AI Weekly: 09/18/23

Good morning and welcome to this week’s edition of AI Weekly! This week is filled with a number of significant headlines. In foundation model news, Google officially announced its plans for its LLM, Gemini, to compete with OpenAI’s ChatGPT. On the other hand, Meta is reportedly developing a new LLM that is poised to compete with GPT-4 and will significantly outperform its current open-source LLM LLaMA-2.

In audio news, Stability AI has announced Stable Audio, a diffusion model that users to generate short music audio clips through simple text prompts.

In large funding news, Databricks has raised over $500 million in a Series I funding round, resulting in a valuation of $43 billion.

In hardware news, Google announced new computing infrastructure specifically designed for generative AI and LLMs that are said to offer improved training and inference performance per dollar: the updated TPU v5e and the new A3 GPU that is competitive with NVIDIA’s H-100 GPU.

Continue reading below for many more disruptive AI headlines from last week!

- ZG

Here are the most important stories of the week:

TEXT

Meta is reportedly developing a new AI language model aimed to compete with OpenAI's GPT-4, as the company intensifies its efforts in the AI race. Link.

• The details of the new AI model are sparse, but it is intended to help other companies build sophisticated text and analysis services.

• This is not Meta's first entry into language models; the company had previously released LLaMA-2, an open-source alternative to commercial LLMs like OpenAI’s GPT-4, Anthropic’s Claude-2, and Google’s Bard.

• A team led by Mark Zuckerberg aims to make the new AI model "several times more powerful" than LLaMA-2 and plans to begin training it by early 2024.

• Zuckerberg is pushing to make the new AI open-source, a move that has reportedly raised legal concerns about issues like copyright infringement and disinformation.

• Meta faces stiff competition not just from OpenAI, but also from Google and Anthropic, indicating a highly competitive landscape for AI technology.

Character.ai, an app that lets users create AI characters, has continued to grow significantly since its launch in May 2023, reaching 4.2 million monthly active users in the U.S., compared to ChatGPT's nearly 6 million. Link.

• Despite high app uninstall rates in the industry, Character.ai has successfully retained many of its early adopters, showing strong retention and user engagement.

• ChatGPT continues to dominate in web-based interactions and globally on mobile, with 22.5 million monthly active users, outpacing Character.ai's 5.27 million.

• Character.ai is popular among younger users, attracting nearly 60% of its audience from the 18- to 24-year-old demographic, compared to only 27% for ChatGPT.

• While ChatGPT experienced a decline in web traffic and average time spent on the site, it is seeing a rebound as the new school year begins, with a 0.4% increase in U.S. website visits in August.

• Character.ai received $150 million in Series A funding earlier this year and was founded by AI experts Noam Shazeer and Daniel De Freitas, giving it substantial resources and expertise for future growth.

Google announced its forthcoming Gemini AI system at Google I/O 2023, aiming to compete with OpenAI's GPT-4 model. Link.

• Google claims that Gemini will surpass ChatGPT, the most popular consumer-facing LLM to date.

• A small group of companies has been given early access to Gemini, according to reports from The Information.

• If Gemini fails to meet the hype, Google risks losing ground in the conversational AI sector, where ChatGPT currently leads.

• The release of Gemini could benefit consumers by increasing competition and potentially driving innovation in the AI industry.

• Google has plans to integrate Gemini into various other aspects of its business, including Android and ChromeOS, increasing its potential impact.

Amazon has introduced generative AI tools to help sellers create product listings with captivating descriptions, titles, and listing details, potentially saving them time and enhancing product information. Link.

• These AI tools utilize LLMs trained on vast amounts of data, and Amazon likely used its own listing data for training.

• The tools aim to extract, improve, and enrich product information on a large scale, with an emphasis on quality, performance, and efficiency.

• Concerns have been raised about the potential for generative AI models to create false information or errors that may not be caught without human review.

• Amazon has implemented warnings for sellers to double-check content generated by the AI to ensure compliance with its listing guidelines.

• Other retailers, including eBay and Shopify, have also embraced generative AI to simplify the process of creating product listings and generating content.

IMAGE/VIDEO

Adobe's Firefly AI tools are now integrated into Photoshop and Illustrator, offering features like AI art generation and color correction. Link.

• The Firefly suite also includes a co-pilot AI chatbot that uses text prompts to assist users in image modification.

• Users receive "Generative Credits" for creating content, with subscription-based allotments that vary depending on their Adobe plan.

• Adobe is using images from its own Adobe Stock service for training the AI, and will provide annual bonuses to stock contributors.

• A "Nutrition Label" metadata system, part of the Content Authenticity Initiative, tags and tracks AI-created or modified content.

• Despite the metadata, the Content Authenticity Initiative acknowledges it's easy to remove but offers a verification site to check authenticity.

Typeset, a generative AI design engine, has been launched to simplify the creation of visual assets like social media posts, presentations, eBooks, and newsletters. Link.

• Unlike traditional design tools that rely on stock templates, Typeset creates designs based on the user's input content.

• Typeset can produce multiple variations of visual assets in minutes, eliminating the need for users to manually reformat, resize, or rearrange elements.

• The design engine's technology follows a heuristic rule engine to achieve consistent design and uses natural language keyword extraction for choosing imagery.

• Typeset has seen rapid growth since its launch, with 100% month-over-month growth in its paying subscriber base and over 10,000 pieces of content created. It is being used by a diverse range of companies, from startups to Fortune 500 firms.

• The company has plans to expand the product with new features like team collaboration, version control, and support for various types of visual content.

SPEECH/AUDIO

Stability AI has announced the launch of its Stable Audio technology, allowing users to generate short music audio clips through simple text prompts. Link.

• Stable Audio is based on diffusion models, similar to Stability AI's image generation technology. However, this model is trained on audio data to generate new audio clips.

• The technology originated from Stability AI’s internal music generation research lab, Harmonai, and is not related to Jukedeck, a startup sold to TikTok by Ed Newton-Rex, the current VP of Audio at Stability AI.

• Stable Audio goes beyond traditional 'symbolic generation' techniques that use MIDI files by working with raw audio samples and was trained on over 800,000 pieces of licensed music, ensuring high-quality output.

• The system uses a text model trained by Stability AI for generating audio based on text prompts with a technique known as Contrastive Language Audio Pretraining (CLAP) and comes with a prompt guide to assist users.

• Stable Audio offers both a free version, which allows for 20 generations of up to 20-second tracks per month, and a $12/month Pro plan that allows for 500 generations of up to 90-second tracks.

INFRA/DEVTOOLS

Databricks, a data analytics and AI software company, has raised over $500 million in a Series I funding round, resulting in a valuation of $43 billion. Link.

• This funding round stands out as many late-stage startups are facing valuation cuts amid a funding slowdown, making Databricks' $5 billion increase in valuation notable.

• Investors in the round include T. Rowe Price, Morgan Stanley, Fidelity, Franklin Templeton (indicating a pre-IPO interest), Capital One Ventures, and Nvidia (reflecting a strategic alignment in AI).

• Databricks has experienced strong revenue growth, with a revenue run rate surpassing $1.5 billion in Q2 and over 10,000 global customers, including 300 generating over $1 million annually.

• Despite the impressive growth, Databricks appears to be taking its time before going public, possibly due to its relatively high effective revenue multiple of 29x.

• The new funding will provide Databricks with additional resources as it competes in the rapidly growing AI market.

DeepMind researchers propose a new method called Optimization by PROmpting (OPRO) that uses LLMs as optimizers, defining optimization tasks in natural language rather than formal mathematics. Link.

• OPRO is highly adaptable and can be guided to solve a variety of problems by modifying the problem description or adding specific instructions.

• The process of OPRO involves iterative optimization using a "meta-prompt" that includes a natural language description, examples, and placeholders. Candidate solutions are generated based on this meta-prompt and are evaluated for quality.

• OPRO showed promise in optimizing LLM prompts for maximum task accuracy, revealing that prompt engineering can significantly affect model performance.

• Experiments demonstrated that OPRO consistently improved the performance of different prompts in language models from both the PaLM and GPT families, proving the method's ability to function across different LLM architectures.

• While OPRO shows potential in systematic prompt optimization, it also highlights the non-intuitive behavior of LLMs, cautioning against anthropomorphizing these models and underscoring the need for further research to understand their inner workings.

Patronus AI, founded by two former Meta responsible AI experts, launches from stealth, announcing a $3 million seed round led by Lightspeed Venture Partners. Link.

• The startup aims to evaluate and test LLMs, particularly focusing on regulated industries with little tolerance for errors.

• The company offers a managed service that identifies potential issues in language models, such as hallucinations where the model fabricates an answer.

• The testing framework involves three main steps: scoring models in real-world scenarios, automatically generating adversarial test suites and stress-testing the models, and benchmarking to find the most suitable model for specific use cases.

• Patronus AI aims to serve as an unbiased third party for evaluating the safety and efficacy of LLMs, ensuring they do not produce business-sensitive or inappropriate outputs.

• With currently six full-time employees, the startup plans to grow in the coming months, emphasizing that diversity and inclusion will be a key focus as the company scales.

A new study shows that AI chatbots like ChatGPT can operate a software company efficiently and cost-effectively with minimal human input. Link.

• Researchers from Brown University and multiple Chinese universities used a version of ChatGPT's 3.5 model in a hypothetical software company named ChatDev to handle various stages of software development.

• The AI bots were assigned specific roles and tasks in the software development process, such as design, coding, testing, and documentation.

• The experiment was conducted across different software scenarios, and analyses showed that ChatDev completed tasks "in under seven minutes at a cost of less than one dollar," on average.

• About 86.66% of the generated software systems were executed flawlessly, demonstrating the potential efficiency and cost-effectiveness of automated software development.

• While promising, the study also noted limitations like errors and biases in the language models, but suggested the findings could assist junior programmers and engineers in the future.

LastMile AI, focused on helping software engineers integrate generative AI into apps, has raised $10 million in seed funding led by Google's AI venture fund, Gradient, among other investors. Link.

• Co-founder and CEO Sarmad Qadri plans to use the funds to expand the team and develop the startup's products, aiming to democratize generative AI by simplifying tooling and workflows.

• LastMile's founders, who previously worked at Meta, are inspired to make AI developer tools more accessible to software engineers rather than just machine learning researchers, addressing the current disconnect in the market.

• The platform offers an AI Workbooks module for experimenting with different models, an AI Workflows tool for building complex workflows, and an AI Templates module for creating reusable development setups.

• Despite competition from similar platforms like LlamaIndex and LangChain, LastMile aims to offer a complete AI app development platform and lower the barrier to entry for non-experts in machine learning.

• The startup sees significant market potential, aiming to solve "last mile" issues enterprises face when incorporating AI into applications and workflows, especially as the market for AI model operations is expected to grow to $16.61 billion by 2030.

HARDWARE

Google Cloud Next 2023 highlighted Google's advancements in generative AI and introduced Duet AI, which is now integrated into Google Workspace and Google Cloud. Link.

• Google announced new computing infrastructure specifically designed for generative AI and LLMs: the updated Tensor Processing Unit (TPU) v5e and a new graphics processing unit (GPU).

• The TPU v5e is said to offer improved training and inference performance per dollar and integrates with various platforms like Google Kubernetes Engine (GKE) and Vertex AI. The A3 GPU supercomputer is based on Nvidia H100 GPUs and offers significant performance improvements.

• Google Kubernetes Engine (GKE) Enterprise was unveiled, which optimizes AI workloads and supports the new TPU v5e and the A3 VM to provide more efficient, scalable computing resources.

• Google introduced Multislice technology, allowing for the scaling of AI models beyond the limits of physical TPUs.

• Google launched Cross-Cloud Network and Google Distributed Cloud (GDC), designed to support high-performance networking for generative AI applications across on-premise and hybrid cloud environments.

Enfabrica, a networking chip startup focused on AI and machine learning workloads, has raised $125 million in a Series B funding round. Link.

• This funding values the company at "five times" its Series A post-money valuation.

• The round was led by Atreides Management and included participation from Sutter Hill Ventures, Nvidia, IAG Capital Partners, Liberty Global Ventures, Valor Equity Partners, Infinitum Partners, and Alumni Ventures.

• Enfabrica's hardware, the Accelerated Compute Fabric Switch (ACF-S), can provide high-speed data movement between GPUs, CPUs, AI accelerator chips, and other devices, reducing GPU compute for large language models by about 50% at the same performance level.

• The company aims to address the growing demand for networking technologies in AI infrastructure.

• While Enfabrica is well-funded, it faces competition from established networking chip players like Cisco, Broadcom, and Marvell who are also targeting the AI infrastructure market.

POLICY/LAW/ETHICS

The Senate Majority Leader, Chuck Schumer, organized a Senate hearing to develop consensus for future AI regulation legislation, which brought together high-profile tech CEOs, civil society leaders, and over 60 senators. Link.

• Elon Musk expressed concerns about the potential grave risks of AI, including the possibility of AI causing harm or even posing a "civilizational risk."

• All attendees agreed that the federal government should have a role in overseeing AI, but there was no consensus on the specific details of regulation.

• Bill Gates highlighted the potential of AI to address societal issues, such as hunger, while some called for substantial investment in innovation to unlock AI's benefits.

• The challenge for Congress is to balance maximizing the benefits of AI with minimizing its potential harms, including job displacement and technology-based discrimination.

• Some lawmakers raised the possibility of creating a standalone agency to regulate AI, while others suggested assigning oversight responsibilities to existing government agencies like the National Institute of Standards and Technology.

The Department of Homeland Security (DHS) has released new AI guidelines, including a commitment not to collect or disseminate data used in AI activities. Link.

• These guidelines are part of the Biden administration's broader efforts to manage the risks associated with AI technology.

• Secretary Alejandro Mayorkas has emphasized the need to use AI responsibly, balancing technological advancements with privacy and civil liberties.

• Chief Information Officer Eric Hysen has been appointed as the department's first Chief AI Officer, responsible for promoting AI innovation and safety.

• New principles dictate that DHS's use of AI must comply with the Constitution and other laws, and must avoid inappropriate considerations like race or gender.

• The guidelines mandate thorough testing of all facial recognition technologies to eliminate unintended bias and call for periodic reviews of existing systems.

The European Union (EU) is planning to grant startups access to its high-performance computing (HPC) supercomputers for training AI models, conditional on compliance with the EU's AI governance program. Link.

• This startup-focused initiative will build upon existing policies allowing industries to access EU supercomputers through a process called EuroHPC Access Calls for proposals.

• The EU has an ongoing AI Act aimed at providing a risk-based framework for AI applications, and is working internationally to establish an AI Code of Conduct.

• EU President Ursula von der Leyen emphasized during her annual 'State of the Union' address that AI poses existential risks to humanity, calling for responsible and ethical AI development.

• Von der Leyen proposed a global framework for AI governance, suggesting the creation of a body similar to the IPCC, focused on providing policymakers with research and briefings on AI risks and benefits.

• The EU also plans to convene the European AI Alliance Assembly in November to involve a broad range of stakeholders, including startups, NGOs, academic experts, and policymakers, in discussions on AI governance.

Eight tech companies, including Adobe, IBM, Nvidia, Palantir, and Salesforce, have made voluntary commitments on AI development as the Biden administration aims to regulate emerging technology. Link.

• The commitments build on earlier agreements by leading AI companies like Microsoft and Google, focusing on driving the "safe, secure, and trustworthy development of AI technology," according to the White House.

• Companies are agreeing to outside testing of AI systems before public release and implementing protocols for identifying AI-generated content. They are also committing to cybersecurity and threat safeguards, as well as sharing AI risk management information.

• The new round of commitments includes a broader range of companies, not just those training large AI systems but also those in "business to business" or creative sectors.

• Biden administration officials are in the process of crafting executive actions on AI and are also talking with lawmakers about developing legislation to regulate the technology.

• Senate Majority Leader Chuck Schumer is hosting an "AI Insight Forum" with leading tech CEOs to discuss AI regulation, following closely on the heels of the White House meeting focused on AI commitments.

The US Copyright Office Review Board denied copyright protection to an AI-generated artwork that won a Colorado State Fair art contest, citing the lack of human authorship required for copyright registration. Link.

• The artwork, created by Artist Jason M. Allen using the Midjourney image synthesis service, won top prize in the fair's "Digital Arts/Digitally Manipulated Photography" category.

• Allen was asked to "disclaim" the AI-generated portions of the work when applying for copyright, meaning he had to formally renounce any claim to that content as his own creation.

• Allen refused to disclaim the AI-generated content, despite detailing that he had made "numerous revisions and text prompts at least 624 times" using the image synthesis model and edited the image with Adobe Photoshop.

• The Copyright Office upheld its decision on appeal, stating that the entire image was not eligible for copyright due to the significant presence of AI-generated content.

• In his appeal, Allen argued that denying copyright protection for AI-generated work would create a "void of ownership," but the office rejected this reasoning, maintaining the requirement for human authorship.

OTHER

OpenAI is opening its first office in the European Union (EU) in Ireland, following offices in San Francisco and London. Link.

• The company is hiring for nine positions in Dublin, including an associate general counsel for the EMEA region, a policy and partnerships lead for global affairs, and a privacy program manager.

• OpenAI's move to open an office in Ireland reflects its efforts to address privacy concerns and regulatory challenges in Europe, particularly regarding its ChatGPT chatbot.

• The EU AI Act, which aims to govern AI applications based on their perceived risks, is on the horizon, making Europe a major focal point for AI companies.

• OpenAI CEO Sam Altman previously met with EU regulators to discuss AI regulation, emphasizing the importance of responsible AI development.

• The company's move into the EU market demonstrates its commitment to addressing regulatory headwinds and privacy issues in the region.

Google has launched the Digital Futures Project, an initiative to support researchers and public policy solutions related to AI. Link.

• Google.org, the company's charitable arm, is creating a $20 million fund to provide grants to think tanks and academic institutions working on AI expertise.

• The initiative aims to address the potential societal challenges and opportunities of AI, including fairness, bias, misinformation, security, and workforce transitions.

• Google's fund will support independent thinkers examining topics such as AI's impact on global security, labor, government productivity, and responsible AI innovation.

• Initial recipients of grants from the Digital Futures Fund include organizations like the Aspen Institute, Brookings Institution, MIT Work of the Future, and the Leadership Conference Education Fund.

• The fund will support organizations worldwide, not limited to the U.S., as part of Google's efforts to promote responsible AI development and governance.

TimeOS has introduced TimeAI, a Chrome extension that transforms standard calendars and note-taking software into proactive AI assistants, offering features like meeting summaries and real-time translations. Link.

• Unlike competitors, TimeAI proactively analyzes user behavior, workflows, and time-related data to assist in time management and task delegation, seamlessly integrating with popular work tools.

• TimeAI provides pre-meeting prompts with options like "Join meeting," "Get me ready," and "Send my AI instead," the latter allowing an AI bot to attend and transcribe meetings on your behalf.

• The platform aims to reduce meeting costs, which can be substantial, by minimizing unnecessary attendance and context switching, thereby boosting productivity and mental well-being.

• TimeAI is not a standalone tool; it integrates with popular communication and task management platforms like Zoom, Google Meet, Microsoft Teams, Asana, and Monday.com.

• Described as a "holistic platform," TimeAI aims to improve decision-making and overall productivity, offering a wide range of features that make it more than just a scheduling or transcription service.

Pre/Dicta, an AI-powered database launched in 2022, predicts judges' rulings in civil cases with 86% accuracy using around 120 datapoints such as judges' educational background, net worth, and decision history. Link.

• The tool aims to help lawyers and plaintiffs make more informed decisions on whether to proceed with litigation, potentially turning the art of judicial forum shopping into a precise science.

• Widespread adoption of such AI tools could reshape legal funding and reduce court backlogs by steering cases toward alternative resolution methods, especially for plaintiff attorneys working on commission.

• The database is comprehensive, covering all state and federal civil litigation cases, but it doesn't intend to predict outcomes of criminal cases and jury trials.

• CEO Dan Rabinowitz notes that the system even predicts the decisions of newly appointed judges with an 81% accuracy rate, despite the lack of case data for these judges.

• While the tool is seen as extremely valuable in deciding whether to litigate, some experts caution that its predictive power may be close to reaching an accuracy ceiling and that it could raise concerns if accuracy rates approach 98% or higher.

Pixis, an AI-powered marketing campaign management platform, raised $85 million in a Series C1 round led by Touring Capital, with participation from Grupo Carso, General Atlantic, Celesta Capital, and Chiratae Ventures. Link.

• The funding brings Pixis' total raised to $209 million, and the company has achieved $50 million in annual recurring revenue this quarter with over 200 brand customers, including DHL, Joe & The Juice, Sears, and Swiggy.

• Pixis expects to achieve profitability in Q4 after growing 140% year-over-year in 2023.

• Pixis was initially focused on generative AI for art and gaming but pivoted to create a full-stack marketing platform for generating assets, targeting customers, and monitoring campaign performance.

• Pixis' platform consists of three core pillars: Targeting, creative content generation, and performance optimization, all powered by AI.

• Pixis plans to use the latest funding to support R&D, expand infrastructure, hire more staff, and potentially make acquisitions in the future.

Helsing, a "Defense AI" startup, has raised €209 million ($223 million) in a Series B funding round led by General Catalyst, backed by Spotify founder Daniel Ek. Link.

• Saab, a Swedish heavy industry and defense group, joined as a strategic investor, strengthening their existing partnership with Helsing.

• This investment could make Helsing the largest European AI company and the largest European defense tech unicorn, with a post-money valuation of over €1.7 billion.

• Helsing's AI platform aims to enhance defense and national security for liberal democracies by improving efficiency using real-time data.

• Helsing has secured contracts with governments in Europe, including providing AI-enabled electronic warfare capabilities for the Eurofighter jet fighter and supporting the Future Combat Air System (FCAS) program.

• The funding round reflects confidence in Europe's role in AI and resilience, according to Helsing's co-founders Torsten Reil and Gundbert Scherf, with strong support from General Catalyst and Saab Defense.

Treefera, an AI-enabled data platform startup, has raised $2.2 million in a pre-seed funding round led by Concept Ventures, with participation from Twin Path Ventures, January Ventures, and Greg Lavender, Intel’s CTO. Link.

• The company aims to address the challenge of accurately assessing the value of carbon credits, especially those related to forests, by providing transparent forest data.

• Treefera sources data from high-resolution satellite imagery and drone-based lidar to offer insights into tree size and health.

• The startup plans to use AI to expand and scale its forest data, filling spatial and temporal gaps and providing information on risk and uncertainty.

• The voluntary carbon market has faced issues with transparency and trust, and Treefera aims to provide independent data sources for carbon credit buyers.

• The next few years will be critical for both Treefera and the broader voluntary carbon market to improve transparency and trustworthiness in the carbon credit industry.