The Artificial Investor - Issue 25: June 2024 recap
My name is Aris Xenofontos and I am an investor at Seaya Ventures. This is the monthly version of the Artificial Investor that covers the top AI developments of the previous month.
$1.5 billion for LLM developers, a $6 billion IPO and Nvidia’s wild ride despite increasing competition. Apple’s entry in the Gen AI arena amidst stronger voices about AI bubble. The expansion of the West/East Tech Cold War and the weaponisation of AI. New developments in multi-modal technology and Gen AI architectures that are alternatives to transformers. All this and much more in today’s issue about June 2024.
😎 Before we dive in today’s issue, as we are entering the summer season and the flow of news has declined notably, we are switching to the monthly version for July and August, and will back with our weekly version in September.
If you prefer to listen to this issue, click here for the audio version.
Powered by ElevenLabs
🚀 Jumping on the bandwagon
We have tracked nearly 90 funding rounds worth an aggregate $4.3 billion during June 2024, a 7.5% increase comparing to the previous month.
Recent mega rounds include an agregate of $1.5 billion raised by three LLM developers, Mistal AI, Cohere and Zhipu AI, which demonstrate the need for substantial resources to develop, train and deploy such models. A couple of developers of specialist LLM-based solutions, such as drug development (Formation Bio) and software development (Poolside.ai), also raised large sums of money on the back of strong commercial traction. Finally, players focusing on the hardware layer continue to attact funding, regardless of whether they develop sector-agnostic AI-related components, such as chips (Black Semiconductor) or sector-specific end-to-end solutions, such as Gen AI-powered autonomous trucks (Waabi).
In terms of exits, OpenAI has made two small acquisitions to strengthen its agent and data retrieval capabilities. Tempus AI, a precision medicine company, saw its stock market debut with shares jumping 15%, following its IPO that raised $410 million, at an implied valuation of $6.1 billion.
📈 On salmon-pink paper
Last, but not least. June was marked by Apple’s annual developer conference, WWDC 2024, where Apple Intelligence was revealed. We covered the topic extensively in Issue 23 - The business case for AI; in summary, Apple is: i) revamping Siri, which will soon be able to give users more control over their apps and answer more complex questions using OpenAI’s ChatGPT, ii) launching AI features in its native apps (Mail, Messages, Photos, etc.), such as the ability to generate responses automatically; also powered by OpenAI’s GPT models, iii) building its own safety-first AI infrastructure that consists of device chips to run small models, a proprietary private cloud to run medium-size models and user-consent-based outsourcing of complex queries to OpenAI’s cloud. OpenAI is paying Apple with access to state-of-the-art models and Apple is paying OpenAI with access to hundreds of millions of potential premium customers. Investors seemed to buy into the potential of Apple’s AI strategy to push device upgrades, thus boost future revenues; the stock was up 7%. Talking about latecomers, it was “leaked” that Amazon is planning to revamp its Alexa product (and potentially charge $5+ monthly for it) and that it’s working on its own AI agent model, Metis.
Follow the money. AI darling, Nvidia became the most-valued company reaching a $3.2 trillion valuation. This came as investors started digesting Nvidia’s announcement that its Datacentre revenues exceeded $100 billion on an annualised basis. At twice the levels of six months ago (but still at 4% of Nvidia’s relevant revenues) stood OpenAI’s annualised revenues in Jun’24 - $3.4 billion. We wrote about what these figures tell us about the current state of the Gen AI market in Issue 24: The size & vendor landscape of the Gen AI market. In an attempt to capture some of the AI upside, Cisco announced a $1 billion AI fund.
Food for thought. David Cahn from Sequoia published an update to his blog describing the AI Infrastructure bubble, with an attempt to quantify it. Linking to our analysis in Issue 24, the size of the Hardware Layer (primarily Nvidia’s revenues) seems to be too high on a relative basis comparing to the Application Layer. Cahn claims it should result in an Application Layer of four times the size in a normalised market, because the AI applications that use this hardware spend an equal amount of money in other infrastructure costs, such as energy (hence the first 2x), and applcations should be making at least a 50% gross margin to be sustainable (hence another 2x). This means that by the end of 2024 where Nvidia’s annualised Datacentre revenues are expected to reach $150 billion, there should be $600 billion ofAI application revenues. However, the real figure is not expected to be above $10 billion. Hence, there is a $590 billion AI application gap.
Early in the trend or hype? Recent market research reveals significant insights into the integration and impact of GenAI in corporations. The 2024 CAIO Report from PYMNTS Intelligence indicates that only 30% of enterprises are fully leveraging strategic GenAI applications and 39% of CFOs are reporting very positive ROI. Challenges persist in integrating GenAI into existing workflows, particularly in companies that have not fully embraced these high-impact applications. Concurrently, a Bain report highlights a widespread adoption of GenAI, with 87% of companies actively developing or deploying initiatives in areas like software development and customer service, yet facing implementation challenges and realistic assessments of the technology's value. On average, companies are investing about $5 million annually in Gen AI, with 20% of large companies investing up to $50 million per year. In the meantime, CIOs complain about the in-house human resources that AI assistants, such as Microsoft's Copilot, require.
⚔️ A double-edged sword
The Middle goes further East. The US/China Tech Cold War is well underway, so it came at no surprise to us that the US president issued a draft executive order (which means it will not need to go through Congress) for banning or requiring notification of certain Tech investments, such as AI, semiconductors, microelectronics and quantum computing, made in China, Macau and Hong Kong. The US is also expected to work on closing its current regulatory loopholes, as Chinese firms, like ByteDance and China Telecom, have been accessing Nvidia’s banned chips through licensing agreements with US cloud providers. The Tech Cold War is allegedly affecting US innovation, as American AI startups founded by China-born entrepreneurs cannot benefit anymore from access to China-based talent. And what’s more, things are expected to escalate: the Middle East also seems to be joining “Cold War 2”. The US Commerce Department has expanded its chip export restrictions, originally only for China, to the UAE, Saudi Arabia and Qatar.
Leading the pack. Colorado has enacted the first comprehensive law in the U.S. for regulating the development and deployment of AI systems. The Colorado Artificial Intelligence Act focuses on high-risk AI systems and their potential for algorithmic discrimination, defining high-risk as those making consequential decisions about areas such as education, employment, financial services, housing, health care and legal services. Developers are required to provide explanatory documentation to deployers, and deployers are obligated to provide information to consumers on how they manage known or foreseeable risks of algorithmic discrimination.
Over before it began. The AI legal battles continue, with music record labels now taking the lead. A number of labels, including Universal, Sony and Warner, filed a lawsuit of $150,000 per song against Suno and Udio, the text-to-music AI applications, for alleged copyright infringement. The record labels claim that using targeted prompts with specific characteristics of popular sound recordings can cause the AI music generators to produce music files that strongly resemble the copyrighted recordings related to the description in the prompt. Both companies have claimed that the record labels are suing a series of “John Does,” who they claim are the specific individuals that helped scrape the music for the AI music generators. Music is expected to follow the steps of text, where OpenAI has signed a number of partnerships with leading publishers, and image, were data players such as Shutterstock, an online image library, have made various licensing deals with AI applications. Specifically, Shuttertock’s AI deals brought in $104 million of revenues in 2023.
A wolf in sheep’s clothing. This may be a topic that will keep us busy in the future: the weaponisation of AI technology. Videos demonstrating Chinese robot dogs that resemble Boston Dynamics’ robots, but with a machine gun mounted on them, went public this month. The US Pentagon has carried out similar experiments in the past. Among others, this has implications for public and private AI investors, whose investment strategy excludes the Defence sector, as seemingly innocent technologies can be easily repurposed for military purposes.
🧩 Laying the groundwork
🔢 New model launches
Google launched a video-to-audio model, Deepmind V2A, which uses a diffusion-based approach to generate soundtracks for a wide range of traditional footage. The Big Tech also launched two new versions of Google Gemini, Gemini 1.5Flash, a smaller multimodal model that is 40% faster than OpenAI’s GPT-3.5 Turbo and has a 1 million context window, and Gemini Pro, a model with a larger context window reaching 2 million tokens.
Meta launched an improved version of its coding model (Code Llama) called Meta LLM Compiler and has been allegedly experimenting with multi-token prediction techniques (instead of the established single-next-token prediction techniques) in order to address limitations of LLMs, such as complex decision-making and the high volume of training data required.
Apple has finally joined the open source movement and released 20 Core ML models (and 4 datasets) in Huggingface, an open-source model and data repository. At the same time, Nvidia launched a new model called Nemotron-4 340B that specialises in synthetic data generation used to train LLMs, which support 50+ languages and 40 programming languages.
There were also several releases in the image/video model space, including Luma’s Dream Machine, a fast text-to-video model that produces 120 frames in 120 seconds, Stability’s Stable Diffusion Medium, a smaller (2B parameters) text-to-image model that can run locally on consumer PCs (5GB GPU RAM needed), and Runway’s Gen-3 Alpha, a model capable of generating 10-second video clips featuring emotional expressions and camera movements.
In other areas, DeepSeek released the second version of its coding model, Coder v2, which boasts 236B parameters, supports 300+ programming languages and outperforms closed-source models like GPT-4 Turbo, Claude 3 Opus and Gemini 1.5 Pro. Galileo released its new LLM evaluation model, Luna, which is 97% cheaper and 11x faster than GPT-3, and can identify model hallucinations with 95% accuracy.
💾 Hardware
Competition for Nvidia is picking up, with AMD unveiling its latest artificial intelligence processors, MI325X, and news breaking about ByteDance (TikTok’s owner) partnering with Broadcom to develop an advanced 5nm AI chip, manufactured by TSMC in Taiwan. Seminanalysis reported that OpenAI, xAI and Meta are building GPU clusters consisting of 100,000+ GPUs, which would require over $4 billion of capex and 150MW+ in datacenter capacity.
🧪 Scientific Breakthroughs
Researchers have developed a novel architecture that eliminates matrix multiplications (MatMul) from language models, while maintaining strong performance at large scales. MatMuls are an essential component for LLMs using the Transformer architecture, but are expensive, increasing memory usage and latency during training and inference. This new architecture suggests replacing traditional 16-bit floating point weights with 3-bit ternary weights and MatMuls with additive operations. Comparative evaluations show that the MatMul-free LM outperformed its Transformer counterpart on advanced benchmarks, with lower memory usage and latency. The architecture has the potential to make language models less dependent on high-end GPUs and enable powerful models to run on less expensive processors.
JEPA is a new model architecture proposed by Yann LeCun that is designed to address the limitations of LLMs, such as the lack of common sense and memory. The architecture includes components like encoders, a predictor module, latent variables, and the ability to combine multiple JEPAs for more complex prediction. The model works by taking pairs of related inputs, transforming them into abstract representations, and using a predictor module to make predictions while handling uncertainty in the process.
A new architecture based on the state space duality (SSD) framework has been used to design the second version of the Mamba model. Mamba-2 is two to eight times faster than its predecessor and competitive with Transformers on language modeling.
Hugginface developed the largest ever open dataset that can be used for pre-training large language models, FineWeb. The dataset, the largest of its kind to date, consists of 15 trillion tokens (44TB disk space) and is derived from 96 snapshots of CommonCrawl, a non-profit crawler that scrapes the Web and publishes its results every 1-2 months.
OpenAI researchers have published a paper describing their attempt to understand better how GPT-4 works. The model was decomposed into 16 million interpretable features, such as human Imperfection, price increase and training logs. The researchers used a proprietary large-scale autoencoder training methodology.
Researchers developed an alternative path to an autonomous driving future: musculoskeletal humanoid robots. The musculoskeletal humanoid has sensors and a flexible body structure and succeeded in using the pedals and steering wheel of a conventional car.
Reseachers developed an alternative way to evaluate and bechmark models based on skill-acquisition efficiency, emphasizing adaptation and learning. The relevant dataset, ARC-AGI, is designed to evaluate a model’s ability to solve novel reasoning tasks.
🍽️ Fun things to impress at the dinner table
Child’s play. Retailer Toys R Us unveiled a new video made with OpenAI’s non-public generative AI video model Sora, which narrates the origin story of the company’s founder. The video was made in a few weeks and was screened during the 2024 Cannes Lions Festival.
Face off. Japanese scientists have developed a way to attach living skin to robot faces, allowing for more realistic facial expressions. The artificial skin prototype is soft, self-healing and can move with the mechanical components of the robot without tearing or peeling away. The prototype face is made in the lab using living cells and is designed to smile and make other facial expressions.
Always green. Google is expanding Green Light, a product that analyses data from internet-connected vehicles and navigation apps to adjust traffic light timing. Data from cities like Abu Dhabi and Seattle suggest a 30% reduction in stop-and-go traffic.
Bacon ice cream? McDonald's is planning to end its AI-powered automated drive-thru order-taking experiment at over 100 locations. Customers reported receiving items they didn’t order, including nine sweet teas added to one customer’s order and an ice cream cone topped with bacon given to another customer.
Ghost logo. How do you come up with the design of an icon for something that does everything but looks like nothing? Apple recently joined the competition to find an iconic representation of AI, with a circular shape made up of seven loops as its attempt. Early AI icons, which included robots, magic wands and wizard hats used by companies such as OpenAI, Microsoft and Meta, are considered inadequate.
Robocop. The Chula Vista Police Department (CVPD) has been using a full-blown police drone program, having dispatched drones nearly 20,000 times since 2018. The drones are only deployed in response to 911 calls or lawful searches and often monitor minor issues such as shoplifting and loud music.
Can’t fool the brain. A study analysing brain scans of 43 participants revealed different brain responses to human and AI voices. Human voices triggered memory and empathy areas, while AI voices activated regions for error detection and attention regulation. When asked, people struggle to distinguish between human and AI-generated voices, only identifying them correctly about half the time.
Government efficiencies. This year’s local elections in Cheyenne have something different comparing to previous years. One of the six individuals running for mayor is an AI candidate called VIC. VIC, developed using OpenAI, will leverage data-driven decision-making and ensure transparency, efficiency and innovation in city administration. A human who has been approved as a candidate, is behind VIC.
💲Show me the money
We have tracked nearly 90 funding rounds worth an aggregate $4.3 billion during June 2024, a 7.5% increase comparing to the previous month.
In terms of exits, OpenAI has acquired Multi and Rockset.The former company is focused on screenshare and collaboration technologies, which is speculated that will be used to allow its AI models to take over a user's computer and perform actions based on text or voice prompts. The latter is a real-time analytics database and will strengthen OpenAI’s retrieval infrastructure.
Tempus AI, a precision medicine company, saw its stock market debut with shares jumping 15%, following its IPO that raised $410 million, at an implied valuation of $6.1 billion. The company grew its revenues from $321 million in 2022 to $532 million in 2023. Tempus has raised $1.3 billion in funding over nine rounds.
Below is a list of June’s AI funding rounds:
Mistral AI raising $640M for its generative AI models solution
Cohere raising $450M for its large language models solution
Zhipu AI raising $400M for its generative AI technology solution
Poolside.ai raising $400M for its AI-powered software development tools
Formation Bio raising $372M for its AI-driven drug development solution
Black Semiconductor raising $273M for its graphene-based chip technology solution
Waabi raising $200M for its generative AI-powered autonomous trucks solution
Emergence raising $97.2M for its agent-based system for knowledge worker tasks
Pika raising $80M for its AI-generated video creation solution
Stability AI raising $80M for its generative AI media creation solution
Vilya raising $71M for its macrocyclic drug development solution
Klarity raising $70M for its automated document review solution
Finbourne raising $70M for its financial data management platform solution
Axelera AI raising $68M for its AI hardware acceleration technology
Prewave raising $67M for its AI-powered supply chain intelligence platform
Genspark raising $60M for its AI-powered search engine solution
GrayMatter raising $45M for its AI-powered manufacturing robots
Seven AI raising $36M for its AI cybersecurity threat detection solution
Decagon AI raising $35M for its generative AI customer service agents
Allora Labs raising $35M for its decentralized AI network
Sensi.AI raising $31M for its AI-powered senior care monitoring solution
Orby AI raising $30M for its generative AI business workflow automation platform
CuspAI raising $30M for its generative AI-driven search engine for new materials
Norm AI raising $27M for its AI platform for regulatory compliance automation
Neural Concept raising $27M for its 3D Deep Learning product engineering solution
Humata Health raising $25M for its AI-powered prior authorizations solution
Bitsensing raising $25M for its high-resolution radar technology
Constructor raising $25M for its AI-powered product discovery and search platform solution
Rocketlane raising $24M for its PSA and client onboarding platform
Maxa AI raising $21M for its AI-powered business data transformation platform
Kinetic Automation raising $21M for its robotic system for EV diagnostics and repair
Anterior raising $20M for its AI health insurance approval solution
Speak raising $20M for its AI-powered language learning app
Vantage Discovery raising $16M for its personalized online shopping search engine solution
Dust raising $16M for its enterprise AI assistants connected to internal data
SewerAI raising $15M for its AI-powered sewer infrastructure inspection solution
Compredict raising $15M for its AI-powered solutions for software-defined vehicles
MagicSchool AI raising $15M for its generative AI tools for educational environments
Windborne raising $15M for its AI-based weather modeling solution
Autify raising $13M for its AI agent for software quality assurance
Omi raising €13M for its AI visuals generation solution
Illumex raising $13M for its generative AI data organization solution
Alexi raising $11M for its AI-powered legal research platform solution
Particle raising $10.9M for its AI news-reading app solution
DeepJudge raising $10.7M for its AI-driven legal search technology
Tektonic AI raising $10M for its GenAI agents for enterprise processes solution
CampusAI raising $10M for its AI learning and collaboration platform solution
GPTZero raising $10M for its AI-generated content detection solution
Carv raising $10M for its AI recruitment platform solution
Zingly.ai raising $10M for its AI customer service improvement solution
Tandem Health raising $9.5M for its AI-driven healthcare co-pilot solution
Hona raising $9.5M for its client engagement software for law firms
Unify raising $8M for its AI-powered LLM routing platform solution
Synthflow raising $7.4M for its AI voice assistant solution for SMEs
EnFi raising $7.5M for its AI FinTech lending solution
FirmPilot raising $7M for its AI marketing engine for law firms solution
Soot raising $7.2M for its visual data representation solution
QpiAI raising $6.5M for its quantum computing and AI solutions
Inventive raising $6.5M for its embedded AI Analyst agent for SaaS products
Linq raising $6.6M for its AI financial analysis and research solution
Brightwave raising $6M for its AI-powered financial research assistant solution
Eyebot raising $6M for its self-serve vision-testing kiosks solution
Cartwheel raising $5.6M for its generative automation platform solution
Kelvin raising $5.1M for its AI-powered home energy audits solution
Botsync raising $5.2M for its autonomous mobile robots solution
BlueFlame AI raising $5M for its AI platform for alternative investment managers
Created by Humans raising $5M for its licensing platform for creators' work to AI models
Liminal raising $5M for its horizontal security for generative AI solution
Hoop raising $5M for its AI task management platform solution
Butterflies raising $4.8M for its social network for human and AI interactions
Jump raising $4.6M for its AI financial advisor automation solution
InScope raising $4.3M for its AI financial reporting and auditing solution
Flow Computing Oy raising €4M for its Parallel Processing Unit solution
Greptile raising $4M for its AI-fueled code base solution
Spike raising $3.2M for its AI solutions for healthcare and digital health businesses solution
Field Materials raising $3.5M for its AI material purchasing solution
Caju AI raising $3M for its generative AI customer engagement solution
TrackLight raising $3M for its AI-assisted fraud detection and prevention solution
Maxim AI raising $3M for its AI platform for evaluation and observability
Allora Labs raising $3M for its decentralized AI network
Human Native AI raising £2.8M for its ethical data for AI training solution
Zeta Labs raising $2.9M for its AI browser control agent solution
qeen.ai raising $2.2M for its GenAI product discovery solution
Tomato.ai raising $2.1M for its speech AI solutions
Backstroke raising $2M for its generative AI messaging platform for B2C brands
Fluently raising $2M for its AI-powered English coaching app
Helsinki-based startup raising €1.7M for its AI-powered knowledge search platform solution
Jupus raising €1.3M for its AI software for the legal sector solution