AI Weekly: 06/12/23

Good morning and welcome to this week’s edition of AI Weekly!

In this week’s news, Cohere led the charge with $270m in Series C funding to fuel their goal of being the leading LLM platform for enterprises. Poetically, an earlier-stage competitor, Contextual.ai, also announced recent funding this week after they raised $20m from a number of high-profile investors.

Separately, Zoom announced their plans to roll out a meeting summarizer, raising questions about the long-term viability of other platforms in the space providing a similar solution.

In our weekly regulation update, the U.S. Senate is taking its next steps to responsibly address the AI platform shift as Senate Majority Leader Chuck Schumer has announced a series of closed-door information sessions to educate senators on AI, covering topics such as capabilities, competition, and national security. Enjoy much more AI news below!

- ZG

Here are the most important stories of the week:

TEXT

Cohere, a startup developing an AI model ecosystem for the enterprise, has raised $270 million in its Series C funding round. Link.

• The funding values Cohere at around $2.1 billion to $2.2 billion, significantly lower than earlier reports suggesting a valuation of over $6 billion.

• The capital will be used to further develop Cohere's AI platform, focused on enterprise customers, enabling them to leverage their preferred cloud provider while maintaining data privacy and simplifying implementation.

• Cohere's platform is cloud agnostic and can be deployed within public clouds, existing clouds, virtual private clouds, or on-site, offering custom large language models (LLMs) based on proprietary data.

• The startup works with companies like Jasper and HyperWrite for copywriting generation tasks and has partnerships with organizations like LivePerson to build fine-tuned LLMs for improved explainability.

• Cohere plans to explore growth areas in search and retrieval, expanding AI capabilities to access real-time information and take actions on behalf of customers, such as booking flights or scheduling meetings.

Meta showcased its work on generative AI technologies for consumer products like Instagram, WhatsApp, and Messenger, as well as internal use at the company. Link.

• The AI technologies in development include AI chatbots for Messenger and WhatsApp, AI stickers, and photo editing tools for Instagram Stories.

• Meta plans to release AI agents on Messenger and WhatsApp, with potential expansion to other applications and smart glasses.

• The AI agents aim to provide unique personas and skill sets to help and entertain users.

• Meta is also experimenting with AI that allows users to modify their photos for Instagram Stories using text prompts.

• The company is hosting an internal AI hackathon focused on generative AI and aims to bring new AI products to users in the future.

Microsoft will offer U.S. government departments access to its artificial intelligence models, including OpenAI's GPT-4, through its Azure OpenAI Service. Link.

• Azure Government customers can utilize these AI models to accelerate content generation, create summaries, optimize search, and simplify code.

• Azure Government provides stringent security and compliance standards required by government agencies to protect sensitive data.

• While specific government customers are not named, the Department of Defense, the Department of Energy, and NASA are among Azure Government's customers, with the Defense Technical Information Center confirming its use of the new tools.

• The service grants government access to GPT-4, GPT-3, and Embeddings for enhanced search and recommendation features, while Azure Government maintains its own encrypted network for utmost confidentiality.

• This announcement follows OpenAI CEO Sam Altman's testimony advocating for a regulatory agency for AI companies and licensing requirements, and Microsoft is OpenAI's primary corporate partner, with over 4,500 customers utilizing the Azure OpenAI Service.

Jerry Liu, former Uber research scientist, launched an open source project called LlamaIndex to overcome limitations in working with private data using LLMs like GPT-3 and GPT-4. Link.

• LlamaIndex has now become a company, offering a framework to help developers leverage LLM capabilities on their personal or organizational data.

• The LlamaIndex framework enables developers to connect various data sources, such as PDFs, PowerPoints, apps like Notion and Slack, and databases like Postgres and MongoDB, to LLMs.

• The framework includes connectors for data ingestion, data format handling, and structuring data for LLM usage.

• LlamaIndex features a data retrieval and query interface that provides context and knowledge-augmented output based on LLM input prompts.

• LlamaIndex recently raised $8.5 million in seed funding, which will be used to build an enterprise solution on top of the open source project, including protection-grade data connectors and domain-specific data indexing capabilities.

Twilio is integrating Google Cloud's generative AI into Twilio Flex, its customer experience platform. Link.

• The integration aims to provide personalized services without compromising privacy.

• Generative AI prompts customer service representatives with the next best action based on the customer's history and needs.

• After the call or chat, the AI builds the customer's profile with after-call summaries and sentiment analysis.

• Toyota is already utilizing Twilio's features for its Toyota Connected Destination Assist navigation feature.

• Generative AI is transforming customer service, allowing agents to focus more effectively on individual customer needs and is expected to have a significant impact on how companies interact with customers.

LinkedIn has introduced an AI-based tool to suggest different copies of ads for marketers. Link.

• The tool uses data from a marketer's LinkedIn page and Campaign Manager settings to generate suggestions for the introductory text of the ad.

• OpenAI models are utilized to create different copy suggestions.

• The feature is currently in the testing phase and available to some customers in North America.

• LinkedIn plans to add new functionalities, expand geographically, and support more languages in the future.

• The company has been leveraging generative AI in various ways, including conversation starters, profile bios, and recruitment posts.

Byju's, the EdTech giant, has unveiled three transformer models as part of its "Wiz" suite to enhance its learning and personalization experience. Link.

• The predictive AI model, Badri, identifies knowledge gaps and offers recommendations to support continual learning.

• MathGPT resolves math problems using analogies and visual aids, while TeacherGPT provides personalized guidance and grades student responses.

• Byju's AI model contextualizes instruction based on students' interests, making learning engaging and adaptable.

• The transformer models have been trained on Byju's student base and calibrated to operate within curriculum boundaries.

• Byju's plans to leverage AI in all aspects of its platform, and the models will positively impact the company's bottom line and contribute to students' educational progress.

Instagram is rumored to be working on an AI chatbot, according to leaked images by app researcher Alessandro Paluzzi. Link.

• The AI chatbot on Instagram would be able to answer questions, give advice, and help users write messages.

• Other social platforms, like Snapchat, have launched AI chatbots with mixed results, facing challenges with age-gating and functionality.

• It is unclear which AI tools Instagram would use to power its chatbots, but Meta (Instagram's parent company) has been utilizing generative AI in other aspects of its business, such as ad sales.

• Instagram is also rumored to be developing a text-based app codenamed Barcelona, which would serve as a Twitter competitor and integrate with Instagram.

• Instagram has not provided any official comments or confirmation regarding these rumors.

Contextual AI, co-founded by Douwe Kiela, has emerged from stealth mode with $20 million in seed funding to develop the "next generation" of LLMs for enterprise applications. Link.

• The company aims to address limitations of current LLMs in terms of making up information with high confidence and difficulties in removing or revising their knowledge base.

• Contextual AI's technology is based on retrieval augmented generation (RAG), which augments LLMs with external sources to improve their performance and generate context-aware responses.

• RAG can enhance attribution, customization, and performance of LLMs while enabling faster processing, lower latency, and lower costs.

• The company plans to build an integrated solution specifically designed for enterprise use cases, ensuring accurate, reliable, and traceable answers from generative AI.

• Contextual AI is in talks with Fortune 500 companies to pilot its technology and will utilize the seed funding for product development, including training LLMs, and expects to expand its team to nearly 20 people by the end of 2023.

IMAGE/VIDEO

Zoom has introduced a trial feature called Zoom IQ that uses AI to generate meeting summaries for users. Link.

• Meeting summaries can be sent through Zoom Team Chat or email without recording the entire meeting.

• Users can compose messages in Team Chat using AI, with the ability to customize tone and length before sending.

• Zoom plans to launch additional AI-powered features through its partnership with OpenAI and Anthropic, including AI-generated emails and summarized chat threads.

• Zoom collects data from user interactions with Zoom IQ features, which can be used to train Zoom's AI models.

• The trial of Zoom IQ's meeting summaries and chat compose features is available for a limited time to select Zoom subscription plans, with pricing information to be revealed in the future.

CODE/DEVTOOLS

Google has unveiled a major update to its AI chatbot Bard, enhancing its logic and reasoning abilities. Link.

• The latest version of Bard can now write and execute code independently, allowing it to solve problems more deeply than previous AI systems.

• The new technique called "implicit code execution" enables Bard to identify computational prompts and run code in the background.

• Bard's improved capabilities should result in more accurate responses to mathematical tasks and coding questions.

• Traditional language models like ChatGPT and Bard have been limited in their reasoning and math abilities, focusing more on language and creative tasks.

• Google claims that the new method has increased Bard's accuracy in coding and math problems by approximately 30% during internal tests. However, accuracy still remains a challenge for AI chatbots, and unreliable information generated by AI is known as hallucinations.

Google has announced the general availability of generative AI services based on Vertex AI, its machine learning platform as a service (ML PaaS) offering from Google Cloud. Link.

• Developers can now access new tools and models, including the world completion model driven by PaLM 2 and the Embeddings API for text, through the Vertex AI platform.

• Generative AI Studio, which allows fine-tuning and deployment of customized models, is now available, offering enterprise-grade data governance, security, and safety features.

• The Model Garden provides access to base models and offers over 60 models, with the ability to add more in the future. The Codey model for code completion, code generation, and chat is now in public preview.

• Vertex AI provides a comprehensive set of tools for tuning, launching, and managing models in production, including Reinforcement Learning with Human Feedback (RLHF).

• Google highlighted case studies of customers utilizing generative AI, such as GA Telesis, GitLab, Canva, Typeface, and DataStax. Additionally, Google has made Enterprise Search on Gen App Builder easier to use, enabling the creation of chatbots and search engines using generative AI and semantic search technologies.

Instabase, a startup providing an apps platform for analyzing unstructured data, has raised $45 million in a Series C round, valuing the company at $2 billion. Link.

• The funding round was led by Tribe Capital, with participation from Andreessen Horowitz, New Enterprise Associates, and Spark Capital.

• Instabase's platform enables organizations to apply AI innovations to understand unstructured data and build applications.

• The platform offers tools for content understanding, processing documents and data for companies to query and analyze.

• Companies can build their own apps or choose pre-built apps from Instabase's marketplace for tasks like income verification, invoice processing, and more.

• Instabase competes with cloud-based document processing and workflow automation tools, but its advantage lies in minimizing the need for time-intensive data science work.

Onboard data company Shift5 has raised $33 million in Series B funding, led by Moore Strategic Ventures, with participation from Booz Allen Hamilton’s venture capital arm, Teamworthy Ventures, and JetBlue Ventures. Link.

• This brings the total funding for Shift5 to $108 million, with the Series B round now totaling $83 million.

• Shift5 defends commercial transportation systems and military platforms against operational failures and OT cybersecurity risks.

• The company has experienced significant growth, doubling its annual recurring revenue and customer base in the last year, across the Department of Defense (DoD) and commercial aerospace and rail sectors.

• The funding will be used to support Shift5's infrastructure as it expands its commercial and federal businesses to address the challenge of onboard observability.

• Shift5 recently partnered with JetBlue Airlines to bring cybersecurity and data observability to commercial aircraft and obtained its first cross-platform Authority to Operate (ATO) Certification from the DoD. The company also received its first investment aligned with Booz Allen Hamilton's National Cyber business.

Granica, an AI efficiency platform startup, has raised $45 million in funding from investors including New Enterprise Associates, Bain Capital Ventures, and individual investors like former Tesla CFO Deepak Ahuja and Eventbrite co-founder Kevin Hartz. Link.

• Granica has emerged from stealth mode with a platform designed to help customers of Amazon Web Services (AWS) and Google Cloud Storage eliminate redundant and low-value data.

• The platform is claimed to be the industry's first AI efficiency platform, leveraging data reduction and data privacy to derive maximum value from growing volumes of training data.

• By reducing the size and cost of petabyte-scale AI training data by up to 80% using compression and deduplication algorithms, Granica enables more AI data to be captured, stored, and used cost-effectively.

• The service also ensures the privacy of sensitive information in object data, improving data security posture while enabling safe use in AI and analytics.

• The company offers an outcome-based pricing model, charging users a small percentage of the savings generated each month, ensuring that the deployment of the solution delivers upside only and helps organizations save costs.

ROBOTICS

Segway-Ninebot is partnering with Drover AI and Luna Systems to integrate their computer vision technology into its AI-enabled e-scooters. Link.

• The partnerships come as a pivot for Segway, which previously offered a vertically integrated solution with its AI-powered scooter.

• Drover and Luna have been leading the camera-focused scooter advanced rider assistance systems (ARAS) movement by providing attachable IoT modules to various scooter operators.

• Segway decided to focus on its core competencies of building hardware and working with partners for software development, as it realized the challenges of scaling AI-based scooters in multiple cities.

• Customers can choose to implement either Drover's or Luna's scooter ARAS algorithms on Segway's S90L models, or retrofit Segway's modular AI system called Pilot Edge for scooter ARAS capability.

• Segway's partnerships allow Drover and Luna to focus on software development while avoiding the challenges and costs associated with hardware production and inventory management.

POLICY/LAW/ETHICS

The US Senate is working on a plan to regulate AI after recognizing its potential impact on society. Link.

• However, the process of introducing a comprehensive bill and passing legislation is expected to take months, in contrast to the rapid adoption and investment in generative AI.

• Senators plan to be briefed on AI facts during the summer before considering legislation in the following months.

• Senate Majority Leader Chuck Schumer has announced a series of closed-door information sessions to educate senators on AI, covering topics such as capabilities, competition, and national security.

• Schumer previously proposed shaping the bill around four principles: transparency and democratic values.

• Lawmakers are still undecided on how to coordinate various legislative proposals, with options including forming a select committee or having different committees work on different pieces of legislation.

The UK government plans to host a global AI summit this fall to discuss safety measures for evaluating and monitoring significant risks from AI. Link.

• The aim is to convene key countries, leading tech companies, and researchers to coordinate international action and develop a shared approach to mitigate AI risks.

• The UK government is positioning itself as an AI industry ally, which may affect international efforts to establish AI regulations.

• The announcement comes after Prime Minister Rishi Sunak met with AI executives and highlighted developments of global companies expanding their AI work in the UK.

• The EU and the US are already involved in discussions and initiatives for AI regulation, with the EU working on a draft legislative framework and an AI Code of Conduct.

• The UK's position on international AI rules and alignment with the US AI industry may influence global AI regulations and the transatlantic collaboration.

EvenUp, a startup founded by Rami Karabibar and Ray Mieszaniec that uses AI to generate legal documents for assessing personal injury cases, has raised $50.5 million in a Series B funding round led by Bessemer Venture Partners, bringing its total raised to $65 million. Link.

• The platform aims to level the playing field in personal injury cases by providing lawyers with AI-generated letters arguing for proposed compensation based on raw case files.

• EvenUp covers various categories of personal injury cases, including motor vehicle accidents, police brutality, and natural disasters.

• The platform extracts relevant information from documents and organizes them into templated demand packages for personal injury claims.

• EvenUp is one of the first startups to focus on AI-aided personal injury litigation, aiming to maximize the value of each claim and challenge the existing settlement mills.

The European Union is urging signatories to its Code of Practice on Online Disinformation to label deepfakes and other AI-generated content. Link.

• The EU wants technology to be implemented to recognize AI content and clearly label it for users.

• The current version of the Code does not require identifying and labeling deepfakes, but the EU is seeking to change that.

• The EU sees two discussion angles for including mitigation measures: services integrating generative AI and services with potential to disseminate AI-generated disinformation.

• The EU wants platforms to implement labeling for deepfakes and other AI-generated content immediately.

• Twitter recently withdrew from the voluntary EU Code, while other issues discussed included war propaganda, fact-checking, election security, and data access for researchers.

OTHER

Salesforce has unveiled two new generative AI products, Marketing GPT and Commerce GPT, at its Connections conference. Link.

• These products will power Salesforce's Marketing Cloud and Commerce Cloud, enabling enterprises to streamline workflows and deliver personalized campaigns and shopping experiences at scale.

• Marketing GPT utilizes the Salesforce Data Cloud and Einstein GPT assistant, allowing users to interact with their Marketing Cloud system using natural language and perform tasks like querying data profiles and generating personalized emails.

• The Marketing Cloud will also receive AI-driven segment intelligence and rapid identity resolution capabilities for comprehensive campaign performance analysis and personalized experiences.

• Commerce GPT focuses on creating personalized shopping experiences by using Data Cloud and Einstein GPT to generate dynamic product descriptions and translate them into different languages.

• Salesforce sees generative AI as a significant area of focus and plans to introduce more GPT innovations in other Customer 360 products like Sales and Service Cloud in the near future. CEO Marc Benioff believes generative AI will revolutionize the world in unprecedented ways.

CentML, an AI startup based in Toronto, has emerged from stealth and aims to address the global shortage of GPUs required for training and inference of generative AI models. Link.

• The company extends the utilization of existing AI chip supply and legacy inventory without compromising accuracy, thus increasing access to compute in the GPU marketplace.

• CentML raised a $3.5 million seed round led by Radical Ventures in 2022 and recognizes the challenges smaller companies face in accessing GPUs dominated by larger players like Nvidia.

• By optimizing models on a wide range of available GPUs, including lower-cost alternatives, CentML serves a larger part of the market and offers more affordable options.

• The cost of inference is growing significantly, and CentML uses an open-source compiler to tune optimizations specifically for a company's inference pipeline and hardware.

• CentML's offering improves efficiency in the market, enables complex models to run on legacy hardware, and increases the overall supply of AI chips available for use.

Teaser AI is a new dating app developed by the team behind Dispo, offering a chat feature with AI versions of potential matches before swiping. Link.

• The app aims to use AI conversations as icebreakers to facilitate real-life meetings rather than replacing human interaction.

• Users answer personality questions, and the AI attempts to imitate their personalities based on their responses and chat interactions.

• The AI-generated conversations can be unpredictable, sometimes producing strange or inaccurate statements.

• Teaser AI limits the conversation to five back-and-forth messages and prevents unsavory or overly intimate discussions.

• The app focuses on addressing issues of ghosting and aims to encourage actual conversations and connections among users.

The team behind Replika, an AI friendship bot, has launched Blush, an AI dating sim designed to help users build relationship and intimacy skills. Link.

• Blush's chatbot is designed for interactions beyond erotic conversation, focusing on helping users practice emotional intimacy and navigate complex relationship issues.

• The app is structured like a conventional dating app, introducing users to over 1,000 AI "crushes" that can help them practice dating and intimacy.

• Blush is available in the App Store, with a premium version costing $99/year, offering more NSFW conversations with avatars.

• The goal of Blush is to destigmatize AI, intimacy, and romance, providing a different approach to building relationship skills rather than focusing on sex bots.

• The team worked with a psychotherapist specializing in couple's therapy and trauma to develop the avatars' backstories and ensure appropriate responses to conflict and real-life relationship issues.

Strategic finance platform Mosaic has raised $26 million in a Series C funding round led by OMERS Ventures, with participation from existing investors Founders Fund, General Catalyst, and Friends and Family Capital. Link.

• Mosaic aims to empower SMB and mid-market finance leaders with AI-driven decision-making functionality to balance high growth with capital efficiency.

• The funding will be used to expand Mosaic's investment in AI and develop features such as automated forecasts, model roll-forward, and metric analysis.

• Mosaic's platform eliminates tedious manual tasks and complex implementations of legacy FP&A solutions, enabling finance teams to focus on strategic tasks.

• The company's fast time-to-value through real-time analytics and dynamic planning features makes data instantly actionable.

• Mosaic has experienced impressive growth, achieving a 300% increase amid the current macroeconomic downturn, and has been recognized by GGV Capital, Crunchbase, and Nasdaq.

GPT Tip of the Week

Prompt: What will be the multitude of characteristics for winning transformer models in the EdTech space?

Response: In the EdTech space, winning transformer models are likely to possess a multitude of characteristics that cater to the specific needs and challenges of education. Here are some key characteristics that can contribute to the success of transformer models in the EdTech domain:

1. Language Understanding and Generation: Winning transformer models should excel in understanding and generating natural language text, enabling them to comprehend student queries, provide accurate responses, and generate high-quality educational content.

2. Adaptability: The ability to adapt and personalize the learning experience to individual students' needs is crucial. Winning models should be capable of adjusting their responses, content, and teaching style based on the learner's level of knowledge, learning pace, and preferences.

3. Domain Expertise: EdTech transformer models should possess a strong foundation of domain-specific knowledge across various educational subjects, such as mathematics, science, literature, history, etc. This expertise enables them to provide accurate and comprehensive information to students.

4. Pedagogical Effectiveness: Effective models should go beyond simply providing information and act as effective teachers. They should be able to explain complex concepts in a simplified manner, guide students through problem-solving processes, and offer constructive feedback and suggestions.

5. Multimodal Capabilities: Integrating multimodal capabilities, such as image recognition, video analysis, and speech recognition, can enhance the learning experience. Winning models should be able to process and interpret different types of media, enabling them to support a wide range of educational content.

6. Continual Learning: The ability to continuously learn and improve is vital in the EdTech space. Models should be capable of integrating new knowledge and insights, staying up-to-date with the latest educational trends, curriculum changes, and research findings.

7. Ethical Considerations: Transformer models in EdTech should prioritize student privacy, data security, and ethical use of data. Winning models should adhere to strict privacy guidelines, maintain confidentiality, and ensure that the data collected from students is used responsibly and transparently.

8. Scalability and Accessibility: The models should be scalable to handle a large number of users simultaneously, ensuring smooth user experiences even during peak usage periods. Additionally, they should prioritize accessibility by considering diverse learning needs, such as providing support for learners with disabilities or language barriers.

9. User Engagement and Interactivity: Engaging and interactive models can foster active learning and maintain student motivation. Winning models should employ interactive elements, gamification techniques, and adaptive learning strategies to create an immersive and enjoyable learning environment.

10. Collaborative Learning and Social Interaction: Encouraging collaboration among students and facilitating social interaction is important for effective learning. Transformer models should enable group activities, discussions, and peer-to-peer interactions to foster a sense of community and enhance collaborative learning.

It's important to note that the characteristics mentioned above are based on the current understanding of EdTech requirements. The field is dynamic, and the specific characteristics of winning transformer models may evolve as new research, technological advancements, and educational practices emerge.