AI Weekly: 06/05/23

Good morning and welcome to this week’s edition of AI Weekly! It’s shocking to say, but we had a relatively slower week in the AI world. For the first week in what feels like forever, there were no headlines about OpenAI, Google, Anthropic or Stability AI coming out with a new model. However, MIT researchers have developed a language model that they claim outperforms larger models by up to 500 times in certain language understanding tasks.

In other news, Instacart has released a new feature called “Ask Instacart,” which provides users with personalized product recommendations and information on food preparation, dietary considerations, and more.

Also, the European Union has announced plans to work with the US to draft a voluntary Code of Conduct for AI with the goal of establishing standards for applying AI while waiting for legislation on AI regulation to be passed in respective countries and regions. Read about that and more of this week’s AI news below!

- ZG

Here are the most important stories of the week:

TEXT

Researchers at MIT CSAIL have developed a language model called SimPLE that outperforms larger models by up to 500 times in specific language understanding tasks. Link.

• SimPLE is a scalable, self-learning model that doesn't rely on human-generated annotations for training, using a technique called self-training.

• The model focuses on contextual entailment, which improves its performance in natural language understanding (NLU) tasks.

• MIT's approach challenges the notion that larger models are inherently superior and highlights the potential of smaller models as powerful and environmentally sustainable alternatives.

• The researchers used textual entailment, the connection between two sentences, to enhance the model's comprehension of diverse language tasks.

• SimPLE can adapt to various tasks without additional training, reducing the need for human annotations and potential privacy risks associated with sensitive data.

OpenAI is addressing AI "hallucinations" by developing a new training method for AI models to reduce misinformation and fabrications. Link.

• The rise of generative AI and concerns about misinformation coincide with OpenAI's research efforts.

• AI hallucinations occur when models generate false information, which can have serious consequences, such as false claims in promotional videos or bogus citations in legal filings.

• OpenAI's proposed approach, called "process supervision," aims to reward AI models for correct steps of reasoning rather than just the final outcome, potentially improving explainability and aligning AI with human-like thinking.

• OpenAI has released a dataset of 800,000 human labels used to train the model discussed in the research, but some experts express skepticism and call for more transparency and accountability in the field of AI.

Character.AI, an AI app offering customizable AI companions, claims to have gained over 1.7 million installs in less than a week since its global mobile launch. Link.

• The Palo Alto-based startup, backed by a16z, was founded by Noam Shazeer and Daniel De Freitas, former Google AI researchers who led the development of LaMDA.

• Within the first 48 hours of the app's launch, it received over 700,000 Android installs, surpassing popular entertainment apps like Netflix, Disney+, and Prime Video.

• The app's web experience, which preceded the mobile launch, attracted over 200 million monthly visits and an average visit duration of 29 minutes.

• Users quickly become engaged, with engagement rates jumping to over 2 hours on average after the first message sent to a character. Over 10 million custom AI characters have been created to date.

• Character.AI recently introduced its premium service, c.ai+, and entered a strategic partnership with Google Cloud for building and training its AI models using Google Cloud's Tensor Processor Units.

Instacart is launching a new AI search tool called "Ask Instacart," powered by OpenAI's ChatGPT. Link.

• The tool is embedded in the search bar of the Instacart app and provides I

• Ask Instacart incorporates personalized question prompts and reminds users based on their shopping history, aiming to help customers save time and discover new products.

• Users can ask specific questions about side dishes, similar fish options, grilling chicken sauces, and dairy-free snacks for kids.

• The new feature makes Instacart more of a one-stop-shop for food preparation, eliminating the need to use external search engines.

• Instacart is committed to incorporating AI features responsibly, focusing on relevant food-related questions with its specialized model.

UpdateAI, a Pasadena-based company, has raised $2.3 million in funding for its ChatGPT-powered Customer Success platform. Link.

• The funding round was led by IdealabX, with participation from Stage Venture Partners, Zoom Ventures, a16z's scout fund, Howard Morgan, and CS Angel.

• The funds will be used to expand operations and grow the company's business reach.

• UpdateAI's Customer Success platform utilizes ChatGPT and proprietary conversational AI technology to generate Smart Summaries™️ of customer meeting notes and automate follow-up workflows.

• The platform automates tasks such as sending follow-up emails to customers and integrating notes with systems like Salesforce.

• The company's long-term vision is to enable Customer Success teams to utilize conversational data in a cross-functional manner, elevating their role within organizations.

IMAGE/VIDEO

Video editing startup Dumme, backed by Y Combinator, has hundreds of video creators testing its AI-powered product and a waitlist of over 20,000 pre-launch. Link.

• Dumme uses AI to create short-form videos from YouTube content, aiming to save editing time and outperform human editors in menial video editing tasks for platforms like YouTube Shorts, TikTok, and Instagram Reels.

• The company repurposed its AI models originally developed for building a video search engine to edit video clips, and its app went viral, crashing their servers.

• Users paste a YouTube video link into Dumme, and AI processes the video, transcribing and analyzing spoken words, frames, and emotions to generate highlight clips.

• Dumme works with existing AI models like GPT-4, Whisper, and its own in-house models, and it collaborates with CoreWeave for AI processing.

• Dumme is targeting YouTube creators, podcasters, and agencies, aiming to provide a more affordable and time-saving solution compared to outsourcing editing work to contractors in cheap jurisdictions. Pricing models are being considered, including per-minute processing fees and monthly subscriptions.

Genealogy company MyHeritage has launched a new mobile app called Reimagine, which uses AI to help users preserve old photos by improving their resolution, fixing scratches and creases, restoring color, and animating faces. Link.

• The app allows users to import printed photos stored in albums and automatically crops and saves them as stand-alone images.

• Users can also add details such as names, dates, and places to make the photos easier to find.

• Reimagine combines AI technology from MyHeritage's Deep Nostalgia with other photo-editing tools specifically designed for preserving old photos.

• The app is available for iOS and Android and monetizes through weekly and monthly subscriptions that provide unlimited access to features like photo scanning, colorization, and enhancement.

• Early tests of the app show promising results in enhancing and improving photos, but there are some limitations with glare correction and color accuracy.

CODE/DEVTOOLS

Blink is launching Blink Copilot, a no-code security operations workflow tool that generates low-code workflows based on user prompts. Link.

• Blink CEO Gil Barak describes three distinct periods in security operations automation: coding workflows, low-code approaches, and the new era of describing workflows and having the platform build them.

• The tool works with large language models from providers like Microsoft, Google, and OpenAI to understand prompts and generate relevant workflows.

• Blink Copilot automates security operations, similar to how Ansible Lightspeed automates IT operations.

• The tool offers a library of over 7,000 workflow components and allows users to modify, review, and edit workflows as needed.

• While the simplicity of the tool makes it easy to use, there should still be humans involved in reviewing and ensuring the implications of the workflows are understood. Blink is adding guardrails to address this concern.

8Flow.ai, an enterprise-grade, self-learning workflow automation engine for customer support agents, has raised a $6.6 million seed round led by Caffeinated Capital, with participation from BoxGroup, Liquid2, HNVR, Trilogy, and notable angel investors. Link.

• The tool integrates with platforms like Zendesk, ServiceNow, and Salesforce Service Cloud to assist agents in their daily tasks.

• 8Flow captures data and feedback from agents to optimize and improve the user interface and experience.

• The Chrome extension of 8Flow learns common steps for each agent and offers one-click actions, automating tasks that involve copying and pasting data between applications.

• The tool allows agents to focus on decision-making steps rather than repetitive manual tasks, increasing efficiency and accuracy.

Predibase, an end-to-end machine learning platform, has announced a $12.2 million expansion to its Series A funding round, bringing the total funding to $28.45 million. Link.

• The company's low-code, declarative ML platform for developers is now generally available.

• Users can train models and deploy their own large language models on the platform, including Predibase's LudwigGPT LLM.

• Predibase has also introduced its Data Science Copilot, a system that provides developers with recommendations to improve model performance.

• The company's customers include Paradigm and Koble.ai.

• Predibase plans to use the funding to expand its go-to-market functions and further develop its platform, focusing on its developer-centric approach in a crowded market.

HEALTHCARE

Gert-Jan Oskam, a man who was paralyzed from the hips down due to a motorcycle accident, has regained control over his lower body through a combination of AI-driven devices and implants. Link.

• Researchers in Switzerland developed implants that created a "digital bridge" between Oskam's brain and spinal cord, allowing him to stand, walk, and even show signs of neurological recovery.

• The brain-spine interface used an AI thought decoder to read Oskam's intentions and match them to muscle movements.

• The researchers implanted electrodes in Oskam's skull and spine and used machine learning to observe brain activity and connect it to the spinal implant.

• Oskam has achieved significant progress in walking, standing, and navigating steps and ramps, and has retained some improvements even when the implant was switched off.

• Although there are limitations and the treatment is invasive, the researchers are optimistic about making the technology more accessible and effective for patients worldwide.

Hyro, an AI healthcare startup founded by Israel Krush and Rom Cohen, has raised $20 million in a Series B funding round led by Liberty Mutual, Macquarie Capital, and Black Opal, bringing its total funding to $35 million. Link.

• Hyro offers AI-powered text and voice conversational solutions for the healthcare sector, automating routine tasks and messages to alleviate administrative burnout.

• The funding will be used to expand Hyro's go-to-market teams and research and development efforts.

• The healthcare industry is facing staffing shortages and outdated processes, making it a key vertical for Hyro's technology.

• Hyro differentiates itself with features such as AI-powered search for up-to-date information, smart routing to determine the appropriate response, and the ability to create feedback loops for continuous improvement.

• The company has gained numerous healthcare networks and providers as clients, and its annual recurring revenue has doubled since its market launch in 2019. Hyro plans to expand to other industries and enhance customization options and AI skills supported by the platform.

POLICY/LAW/ETHICS

The European Union (EU) plans to work with the US to draft a voluntary Code of Conduct for AI and hopes to involve governments in other regions like Indonesia and India. Link.

• The aim is to establish standards for applying AI while waiting for legislation on AI regulation to be passed in respective countries and regions.

• The EU intends to draft the AI Code of Conduct within weeks and wants AI giants to voluntarily abide by the standards.

• The EU-US Trade & Tech Council (TTC) is serving as a platform for discussions on AI rule-making and cooperation on tech governance and trade issues.

• The EU is urging for the fast development of the voluntary code, as it recognizes the urgent need to address the risks and benefits of AI in real-time.

• Experts and stakeholders at the TTC panel highlighted the importance of AI auditing, transparency, accountability, and comprehensive assessment of AI risks and harms. They emphasized the need to protect individuals from current AI-driven harms and to involve civil society and marginalized communities in the standard-setting process.

Judge Brantley Starr has implemented a requirement in his courtroom that attorneys must confirm that no part of their filing was generated by artificial intelligence, or if it was, that it was reviewed by a human. Link.

• Attorney Steven Schwartz used ChatGPT, an AI language model, to assist him in a federal filing, but the AI generated fictitious cases and precedent. Schwartz now regrets using AI in this manner.

• Judge Starr's decision to introduce the "Mandatory Certification Regarding Generative Artificial Intelligence" is not explicitly stated to be a response to Schwartz's filing, but it may be a reaction to it.

• The certification requires attorneys to confirm that no part of their filing was drafted by AI or that any AI-generated language was verified by a human for accuracy.

• The certification covers quotations, citations, paraphrased assertions, and legal analysis, as AI systems can be prone to hallucinations and bias.

• It is suggested that other judges may adopt similar rules to ensure the clear declaration and accuracy of AI usage in legal proceedings.

UK Prime Minister Rishi Sunak is seeking to update the government's approach to regulating artificial intelligence due to concerns about the risks posed by the technology. Link.

• Sunak is pushing for an international agreement on AI development, potentially leading to the establishment of a global regulator. There are also calls for a separate bill to create the UK's first AI-focused watchdog.

• The government's previous AI white paper, published in April, focused more on the benefits of AI than on regulation. However, recent developments, including advancements in AI technology and warnings from experts, have prompted a shift in government thinking.

• Sunak met with top executives from the AI industry, acknowledging the "existential risks" associated with AI.

• British officials will discuss AI's implications for intellectual property and disinformation at a G7 meeting.

• Some MPs are advocating for an AI bill that sets conditions for companies developing AI in the UK and the creation of an AI-specific regulator. The government aims to remain agile in regulation, avoiding a product-by-product approach due to the rapid pace of technological change.

The Chinese Communist Party (CCP) has raised concerns about the potential risks of AI advancements, particularly in relation to political and social harm. Link.

• Chinese leader Xi Jinping led a meeting addressing these concerns and called for enhanced national security measures.

• The focus was on preparing for worst-case scenarios and extreme cases related to national security.

• The meeting highlights the urgency of safeguarding political security and improving the governance of internet data and AI, while advocating for a risk monitoring and early warning system.

• Technology leaders, including executives from Google and Microsoft, stress the severity of AI threats and the need for global prioritization and regulation.

• They propose a similar approach to the management of pandemics and nuclear war for addressing AI risks.

OTHER

During a recent conversation, Sam Altman discussed both practical developer concerns and broader questions surrounding OpenAI's mission and the societal implications of AI. Link.

• OpenAI is currently limited by GPU availability, which affects the reliability and speed of their API.

• Near-term roadmap for 2023 includes cheaper and faster GPT-4, longer context windows, extended finetuning API, and a stateful API.

• ChatGPT plugins are unlikely to be released soon, as they don't have a product-market fit yet.

• OpenAI aims to avoid competing with their customers, except for ChatGPT, which they plan to develop as a super smart assistant for work.

• OpenAI supports regulation for future models but believes existing models are not dangerous. They also consider open-sourcing GPT-3 and emphasize the importance of open source.

• The scaling laws for model performance still hold, indicating that larger models will continue to yield better performance. However, the rate of scaling may slow down compared to previous years.

• The continuation of scaling laws suggests shorter timelines for AGI development, as most of the remaining work may involve scaling up existing methods rather than major breakthroughs.

SigmaOS, a web browser startup, has launched its own AI assistant called Airis to a limited group of users. Link.

• Airis is unique because it understands the context of a webpage and provides answers accordingly.

• Airis can be activated by right-clicking on a text selection or by pressing A to open it in split-screen mode for contextual conversations.

• Airis can rewrite or reword sentences and paragraphs using various prompts, including making it simpler, adding humor, or translating it into different languages.

• Unlike other assistants like ChatGPT, Airis allows users to assign multiple rewriting or translation tasks on different pages simultaneously.

• Other browsers, such as Microsoft Edge and Opera, are also experimenting with integrating AI assistants, and SigmaOS plans to train Airis to understand entire websites in the future.

Lightmatter, a photonic computing startup, has received $154 million in a Series C funding round to develop and deploy its technology from SIP Global, Fidelity Management & Research Company, Viking Global Investors, GV, HPE Pathfinder and existing investors. Link.

• Lightmatter aims to revolutionize the AI computation market with a hardware-software combo that offers improved performance and energy efficiency.

• Traditional chips used for AI computations are reaching the limits of density and speed, leading to challenges in cooling and increased energy consumption.

• Lightmatter's approach utilizes optical flow and microscopic optical waveguides to perform computational processes, combining analog and digital methods.

• The use of multiple colors in the optical computing process enables increased computational power.

• The company's full stack includes Envise (computing hardware), Passage (interconnect), and Idiom (software platform), which seamlessly integrate with popular machine learning frameworks like PyTorch and TensorFlow.

Sana, an AI-powered learning and knowledge platform, has raised an additional $28 million in an investment round led by NEA, bringing its total Series B funding to $62 million. Link.

• The platform combines enterprise search, a learning management system, meeting tools, and a knowledge management system into a single product, with Sana AI as its latest release.

• Sana AI acts as an omnipresent assistant, allowing users to search across company apps, take actions based on natural language commands, generate real-time meeting summaries, and create learning courses.

• By leveraging AI to capture, organize, and access knowledge, Sana aims to enhance organizational productivity and enable teams in various domains.

• NEA was attracted to Sana's commercial performance, tripled year-over-year growth, and customer advocacy from notable companies like Merck, Kry/Livi, and Svea Solar.

• The funding will support Sana's product development and commercial expansion across its Stockholm, London, and New York offices, with founder and CEO Joel Hellermark leading the company.

MindsDB, a rapidly growing open-source AI project, has secured an additional $25 million in financing led by new investor Mayfield, with participation from TQ Ventures and existing investor Benchmark Capital. Link.

• This brings MindsDB's total seed capital raised to $41.5 million and total funds raised to $50 million.

• MindsDB offers a Cloud for Serving Artificial Intelligence Logic, allowing organizations to transition their AI-powered applications from prototyping and experimentation to production quickly and at scale.

• The platform provides a fast and scalable solution for shipping AI applications.

• The funding will support MindsDB's continued growth and development in the AI space.

Deep Sentinel, a security provider, has raised $15 million in a funding round led by Intel Capital. Link.

• Other participants in the round include Shasta Ventures, Slow Ventures, UP2398, and The Syndicate led by Jason Calacanis.

• Deep Sentinel offers smart home security systems that combine high-definition cameras, AI technology, and human intervention.

• The systems are designed to detect potential threats in real-time and alert trained security guards who can communicate with homeowners and, if necessary, law enforcement.

• The funding will help Deep Sentinel increase the affordability, availability, and reliability of its security solutions.

• Deep Sentinel has integrated Intel's NUC into its product line, enhancing its live guard surveillance service and supporting up to six PoE cameras.

GPT Tip of the Week

Prompt: What are different ways to use Adobe’s generative fill to make everyday life more efficient?

Response: Adobe's Generative Fill feature in Photoshop can be utilized in various ways to make everyday life more efficient. Here are a few examples:

1. Content Creation: Generative Fill can help generate content quickly and easily, saving time and effort. For example, if you need images for a presentation or social media post, you can use Generative Fill to create relevant visuals based on simple text prompts. This eliminates the need to search for or create images from scratch.

2. Design Iteration: When working on design projects, Generative Fill enables rapid iteration. You can experiment with different concepts and variations by adding, extending, or removing content from images. The non-destructive editing capability allows you to explore creative possibilities without permanently altering the original image.

3. Graphic Design: Generative Fill can assist graphic designers in creating unique designs more efficiently. Whether it's generating textures, patterns, or visual effects, you can leverage Generative Fill to quickly generate elements that match your vision. This helps streamline the design process and frees up time for other tasks.

4. Image Editing: If you need to modify images for various purposes, Generative Fill can be a valuable tool. It can assist in tasks like object removal, image extension, or perspective correction. By providing text prompts, you can describe the desired changes, and Generative Fill will generate the modified content accordingly.

5. Content Generation for Marketing: Generative Fill can be used to generate content for marketing materials. Whether it's creating banners, advertisements, or social media graphics, you can leverage the AI-powered capability to quickly generate visually appealing and on-brand content. This can accelerate the content creation process for marketing campaigns.

6. Idea Exploration: Generative Fill encourages ideation and exploration by allowing you to generate content at a transformative rate. You can quickly experiment with different ideas and concepts, leveraging the speed and ease of generative AI. This can be particularly useful when brainstorming or conceptualizing new projects.

Remember that while Generative Fill provides powerful automation and efficiency, it's important to ensure that the generated content aligns with your creative vision and specific requirements. It's a tool to enhance and expedite workflows, but human judgment and artistic expertise remain crucial in achieving the desired outcome.