AI Weekly: 09/04/23

Good morning, happy Labor Day, and welcome to this week’s edition of AI Weekly! In this week’s news, OpenAI launched ChatGPT Enterprise to allow businesses to securely use their staple LLM. Meanwhile, AI21 Labs, a Tel Aviv-based startup on the foundation layer, has raised $155 million in a Series C funding round led by Walden Catalyst.

In the model fairness and bias realm, Meta has released an AI benchmark called FACET that aims to address biases and disparities present in AI models and encourages researchers to benchmark fairness in vision and multimodal tasks.

Also, DoorDash is introducing AI-powered voice ordering technology for restaurants to answer customer phone calls and provide curated recommendations. Enjoy reading more about last week’s AI news below!

- ZG

Here are the most important stories of the week:

TEXT

OpenAI is launching "ChatGPT Enterprise" to tap into the growing AI adoption trend in corporate America. Link.

• ChatGPT Enterprise promises "enterprise-grade security and privacy" and a powerful version of ChatGPT for business clients seeking to leverage generative AI.

• The new offering aims to provide AI assistance for various tasks, customized for organizations, and safeguarding company data.

• Early adopters of ChatGPT Enterprise include Block (fintech startup), Estee Lauder Companies (cosmetics giant), and PwC (professional services firm).

• OpenAI's announcement comes amid the usage of ChatGPT by employees from over 80% of Fortune 500 companies, prompting concerns about data privacy and security.

• ChatGPT Enterprise addresses these concerns by not training on business data or conversations and ensuring models don't learn from usage. Pricing details are available through the sales team.

Researchers from the semiconductor blog SemiAnalysis claim Google's upcoming AI model, Gemini, will outperform OpenAI's GPT-4 due to Google's extensive GPU infrastructure. Link.

• The assertion led to heated online debates about whether more computing power equates to a better AI model.

• OpenAI CEO Sam Altman responded on X-formerly-Twitter, mocking the claims made in the blog post and referring to it as "internal marketing/recruiting chart."

• One of the blog authors, Dylan Patel, posted a meme in response, highlighting Google's alleged dominance in AI.

• Debates centered on the complexity of AI model performance, which involves factors like training process, data quality, and tasks.

• While the arguments continue, Google's history of AI research and development, along with substantial resources, suggest Gemini could potentially introduce noteworthy advancements.

Google is expanding its AI-powered search experience, known as Search Generative Experience (SGE), to countries outside the U.S., starting with India and Japan. Link.

• SGE introduces conversational mode to Google Search, allowing users to ask questions and receive AI-generated answers, similar to a chatbot.

• The feature has been updated with support for videos, images, local information, travel recommendations, summaries, definitions, coding queries, and integrated ads.

• The expansion brings localized customizations: In Japan, it supports local language, while in India, it supports English and Hindi with voice input functionality.

• A new feature in SGE allows users to click on an arrow icon next to AI-generated information to access relevant web pages backing up the response.

• SGE has gained popularity among younger users, with high satisfaction scores among those aged 18-24, who prefer conversational-style questioning.

IMAGE/VIDEO

Google DeepMind, in collaboration with Google Cloud, introduces SynthID, a tool for watermarking and identifying AI-generated images, specifically from Google's image-generating model. Link.

• SynthID embeds a digital watermark within image pixels, invisible to the human eye but detectable by algorithms. It supports Imagen, Google's text-to-image model, available exclusively in Vertex AI.

• SynthID aims to address concerns of AI-generated content's authenticity, helping to prevent the spread of misinformation and enabling identification of AI-generated media.

• The tool combines two AI models for watermarking and identification, surviving modifications like filters, color changes, and compression.

• While not foolproof, SynthID offers a technical approach to responsibly handle AI-generated content and could extend to other media types beyond images.

• Efforts to standardize watermarking for AI-generated content are ongoing, with regulations and initiatives from tech firms like Microsoft, Shutterstock, and OpenAI to enhance transparency and authenticity.

SPEECH/AUDIO

DoorDash is introducing AI-powered voice ordering technology for restaurants to answer customer phone calls and provide curated recommendations. Link.

• The move comes in response to customer preferences for ordering takeout via phone, but a significant portion of calls go unanswered, leading to potential revenue losses.

• DoorDash's new system aims to increase sales by answering all customer calls and capturing unmet demand.

• The technology combines AI with live agents to ensure customer calls are answered quickly, even during peak times.

• The system offers personalized voice ordering experiences in multiple languages, quick reordering for returning customers, and the option for live agents to assist as needed.

• Restaurants can also leverage DoorDash Drive, the company's white-label solution, for facilitating delivery of phone orders and providing end-to-end order tracking for customers.

CODE/INFRA

Google's Duet AI, a collection of generative AI features, is expanding to new products and services within Google Cloud. Link.

• Duet AI, still in preview, can assist with code refactoring by making small changes without altering overall external behavior.

• Developers can use a Duet AI-powered chat window in their preferred software development environment to execute natural language prompts for code improvements.

• Duet AI can provide "how to" information about infrastructure configuration, deployment suggestions, cost, and performance optimization in the Google Cloud Console.

• Duet AI is integrated into Cloud Workstations for writing code with best practices, and Application Integration for generating flows using existing APIs and creating documentation.

• Customization of Duet AI is possible for select enterprises, allowing the integration of organization-specific knowledge to generate context-aware code suggestions.

• Duet AI is also being integrated with other Google services such as Apigee, BigQuery, Looker, AlloyDB, Cloud SQL, and Cloud Spanner to enhance various capabilities.

Arize AI, a B2B machine learning software provider, has launched industry-first capabilities to optimize the performance of LLMs deployed by enterprises. Link.

• The new capabilities include a "Prompt Playground" for selecting and iterating on prompts, as well as a retrieval augmented generation (RAG) workflow to help organizations understand what data to include in LLM responses.

• Prompt analysis and iteration workflows allow teams to uncover poorly performing prompt templates, iterate on them, and verify improved LLM outputs before deployment.

• Arize also offers insights into the private or contextual data that influences LLM outputs, helping teams understand the "secret sauce" that influences prompts.

• The company's solutions can be deployed on premises for security reasons and are SOC-2 compliant, providing end-to-end observability and troubleshooting to help organizations optimize LLMs post-deployment.

IBM and Salesforce are partnering to deliver AI solutions to customers. Link.

• Salesforce brings popular CRM software and AI apps (Sales GPT, Service GPT, Salesforce Einstein, Slack GPT, and Marketing GPT).

• IBM provides industry expertise and innovative delivery models through its IBM Consulting arm of 160,000 consultants.

• The partnership includes using IBM Garage, an operating model for business transformation, to integrate Salesforce AI solutions.

• Customers can also adopt IBM's WatsonX enterprise AI platform for finding and fine-tuning AI models, as well as Data Classifier for mapping internal data.

• The collaboration aims to help businesses connect with customers on a new level using AI, data, and CRM technologies.

Context.ai, a platform that has developed a service to help companies understand and measure the performance of LLMs like ChatGPT, has raised a $3.5 million seed round co-led by Google's venture arm, GV, and Theory Ventures. Link.

• The startup aims to provide insights into how users are interacting with LLMs and how well the models are performing in providing accurate and useful responses.

• Context.ai analyzes chat transcripts generated by LLMs using NLP and groups conversations by topic to determine user satisfaction.

• The startup offers a way for companies to assess the effectiveness of their LLMs in delivering customer support and answering queries.

• The company ensures security and privacy by stripping out personally identifiable information (PII) and not using the content for model building or marketing purposes.

Superframe, an AI-powered software company, has raised $5 million in seed funding from over 40 angel investors, including data and AI experts, Salesforce consultants, and general operating experts. Link.

• The company has launched its first product, an AI assistant for managing complex Salesforce implementations, aimed at saving companies time and money by streamlining configuration changes.

• Accuracy is highlighted as Superframe's key differentiator in the AI market, aiming to build customer trust through reliable and unique offerings.

• Superframe plans to address pain points in go-to-market tools, such as Salesforce, Marketo, and HubSpot, by using OpenAI's language models to provide instant, accurate answers to users' queries and propose configuration changes.

• The goal is to help users navigate complex systems and business processes while enabling them to rely on their expertise and clearing backlogs.

• Superframe is in beta testing with selected customers and plans to launch publicly in early 2024, using the seed funding for product development and expanding its engineering team.

HEALTHCARE

QuantHealth, a Tel Aviv-based startup that has developed an AI-powered platform for drug discovery and clinical trial optimization, has raised $15 million in a Series A funding round, bringing its total raised to $20 million. Link.

• The platform utilizes extensive integrated datasets covering over 350 million patients and more than 700,000 biomedical graphs and clinical trials.

• QuantHealth's model can predict clinical trial outcomes with 86% accuracy on binary endpoint metrics, helping identify potential risks and optimize trial designs.

• The company aims to address the declining success rates of clinical trials in the pharmaceutical industry and improve the efficiency of drug development.

• QuantHealth collaborates with pharma, biotechs, clinical research organizations, and regulators in the U.S. and Europe, and plans to expand its team and platform capabilities with the funding.

POLICY/LAW/ETHICS

OpenAI responded to class-action lawsuits from authors alleging that ChatGPT was illegally trained on pirated copies of their books. Link.

• OpenAI asked the court to dismiss all claims except one alleging direct copyright infringement.

• OpenAI argued that authors misconceive copyright scope, failing to consider limitations and exceptions like fair use that accommodate innovations in AI language models.

• OpenAI claimed its use of copyrighted materials for innovation and transformative purposes does not violate copyright, contrasting with direct profit-seeking plagiarists.

• OpenAI defended that copyright protects expressions, not underlying ideas, facts, or building blocks of creative content, citing Google Books case.

• OpenAI also aimed to dismiss claims of DMCA violation, vicarious copyright infringement, and other allegations, highlighting contradictions and legal insufficiencies in authors' claims.

Meta has released an AI benchmark called FACET (FAirness in Computer Vision EvaluaTion) designed to evaluate the "fairness" of AI models in classifying and detecting things in photos and videos, particularly people. Link.

• FACET contains 32,000 images with 50,000 labeled people, including demographic and physical attributes, as well as occupation and activity labels, enabling evaluations of biases against different classes.

• The benchmark aims to address biases and disparities present in AI models and encourages researchers to benchmark fairness in vision and multimodal tasks.

• FACET is designed to provide deep evaluations of biases in AI models related to gender presentation, skin tone, hair attributes, and other characteristics.

• Meta claims that FACET is more thorough than previous benchmarks, as it can address specific questions about biases in AI models' classifications.

• The benchmark is based on images sourced from Segment Anything 1 Billion and was labeled by "trained experts" from various geographic regions.

Nearly 20% of the top 1000 global websites are blocking crawler bots used by AI services to gather web data, according to Originality.AI's data. Link.

• This self-initiated action by websites highlights the absence of clear legal or regulatory rules for AI's use of copyrighted content.

• OpenAI introduced the GPTBot crawler, which several high-profile news sites, including NY Times and Reuters, began blocking. The Common Crawl Bot is also blocked by around 6.77% of the top 1000 sites.

• Large language models and generative AI have rekindled debates about data scraping, as AI companies use crawlers to collect data for training and chatbots.

• Publishers are more aggressively blocking crawlers due to concerns about handing over data to AI companies without compensation.

• AI companies like OpenAI are seeing rapid commercialization, which media companies view with caution and are considering licensing their data to AI firms.

• The balance between embracing AI and resisting it poses challenges for news companies, particularly in terms of ethics and trust.

• If more websites block AI crawlers, AI products could face challenges in refining and updating their offerings.

X's updated privacy policy includes collecting biometric data, job history, and education history from users. Link.

• Another section of the policy reveals that X plans to use collected and publicly available data to train its machine learning and AI models.

• The policy change was noticed by Alex Ivanovs, known for finding updates in tech companies' terms of service.

• X owner Elon Musk has ambitions in the AI market with xAI, a separate company that aims to use public tweets for training AI models.

• Musk previously accused tech giants of using Twitter data for AI model training and filed suits against entities scraping Twitter data.

• Musk clarified that X will use just public data, not private messages, for training AI models, responding to a post on X.

OTHER

AI21 Labs, a Tel Aviv-based startup, has raised $155 million in a Series C funding round led by Walden Catalyst, with participation from Pitango, SCB10X, b2venture, Samsung Next, Amnon Shashua (Intel-owned Mobileye founder), Google, and Nvidia. Link.

• The funding brings AI21 Labs' total raised to $283 million and values the company at $1.4 billion.

• AI21 Labs, founded in 2017 by Amnon Shashua, Yoav Shoham, and Ori Goshen, focuses on text-generating AI tools and platforms.

• The company's flagship product is AI21 Studio, a developer platform for building text-based business apps using AI21's proprietary text-generating AI models.

• The company also offers Wordtune, a multilingual reading and writing AI assistant.

• AI21 Labs competes with companies like Google, AWS, Microsoft, OpenAI, and others in the generative AI space, and its solutions are developed on large language models, offering refined control and up-to-date training data for accurate results.

Baidu has launched ERNIE Bot to the public in various app stores and its website, focusing on the Chinese market. Link.

• Baidu aims to gather more human feedback to iterate and improve ERNIE Bot's user experience.

• ERNIE Bot can summarize novels, generate suggestions for expanding stories, create images and videos from text inputs, and handle data analysis and visualization.

• China now requires approval from authorities to release generative AI experiences, with Baidu among the first to receive approval, as it aligns with China's goal to control content while promoting competition.

• The US Copyright Office seeks public input on issues related to generative AI, including copyright eligibility of AI-generated content, liability for infringement, and defining human authorship for AI-created works.

• The office is addressing the challenge of determining the level of human authorship required to register copyrights on AI-generated content, with several cases hinting at boundaries. The public comment period is open until November 15th.

Floworks, an AI assistant startup, has raised a $1.5 million seed round led by Y Combinator and Sense AI with participation from Gaingels, Entrepreneur First and ThinKuvate. Link.

• Co-founders Sudipta Biswas and Sarthak Shrivastava developed Floworks to be an AI assistant enhancing worker experience by interacting with software tools used daily.

• The AI assistant supports applications such as HubSpot, Salesforce, Google Docs, Google Calendar, and Gmail.

• Users sign into their desired applications through the Floworks web app and interact with the assistant via Slack, using natural language to perform tasks.

• The assistant triggers approvals for actions, asking for additional information or clarification if needed.

• Floworks aims to differentiate itself by enabling interoperability across multiple applications and plans to expand revenue opportunities by partnering with channel partners.

Martian Lawyers Club, a platform that aims to enhance game personalization using generative AI, focusing on core game systems rather than content, has raised a $2.2 million pre-seed funding round led by Fly Ventures, with participation from System.One, Amar Shah, and Dhyan Ventures. Link.

• Co-founders Kamen Brestnichki and Levi Fussell met at the University of Edinburgh and are leveraging their expertise in machine learning, computer graphics, and game development.

• MLC's approach centers on creating games that offer interactive and dynamic experiences, responding to player inputs in real-time.

• The company plans to offer an SDK that provides a sandbox experience for developers to design game interactions while leveraging generative AI to create dynamic game code snippets.

• MLC is working on its first game, a collectible card game, and aims to advance game personalization by redefining traditional game development paradigms.

Refiberd, a startup addressing the issue of textile waste in the fashion industry by using hyperspectral cameras and AI to accurately sort textiles for recycling, has raised a $3.4 million seed round led by True Wealth Ventures, with participation from Better Ventures, the Schmidt Family Foundation, and others. Link.

• The fashion industry discards over 14 million tons of clothing each year, leading to environmental concerns. Chemical recycling of textiles has shown promise, but sorting and identifying different materials poses a challenge.

• Refiberd's approach involves using hyperspectral cameras to capture images of textiles, analyzing them with AI to differentiate between various materials and blends. They have created a sample library of over 10,000 entries and use generative AI to fill in gaps where data is lacking.

• The funding will be used to roll out initial pilots aimed at textile companies, chemical recyclers, mechanical recyclers, and textile sorters.

• Refiberd's technology has the potential to significantly reduce textile waste and contribute to environmental sustainability by improving textile recycling processes.