AI Weekly: 03/27/23

Good morning and welcome to this week’s edition of AI Weekly! In a week full of a few huge announcements, OpenAI is once again on the front page. The company released ChatGPT Plugins, granting the model access to interact with numerous third-party applications’, including their knowledge sources and databases. This will be the start of what seems to be OpenAI’s version of the App Store, so I’m sure we will see many more integrations in the coming weeks.

In other news, Google is rolling out its ChatGPT competitor Bard, so it will also be interesting to see how they respond to the latest OpenAI developments.

More on that and some exciting announcements from Runway, Adobe, and a number of other companies below!

- ZG

Here are the most important stories of the week:

TEXT

OpenAI has launched Plugins for ChatGPT that grant it access to third-party knowledge sources and databases, including the web. Link.

• The plugins are available in alpha to ChatGPT users and developers on the waitlist, with initial prioritization for premium ChatGPT Plus subscribers and a small number of developers.

• The web-browsing plugin is the most intriguing, as it allows ChatGPT to draw data from around the web to answer questions, citing its sources in responses.

• OpenAI admits that a web-enabled ChatGPT might perform undesirable behaviors but has implemented safeguards to prevent this.

• OpenAI released a code interpreter for ChatGPT that provides the chatbot with a Python interpreter in a sandboxed environment, with disk space for uploading and downloading results.

• OpenAI has open sourced a "retrieval" plugin that enables ChatGPT to access snippets of documents from various sources by asking questions in natural language.

Google announces limited public rollout of ChatGPT competitor, Bard, with a waitlist for users in the U.K. and U.S. Link.

• Bard, a chatbot based on Google's LaMDA, enables users to ask questions and refine answers through follow-up queries.

• The AI chatbot aims to boost productivity, accelerate ideas, and fuel curiosity for users, with applications in various fields.

• Google's Bard works well for "NORA" queries or questions with no single correct answer, addressing accuracy, information sources, and ethical concerns.

• Unlike Microsoft's Bing chatbot, Bard doesn't display footnotes with web sources but offers options to view more answers or rate responses.

• Bard's limited release marks the beginning of a long process, with potential debates around plagiarism, third-party website relationships, and regulatory concerns.

Ubisoft employs an AI "Ghostwriter" tool to write dialogue for NPCs in games to enhance realism and reduce manual work. Link.

• The tool doesn't replace video game writers but helps create first drafts of barks, allowing scriptwriters to focus on the main narrative.

• Ghostwriter generates variations of interactions for scriptwriters to choose from, who then edit them as needed.

• Pairwise comparison is used for evaluation and improvement, enabling the tool to become more effective and accurate over time.

• The AI tool, Ernestine, allows narrative designers to create their own AI system tailored to their design needs.

• Ubisoft faced criticism from some game developers, but emphasized that the tool was created in collaboration with writers for short NPC lines.

IMAGE/VIDEO

Runway announced its new generative video model, Gen-2, which focuses on generating videos from scratch. Link.

• Runway previously developed the open-source text-to-image model Stable Diffusion and AI video editing model Gen-1.

• Gen-1 transformed existing footage, while Gen-2 aims for full video generation based on textual input.

• Runway's Gen-2 demo clips are short, unstable, and not photorealistic, with limited access for now.

• Despite being a small team, Runway is producing work comparable to larger companies like Meta and Google in text-to-video AI.

Adobe introduces Firefly, a generative AI model family focused on images and text effects for Creative Cloud, Document Cloud, Experience Cloud, and Adobe Express workflows. Link.

• Adobe's decade-long AI innovation history includes Adobe Sensei, which powers features like Neural Filters, Content Aware Fill, Attribution AI, and Liquid Mode across Adobe applications.

• Adobe Firefly aims to provide intuitive, powerful generative AI tools, increasing productivity and creative expression for creators across skill levels.

• Firefly will enable creators to generate content, such as images, audio, vectors, videos, and 3D, along with creative ingredients like brushes, color gradients, and video transformations, with speed and ease.

• Adobe launched a Firefly beta, showcasing high-quality image and text effect generation; integration into Adobe Express, Experience Manager, Photoshop, and Illustrator is planned.

• Adobe's approach to generative AI includes efficiency, creator compensation, advocating for open standards, and providing Firefly via APIs for custom workflows and automations.

Runway enters a multi-year strategic research partnership with AWS as its preferred cloud provider to scale high-performance computing clusters and leverage research infrastructure. Link.

• The collaboration enables in-house model development and training, accelerating the deployment of new models and products for better user experiences across the Generative Suite.

• Runway's Gen-2, a multimodal AI system for video generation, was trained on AWS in collaboration with NVIDIA, representing a significant improvement in state-of-the-art AI systems.

• Matt Garman, SVP at AWS, praises Runway's pioneering generative models and looks forward to changing how creatives harness AI for their vision.

• Gen-2 can realistically synthesize new videos using either image or text prompts.

• The partnership aims to scale training and capacity for groundbreaking video generation models and explore the potential of generative AI together.

Levi Strauss & Co. announces partnership with Lalaland.ai to use AI-generated models for showcasing clothing. Link.

• AI-generated models will supplement human models and increase the number and diversity of models in a sustainable way.

• Dr. Amy Gershkoff Bolles reassures that AI will not fully replace human models, but enhances the consumer experience.

• The partnership aims to create a more personal and inclusive shopping experience for customers.

• Currently, the company typically features one model per product; AI-generated images will offer more variety.

• The growing use of AI could lead to human models being needed only for in-person events in the future.

Nvidia and Google have partnered to launch the L4 platform, optimized for video-focused applications. Link.

• The L4 platform, designed to accelerate AI-powered video performance, is available in private preview on Google Cloud.

• Google is integrating the L4 platform into Vertex AI, its managed machine learning service for enterprises.

• Nvidia's hardware partners, including Asus, Cisco, Dell, Hewlett Packard Enterprise, and Lenovo, will offer L4 later this year.

• Nvidia also announced L40, H100 NVL, and Grace Hopper for Recommendation Models as AI-focused hardware solutions.

• Nvidia's DGX Cloud platform, providing companies access to infrastructure and software for AI, is partnering with cloud service providers like Oracle Cloud Infrastructure, Microsoft Azure, and Google Cloud.

CODE/DEVTOOLS

With Copilot X, GitHub Next aims to transform Copilot into an AI assistant available throughout the entire development lifecycle. Link.

• GitHub Copilot, an AI pair programmer, writes 46% of code and helps developers code up to 55% faster.

• Copilot X will adopt OpenAI's GPT-4 model and feature chat and voice interfaces for better developer interaction.

• Copilot will be integrated into pull requests, the command line, and documentation to provide AI-generated assistance across multiple platforms.

• OpenAI's GPT-4 model improves logical reasoning and code generation, allowing Copilot to better understand codebases and provide tailored suggestions.

• GitHub Copilot X aims to enhance developer productivity and happiness, enabling faster and higher-quality software development.

Mozilla has launched Mozilla.ai, a startup aiming to build a trustworthy and independent open-source AI ecosystem, with a $30 million investment. Link.

• With AI products facing issues such as misinformation, deepfakes, and copyright problems, Mozilla aims to influence the industry through product development.

• Mozilla has a history of being a small but influential player in the market, particularly with Firefox, and its expertise gives it an outsize voice in tech discussions.

• Mozilla.ai plans to focus on building safety and transparency tools for developers, rather than competing directly with GPT-4.

• Mozilla sees AI as part of Firefox's future, with CEO Mitchell Baker suggesting that disruption in the browser space is more likely now than six months ago.

• The announcement comes as tech giants keep their AI experiences closed, and Mozilla aims to champion a better, open model for AI development.

CodiumAI, founded by CEO Itamar Friedman, emerged from stealth with an $11 million seed investment. Link.

• The company addresses the lack of code logic checkers by using generative AI to build tests automatically.

• Developers can interact with the generated code and request more specific tests.

• CodiumAI's custom large language model, TestGPT, is used to generate code analysis, test plans, and test code.

• The company has been working with 1,000 early users and is now opening the free product to the public.

• The seed investment was co-led by Vine Ventures and TLV Partners, with participation from industry angel investors.

Aspecta has raised a $3.5 million seed round led by ZhenFund to provide an AI-powered profile builder for developers, creating LinkedIn-like identity pages. Link.

• The platform reviews code quality in projects and considers social endorsement and network analysis.

• Aspecta is integrated with GitHub, Google, Stackoverflow, and MetaMask and is onboarding Twitter.

• The startup has attracted 130,000 people to its waitlist

• Aspecta plans to become a decentralized autonomous organization (DAO) to manage the identity ecosystem.

• The platform also targets content creators and aims to examine their achievements for presentation.

Numbers Station, a startup that aims to bring foundation models to enterprise use cases for data transformation and record matching, has raised $17.5 million in a Series A funding round led by Madrona. Link.

• The round also had participation from Norwest Venture Partners, Factory and a number of angel investors, including Cloudera co-founder Jeff Hammerbacher.

• Numbers Station translates natural language queries into SQL commands using foundation models.

• The platform currently offers three capabilities: SQL transformation, AI transformation, and record matching.

• The startup personalizes and adapts models to fit specific organizations, using fine-tuning and feedback techniques.

• The company's long-term vision is to build a data intelligence platform for analyzing as well as transforming data.

Tavus, founded by Hassaan Raza and Quinn Favret in 2020, raises $6.1 million seed round led by Sequoia, with participation from Accel Partners, Index Ventures, Lightspeed Ventures, and YC Continuity. Link.

• The platform allows users to create AI video templates and generate multiple personalized video versions using a web-based editor.

• Tavus integrates ChatGPT-powered snippets for more personalized introductions, using information from LinkedIn profiles.

• The technology caters to various users like recruiters, university deans, and C-level executives, with applications across the customer journey.

• Companies can access Tavus through a web dashboard or integrate it into their systems via APIs or natively for event-driven workflows.

• Tavus offers a basic intro plan for smaller businesses at $275 per month, while custom "business" plans are tailored according to a company's needs and usage.

HEALTHCARE

Bionic Health, a Durham-based startup, raised $3 million in seed funding to develop an AI health clinic. Link.

• The clinic offers bloodwork, diagnostics, and personalized insights based on data from OpenAI's GPT-4 and other machine learning models.

• The initial focus is on preventative health services, with potential future expansion into all aspects of doctor-patient care.

• The clinic will have human doctors working alongside AI to provide feedback and improve the system.

• Co-founders Robbie Allen and Dr. Jared Pelo have experience in creating successful AI startups.

• Bionic Health aims to address healthcare issues such as the shortage of general practitioners and specialists.

Unlearn.AI has raised $15 million from Radical Ventures and Whittington Ventures to expand partnerships and accelerate regulatory approval for its machine learning platform creating "digital twin" patient profiles in clinical trials. Link.

• The investment from Radical Ventures and Wittington Ventures values the startup at $265 million.

• Mira Murati, CTO at Microsoft-backed OpenAI, joined Unlearn.AI's board.

• Unlearn uses generative AI to build digital twins, aiming to speed up clinical drug trials with neural networks.

• Unlearn's software could lower drug development costs by reducing the number of required clinical trial enrollments.

• The startup focuses on Phase III clinical trials and aims to gain approvals from regulators like the US Food and Drug Administration to commercialize its technology globally.

MVision AI, a Finnish medical imaging and AI startup, secures $5.85 million in post-seed financing. Link.

• The company's AI-powered automatic segmentation tooling streamlines radiotherapy treatment planning workflow.

• MVision's automation enables same-day treatment planning, compared to the 2-3 weeks wait time in manual treatment planning.

• The auto-segmentation tool allows personalized treatment plans, minimizes radiation exposure to healthy organs, and aims to standardize global cancer care.

• MVision AI has CE-marked and FDA-cleared auto-segmentation features for radiotherapy planning and plans to release more features for MR images.

• The company aims to build a globally scalable cloud-based platform for cancer treatment and expand its solution to more patients worldwide.

OTHER

Microsoft Loop launches in public preview as a hub for managing tasks and projects across Microsoft 365 apps and services. Link.

• Loop is available to users with a Microsoft Account or Azure Active Directory account, with companion apps for iOS and Android coming soon.

• Loop competes with Google Workspace Spaces and consists of Loop components, Loop pages, and Loop workspaces.

• Loop components are JavaScript widgets that enable real-time collaboration across apps like Outlook, Microsoft Teams, and Word.

• Edits to Loop components are automatically synced across Microsoft 365 apps, with plans to add components for business workflows and custom development.

• Microsoft is integrating the AI-powered Microsoft 365 Copilot system into Loop, which will provide suggestions for brainstorming and collaboration.

Workera raised $23.5 million in a Series B round led by Jump Capital for its upskilling platform. Link.

• 40% of companies are actively upskilling, automating, or enhancing work through tech, according to a PwC survey.

• Workera aims to measure skills and offer personalized learning for employees in the workplace.

• The platform uses a data-driven approach with a "skills dataset" to measure more skills in less time.

• Workera uses generative AI to create questions for skills assessments and summarize information about a skill.

• Clients include Samsung, Siemens Energy, the U.S. Air Force, and Belcorp.

IPRally, an AI-powered patent search engine using a graph neural network, has raised €10 million ($10.8 million) in a Series A funding round led by Endeit Capital. Link.

• IPRally's technology provides technical and semantic understanding around patent data.

• The platform delivers explainable results, allowing users to understand the relationships between concepts.

• IPRally has clients such as Nvidia, Dolby, and Bentley and aims to expand its scope beyond patent search.

• The latest funding round also saw participation from Join Capital, Spintop Ventures, and Icebreaker.vc.

Pilot AI, a platform that uses AI to analyze sales calls and input data into CRM systems, has recently raised $2.2 million in seed funding. Link.

• The company was part of Y Combinator's summer 2022 batch.

• Pilot AI generates call notes, summaries, questions, answers, and next steps, saving sales reps time.

• The startup claims to be more streamlined and affordable than competitors in the market.

• Pilot AI has a growing customer base, including small sales organizations and larger industry leaders.

• The pandemic has accelerated the adoption of AI in sales, with businesses focusing on improving workforce productivity.

GPT Tip of the Week

Prompt: What are different types of defensibility in AI startups to look for as big players like OpenAI and Google continue to make strides?

Response: There are various types of defensibility that AI startups can aim for as big players like OpenAI and Google continue to advance. Here are a few to consider:

1. Data defensibility: This involves having access to proprietary or unique data sets that are difficult for competitors to replicate, and can provide a competitive advantage in training AI models.

2. Technical defensibility: This refers to the development of unique algorithms, models or architectures that are difficult for competitors to copy or reproduce, and which can provide a competitive advantage in terms of accuracy, speed or efficiency.

3. Intellectual property defensibility: This involves the creation of intellectual property in the form of patents, trademarks, or copyrights that can protect the startup's technology and provide a barrier to entry for competitors.

4. Network effects defensibility: This refers to the creation of a platform or ecosystem with a network of users or partners that becomes more valuable as more users or partners join, making it difficult for competitors to catch up.

5. Brand defensibility: This involves creating a strong brand that is associated with quality, innovation, and trustworthiness, making it difficult for competitors to establish a similar reputation.