AI Weekly: 07/17/23

Good morning and welcome to this week’s edition of AI Weekly! In this week’s news, Anthropic has released Claude 2, Stability launched a new sketch-to-image model, and a Madrid-based startup called Clibrain has released its new LLM specifically optimized for Spanish.

In policy/regulation/ethics news, the FTC is investigating OpenAI for possible violations of consumer protection law, China has introduced new rules to regulate generative AI, and Kenyan workers who were involved in content moderation for ChatGPT have filed a petition calling for investigations into Big Tech outsourcing and the conditions of work in Kenya.

Also, Chipotle has created an AI-driven robot called “Autocado” to automate the tedious tasks of creating guacamole, such as cutting, coring, and peeling avocados. Enjoy reading about more of last week’s AI news below!

- ZG

Here are the most important stories of the week:

TEXT

Google has launched an AI-based chatbot called Google Bard and an AI-first notebook called NotebookLM. Link.

• NotebookLM is different from traditional chatbots as it uses AI to enhance understanding of existing content and allows users to ground the language model in their notes and sources.

• When a document is uploaded to NotebookLM, it generates a summary, key topics, and questions for better comprehension.

• NotebookLM can generate glossaries, summaries of specific aspects, and potential questions for investors based on pitch decks.

• The product is currently being tested by a small group of users in the US, and Google intends to refine the technology based on user feedback before expanding its availability.

• Google also announced the launch of certain Google Bard features, including the ability to use images in prompts, which were initially introduced at Google I/O.

Anthropic is releasing an updated version of its chatbot, Claude, called Claude 2, promising improved performance, more detailed answers, and fewer harmful responses. Link.

• Claude 2 is designed to be more conversational and explain its thinking, aiming to mimic human-like interactions.

• The new version can process 100,000 tokens, a significant upgrade from the previous version's 9,000 tokens, enabling it to understand and process over 70,000 words at once.

• Claude 2 has the capability to process large amounts of information, including hundreds of pages of documents, and generate longer responses, such as short stories.

• Anthropic highlights Claude 2's intelligence by mentioning its performance on the multiple-choice portion of the bar exam, scoring 76.5% compared to the previous version's score of 73%. It also claims improved coding skills, with a score of 71% on the Codex HumanEval.

• Anthropic emphasizes that Claude 2 is significantly safer, with a reduced likelihood of issuing offensive or dangerous responses, stating that manual analysis shows it is twice as good at giving safe responses compared to the previous version.

Clibrain, a Madrid-based AI startup, has released Lince Zero, a Spanish-language-tuned LLM designed to optimize generative AI for Spanish speakers. Link.

• Lince Zero is based on open source technologies and offers linguistic quality and comprehension of Spanish nuances through training on a dedicated corpus of Spanish language data.

• Clibrain's goal is to provide a high-quality model with a linguistic edge for Spanish queries, targeting the close to half a billion native Spanish speakers globally.

• The startup plans to develop a more powerful foundational model called Lince, which will be available via API to paying customers and embedded into its communication and productivity apps.

• Clibrain aims to offer proprietary models in the future, including multimodal models that can respond to images and audio.

• The company was founded in April, spun out of a car-backed loan entity, and is bootstrapping development with funds from previous startup exits, potentially considering external fundraising in the future.

IMAGE/VIDEO

Meta claims to have achieved a breakthrough with CM3Leon, an AI model that excels in text-to-image generation and can generate captions for images. Link.

• CM3Leon is a transformer model that leverages attention mechanisms, making it more efficient and faster to train than previous transformer-based methods.

• Meta used supervised fine-tuning to improve CM3Leon's performance on image generation and caption writing, enabling it to follow text instructions and answer questions about images.

• CM3Leon outperforms other models in generating images with complex objects and following diverse text prompts.

• The model shows potential in editing existing images and generating captions, surpassing specialized image captioning models.

• Meta acknowledges the challenge of bias in generative AI models but does not address it directly, emphasizing the need for transparency in the industry's progress.

Shutterstock plans to expand its deal with OpenAI, providing the startup with training data for its AI models. Link.

• Over the next six years, OpenAI will license data from Shutterstock, including images, videos, music, and metadata, while Shutterstock gains priority access to OpenAI's latest technology and editing capabilities.

• OpenAI will work with Shutterstock to bring generative AI capabilities to mobile users through Giphy, which Shutterstock acquired from Meta.

• Shutterstock's CEO sees the partnership as a commitment to driving AI tech innovation and positioning Shutterstock as the preferred data and distribution partner in generative AI.

• Stock content galleries and generative AI startups have had a contentious relationship, with concerns over copyright infringement and compensation for artists and photographers.

• Unlike Getty Images, Shutterstock has embraced generative AI, partnering with OpenAI and establishing licensing agreements with other companies, while also offering compensation to artists through a contributor fund.

Stability AI, the company behind Stable Diffusion, is launching a sketch-to-image service called Stable Doodle. Link.

• Stable Doodle utilizes the Stable Diffusion XL model and a conditional control solution called T2I-Adapter developed by Tencent's Applied Research Center.

• The service allows users to generate high-quality original images from sketches by providing a prompt to guide the image generation process.

• Stable Doodle aims to offer more precise control over image generation compared to other sketch-to-image AI tools.

• It targets professionals and novices alike, enabling anyone with basic drawing skills to create visually pleasing images quickly.

• Stable AI plans to expand the use cases of Stable Doodle, including real estate applications, and sees it as a tool to enhance efficiency and productivity for designers and illustrators.

Kakao Brain, the AI division of South Korean tech giant Kakao, has announced updates to its AI image generator Karlo and its KoGPT large language model. Link.

• Karlo 2.0 is an AI-powered tool that can generate images from text prompts in seconds, supporting both English and Korean inputs.

• Kakao Brain has also established a 10 billion won ($7.7 million) fund called Karlo 100X to support AI startups focused on image generation technology.

• The updates reflect the pressure faced by incumbent tech companies to keep up with AI advancements from newcomers like OpenAI and Midjourney.

• Karlo 2.0 is trained on 300 million texts and images, offering better image quality and clarity, and includes a publicly-available API for developers.

• KoGPT 2.0, to be released in Q3 2023, will be trained on 30 billion parameters and will include additional languages such as English and Japanese.

SPEECH/AUDIO

Meta AI has developed a dataset that clusters speech at the "utterance level," aiming to improve the performance of automatic speech recognition (ASR) tools. Link.

• Traditional ASR training datasets organized by demographic limit the variation of pronunciations, hindering understanding across diverse users.

• The new dataset includes 27,000 command utterances collected from 595 US volunteers, focusing on seven main themes like music, capture, and messaging.

• The dataset was used to train a model and evaluate its performance on other datasets, showing promising results with a 10% increase in ASR performance using the clustering method.

• The gains were particularly notable in accent inclusivity and the age 66-85 demographic, which is traditionally underrepresented in voice command applications.

• Meta AI sees the clustering algorithm as part of its broader efforts towards responsible AI and is exploring adapting the system to other languages in the future.

CODE/DEVTOOLS

Pinecone has announced a partnership with Microsoft Azure to accelerate generative AI applications for Azure customers. Link.

• Pinecone, a leading provider of vector database technology, has experienced rapid adoption, with over 100,000 free users and 4,000 paying customers.

• Vector databases enable developers to access domain-specific information not available on the internet or traditional databases, providing better context and accuracy for generative AI models.

• While some experimentation involves larger context windows for LLM prompts, Bob Wiederhold, Pinecone's President and COO, believes that manageable amounts of information yield better results.

• Large enterprises may experiment with training their own models or fine-tuning existing models, bypassing the need for vector databases, but these approaches require expertise and can be costly.

Vellum.ai has closed a $5 million seed round with participation from Rebel Fund, Eastlink Capital, Pioneer Fund, Y Combinator, and several angel investors. Link.

• The startup, founded in January of this year, already has 40 paying customers and sees revenue growth of 25-30% per month.

• Vellum was created by founders who previously worked with GPT-3 at Dover and recognized the growing demand for tooling in generative AI prompting.

• The company provides tools for AI prompters, including side-by-side model output comparison, company-specific data search, and testing and version control.

• Prompt engineering plays a crucial role in refining language models and ensuring consistent and accurate results, requiring a dedicated tooling framework.

HARDWARE/ROBOTICS

Chipotle introduces "Autocado," a robot designed to automate the tedious tasks of creating guacamole, such as cutting, coring, and peeling avocados. Link.

• The decision to automate comes amid labor shortages in the restaurant industry, but Chipotle states that the robot will work collaboratively with employees to speed up guacamole production.

• Autocado is designed specifically for Chipotle's needs and aims to reduce guacamole preparation time by half, potentially saving the company millions in annual food costs.

• This move reflects a larger trend of restaurants automating tasks, with examples such as White Castle's Flippy 2 robot and the introduction of AI-driven drive-thrus.

• An employee loads up to 25 pounds of avocados into Autocado, which then slices, peels, and collects the avocados into a bowl.

• Chipotle partnered with Vebu Labs, a robotics startup, to develop Autocado and plans to invest in other robotic solutions, such as dishwashing robots and a robotic kitchen assistant called "Chippy" for making tortilla chips.

The California Public Utilities Commission (CPUC) is set to vote on whether to allow Cruise and Waymo to expand their paid autonomous ride-hailing services in San Francisco to operate 24/7. Link.

• The vote is considered the most significant regulatory moment for autonomous vehicles in the US in recent years.

• City officials have raised concerns about incidents involving autonomous vehicles, such as blocking traffic, obstructing emergency vehicles, and interfering with firetrucks. The San Francisco Fire Chief and other agencies have logged complaints with the CPUC.

• Activist group Safe Street Rebel has called for placing orange traffic cones on the hoods of robotaxis to disable them, expressing concerns about traffic, surveillance, and potential privacy violations.

• Waymo and Cruise have defended themselves, stating they are working with city officials to improve safety and that no serious injuries or deaths have occurred due to autonomous vehicles in San Francisco.

• The spread of robotaxis is testing San Francisco's willingness to continue being a testbed for futuristic experiments, as public demands for better and safer experiences in cities grow.

HEALTHCARE

Google has been testing its Med-PaLM-2 chatbot at the Mayo Clinic since April, aiming to assist with health-related questions, offer diagnoses, and perform tasks like summarizing medical documents and organizing patient data. Link.

• The chatbot is based on Google's Med-PaLM-2 language model, trained specifically on medical demonstrations and questions from medical-licensing exams, making it better suited for health-related conversations.

• Google's analysis found that the chatbot demonstrated correct comprehension of situations and performed as well as actual physicians, sometimes generating preferred responses.

• Although the chatbot still has accuracy issues, mainly in providing irrelevant answers, Google sees potential for expanding AI applications in the medical world.

• Privacy concerns arise with the use of AI in healthcare, but Google assures users that health data will be encrypted and used exclusively by the chatbot locally, with no access by Google as a corporation.

London-based startup Causaly has raised $60 million in a Series B funding round led by ICONIQ Growth, bringing its total funding to $93 million. Link.

• Causaly has developed an AI platform to help accelerate drug development and testing, working with major pharmaceutical companies and research institutions.

• Its platform focuses on various stages of drug development, including target identification, biomarker determination, and understanding disease pathophysiology.

• By using its platform, Causaly aims to reduce the time it takes to develop a drug from 10-15 years to several years, as well as minimize false starts and dead ends in the discovery process.

• The company's dataset and AI capabilities enable faster modeling, computations, and analysis, improving the efficiency and success rate of drug discovery.

• Causaly's approach stands out as it provides tools and support to biomedical teams rather than developing its own therapeutics, emphasizing collaboration and productivity gains in the industry.

POLICY/LAW/ETHICS

The Federal Trade Commission (FTC) is investigating OpenAI for possible violations of consumer protection law, including its handling of personal data and potential to provide inaccurate or misleading information. Link.

• The probe could complicate OpenAI's relationship with policymakers and draw attention to concerns about generative AI's impact on jobs, national security, and democracy.

• The FTC has issued a 20-page investigative demand to OpenAI, requesting information on data acquisition, accuracy of statements generated by ChatGPT, and risks of harm to consumers.

• The request also seeks testimony, complaint records, details of data leaks, and descriptions of OpenAI's algorithm testing and manipulation processes.

• The FTC's investigation is a significant example of direct US government regulation of AI, as lawmakers are working on legislation in the sector.

• OpenAI has previously faced criticism for algorithmic bias, privacy concerns, and hallucinations by ChatGPT, with some critics filing a complaint to the FTC.

China has introduced new rules to regulate generative AI and become one of the first countries to do so. Link.

• The rules, set to take effect on August 15, apply to AI services available to the general public in China, with exemptions for research institutions and services for overseas users.

• The updated guidelines have relaxed some provisions from the initial draft, indicating China's interest in fostering the growth of the AI industry to stimulate economic growth and job creation.

• The rules encourage innovative use of generative AI and support the development of secure and trustworthy chips, software, tools, computing power, and data sources.

• Providers of generative AI services that can influence public opinion or mobilize the public are required to conduct security reviews and register their algorithms with the government.

• Chinese companies, including Baidu, Alibaba, and JD.com, are waiting for regulatory approval before launching their own generative AI services, which has led to predictions of increased competition between the US and China in emerging technologies.

Kenyan workers who were involved in content moderation for ChatGPT have filed a petition calling for investigations into Big Tech outsourcing and the conditions of work in Kenya. Link.

• The workers allege exploitation, lack of psychosocial support, and exposure to harmful content that resulted in severe mental illness.

• They are seeking regulations to protect workers engaged in outsourcing harmful and dangerous technology and legislation to regulate such outsourcing.

• Sama, the company at the center of the litigations, has clients including Google and Microsoft and focuses on computer vision data annotation.

• OpenAI acknowledges the challenging nature of the work and claims to have established ethical and wellness standards for its data annotators.

• Sama is open to working with the Kenyan government and welcomes third-party audits to ensure fair wages and a dignified working environment.

OTHER

Elon Musk has announced the formation of a new AI startup called xAI, with a focus on understanding reality and the true nature of the universe. Link.

• The company's website provides limited information about its mission and plans, but mentions close collaboration with X Corp, Tesla, and other Musk-owned companies.

• xAI's team includes employees with backgrounds in DeepMind, OpenAI, Google Research, Microsoft Research, Tesla, and the University of Toronto, who have contributed to breakthroughs in the field of AI.

• Elon Musk previously co-founded OpenAI but left in 2018 due to potential conflicts of interest with Tesla's AI operations.

• Musk had expressed plans to start his own AI firm, including the development of "TruthGPT" for maximum truth-seeking AI.

• xAI is planning a Twitter Spaces discussion to introduce the team and allow for questions from listeners.

The Associated Press (AP) has reached a two-year news-sharing deal with OpenAI, allowing access to select news content and technology. Link.

• This agreement is significant as it represents one of the first official news-sharing agreements between a major U.S. news company and an AI firm.

• OpenAI will license a portion of the AP's text archive dating back to 1985 to train its AI algorithms, while the AP will gain access to OpenAI's technology and expertise.

• Technical details of the sharing arrangement are still being worked out between the two companies.

• The AP has been an early adopter of automation technology, using it for corporate earnings reports, local sports coverage, news gathering, production processes, and more.

• The partnership with OpenAI aims to explore responsible use cases for generative AI in news products and services, with the goal of developing standards and best practices for the industry. The AP emphasizes the importance of protecting intellectual property and ensuring fair compensation for content creators in the use of AI technology.

eBay has completed its acquisition of Certilogo, a company specializing in AI-powered digital IDs and authentication for apparel and fashion goods. Link.

• Certilogo's technology helps brands manage the lifecycle of their garments while providing consumers with a way to verify authenticity and access reliable information about items.

• The acquisition demonstrates eBay's focus on enhancing secondhand fashion authentication on its marketplace.

• eBay aims to offer secure, connected product solutions to brands and protect customers from counterfeits through digital product passports.

• The purchase reinforces eBay's position as a trusted destination for pre-loved apparel, giving customers confidence in making sustainable purchase choices.

• This acquisition follows eBay's previous investments in authentication and fraud detection capabilities, including the acquisition of Sneaker Con Digital's authentication business and AI-powered fraud detection company 3PM Shield.

Prolific, a London-based startup that addresses the challenge of finding comprehensive cross-sections of people to respond to questions, has raised £25 million ($32 million) in funding to expand its operations. Link.

• The funding round was co-led by Partech and Oxford Science Enterprises (OSE).

• Prolific was founded in 2014 and provides a network of 120,000 human participants to inform and stress test AI models for organizations like Google, Stanford University, and the European Commission.

• Prolific pays its volunteer freelance participants a minimum of £6 ($8) per hour and has paid out approximately $100 million to them to date.

• While Prolific focuses on providing human intelligence to improve AI, it has no plans to expand its network beyond AI applications but faces potential competition from companies like Amazon, Nielsen, and OpenAI in the future.

Pano AI, a company that provides an AI-powered early fire detection system that sends warnings to firefighters, enabling them to respond faster, recently closed a Series A extension funding round, raising $17 million at a slightly higher valuation than its previous round. Link.

• The company mounts high-definition remote-controllable cameras that scan for smoke and fire, which are analyzed by an AI algorithm.

• By detecting fires early and providing accurate alerts, emergency responders can deploy resources more efficiently and prevent unnecessary deployments.

• The company's customers include power utilities, private landowners, and government fire agencies, with notable customers such as PacifiCorp, Xcel Energy, and Portland General Electric.

• Pano AI currently employs around 45 people and is actively monitoring over 6 million acres of land, with plans to expand its deployments in the US, Australia, and potentially Europe.