AI Weekly: 05/22/23

Good morning and welcome to this week’s edition of AI Weekly! We had another big week of news headlined by Meta’s decision to open-source their LLM, LLaMA, Zoom’s partnership with Anthropic to integrate Claude into their platform, and a new research study that has shown AI’s ability to detect early signs of Alzheimer’s disease.

In other news, Apple released a new accessibility feature called Personal Voice, which allows individuals with ALS to create a digitized version of their own voice, while OpenAI has launched an official mobile app for ChatGPT.

Oh, and one company is building AI robots to do our work around the house…or take over the world. More on that and other disruptive AI happenings below!

- ZG

Here are the most important stories of the week:

TEXT

OpenAI has launched an official iOS app for ChatGPT, allowing users to access the AI chatbot on their mobile devices. Link.

• The app is free, ad-free, and supports voice input, initially available to users in the United States.

• Users can interact with ChatGPT to ask questions, seek advice, find inspiration, conduct research, and more.

• The app syncs user history across devices and integrates with Whisper, OpenAI's speech recognition system, enabling voice input.

• ChatGPT Plus subscribers can access GPT-4's capabilities through the app, along with early access to new features and faster response times.

• The rollout begins in the U.S. and will expand to other countries in the coming weeks, with an Android version also in development.

Meta is choosing to make its large language model, LLaMA, open source, in contrast to the strategy of its competitors Google, Microsoft, and OpenAI. Link.

• Yann LeCun, Meta's chief AI scientist, believes that keeping powerful models closed is a "huge mistake" and that the winning platform will be the open one.

• Meta began experimenting with open-source by making its large language model's underlying code available to selected individuals via email, but it was leaked almost immediately.

• Critics argue that making such technology widely available could lead to the dissemination of problematic or dangerous content.

• LeCun disagrees and believes that while people can create such content, it can be prevented from spreading, and progress is faster in an open ecosystem.

• The debate between open-source and closed-source approaches in the AI industry continues, with the outcome yet to be determined as companies stake their positions in the growing AI arms race.

Sudowrite, an AI-focused startup, has launched Story Engine, an AI tool for writing long-form stories. Link.

• The tool is based on OpenAI's GPT-3 model and aims to generate complete novels in just a few days.

• Sudowrite claims that the generated stories are true to the author's vision and can include plot points, character arcs, and descriptions.

• The tool is designed to help writers collaborate with AI and streamline the writing process.

• The launch of Sudowrite and its AI-generated storytelling has sparked controversy, especially amid ongoing strikes by writers demanding job security and protection against AI replacing human writers.

• Critics argue that AI-generated stories cannot fully reflect the author's vision and undermine the craft and creativity of human writers.

IMAGE/VIDEO

Zoom has entered into a strategic partnership with Anthropic, an AI company specializing in AI safety research. Link.

• The collaboration aims to integrate Anthropic's AI assistant, Claude, into the Zoom platform, including the Zoom Contact Center.

• Claude can serve as a customer service agent, sales representative, document parser, search tool, coach, and administrative assistant.

• When integrated with Zoom, Claude will guide customer service interactions to improve resolutions and customer experiences.

• Claude will also function as a self-service tool, intelligently assisting users to achieve the best outcomes and enhance productivity in the Zoom Contact Center.

• Zoom will utilize Claude to provide insights for managers to coach customer service representatives and improve the quality of interactions.

Researchers have developed a new AI system called DragGAN that allows users to manipulate images by dragging elements of a picture to change their appearance. Link.

• The system enables users to modify the dimensions of objects, change facial expressions, rotate subjects as if they were 3D models, and adjust reflections and heights in landscapes.

• The research team has provided video demonstrations of the image manipulation capabilities, showcasing the flexibility and precision of the DragGAN model.

• The user interface of DragGAN is similar to traditional image-warping techniques but utilizes AI to generate new content rather than smudging existing pixels.

• The model is capable of generating occluded content, such as the teeth inside a lion's mouth, and can deform objects based on their rigidity, like bending a horse leg.

• Although this is currently a research paper and the complete evaluation of the technology is not possible, it demonstrates progress in making image manipulation more accessible and introduces new possibilities for editing images.

Stability AI has released StableStudio, an open source version of DreamStudio, their commercial AI-powered design suite. Link.

• Stability AI aims to foster community-driven development and create a world-class user interface for generative AI through open-source collaboration.

• StableStudio is an image generation platform aligned with rival generative image platforms, replacing API calls on the back end with a plug-in system.

• DreamStudio will continue as Stability AI's preferred, managed instance of StableStudio, staying up to date whenever possible.

• The release of StableStudio reflects Stability AI's commitment to open-source and transparency, allowing the community to contribute to its development.

• Stability AI has faced challenges in monetizing its efforts and generating revenue, prompting a renewed push for original research and potential fundraising to the tune of a $4 billion valuation.

SPEECH/AUDIO

Apple has announced a new accessibility feature called Personal Voice, which allows individuals with amyotrophic lateral sclerosis (ALS) to create a digitized version of their own voice. Link.

• The feature can be used on iPhones, iPads, and Macs, requiring only 15 minutes of spoken phrases for training.

• Personal Voice processes the voice data locally, ensuring privacy and data security.

• Users can turn their typed text into synthesized speech in their own voice, improving communication as ALS progresses.

• The development of Personal Voice involved close collaboration with the ALS community, including Team Gleason, a nonprofit organization founded by former NFL player Steve Gleason.

• The feature is part of Apple's commitment to accessibility and was announced as part of Global Accessibility Awareness Day.

CODE/DEVTOOLS

ClearML has launched ClearGPT, an enterprise-grade generative AI platform designed to deploy and utilize large language models securely and efficiently at scale. Link.

• ClearGPT allows customization of LLMs to align with an organization's specific data, ensuring comprehensive protection of intellectual property, compliance, and knowledge.

• The platform enables organizations to develop internal and external products faster, outperform competitors, and generate new revenue streams using AI.

• ClearGPT securely handles the end-to-end workflow, including data gathering, model training, and quality assessment, within the organization's secure network.

• The platform eliminates concerns regarding security, performance, cost, data governance, and customization associated with existing LLM solutions.

• ClearGPT provides a secure environment within the organization's network, ensuring complete control, data privacy, and compliance with standards such as ISO 27001, SOC2, HIPAA, and GDPR.

• The platform offers real-time feedback, extensive customization options, and reduced operational costs compared to other LLM solutions.

• ClearGPT supports enterprise-grade LLMs, empowering organizations to develop automated AI-driven processes and gain new business insights.

• ClearML believes ClearGPT will revolutionize the enterprise AI landscape by driving innovation, productivity, and efficiency at scale, while maintaining control and ownership of data and models.

Meta has developed a generative AI tool for coding called CodeCompose, similar to GitHub's Copilot. Link.

• CodeCompose is used internally by Meta's teams for code suggestions in Python and other languages while coding in IDEs like VS Code.

• CodeCompose models have 6.7 billion parameters, slightly fewer than Copilot's model, and are fine-tuned on Meta's first-party code, including internal libraries and frameworks.

• CodeCompose makes suggestions such as annotations, import statements, and complete lines or chunks of code as users type.

• Meta claims that thousands of its employees accept CodeCompose's suggestions weekly, with an acceptance rate of over 20%.

• Meta acknowledges the controversies surrounding code-generating AI tools, such as copyright infringement and security vulnerabilities, but emphasizes that developers are not obligated to follow CodeCompose's suggestions, and security was a major consideration in its development.

Together, a startup focused on open source generative AI, has raised $20 million in a seed funding round led by Lux Capital. Link.

• The company aims to create an open ecosystem for generative AI models and services to help organizations incorporate AI into their production applications.

• Together's cloud platform offers scalable compute at lower prices than dominant vendors like Google Cloud, AWS, and Azure.

• The startup's first projects include RedPajama, which aims to foster a set of open source generative models, and GPT-JT, a fork of the open source text-generating model GPT-J-6B.

• Together's platform is designed to simplify the training, fine-tuning, and productization of open source generative models, addressing the challenges faced by enterprises.

• The company emphasizes transparency, control, and privacy as key advantages of its open source approach, attracting interest from regulated enterprises.

HEALTHCARE

Researchers at the University of Alberta have developed a machine learning model that shows promise in diagnosing Alzheimer's dementia at earlier stages. Link. Link.

• The model achieved 70 to 75 percent accuracy in distinguishing Alzheimer's patients from healthy individuals.

• Early detection of Alzheimer's is challenging, but it allows patients to take action and start treatment sooner.

• The goal is to create a simple screening tool that can be used via smartphones, aiding in early detection and providing convenient access to telehealth services.

• The model analyzes speech characteristics rather than specific words, focusing on acoustic and linguistic speech features that transcend language.

• The model can potentially be used across different languages, benefiting a broader range of individuals.

• The tool's user experience is straightforward: users speak into the tool, and it provides a prediction of whether they have Alzheimer's or not, which can then be shared with healthcare professionals for further assessment and action.

POLICY/LAW/ETHICS

OpenAI CEO Sam Altman and other tech executives are urging Congress to regulate generative AI to limit its potential dangers. Link.

• Altman's regulatory wish list includes the establishment of a licensing agency, testing of potentially dangerous AI models before deployment, and independent audits.

• The executives appeared before a Senate Judiciary subcommittee to discuss the potential misuse of generative AI.

• There is a consensus among the witnesses that international bodies should help set standards and monitor AI.

• However, AI regulation may face challenges in Congress, as past efforts to pass tech-related legislation have been difficult.

• Concerns raised by lawmakers about generative AI include election misinformation, job disruption, non-English language weaknesses, copyright issues, and harmful content.

France's privacy regulator, CNIL, has published an action plan for AI, focusing on privacy-friendly AI systems and the impact of AI on individuals. Link.

• CNIL has set up a dedicated Artificial Intelligence Service to scope the tech and produce recommendations for privacy-friendly AI systems.

• The regulator aims to establish clear rules to protect personal data and contribute to the development of privacy-friendly AI systems.

• CNIL will prioritize areas such as the protection of publicly available data from scraping and the fairness and transparency of data processing underlying AI tools.

• The action plan aligns with the ongoing work in the EU to develop a risk-based framework for regulating AI.

• CNIL has received complaints against OpenAI's ChatGPT and is working on clarifications regarding the legal framework for training and using generative AIs.

OTHER

Scientists from Stanford University have developed an electronic skin that enables amputees to experience a human-like sense of touch through prosthetic limbs. Link.

• The artificial skin contains sensors for temperature, pressure, and strain, which convert these signals into electrical impulses that can communicate with the brain, resembling the way nerve impulses function.

• The wearable electronic circuit, called a monolithic e-skin, offers the potential for remote control of robotic limbs while feeling objects and sensations.

• The research team overcame the challenge of integrating skin-like materials into complex circuits that generate nerve-like pulse trains while maintaining low operating voltage for safety on the human body.

• The e-skin prototype, with the thickness of a paper, successfully combines the desired electrical and mechanical features of human skin in a soft and durable form.

• Future plans involve enhancing scalability and developing an implantable chip for wireless communication through the body's peripheral nerve, expanding the capabilities of the technology.

Open-source startup Union AI has raised a $19.1 million Series A funding round led by NEA and Nava Ventures and has announced the general availability of its Union Cloud service. Link.

• Union AI's platform is built around Flyte, an open source tool for building production-grade workflow automation platforms with a focus on data, machine learning, and analytics.

• Flyte was originally developed at Lyft to address the challenges of putting machine learning models into production and has since been open sourced.

• Union AI aims to provide infrastructure tooling specifically tailored for machine learning teams, recognizing the differences between software deployments and AI model deployments.

• The Flyte platform is already being used by companies such as HBO, LinkedIn, Spotify, Stripe, and ZipRecruiter, delivering improved ML productivity and business gains.

• Union AI's platform also includes Pandera, a data testing framework, and Union ML, a framework for building and deploying models using existing tools, with Union Cloud providing enterprise features and data control.

Sanctuary AI, a Vancouver-based firm, has unveiled its humanoid robot called Phoenix, standing 5'7" tall and weighing 155 pounds. Link.

• The robot is designed to augment or replace humans and can lift payloads up to 55 pounds and travel up to three miles per hour.

• Phoenix features complex hands with 20 degrees of freedom and proprietary haptic technology to mimic the sense of touch.

• In a limited pilot deployment at a retail store, the fifth-generation system completed various retail-related tasks, including picking, packing, cleaning, and labeling.

• Sanctuary AI aims to create general-purpose robots that can perform a broad range of work tasks using its Carbon-based AI control system.

• The company has secured significant funding, including a Series A round of nearly $60 million and a $30 million investment from the Canadian government.

Procurement software startup Zip has raised $100 million in a Series C funding round led by Y Combinator, valuing the company at $1.5 billion. Link.

• The funding will be used to develop new applications of generative AI and drive adoption of Zip's procurement platform.

• Zip's platform streamlines the end-to-end intake-to-pay process by routing purchase requests for approval across various teams and integrating with enterprise resource management software.

• The company plans to leverage AI to further improve efficiency and provide insights for customers.

• Despite competition, Zip has experienced significant growth, doubling its customer roster and tracking $937 million in total savings to date.

• The Series C funding will support the expansion of Zip's footprint, including opening a new office in Dallas alongside its existing San Francisco headquarters and Toronto office.

GPT Tip of the Week

Prompt: What are the different factors that may differentiate throughout different open-source LLMs?

Response: Open-source language models (LLMs) can differ in various factors depending on the design choices made by their developers and the specific goals and priorities of each project. Here are some factors that may differentiate different open-source LLMs:

1. Model Architecture: Different LLMs may use distinct architectural designs. They can vary in terms of their underlying neural network architecture, such as recurrent neural networks (RNNs), transformers, or hybrid models. These architectural variances can impact the model's performance, computational requirements, and training dynamics.

2. Training Data: The training data used to train an LLM can significantly influence its behavior and capabilities. Open-source LLMs can be trained on diverse datasets, including publicly available text corpora, books, websites, and other sources. The size and quality of the training data, as well as any preprocessing techniques applied, can vary between LLMs.

3. Training Objectives: Different LLMs can have specific training objectives based on their intended use cases. For example, some LLMs may focus on general language understanding and generation, while others might prioritize specific domains like scientific literature, legal texts, or medical records. The training objectives impact the model's specialization and the type of language it is proficient in.

4. Model Size: LLMs can vary in terms of their model size, measured in terms of the number of parameters. Larger models tend to have more capacity for capturing complex language patterns but require more computational resources during training and inference.

5. Ethical Considerations: Open-source LLMs may have different ethical considerations and guidelines incorporated into their development. For example, some projects might prioritize minimizing biases in the model's outputs or have mechanisms to handle sensitive or inappropriate content.

6. Developer Community: Each open-source LLM project has its own developer community, which can influence the model's direction, updates, and the availability of additional resources, such as pre-trained models, fine-tuning techniques, or specialized tooling.

7. Licensing and Governance: Open-source LLMs may have different licensing agreements that dictate how the model can be used, modified, or redistributed. Furthermore, the governance structure of the project, including decision-making processes and community participation, can vary and impact the long-term sustainability and evolution of the LLM.

8. Integration and Accessibility: Different LLMs might provide varying levels of integration and accessibility options. Some projects focus on easy-to-use APIs and developer-friendly interfaces, while others prioritize research and experimentation, offering more flexibility but requiring additional technical expertise to leverage effectively.

It's important to note that these factors are not exhaustive and the specific differences between open-source LLMs will depend on the individual projects and their objectives. It is recommended to refer to the documentation and resources provided by each LLM project to understand their unique characteristics and considerations.