AI Weekly: 08/14/23

Good morning and welcome to this week’s edition of AI Weekly! We have some major headlines this week, starting off with NVIDIA’s announcement about their new GH200 super chip, a more powerful processor for generative AI workloads that will be released in the second quarter of 2024. Anthropic has also released their newest version of Claude — Claude Instant — which contains improved performance in areas such as math, coding, reasoning, and safety.

In AI music news, Google and Universal Music Group are reportedly negotiating a partnership to license artists' voices for AI-generated music, opening up a large new set of opportunities for artists to monetize their likenesses.

In healthcare news, scientists at the Francis Crick Institute and UCL Queen Square Institute of Neurology in London have created an AI tool that can classify four subtypes of Parkinson's disease with up to 95% accuracy. Excitingly, this tool’s potential to predict Parkinson's subtypes using images could revolutionize drug testing and personalized treatment.

Oh, and you guessed it: more authors are up in arms about their books being used to train LLMs.

Continue reading for more on the above and other AI news below!

- ZG

Here are the most important stories of the week:

TEXT

Anthropic, a startup co-founded by former OpenAI executives, has released an updated version of its text-generating model, Claude Instant. Link.

• The new version, Claude Instant 1.2, incorporates strengths from Anthropic's flagship model, Claude 2, resulting in improved performance in areas such as math, coding, reasoning, and safety.

• Internal testing shows that Claude Instant 1.2 outperforms its predecessor, achieving higher scores in coding and math benchmarks.

• Claude Instant 1.2 generates longer and better-structured responses, enhances quote extraction, multilingual capabilities, and question answering.

• The new version exhibits reduced text hallucinations and increased resistance to jailbreaking attempts, which involve bypassing safety features.

• Claude Instant 1.2 features a context window size equivalent to Claude 2's, improving its ability to remember recent conversation content.

OpenAI is continuously enhancing ChatGPT despite ongoing debates about its performance changes. Link.

• Logan Kilpatrick, OpenAI's first developer advocate, announced a set of significant updates for ChatGPT.

• New features include example prompts, suggested replies, follow-up questions, default GPT-4 setting for ChatGPT Plus subscribers, and support for multiple file uploads.

• User reactions have been mixed, with some praising the additions and others expressing concerns about distractions or issues with ADHD.

• OpenAI's recent trademark application for GPT-5 hints at future developments, including capabilities like human speech and text production and language translation.

• The adoption and impact of these features remain to be seen as users incorporate them into their workflows and assess their usefulness.

Bud Financial has launched Bud.ai, a generative AI platform designed to enhance customer engagement for banks and financial services organizations. Link.

• The platform's improved logic core generates real-time insights for consumer and corporate users, enabling hyper-personalization and matching customers with suitable financial products.

• By integrating its LLM technology, Bud.ai allows financial services organizations to gain insights from unstructured data and improve customer experiences.

• Bud introduced Jas, a personalized generative chat interface built on the Bud.ai core, offering consumers access to a trained AI assistant for financial planning and credit product recommendations.

• The AI core enhances Bud's flagship products, Assess and Engage, leading to increased engagement and operational efficiency for clients.

• Bud plans to develop specialized AI solutions for areas like risk management, fraud detection, and customer support while prioritizing data security and ethics. The company aims to collaborate with Google to shape the future of financial services.

SPEECH/AUDIO

Google and Universal Music Group (UMG) are reportedly negotiating a partnership to license artists' voices for AI-generated music. Link.

• Google may pay UMG licensing fees to enable users to create AI-generated songs with musicians' voices and melodies.

• The agreement is aimed at developing a legitimate tool for fans to create AI-generated tracks while compensating copyright owners.

• Similar discussions are underway with Warner Music for an AI music tool.

• Concerns about audio deepfakes and generative AI's potential impact on artist reputation and ownership have risen within the music industry.

• The agreement aims to address the legal issue of AI-generated songs, leaving questions about ethical implications and artist profit.

CODE/DEVTOOLS

Nvidia is partnering with AI startup Hugging Face to introduce a service called "Training Cluster as a Service" aimed at simplifying the creation of custom generative AI models for enterprises. Link.

• The service will be powered by Nvidia's DGX Cloud, an AI "supercomputer" in the cloud, offering access to cloud instances with Nvidia H100 or A100 GPUs, AI Enterprise software, and consultations with Nvidia experts.

• The service will integrate DGX Cloud infrastructure with Hugging Face's platform of over 250,000 models and 50,000 datasets.

• Hugging Face's CEO, Clément Delangue, emphasizes the collaboration's aim to bring advanced AI supercomputing to enable companies to take control of their AI projects with open source software and speed.

• Nvidia's push into cloud services aligns with the growing demand for AI cloud training infrastructure, with Tractica forecasting AI to contribute significantly to total public cloud services revenue by 2025.

• The collaboration benefits Nvidia's expansion into cloud services and Hugging Face's aim to provide open-source AI tools.

Nvidia has introduced a platform called AI Workbench that allows users to create, test, and customize generative AI models on a local PC or workstation before scaling to data centers and public clouds. Link.

• The platform aims to simplify the customization of large AI models and streamline the deployment process for enterprise-scale AI projects.

• AI Workbench enables developers to fine-tune and test models from repositories like Hugging Face and GitHub using proprietary data and access cloud computing resources when scaling is needed.

• The goal is to address challenges and time-consuming tasks associated with customizing large AI models and deploying them into production.

• Nvidia's platform provides a unified workspace that integrates models, frameworks, SDKs, and libraries, including those for data preparation and visualization.

• Nvidia's AI Workbench offers a decentralized approach to model fine-tuning, occurring on local machines, which may appeal to developers looking for more flexibility beyond reliance on specific cloud services.

AI/ML development platform Weights & Biases has raised $50 million in a strategic round led by Nat Friedman (ex-GitHub CEO) and Daniel Gross (former Y Combinator partner). Link.

• Existing investors Coatue, Insight Partners, Felicis, Bond, BloombergBeta, and Sapphire also participated, valuing the company at $1.25 billion.

• Weights & Biases is preparing to launch a new product called "Prompts," which helps users monitor and evaluate the performance of LLMs similar to OpenAI's GPT-4.

• The investment follows a previous Series C round of around $135 million, bringing the startup's total raised to $250 million.

• Weights & Biases focuses on MLOps (machine learning operations), offering tools for data scientists to create, test, and deploy machine learning models efficiently.

• The platform's products are designed with partner and customer input and emphasize tools to interrogate datasets used to train models, helping to identify issues and biases before production.

Virtualitics, a startup that has developed a data visualization platform that uses 3D visualizations, knowledge graphs, and AI to expose relationships between different data points, has raised $37 million in a Series C funding round led by Smith Point Capital, bringing its total raised to $67 million. Link.

• The platform can generate annotations and explanations, which can be embedded in reports and dashboards and shared across an organization.

• Virtualitics aims to simplify data analysis beyond traditional methods and does not require deep technical expertise.

• The platform's visualizations can be viewed in VR and AR.

• The company's year-over-year revenue has increased by 370% with recently acquired government sector customers contributing to the growth.

• The funds will be used to accelerate the integration of AI technology into the platform and expand the business.

Flower, a startup that aims to decentralize AI training by using a platform that allows developers to train models on data distributed across devices and locations, has raised $3.6 million in a pre-seed round to expand its team and develop open source software for its framework. Link.

• Daniel Beutel, co-founder of Flower, believes that the reliance on public data for training AI is limiting the AI field's progress.

• The startup uses federated learning, a technique where AI algorithms are trained across decentralized devices without exchanging data samples.

• Flower's approach ensures that data doesn't leave the source device or location during training, respecting privacy and compliance concerns.

• Flower launched FedGPT, a federated approach to training large language models (LLMs) like OpenAI's ChatGPT and GPT-4, allowing organizations to build LLMs using internal data without sharing them with providers.

• Flower's community has grown to over 2,300 developers, and the startup has garnered interest from Fortune 500 companies and academic institutions.

Rasgo has launched Rasgo AI, a self-service analytics solution that integrates GPT-4 into enterprise data warehouse (EDW) environments. Link.

• Rasgo AI uses GPT-4 for "intelligent reasoning," enabling it to act as a knowledgeable business analyst for data warehouses, assisting with data analysis and decision-making.

• Unlike other GPT integrations, Rasgo's AI offers more than just natural language chat interfaces, focusing on intelligent reasoning and reasoning tasks.

• The platform uses AI Guardrails to map data structures into familiar business terms, enhancing the efficiency and accuracy of data analysis while ensuring data security.

• The platform's goal is to enable knowledge workers to focus on strategic decision-making by offloading low-value tasks to AI, resulting in gains in enterprise value.

• Rasgo's AI is designed to collaborate with human data teams, autonomously assessing tables in the data warehouse and discerning which tables need intelligent reasoning. The goal is to optimize accuracy and instill trust in organizational processes.

Deepset, a platform for building enterprise applications powered by LLMs, has raised $30 million in a funding round led by Balderton Capital, with participation from GV and Harpoon Ventures. Link.

• The funds will be used to expand Deepset's products and services and grow its team from around 50 people to 70 to 75 by the end of the year.

• Deepset aims to address the challenges faced by data science teams that are overworked and struggling to cope with the growing demands of product teams and end-users in the enterprise.

• The company's platform leverages LLMs to build applications and products that streamline AI development within enterprises.

• Deepset's offerings include the open-source framework Haystack for building NLP back-end services and Deepset Cloud, an enterprise LLM platform for AI teams, which allows customers to try out different LLMs, embed them into applications, and continuously monitor their performance.

• The company sees vendors in the MLOps space as its main rivals and aims to stand out by providing a platform that allows developers to focus on building and deploying NLP back-end services more easily and quickly.

MULTIMODAL

A new breakthrough in AI introduces a model called 3D-LLM, which bridges the gap between language and the 3D world we inhabit. Link.

• 3D-LLM enables AI to perceive and interact with the real world using point clouds (spatial coordinates of objects) and language.

• The model can answer questions about the environment, navigate spaces, and showcase commonsense reasoning similar to ChatGPT's capabilities.

• The model was trained using a unique 3D-text dataset that combined 3D data and language, gathered through ChatGPT's abilities.

• The AI processes scenes by extracting features from various perspectives and reconstructing them in a way that aligns with the model's understanding.

• This development signifies a significant advancement in AI's ability to comprehend and engage with the three-dimensional aspects of our world.

HARDWARE/ROBOTICS

Nvidia is set to release the GH200 super chip, a more powerful processor for generative AI workloads, including large language models, recommender systems, and vector databases. Link.

• The GH200 will feature the same GPU as Nvidia's current H100 offering but with triple the memory capacity, aimed at handling complex AI models more effectively.

• The GH200 systems are expected to launch in the second quarter of 2024, and the price has not been disclosed yet.

• Nvidia's H100 line is priced at around $40,000 and is already popular in the market for high-end AI processors.

• Nvidia holds a dominant position in the generative AI-capable GPU market, with cloud providers like AWS, Azure, and Google utilizing their H100 Tensor Core GPUs.

• Nvidia faces competition from AMD, which plans to increase production of its own AI GPU in the fourth quarter of this year, and also from companies like Microsoft that are exploring AI chip manufacturing.

Waymo and Cruise have been approved to operate their paid robotaxi services 24/7 in San Francisco by the California Public Utilities Commission (CPUC). Link.

• The approval comes after a contentious six-hour public hearing with both support and opposition from residents.

• The companies are permitted to offer robotaxi services throughout the city, charging for rides, and addressing concerns raised by officials and residents regarding traffic and safety.

• The CPUC can take actions to limit vehicle numbers or revoke permits if further incidents occur.

• Waymo and Cruise celebrate the decision as a significant milestone for autonomous vehicles and progress.

• The hearing showcased various opinions from disabled community members and residents, focusing on safety, convenience, and concerns about big tech companies' influence.

Robomart offers AI-driven vans that can be turned into mobile grocery stores, allowing customers to choose items directly from the van. Link.

• The vans can be hailed via an app and customers can select their desired groceries from the van's inventory upon arrival.

• The van automatically detects chosen items and charges the customer's card on file within the app.

• Robomart allows retailers to retain customer information instead of sharing it with larger businesses like Uber or Instacart.

• The original self-driving feature of Robomart's vision is currently operated remotely by humans due to regulations.

• The Global Artificial Intelligence Delivery Robot Market is expected to reach a multi-million dollar valuation by 2030, and Robomart is considered one of the leading players in this space.

HEALTHCARE

Scientists have created an AI tool that can classify four subtypes of Parkinson's disease with up to 95% accuracy. Link.

• The research was conducted by the Francis Crick Institute and UCL Queen Square Institute of Neurology in London, training a computer program to recognize disease subtypes using stem cell images from patients.

• The AI model's accuracy could lead to personalized medicine and targeted drug discovery.

• The AI algorithm analyzed stem cells to differentiate between subtypes involving toxic protein buildup and those linked to dysfunctional mitochondria.

• The AI's potential to predict Parkinson's subtypes using images could revolutionize drug testing and personalized treatment.

• Current Parkinson's diagnosis lacks subtype specificity, limiting tailored treatments and care for patients.

Basys.ai, a company founded in 2022 by Amber Nigam and Jie Sun, is using generative AI and deep learning to automate the process of prior authorization in healthcare. Link.

• Prior authorization involves gaining approval from health insurance companies before medical procedures, and the manual process can lead to delays and high administrative costs.

• The Centers for Medicare & Medicaid Services (CMS) proposed digitizing the prior authorization process to improve efficiency in the healthcare system.

• Basys.ai's platform automates up to 90% of prior authorization requests with high accuracy by using a trained AI engine that doesn't require sensitive data from insurance companies or doctors.

• The company's approach is based on longitudinal patient data from institutions like the Joslin Diabetes Center and Mayo Clinic.

• Basys.ai has raised $2.4 million in pre-seed funding, led by Nina Capital, and is initiating pilots with health insurance companies in Massachusetts and Minnesota.

POLICY/LAW/ETHICS

Changes to Zoom's terms of service caused confusion and concern among customers about the company's potential rights to use content from meetings for training its AI algorithms. Link.

• Users and companies are worried about the possibility of their conversations and content being shared without their consent.

• Zoom modified its terms of service in March, leading to recent uproar and speculation about broad permissions for AI training.

• Zoom clarified its position through a blog post and changes to the terms, stating that customer consent is required for using audio, video, and chat content for AI model training.

• CEO Eric Yuan acknowledged the internal process failure and committed to fixing it to ensure user consent for AI training.

• Despite Zoom's responses, concerns persist, with questions about the scope of consent and advocacy groups like Fight for the Future demanding policy changes.

SEC Chairman Gary Gensler warned that AI could lead to future financial crises. Link.

• Gensler believes that AI's economic dynamics, involving scale and networks, could contribute to future crises.

• He predicts that future U.S. business systems will rely on a few foundational models, increasing the risk of "herding" behavior where companies rely on the same information.

• The SEC proposed a rule to address conflicts of interest in AI technologies used by investment advisers, aiming to prioritize investor interests over brokers' or advisers' interests.

• Gensler emphasized that advisers and brokers are legally obligated to prioritize investor interests, whether using algorithms or not.

• Even if AI provides faulty financial advice, investment advisers remain accountable under their fiduciary duty of care and loyalty to clients.

Cloud word processor Shaxpir's project Prosecraft uploaded and scanned over 27,000 books into a dataset without authors' consent for analysis. Link.

• Authors, including Maureen Johnson and Celeste Ng, expressed anger at having their books used without permission.

• After backlash, Prosecraft's creator Benji Smith took down the website, citing concerns about AI's negative impacts on creators' involvement in their creative process.

• Prosecraft compiled two paragraphs from books, ranking them based on language vividness, length, and passivity.

• Authors worried about AI tools misusing their work and plagiarizing inadvertently.

• Authors' concerns over AI usage have extended to Amazon, where fake AI-written books were attributed to them; Amazon removed these books following backlash.

OpenAI's GPTbot, which scrapes online content for training AI models, is causing content creators to reconsider allowing their content to be scraped due to concerns that future AI models may compete with them. Link.

• Content creators are becoming wary of letting AI companies crawl their websites, as the rise of large language models and AI-driven answers reduces the incentive to share high-quality information.

• OpenAI's GPTbot has prompted online communities to take steps to block it, with creators calling for an "opt-in" approach where permission is sought before scraping content.

• The erosion of trust between content creators and AI companies has led to calls for stricter data collection methods and the need to pay for content used in AI training.

• OpenAI's efforts to respect paywall-protected sources and recent deals to license content from organizations like the Associated Press have raised questions about why not all content is paid for.

OTHER

Jeli, a company focused on incident management, has introduced a beta version that incorporates AI functionality similar to ChatGPT. Link.

• This move aims to enhance incident understanding and tracking by summarizing incidents, providing narratives about the incident progression, and aiding humans in comprehending complex incident situations.

• The "Catch Me Up" feature summarizes ongoing incident conversations, saving time for involved parties.

• Jeli uses OpenAI's ChatGPT and has specifically trained the model for computer incident management, focusing on accurate and relevant responses.

• The company is careful to avoid hallucination issues, providing evidence from source material for the generated data, and encourages human editing of the AI-generated content.

• While the current version doesn't support querying incident reports, the beta aims to refine the AI functionality based on user feedback for future enhancements.

Dropzone AI, a startup focusing on automating security alert investigations, has secured $3.5 million in seed funding led by Decibel Partners, along with participation from Pioneer Square Ventures Fund and notable angel investors. Link.

• The company's autonomous AI security agent uses LLMs to replicate expert security analysts' thought processes, investigate security alerts, and generate detailed reports and recommendations for human analysts.

• The AI agent aims to help security teams handle the increasing volume and complexity of cyberattacks by autonomously investigating alerts and providing comprehensive reports.

• The system is not reliant on pre-programmed playbooks and instead employs LLMs for autonomous reasoning and investigation.

• Dropzone AI's product includes an autonomous alert investigation system and a secondary chatbot capability for ad hoc investigations and information retrieval.

• The company aims to provide oversight and transparency by assisting human analysts with frontline work and generating detailed investigation reports.

11xAI, a company that creates automated digital workers to replace human employees has raised $2 million in a pre-seed funding round led by Project A Ventures and has launched its service. Link.

• The London-based startup aims to help smaller businesses increase productivity and compete with larger companies.

• The company's first AI worker is named Alice, an AI sales development representative. Future plans include creating James for automated talent acquisition and Bob for automated human resources tasks.

• 11xAI envisions a future where businesses hire autonomous workers for all parts of their operations, allowing for an AI-powered workforce.

• The company is focused on preventing bias in its AI models through audits, monitoring, regular bias testing, and diverse datasets.

• Founder and CEO Hasan Sukkar plans to expand the team, enter the U.S. market, and launch two more digital workers.

Chargeflow, a startup that uses AI to combat chargeback fraud in e-commerce transactions, has raised $11 million in a seed funding round led by OpenView Venture Partners. Link.

• Chargeback fraud is a significant issue for businesses, and Chargeflow's AI generates custom "dispute evidence" for each chargeback and submits it to banks or credit card companies.

• Chargeflow, founded by Ariel Chen and Avia Chen, automates and streamlines the chargeback dispute process, reducing the labor cost and time needed to fight chargebacks.

• The company calculates the projected success of chargeback disputes and bills merchants based on successful dispute recoveries.

• Chargeflow plans to use the funding to enhance its technology development efforts, support its growth trajectory in the U.S., and expand its team from 40 to 80 employees.