AI Weekly: 02/13/23

Good morning and welcome to the fifth edition of AI Weekly! Every week I write this newsletter, Google, DeepMind, or OpenAI is launching a new product. While many questions about the overall landscape remain, what is clear is that the pace of AI innovation is picking up. Not only are we seeing new companies emerge, but companies with established market share like Quora and OKCupid are starting to integrate AI into their products. Thus, as incumbents become more capable of adapting, it is critical to consider the new markets that might also be created as a result of this emerging technology. Hopefully this past week’s news will help guide some of that thinking. Happy reading!

- ZG

Here are the most important stories of the week:

TEXT

The Bing app has seen a significant increase in downloads and has surged to the No. 12 spot in the US App Store's Top Charts. Link.

• Microsoft has recently announced plans to add AI-based capabilities to its search engine Bing.

• The update adds a ChatGPT-like experience within the search engine, allowing users to converse with an AI chatbot for complex queries.

• The new version of Bing is in limited preview and users must join a waitlist to try the new AI features.

• Microsoft's Edge browser has also been updated with AI features and has moved up to No. 3 in the US App Store's Top Charts for Utilities.

• The significant increase in Bing's app downloads indicates a high demand for AI experiences and poses a potential threat to Google, who is the default search engine in Safari.

Quora has launched its new AI chatbot app, Poe, to the public. Link.

• Poe allows users to ask questions and get answers from a range of AI chatbots, including those from OpenAI and other companies.

• The app is available only to iOS users but will soon be available on other platforms.

• Users must create an account verified with a phone number and email address to use the app.

• The app includes three AI chatbots with different areas of focus, all with limitations and the potential for inaccuracies.

• Quora plans to offer an API to make Poe more accessible to AI developers, and the app also includes social components such as the ability to create a profile, follow others, and publish model output.

OkCupid, a dating app, is testing a new category of match questions generated by the OpenAI ChatGPT chatbot. Link.

• The app's match questions let users define themselves and show their compatibility with potential matches.

• The chatbot was asked to generate questions for people to ask when dating.

• Six questions generated by the chatbot have been answered over 135,000 times.

• The majority of respondents consider themselves introverts, prefer nighttime, prefer hanging out with friends on the weekends, and value trust in a partner.

Macro, an AI-driven PDF editor, has raised $9.3 million in a seed round led by Andreessen Horowitz with participation from several other investors. Link.

• Macro is a custom PDF editor that uses AI to make documents interactive and hyperlinked.

• The company was founded by Jacob Beckerman, a former investment logic engineer at Bridgewater Associates, who was frustrated with the limited capabilities of standard document apps.

• The software offers a range of features, including cross-reference previews, a filterable side panel of detected errors, and the ability to consolidate edits from multiple files.

• Macro is focused on serving power users in industries such as finance, law, and sales and aims to solve document management problems such as version control and information chaos.

• The company's upcoming plans include the release of a word-processing tool and an inter-company communication tool, with a focus on tailored solutions for different verticals.

Google has invested $300 million in Anthropic, a company founded by former OpenAI employees, for a 10% stake. Link.

• Anthropic has announced that Google Cloud is its "preferred cloud provider" and they will "co-develop AI computing systems".

• The company is developing its own chatbot named Claude, which could potentially rival OpenAI's ChatGPT.

• Anthropic was founded by OpenAI employees who disagreed with the company's direction towards a commercial focus.

• The company prioritizes building "reliable, interpretable, and steerable AI systems.”

Character AI, a company specializing in AI chatbots based on famous figures, is seeking to raise $250 million in funding. Link.

• The funding would value the startup at $1 billion or more.

• Character is in early-stage talks with Sequoia Capital and other top venture-capital firms.

• The company was founded by former Google researchers Noam Shazeer and Daniel De Freitas, who worked on Google's LaMDA AI conversational technology.

• Character's website offers various AI chatbots based on the LaMDA system, allowing users to chat with virtual versions of famous figures like Queen Elizabeth, Elon Musk, William Shakespeare, Donald Trump, and Walter White.

• The company's technology, which is currently free while the startup analyzes user interactions, is considered a competitor to ChatGPT.

• The $250 million it wants to raise is considered a large sum for a startup that has only released a beta product.

IMAGE/VIDEO

Google has announced Dreamix, its AI video generator and editor. Link.

• Dreamix uses diffusion-based technology to modify existing videos with text descriptions or generate new videos from a single image.

• The AI video editor works by adding noise to the source images and passing them to a video diffusion model.

• The model then generates new images based on text prompts and combines them into a video.

• Dreamix provides a flexible solution for video editing and generation, as the source images act as a sketch, allowing for changes to be made.

• Dreamix is not limited to just editing existing videos and can also generate new videos from a single image by making slight changes to generate additional images and then applying them to the video model.

GlossAI recently raised $8 million in a seed round led by New Era Capital Partners, featuring participation from several venture capital firms and private investors. Link.

• GlossAi is a startup that uses AI to automate the process of making content more accessible and shareable across various platforms and formats.

• The company distinguishes itself from other automatic editing and snipping tools by offering a middle ground between purely generated content and intelligent clipping of existing video.

• GlossAi's approach is to take long-form content and turn it into multiple short clips, blog posts, e-books, social media content, and more.

• The company focuses on "verbal content-rich" media, such as webinars, presentations, and interviews, to get more ROI from production costs.

• GlossAi's technology can help corporations and other content producers reach a wider audience, but there are concerns that it could lead to increased spam.

"Nothing, Forever," an AI version of the TV show Seinfeld, has been banned from Twitch for 14 days. Link.

• The ban was caused by Larry Feinberg, a clone of Jerry Seinfeld, making transphobic statements during a stand-up bit.

• The AI for "Nothing, Forever" is trained on classic sitcom episodes and various AI tools, including OpenAI's GPT-3 and Curie models.

• The ban occurred after a switch from the GPT-3 model to the less sophisticated Curie model caused "errant behaviors" and generated inappropriate text.

• The incident highlights the challenges faced by AI when it is trained on biased or hateful material.

• The developers of "Nothing, Forever" are appealing the ban and are working to ensure that similar incidents don't occur in the future.

Galileo AI is an AI platform for creating high-fidelity UI designs. Link.

• It uses text descriptions as inputs and generates UI designs in a fast manner.

• The platform is trained on thousands of designs for better accuracy.

• It can also populate designs with AI-generated illustrations and images.

• Galileo AI uses large language models to understand complex context and generate end-to-end product copy.

• It offers a convenient solution for creating complex UI designs with minimum effort.

Netflix Japan has faced backlash from creating a short anime film using AI-generated artwork for its backgrounds. Link.

• The project, called "The Dog & The Boy," was created in response to a supposed labor shortage in the anime industry.

• The announcement of the use of AI-generated artwork was met with criticism and outrage from people who felt that Netflix was using AI to avoid paying human artists.

• The short film was created by the Netflix Anime Creators Base, in collaboration with Rinna Inc. and WIT Studio.

• There has been a growing demand for new anime productions, but the industry has been plagued with labor abuses and low wages for artists.

• As the use of AI art becomes more prevalent, artists are pushing back against their work being co-opted by algorithms.

SPEECH/AUDIO

Google has developed SingSong, a system that generates instrumental music to accompany input vocals. Link.

• The system builds on recent advancements in source separation and audio generation technology.

• Source-separated (vocal, instrumental) pairs are produced using a state-of-the-art source separation algorithm on a large corpus of music audio.

• AudioLM, a state-of-the-art unconditional audio generation approach, is adapted for conditional "audio-to-audio" generation and trained on the source-separated pairs.

• Featurization strategies are explored to improve the system's generalization to isolated vocals from users.

• In a listening test, listeners showed a preference for instrumentals generated by SingSong compared to a strong retrieval baseline.

3D/ROBOTICS

Robotic painter, FRIDA (Framework and Robotics Initiative for Developing Arts), is a thought experiment and research project exploring the intersection of human and robotic creativity. Link.

• FRIDA uses AI models to generate artistic output and requires human input and actual paint mixing to make it effective.

• The system's imprecision is intentional, as it allows the robot to make mistakes and adjust the remainder of the painting accordingly.

• Speed is not a priority, and each painting takes several hours to complete.

• The project is being run by the Carnegie Mellon University Robotics Institute.

• The artwork generated by FRIDA can be viewed on a CMU-run Twitter account.

San Francisco transportation officials are urging Waymo and Cruise to slow down their expansion of autonomous vehicle services due to safety concerns. Link.

• The two companies are currently the only ones permitted to offer driverless rides in the city.

• The recent incidents involving autonomous vehicles obstructing traffic and emergency responders have raised safety concerns among city officials.

• The San Francisco Transportation Authority supports the expansion of autonomous technology, but wants more transparency and safeguards to be put in place.

• The officials are calling for companies to collect more data about the vehicles' performance and restrict autonomous vehicle operations during peak travel hours.

• Cruise aims to operate its paid robotaxi service in San Francisco 24/7, while Waymo has yet to receive an additional permit from the California Public Utilities Commission to charge for its rides.

HEALTHCARE

As of January 2023, there are over 520 FDA-cleared artificial intelligence (AI) medical algorithms available in the US, with most related to medical imaging. Link.

• The number of FDA-cleared algorithms has increased rapidly in the past decade, with more than half of the algorithms on the US market cleared between 2019 to 2022.

• AI and ML technologies have the potential to transform healthcare by providing new insights from healthcare data.

• The majority of FDA-approved algorithms are related to medical imaging, which is due to AI's high accuracy in pattern recognition.

• AI is expanding into non-clinical areas of healthcare, such as in IT systems for sorting patient data and searching for population health traits.

• AI is becoming more widely adopted in healthcare, with a growing role in data analytics and various applications in medical imaging and other areas of healthcare.

VinBrain, a health-tech startup based in Vietnam that aims to improve patient outcomes using AI diagnostics, expands to over 100 hospitals. Link.

• The company's AI platform, DrAid, uses multimodal data analysis to diagnose patients based on their symptoms, laboratory and radiology reports, medical history, and medical images.

• DrAid is currently used in over 100 hospitals across several countries, including the US, New Zealand, Myanmar, and Vietnam, and analyzes medical images for over 120,000 patients every month.

• VinBrain has developed over 300 AI models to process speech, text, video, and images, including X-ray, CT, and MRI data.

• VinBrain trains its AI models using NVIDIA DGX SuperPOD and adopts NVIDIA GPUs to improve runtime efficiency and deployment.

• The company's goal is to become the top company in multimodality in healthcare data, improving the quality and accessibility of healthcare worldwide.

POLICY

The US Department of Commerce’s National Institute of Standards and Technology (NIST) has released the Artificial Intelligence Risk Management Framework (AI RMF 1.0) to help organizations manage the risks of AI technologies. Link.

• The AI RMF is a voluntary guidance document that is intended to adapt to the evolving AI landscape and be used by organizations to maximize the benefits of AI while protecting against potential harm.

• AI poses different risks than traditional software due to its training on data that can change over time and its influence by societal dynamics and human behavior.

• The framework promotes a change in institutional culture by encouraging organizations to approach AI with a new perspective and to think about, communicate, measure, and monitor AI risks.

• The AI RMF provides a flexible, structured, and measurable process to address AI risks and to help organizations operationalize trustworthy and responsible AI.

• The framework is part of NIST’s larger effort to cultivate trust in AI and to develop additional guidance for responsible AI practices. NIST plans to update the framework periodically and will launch a Trustworthy and Responsible AI Resource Center to help organizations implement the AI RMF.

OTHER

Ushur, a start-up that focuses on automating customer experiences, has raised $50 million in a Series C round led by Third Point Ventures. Link.

• The company provides dashboards that allow users to build AI-powered workflows for tasks such as claims processing, customer support, and appointment scheduling.

• Ushur targets enterprises in heavily regulated industries such as insurance, healthcare, financial services, and banking.

• The startup claims to be differentiated by its dedication to plug-and-play, self-served customer experiences, allowing IT teams to quickly deploy the platform with in-house citizen developers and IT teams.

• Ushur has a strong customer base, with over 50% of clients being Fortune 500 companies and serving three of the world's top six healthcare companies.

• Ushur plans to expand into the public sector within the next 12 months, including pursuing military contracts.

Bret Taylor, former co-CEO of Salesforce, and Clay Bavor, a long-time veteran of Google, are starting a new company together in March. Link.

• Bavor and Taylor share an obsession with recent advances in AI and aim to build a new company to apply AI to solve important business problems.

• Bavor was the Vice President of AR and VR at Google for almost 7 years and led various product and design teams including Gmail, Google Docs, Google Apps for Enterprise, and advertising and search products.

• Taylor has a background in entrepreneurship, founding FriendFeed in 2007, which was later purchased by Facebook, and Quip, an enterprise document collaboration company that was acquired by Salesforce in 2016.

• Bavor recently announced cuts to Google’s R&D projects within the Google Labs group due to Alphabet’s extensive layoffs.

• The nature and focus of the new company is not yet known but it should be interesting to see what the duo will come up with.

Moonhub, an AI hiring platform, has raised a $4.4 million seed round led by GV with participation from Khosla Ventures, Day One Ventures, and AIX Ventures. Link.

• Moonhub is a company founded by Nancy Xu, who studied computer science at Stanford University.

• Moonhub's aim is to revolutionize the hiring process by using AI to identify the best candidates who might otherwise be missed by traditional methods.

• The platform analyzes billions of data points of individuals from the public web to build an AI layer for candidate analysis.

• It is launching an external-facing search tool later this year to allow users to find candidates using natural language queries.

• Moonhub already has dozens of customers and is growing with a diverse team.

GPT Tip of the Week

Prompt: What are the key considerations for creating an effective and efficient data collection process for training AI models?

Response: Here are six key considerations for creating an effective and efficient data collection process for training AI models:

1. Define your data requirements: Start by defining what types of data you need to train your AI model. This will help you determine the most appropriate data sources to collect and the amount of data you need to achieve the desired results.

2. Quality over quantity: High-quality data is more important than large amounts of low-quality data. Ensure that your data is accurate, relevant, and up-to-date, as well as free of any biases or inconsistencies that could negatively impact your model's performance.

3. Data privacy and security: Make sure you have a clear understanding of the privacy and security implications of collecting and storing sensitive data, such as personal information. Ensure that you have the proper permissions and security measures in place to protect this data.

4. Data annotation: Depending on the type of AI model you are building, you may need to manually annotate your data with labels or other relevant information. Make sure you have a plan in place for this process and consider using crowdsourcing or other techniques to make the process more efficient.

5. Data storage and management: Store your data in a secure, centralized repository that makes it easy to access and manage. Consider using a cloud-based solution that can scale as your data needs grow.

6. Continuously monitor and evaluate your data collection process: Regularly review your data collection process to identify any areas for improvement, such as increasing data accuracy or reducing the time required to collect data. Use this information to refine your process and ensure that you are collecting the most relevant and useful data possible.