The Generative AI Investment Gold Rush

Welcome to the 14th issue of the AstroFeather AI newsletter!

[ Daily AI News | Newsletter Signup | View Online ]

This was another exciting week for AI, but in a different way than I've seen in recent months. All the product announcements seem to have been replaced by multi-million (and even billion) dollar acquisitions, funding rounds, and other investments! There have also been some major regulatory updates, and the announcement of several conversational AI assistants for different environments (including the classroom and NASA space missions).

I hope you enjoy reading this week’s updates and if you have any helpful feedback, feel free to respond to this email or contact me directly on LinkedIn [@adideswilliams]

Thanks - Adides Williams, Founder @ AstroFeather

In this week’s recap (10 - 15 min read):

• If You Can’t Beat Them, Buy Them.

• Leading Foundation Models Fall Short of the EU AI Act.

• Product Previews and Launches.

• Company Announcements and News Throughout the Industry.

• Venture Capital Funding.

Must-Read News Articles and Updates

Update #1. If You Can’t Beat Them, Buy Them.

To stay competitive in the AI space, tech companies and consulting firms are opting to buy AI platforms (and services) through startup acquisitions rather than building AI technology themselves.

The latest: Databricks, Thomson Reuters, Ramp (a financial automation company), and Accenture recently announced lucrative deals to acquire startups that provide a range of AI services, including legal research, language model development, and customer support:

• Databricks acquires MosaicML: Databricks paid $1.3 billion to acquire MosaicML, a 2-year-old company specializing in neural networks and generative AI (genAI) tools. With about 60 people on the MosaicML team, the Databricks acquisition works out to about $20 million per employee. According to Databricks, the entire MosaicML team will join Databricks as part of a retention agreement.

• Thomson Reuters Acquires Casetext: Thomson Reuters acquired Y Combinator-backed legal AI company Casetext for $650 million. Casetext develops an AI assistant called CoCounsel that helps with document review, legal research notes, deposition preparation, and contract analysis. Casetext was one of the few companies to have early access to OpenAI's GPT-4 language model, allowing it to develop solutions for the legal industry.

• Ramp acquires Cohere.io: Ramp, known for its financial automation and expense management platforms, recently acquired Cohere.io for its genAI automated customer support platform. Cohere.io's customer support automation product has reportedly helped teams automate workflows and resolve up to 60% of support tickets.

• Accenture acquires Flutura: IT firm Accenture recently announced the acquisition of Bengaluru-based industrial AI company Flutura. Flutura specializes in industrial data science services for manufacturers, and its AI platform helps asset management and reliability engineering teams assess, predict, and improve production and manufacturing facilities. The acquisition aims to strengthen Accenture's industrial AI services and help clients in sectors such as energy, chemicals, metals, mining, and pharmaceuticals.

Behind the news: In addition to acquisitions, several firms like Salesforce, PricewaterhouseCoopers (PwC), Amazon Web Services (AWS), and Dropbox have pledged to invest millions for minority stake in AI startups or establish incubator and accelerator programs. Some notable examples include:

• Salesforce announced that it is increasing its Generative AI Fund to $500 million, with a focus on supporting startups developing "responsible generative AI" and prioritizing "ethical" AI technologies. The fund's expansion also coincides with the launch of Salesforce's AI for Impact Accelerator, which will award $2 million to education, workforce, and climate organizations to "advance the equitable and ethical use of trusted AI." Finally, Salesforce recently pledged a five-year, $4 billion (£3.2 billion) injection to fuel the growth of its UK business to drive the region's next wave of digital transformation "in this new AI era."

• Amazon Web Services (AWS) has introduced the AWS Generative AI Innovation Center, a program that will invest $100 million in support of customers and partners working on genAI initiatives. The program aims to connect AWS-affiliated data scientists, engineers, and architects with customers to accelerate innovation in genAI. Participants will receive workshops, training, and access to AWS products such as CodeWhisperer and Bedrock. The program will initially prioritize customers who have expressed interest in genAI, particularly in industries such as financial services, healthcare, media, automotive, and telecommunications.

• Dropbox has launched Dropbox Ventures, a $50 million venture fund focused on supporting startups in the AI space. The fund aims to provide financial support and mentorship to startups developing AI-powered products that shape the future of work.

• PricewaterhouseCoopers (PwC) will invest $1 billion over the next three years in generative AI (genAI) to automate aspects of its tax, audit, and advisory services. PwC plans to develop and integrate genAI into its own technology stack and client service platforms, as well as advise other firms on how best to use genAI. The investment includes funding for hiring and training, as well as targeting AI software makers for potential acquisitions.

Why it matters:

• It is worth noting that the multi-million to multi-billion-dollar acquisitions and investments mentioned in the above examples were all announced between March and June of this year, signaling a "gold rush" of sorts for generative AI (genAI) startups and their platforms and services.

• As the genAI gold rush continues, this is an opportune time for companies to establish themselves as "AI-first," but developing in-house genAI platforms can be costly. I expect more SaaS companies looking for a competitive edge to begin investing in genAI (and other AI) infrastructure later this year.

Update #2. Leading Foundation Models Fall Short of the EU AI Act.

Global policymakers in the European Union (EU), the United States (US), and China are writing laws and regulatory guidelines for the development, implementation, and commercial availability of AI systems and services. Of note, the EU is finalizing its AI Act as (potentially) the world's first comprehensive set of AI regulations, with a draft of the Act recently approved by the European Parliament.

The latest: Given the importance of the EU AI Act, a team of Stanford researchers evaluated several foundation models for their compliance with the requirements proposed by the EU's forthcoming AI legislation. (*Side note: In general, a foundation model (or base model) is an AI model trained on large amounts of data that can be adapted (or fine-tuned) for other tasks, such as text or image generation).

Study results: The research team scored 10 leading foundation models against 12 key requirements of the EU’s AI Act on a scale of zero (worst) to four (best), with the highest score being 48. Worryingly, almost all foundation models (7 out of 10) achieved an overall compliance score of ~50% or less, and only one model achieved an overall compliance score above 60%:

• Hugging Face's open-source BLOOM model scored 36/48 (75%), making it the top performer in the group.

• Google's PaLM 2 (foundation model for Bard) scored 27/48 (56%) (which puts it roughly the middle of the group).

• OpenAI's GPT-4 (successor to the gpt-3.5-turbo base model for ChatGPT) scored 25/48 (52%).

• Stability.ai's open-source Stable Diffusion (foundation model for several image generators) scored 22/48 (46%).

• Anthropic's Claude (ChatGPT competitor – backed by Google) scored only 7/48 (15%) points, placing it near the bottom.

Study observations – open vs closed source models: Interestingly, while all models showed room for improvement, there were some strengths and weaknesses unique to open source and closed source models.

• Closed-source models performed well in areas such as comprehensive documentation and mitigation of risks associated with deployment.

• Open-source models scored well for disclosure of resource usage and transparency of data sources.

How does this affect generative AI (genAI)? Although foundation models such as GPT-4 or Stable Diffusion are (currently) exempt from the "high risk" category of the AI Act, they will still be subject to several requirements, and non-compliance will result in hefty fines:

• Associated fines for noncompliance: Fines are proposed on a sliding scale, and the most recent draft of the AI Act proposes that the most serious breaches of obligations should result in fines of up to €40 million ($44 million) or 7% of total global revenue (whichever is greater).

• Regulatory framework: Currently, foundation models for chatbots and digital assistants fall under the "limited risk" category of the EU AI Act, which applies to systems that interact with and/or could manipulate human behavior and focuses on transparency obligations. As such, providers of chatbots would be required to inform users that they are interacting with a machine so that they can decide whether to proceed (or, if applicable, request to speak with a human instead).

Behind the news: While the EU appears to be the closest to finalizing laws and regulations for AI, it's worth noting that the US is also moving toward drafting its own regulatory guidelines and oversight for AI services and platforms. Some notable discussions include:

• US AI Regulatory Commission: US Democratic and Republican lawmakers recently introduced legislation to create a 20-member commission focused on developing risk-based guidelines for AI regulation, as well as recommendations for possible new government agencies (or other government structures) to oversee and regulate AI systems. To maintain political bipartisanship, half of the members will be appointed by Democrats and half by Republicans.

• Remove Section 230 immunity for genAI content: Two US. senators, Republican Josh Hawley and Democrat Richard Blumenthal, have introduced a bipartisan bill that would remove Section 230 immunity from social media companies for harmful content created by generative AI. Section 230 is a law that currently protects internet companies from liability for user-generated content. However, the changes proposed by Senators Hawley and Blumenthal would allow lawsuits against social media companies for disseminating harmful material, including deepfake photos and videos generated by AI.

• Senate Hearing: A US Senate subcommittee recently explored the risks and opportunities of AI in a hearing attended by executives from OpenAI (Sam Altman) and IBM (Christina Montgomery), as well as cognitive scientist Gary Marcus. During the session, Altman advocated for the creation of a new government agency that could license (or revoke) AI models "above a certain threshold of capabilities," Montgomery argued for "precision regulation" and the need to "define the highest risk uses of AI," and urged Congress to use existing government agencies to begin regulating AI as soon as possible, and Marcus emphasized the need for "close collaboration between independent scientists and governments" in developing AI laws and regulatory guidance.

• House Hearing: Clément Delangue, CEO of Hugging Face (an open-source data science and machine learning platform), recently testified before the U.S. House of Representatives Science Committee for a hearing on advancing AI innovation for US national interests. In general, Delangue argued that open source is most aligned with American interests because "open science and open source prevent black box systems, make companies more accountable, and help [solve] today's challenges like mitigating biases," as well as drive innovation. Recall also from the Stanford research discussion that Hugging Face's BLOOM open-source AI model was judged to be the most compliant with the EU AI Act, thanks in part to the company's focus on ethical openness (open documentation, safeguards, community moderation, and opt-in/opt-out datasets to respect copyright).

Why it matters: 

• The EU's AI Act is potentially the most important regulatory initiative on AI, as it will impact the deployment of AI services for the EU (and its population of 450 million people), as well as provide a legal framework for other nations to look to as they develop their own AI regulations.

• It is worth noting that there is currently no serious consideration of a US analog to the EU AI Act, or any federal legislation to regulate the development and distribution of AI. While there are state privacy laws that may extend to AI systems that process personal data, there is no substantive federal legislation in place. As a result, there is a real possibility that US policymakers will look to the faster-moving EU AI Act for inspiration in order to avoid establishing conflicting standards (see Brussels Effect).

Update #3. Product Previews and Launches.

Windows 11 Copilot: Microsoft introduced Copilot, a new feature in Windows 11 that uses generative AI to help users. Copilot appears as a column on the right side of the screen and can engage in conversations with users. It offers different conversation styles, from straightforward and factual to more creative but potentially less accurate. Copilot also supports the creation of AI images using OpenAI's DALL-E 2 model and will be able to adjust certain Windows settings and execute commands, potentially simplifying tasks for less technical users.

Microsoft Bing shopping guides: Microsoft is introducing AI-generated shopping guides on Bing that allow users to compare products across categories. The guides will appear at the top of search results, providing summaries and product comparisons. Bing's chatbot will also feature review summaries from retailers such as Amazon and Walmart.

Unity new AI game development tools: Unity announced several updates for game creation:

• Unity Muse: for creating textures, 2D sprites, and character motions from text prompts.

• Unity Sentis: for embedding AI models in Unity projects.

• AI marketplace: to find generative AI tools, AI assistants, and plugins.

MidJourney v5.2 update: With this latest update, Midjourney introduces a new "Zoom Out" feature that simulates zooming with a camera lens. The feature allows users to expand the boundaries of an existing AI-generated image while keeping the original subject in the center. Other enhancements in version 5.2 include an overhauled aesthetics system, a stronger "-stylize" command, a new "high variation mode" for increased compositional variety, and a "/shorten" command to cut non-essential words from prompts.

ElevenLabs Voice Library: ElevenLabs has launched Voice Library, a community platform for discovering and sharing AI-generated voices for use in a variety of projects. The core of Voice Library is Voice Design, which allows users to create synthetic voices that can be shared with the community.

Dropbox’s new AI tools: Dropbox has two new AI features. Dropbox Dash is a universal search bar that can search across different tools and content from third-party platforms, while Dropbox AI can extract information from files, generate summaries, and answer questions based on the content of documents, research papers, and meeting recordings.

Several chatbots and foundation language models were announced for commercial, enterprise, and classroom environments:

• ERNIE Bot 3.5: Chinese tech giant Baidu announced that its latest conversational chatbot, Ernie 3.5, has reportedly outperformed OpenA

  I's ChatGPT (and GPT-4) on some benchmark tests. ERNIE 3.5 features new plugins to expand its capabilities and includes several improvements over Ernie 3.0 in creative writing, reasoning, and code generation.

• Harvard “CS50” Bot: Harvard University unveiled "CS50 bot" as a learning assistant for its flagship course, Computer Science 50 (CS50). "CS50 bot" is designed to guide students to answers, help them find bugs in their code, and provide feedback on their program designs.

• Merlyn Mind large language models (LLMs): Merlyn Mind has launched three education-focused LLMs for the development of AI tutors and other conversational chatbots suitable for the classroom.

• Reka AI "Yasa": Announced closed beta access for its AI assistant, Yasa. According to the company, Yasa is multilingual and multimodal, capable of reading text, images, video, and tabular data.

• NASA chatbot: NASA engineers are reportedly developing a ChatGPT-style assistant that would allow astronauts to accomplish the following: 1) have conversations with space vehicles; 2) receive alerts and interesting findings (from AI-powered robots); 3) conduct experiments; and 4) rerform maneuvers. The AI system will also provide support for future Artemis missions and potential spacecraft issue detection.

Update #4. Company Announcements and News Throughout the Industry.

Adobe eases enterprise fears about AI-generated art: Adobe now offers enterprise customers IP indemnity and will cover any copyright claims related to work created with Adobe Firefly. The move by Adobe should provide peace of mind to enterprise customers who may face litigation over the use of Firefly-generated content, as well as address specific legal concerns around the commercial use of artwork created by generative AI. Adobe reportedly feels confident in offering this indemnification because Firefly has been trained on Adobe Stock images, openly licensed content, and public domain content, reducing the risk of copyright infringement.

Google DeepMind’s Gemini: Google DeepMind is developing a chatbot called Gemini that aims to rival or surpass OpenAI's ChatGPT. Gemini incorporates techniques from AlphaGo, DeepMind's AI system that defeated a professional Go player. The language model will reportedly use innovations in reinforcement learning to tackle tasks that current language models struggle with. Gemini is said to be DeepMind's most ambitious project in this area, with participation and support from top Google executives.

Meta’s AI recommendation models: Meta has announced its plan to develop AI behavior analysis systems that are "orders of magnitude" larger than existing LLMs like ChatGPT and GPT-4. However, Meta has not shared specifics, and the ambition to build behavior analysis models with tens of trillions of parameters has sparked debate about their necessity and possible implications, especially regarding user privacy and targeted advertising.

China's Tech Sector in the global AI competition: China's tech sector is focused on competing with US giants like Google and Microsoft in the global AI race. Chinese entrepreneurs, engineers, and former employees of ByteDance, JD.com, and Google are expected to spend $15 billion on AI technology this year, with a goal of becoming a major player in the field. Despite being an estimated three years behind the US, Chinese startups and talent are reportedly determined to catch up. However, challenges such as US tech sanctions, data regulations, censorship demands, and limited international expansion for Chinese firms remain significant hurdles.

US public schools test AI tutor Khanmigo: Newark City School District is using Khanmigo to test the feasibility of AI-assisted tutoring in the classroom. Khanmigo is based on OpenAI’s GPT-4 (successor to ChatGPT), which allows it to engage in back-and-forth dialogue with students. Khanmigo also helps students solve math problems, improve vocabulary, write stories, and can serve as a debate partner on a range of topics. (*Feel free to check out my full writeup on Khanmigo and AI virtual tutors in AstroFeather Issue No. 13)

The Verge AI survey results: The Verge surveyed 2,000 US adults and found that 57% have heard of ChatGPT, while 25% have heard of MidJourney. According to the survey, most users are using AI for brainstorming, answering questions, and creative tasks (music, design, photos, stories).

Toyota’s genAI design process: Toyota is using genAI to accelerate the design process for new electric vehicles (EVs). Toyota's text-to-design model generates optimized design variations based on engineering constraints and desired performance outcomes.

Update #5. Venture Capital Funding.

Inflection AI raised $1.3 billion to continue development of its “supportive companion” chatbot called Pi.

Runway raised a $141 million extension to its Series C round to develop a multimodal AI video generator platform.

Celestial AI raised a $100 million Series B to continue development of a novel photonics-based architecture for data transfer.

Typeface raised a $100 million Series B to develop brand-focused generative AI tools and services.

Reka raised $58 million to continue development its enterprise-focused multimodal conversational chatbot called Yasa.

Gleamer raised a $29.5 million Series B to build a AI software that helps radiologists diagnose and detect bone trauma lesions in scans.

Thanks for reading this issue of the AstroFeather newsletter!

I’m always looking for ways to improve and would love to hear your constructive feedback about the format and content of the newsletter. You can reply to this email, and I’ll be sure to respond.

See you in the next issue!

If you enjoy AstroFeather weekly content, be sure to share this newsletter!

Adides Williams, Founder @ AstroFeather (astrofeather.com)