools, develop classroom-specific products, integrate generative AI, and launch versions in other languages, such as a German version.

Retool, initially a platform for line-of-business apps, has expanded to offer back-end services and workflow automation. Link.

• The startup is introducing tools for building AI-based apps, including a hosted vector store to add context to large language models.

• Many Retool customers want to incorporate AI into their apps for data analysis, but fine-tuning models with all production data is challenging.

• The company launches Retool Vectors, a hosted vector storage service to simplify the integration of custom data into AI models.

• Retool has also added AI-based actions for tasks like text summarization and image generation in partnership with OpenAI.

• Retool uses its Workflow service to keep production and vectorized databases in sync for up-to-date information.

Hiber, based in Sweden, has integrated generative AI into its 3D world creation tools, enabling users to build worlds with natural language prompts. Link.

• The integration was unveiled during the Google Cloud Next event and is aimed at making 3D web participation more accessible.

• Hiber's existing tools, including Hiber3D Development Kit and a no-code tool, have already led to the creation of over five million 3D worlds on HiberWorld.

• The company leverages Google Cloud services and AI algorithms to lower the costs associated with generative AI.

• Hiber's collaboration with Datatonic and Google Cloud allows users to generate 3D worlds using natural language prompts, streamlining the creative process.

• The long-term roadmap includes the introduction of an AI co-pilot for user assistance and expanding revenue streams through 3D shops for avatars.

Kindo has raised a $7 million seed round led by Riot Ventures to launch its AI productivity platform for businesses. Link.

• The platform allows companies to integrate various AI models, including commercial, open source, or proprietary, into their workflows to improve productivity while maintaining security, compliance, and centralized governance.

• Along with Riot Ventures, the round included Eniac Ventures, RRE Ventures, Marlinspike Partners, Flexcap Ventures, and individual investors like Scooter Braun, Andrew Peterson, and Dave Politis.

• Kindo was founded in September 2022 by industry veterans from tech companies such as Bird Rides, Clover Health, Riot Games, Google, OpenAI, and Square.

• The platform offers visibility and control over AI usage, connects to over 200 SaaS applications, and provides access to various AI models and assistants like OpenAI's ChatGPT and Google Bard.

• Kindo aims to help companies leverage AI safely and responsibly and plans to scale its operations with the seed funding to revolutionize AI integration into workflows.

Automat, a platform that translates videos or descriptions into software automation workflows, recently raised $3.75 million in equity financing to expand and support the launch of new features. Link.

• Some of these new features include a Sandbox feature and standalone document and image data extraction APIs.

• Lucas Ochoa and Gautam Bose, former Google Creative Lab employees, created Lasso, which later rebranded to Automat, to simplify enterprise automation using AI.

• Automat translates videos or descriptions into software automation workflows, enabling it to automate various PC workflows.

• The platform offers two versions: unattended (cloud-based) and attended (works alongside users).

• Customers submit a video or description of the process they want to automate, and Automat builds the automation.