AI Weekly: 03/13/23

Good morning and welcome to this week’s edition of AI Weekly! Strangely, a lot of the news this week revolved around text-based modalities, so it was a quiet week on the speech and image fronts. Regardless, the AI world continues to move quickly. More and more well-established incumbents are continuing to release GPT-powered features to their products in an attempt to maintain market share, Google is of course on the verge of releasing a new model, and we should expect to see the release of GPT-4 in the coming days. Enjoy reading about the latest AI news!

- ZG

Here are the most important stories of the week:

TEXT

Discord is launching a set of new AI experiences to a number of servers, including ChatGPT-enabled technology that allows users to have extended conversations with the bot Clyde, as well as AI-generated conversation summaries. Link.

• The company is updating its AutoMod moderation tool to harness the power of large language models to alert moderators when server rules may have been broken while being aware of the conversation context.

• Discord is open sourcing Avatar Remix, an app that lets users remix each other's avatars through generative image models, and launching an AI incubator to fund developers and startups building AI on Discord.

• Discord is following in the footsteps of Snapchat and Slack, which recently integrated ChatGPT-like features into their products as well.

• Discord believes that its AI experiences will become a fundamental part of the Discord experience and is testing and iterating the product.

Google is progressing towards its goal of building an AI language model that supports 1,000 different languages. Link.

• The Universal Speech Model (USM) is a system that is described as a “critical first step” in realizing Google's language model goals.

• USM is a family of state-of-the-art speech models with 2 billion parameters trained on 12 million hours of speech and 28 billion sentences across over 300 languages.

• USM supports over 100 languages and will serve as the “foundation” to build an even more expansive system.

• One possible application for the technology could be inside augmented-reality glasses like the concept Google showed off during its I/O event last year.

• Google's misrepresentation of the Arabic language during I/O proves how easy it can be to get something wrong.

AI21 Labs, an Israeli startup, has released its latest generative AI model, Jurassic-2, after training on 178 billion parameters. Link.

• Jurassic-2 aims to offer more advanced instructions for highly customized results, as well as improve response times for generation by up to 30%.

• The company plans to integrate Jurassic-2 into its natural language processing (NLP)-as-a-service platform, AI21 Studio, and a series of APIs for developers to use in their custom applications.

• AI21 Labs aims to integrate semantic reasoning with statistical representation, enabling more guided and reliable results with generative AI.

• The Jurassic-2 LLM is available to developers via APIs and is part of AI21's products, including the Wordtune suite of services.

• AI21 Labs has a future model release that will carry forward the spirit of its advanced reasoning system, MRKL.

Salesforce and OpenAI have launched the ChatGPT app for Slack that integrates ChatGPT’s AI technology for delivering instant conversation summaries, research tools, and writing assistance in Slack. Link.

• The app empowers customers with the information they need to move work forward faster, combines knowledge found in Slack with the intelligence of ChatGPT, and is currently in beta.

• According to senior IT leaders, generative AI can help them better serve their customers, operate more efficiently, and combine public and private data sources.

• The ChatGPT app for Slack offers AI-powered conversation summaries to get up to speed faster, AI-powered research tools to find answers on any project or topic, and AI-powered writing assistance to draft messages in seconds.

• The app is built on top of Slack’s trusted platform, and will be available alongside more than 2,600 other integrations in the app ecosystem.

• OpenAI chose Slack as its productivity platform to accelerate employee productivity, connect directly with customers, and support the hyper-growth of the business, and has been piloting the app to engage with customers across sales, service, and engineering teams.

Salesforce announces Einstein GPT, a generative AI platform for CRM. Link.

• Einstein GPT creates personalized content across every Salesforce cloud with generative AI, making every employee more productive and every customer experience better.

• Einstein GPT is open and extensible – supporting public and private AI models purpose-built for CRM – and trained on trusted, real-time data.

• Einstein GPT will integrate with OpenAI to provide Salesforce customers with out-of-the-box generative AI capabilities.

• The new ChatGPT app for Slack integrates OpenAI’s advanced AI technology to deliver instant conversation summaries, research tools, and writing assistance.

• Salesforce Ventures’ $250 million Generative AI Fund will bolster startup ecosystem and development of responsible generative AI.

Privacy-focused search engine DuckDuckGo has launched a beta AI-powered summarization feature called DuckAssist, which can directly answer straightforward search queries for users. Link.

• The sources currently used for DuckAssist are mostly Wikipedia, but DuckDuckGo is experimenting with incorporating other sources to adapt sourcing to the context of the query.

• The feature is currently only available in English and can be accessed through DuckDuckGo's apps and browser extensions.

• The AI models currently being used to power natural language summarization are The Davinci model from OpenAI and the Claude model from Anthropic, with plans to experiment with the new Turbo model OpenAI recently announced.

• DuckAssist uses AI to generate new natural language responses based on specific/relevant sections of Wikipedia articles that DuckDuckGo provides via its own scanning of sources.

• Accuracy is a key concern with generative AI, so DuckAssist has been designed to limit the sources it is summarizing from to increase the probability it will give a correct answer, with a label for every answer stating that it has not been independently checked for accuracy and providing a link to the most relevant Wikipedia article for more info.

Forethought has developed a new tool called SupportGPT that uses generative AI to deliver auto-generated customer service responses without human intervention. Link.

• SupportGPT uses a narrower set of data than more generalized GPT applications to deliver more accurate answers and reduce the likelihood of generating incorrect responses.

• Forethought has developed algorithms to constrain the AI and reduce "hallucinations" where the AI provides incorrect answers.

• SupportGPT Playground is a sandbox where companies can experiment with SupportGPT using their own data.

• Salesforce has also announced a pilot of Einstein GPT, which adds generative AI capabilities across the Salesforce platform.

• SupportGPT represents a big leap forward for customer service, helping to provide more accurate and meaningful answers more of the time.

ThoughtSpot has introduced a generative AI entry that lets users query data using natural language. Link.

• The company has been working toward this approach for years, and previously used AI to turn plain language queries into SQL.

• The company now relies on GPT-3 to enable users to enter a query and get a result.

• ThoughtSpot uses the GPT-3 API to help translate natural language into SQL, but also adds its own layer to deliver a single correct answer.

• The company has built in a feedback loop to improve accuracy and fine-tune answers for the future.

• ThoughtSpot was founded in 2012, has raised over $660 million and a private beta of the new integration with GPT-3 opens today.

HEALTHCARE

AI has been used to analyze data from 31,477 older adults in India to estimate the prevalence rate of dementia. Link.

• The study found that the prevalence rate of dementia in adults in India aged 60 or over could be as much as 8.4 per cent, equating to over 10 million older adults in the country.

• The prevalence rates recorded in similar age groups in the US and UK were found to be 8.8 per cent and 9 per cent, respectively, while dementia rates in Germany and France were found to be between 8.5 and 9 per cent.

• Scientists developed an AI learning model trained on data that consisted of 70 per cent pre-labeled dataset with known cases of dementia.

• The remaining 30 per cent of the data was reserved as a test to see how accurate the AI’s predictive capabilities were.

• The study's findings provide essential information for the long-term planning of public and social care policy in India.

Researchers at MIT's Abdul Latif Jameel Clinic for Machine Learning in Health, Mass General Cancer Center, and Chang Gung Memorial Hospital are developing an artificial intelligence tool for lung cancer risk assessment named Sybil. Link.

• The tool analyzes the LDCT image data to predict the risk of a patient developing lung cancer within six years.

• Lung cancer is the deadliest cancer in the world, resulting in 1.7 million deaths worldwide in 2020, killing more people than the next three deadliest cancers combined.

• The researchers demonstrated that Sybil obtained good to strong scores over the course of six years from diverse sets of lung LDCT scans.

• The imaging data used to train Sybil was largely absent of any signs of cancer because early-stage lung cancer occupies small portions of the lung.

• An exciting next step in the research will be testing Sybil prospectively on people at risk for lung cancer who have not smoked or who quit decades ago.

POLICY/LAW

DoNotPay, a service that uses AI to help customers handle legal services, has been accused of practicing law without a license in a proposed class action lawsuit filed by law firm Edelson on March 3. Link.

• The lawsuit claims that DoNotPay is not a robot, a lawyer, nor a law firm, and that it does not have a law degree or any legal supervision.

• The lawsuit was filed on behalf of a customer who had used DoNotPay to draft legal documents and received substandard results.

• DoNotPay was founded in 2015 as an app to help customers fight parking tickets and has since expanded its services to include fighting corporations, beating bureaucracy, finding hidden money, and suing anyone.

• DoNotPay denies the allegations and plans to defend itself vigorously.

• CEO Joshua Browder has accused Edelson of bullying and said that he was inspired to set up DoNotPay to empower consumers to take on corporations without the need for lawyers.

OTHER

Ken Van Haren and Chris Stanley, frustrated by the amount of time spent on infrastructure versus data science, launched Patterns to streamline AI model engineering. Link.

• Patterns' platform lets users build integrations, automations, and workflows with AI from pre-built connectors and a web-based IDE.

• Patterns has elements of an MLOps platform and faces competition from other vendors in the space, including Galileo, Qwak, Diveplane, Tecton, Arize, Iterative, Comet, and Weights & Biases.

• Despite the competition, Patterns has grown its user base to around 1,500, and the company expects to close a government contract this year.

• Patterns' immediate plans are to grow its headcount, which currently stands at four full-time employees, including Van Haren and Stanley.

• Examples of what can be built with Patterns include a question-answering bot and a language model fine-tuned on a dataset of over 6.5 million Hacker News comments.

Machine learning techniques can predict food insecurity outbreaks more accurately than traditional risk systems currently used across 37 food-insecure countries in Africa, Asia, and Latin America. Link.

• A new study in Science Advances used deep learning to extract relevant text from over 11 million news articles published between 1980 and 2020 to analyze how journalists reported about food insecurity and associated causes.

• Between 2009 and 2020, news indicators "substantially" improved the Integrated Phase Classification (IPC) predictions of food insecurity across 21 countries, up to 12 months ahead of time.

• Traditional food security models rely on risk measures that are often delayed, outdated, or incomplete, and tend to rely on on-the-ground measurements or indexes of food prices, resulting in gaps for poorer countries where data isn't collected or is missing.

• The use of data for evidence-based planning and decision-making is necessary for the success of interventions in humanitarian crises, as timeliness is at the core of humanitarian funding effectiveness.

• Disaster response organizations like NGO Team Rubicon stress the importance of informed preparedness to develop resilience plans and respond to unforeseen hazards in the face of increasing risk of disaster events worldwide due to climate change.

Salesforce Ventures launched a $250 million generative AI fund to support responsible development of generative AI. Link.

• The fund will initially invest in four companies, Anthropic, Cohere, Hearth.AI, and You.com, that have demonstrated an ability to transform application software and impact end users' workflows using responsible and trusted development processes.

• More than two-thirds of leaders will prioritize generative AI over the next 18 months, but ethical and technical concerns remain a top concern.

• The fund accelerates Salesforce's commitment to fostering the next generation of innovation and bringing useful AI services into the world.

• The four companies partnered with Salesforce to bring trustworthy, conversational AI assistant Claude to businesses, empower all businesses to build incredible products with world-leading language AI, build a new category of next-gen products centered on agentic relationship management, and develop a suite of generative AI apps and products to enhance productivity.

• The partnership will pave the way for the future of do-engines and create more matched opportunities and connections.

AI chip startup Mythic announced a $13 million financing round to bring its energy-efficient AI processor to market. Link.

• Mythic developed chip tech that stores analog values on flash transistors, allowing for parallel calculations without stopping.

• The company initially worked on computer vision for high-altitude drones and GPS signal acquisition for the US Air Force.

• Mythic doubled down on computer vision use cases with its first commercial chip, M1076, which attracted Lockheed Martin as a major investor.

• The company had to restructure and prioritize efficiency to survive the tough economic conditions in the semiconductor industry and compete in a crowded field.

• Mythic is returning to its roots with a renewed focus on defense, chasing after more customers with government contracts, like Lockheed.

Startup Humane, founded by ex-Apple employees, raised $100 million in Series C funding for their AI device and cloud services platform. Link.

• Notable investors include Microsoft, OpenAI CEO Sam Altman, SK Networks, LG Technology Ventures, and more.

• Humane has raised $230 million to date and has 200 employees.

• Humane's work is shrouded in mystery, but their patent filings and hiring reveal clues about their project.

• Humane has partnerships with SK Networks, Microsoft, OpenAI, LG, and Volvo to bring their platform and services to market.

• Critics are wary of Humane's huge funding and lack of concrete details, citing examples of high-profile hardware startups that failed to deliver on their promises.

Overhaul, a logistics software company that provides visibility software for freight shipping delays, has raised $38 million in a growth financing round led by Edison Partners, with participation from eGateway Capital, StepStone Group and TRM Ventures. Link.

• Overhaul has a customer base that now totals over 350 companies.

• The company uses real-time operational and behavioral data to anticipate and mitigate freight shipping delays.

• Overhaul's platform uses AI and machine learning models for different purposes, such as detecting when a cargo load's security is at risk and alerting a security operations team.

• Overhaul plans to use the funds for R&D and customer acquisition efforts and aims to achieve profitability in 2023.

Singapore and San Francisco-based fintech startup Growfin has raised $7.5 million in Series A funding to expand its SaaS software for finance departments. Link.

• The company plans to use the funding to continue expanding in the US and Asia, and to double down on building more AI-based technology to expand its platform, including a forecasting tool that predicts trends based on past payment behavior and current receivables data.

• The latest funding comes on the back of 8x growth in customer numbers over the last 12 months, during which Growfin has helped clients collect more than $1 billion in account receivables.

• Growfin's initial product was an AI-powered finance CRM that automates finance accounts receivable workflows and provides the right collaboration capabilities and real-time visibility to sales, customer success, and customers themselves in one place.

• The company's primary users currently scale B2B tech companies in SaaS, adtech, logistics tech, and edtech, and it now has 25 customers, including Intercom, Fourkites, Mindtickle, LeadSquared, and Quick Dry Restoration.

• Growfin's competitors include HighRadius, Upflow, Tesorio, YayPay, and Gaviti.

Stabilite AI is seeking to raise more money at a valuation of about $4 billion. Link.

• It is unclear how much capital the company is seeking to raise, and a final decision hasn't been made on whether to go ahead with the financing.

• This move comes after the company raised $101 million at a valuation of about $1 billion in October 2021.

• Artificial intelligence has become a hot topic in Silicon Valley, with major tech players investing heavily in the technology.

• Stable Diffusion's open-source software sets it apart from competitors, allowing for practical applications like designing video games and advertisements.

• Stable Diffusion competes with OpenAI's Dall-E 2, another digital-image tool.

Neuralink's application to begin human testing for its brain-computer interface was reportedly rejected by the FDA in 2022 due to concerns about safety. Link.

• Concerns included the implant migrating through the brain, overheating, battery failure, and brain damage during removal.

• The FDA's caution is warranted as Neuralink is proposing an entirely new in-body electronic system and even a new method of implantation.

• The company has been cited for cruelty in its animal testing, which may have exceeded ethical boundaries.

• People in leadership have left the company, and one co-founder is leaving to form a new brain implant company.

• While Neuralink's progress is difficult to gauge due to its secrecy, the company must first prove that the implants are safe at a basic level before pursuing lofty applications such as restoring sight and mobility.

GPT Tip of the Week

Prompt: How can I improve the accuracy of my AI models?

Response: There are several ways to improve the accuracy of your AI models, including:

• Collecting and using more high-quality data to train your models.

• Using more complex and sophisticated models, such as deep neural networks.

• Tuning the hyperparameters of your models to optimize their performance.

• Regularizing your models to prevent overfitting and improve their generalization.

• Ensembling multiple models together to create a more robust and accurate prediction.

By implementing these techniques, you can significantly improve the accuracy and reliability of your AI models.