The Artificial Investor - Issue 26: July 2024 recap
My name is Aris Xenofontos and I am an investor at Seaya Ventures.
?? We hope you are taking some time off or disconnecting somehow and enjoying the summer season. This is the monthly version of the Artificial Investor that covers the top AI developments of the previous month.
Tech stock market correction, an earning season with strong results across the AI value chain, some unprecedented international collaborations driven by AI, the emerging trends of small and visual language models, the most powerful AI training cluster in the world, and a half-billion-dollar AI exit. All this and much more in today’s issue about July 2024.?
Let’s dive in.
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 50 funding rounds worth an aggregate 5.1 billion dollars during July 2024, an 19% increase compared to the previous month.?
In terms of exits, Graphcore, an Nvidia competitor, was acquired by SoftBank for an undisclosed amount, rumoured to be around 500 million dollars. Augmedix, an ambient AI company, was acquired by Commure for 139 million dollars in cash. Cyabra, an anti-disinformation AI startup, plans to list on the Nasdaq via a SPAC merger, valuing the company at 70 million dollars. LoudnClear.ai, a multi-layered conversational AI platform for customer service teams, was acquired by Forwrd.ai for an undisclosed amount. Leonardo.ai, a generative AI content company, was acquired by Canva for an undisclosed amount.
Recent mega rounds include an aggregate of 1.3 billion dollars raised by three LLM developers, Baichuan, Cohere and Magic, which confirms the need for substantial resources to develop, train and deploy such models, amidst recent criticism that claims the LLM S-curve is tailing off. A couple of developers of verticalised LLM-based solutions focused on content-heavy domains, such as law and compliance (Clio, Harvey) have piled up another 1 billion dollars. The hardware layer remains very attractive, as well as capital intensive, collecting nearly another 1 billion dollars, both in sector-agnostic AI-related components, such as processors (Groq) and sector-specific end-to-end solutions, such as robotics (Skild AI, Applied Intuition).
?? On pink paper
Supply-driven market. Despite the recent correction in the prices of public Tech stock (Nasdaq is down 10% from its peak), the underlying businesses across the AI hardware supply chain continue to perform very well. AMD, Nvidia’s no.1 competitor, reported second-quarter earnings that beat Wall Street expectations for revenue and showed continued growth in sales of the company's AI chips. AMD's Data Center segment, which includes GPUs (AI chips competing with Nvidia), saw sales soar 115% year over year to 2.8 billion dollars, surpassing analyst expectations, with revenue now expected to exceed 4.5 billion dollars in 2024. Further down the supply chain, TSMC, the Taiwanese provider of essential chip components, reported 20.82 billion dollars of net revenues for the second quarter of 2024, up 40% vs. a year ago. TSMC held 62% of global foundry market share in the first quarter, according to Counterpoint Research data, and reached the 1 trillion dollar valuation mark.
Clear skies. Big Tech performance remains strong on the Cloud infrastructure side too. Overall, Microsoft reported better-than-expected revenue (64.7 billion dollars) for its fiscal fourth quarter. The company's top segment, Intelligent Cloud, which largely consists of Microsoft Azure, generated 28.5 billion dollars in revenue, up 19% from the previous year; AI added about an extra 1.5 billion dollars in the quarter. IBM's second-quarter sales increased 2% to 15.8 billion dollars, slightly ahead of the analyst estimates. The company’s AI business has seen a significant increase in bookings, surpassing 2 billion dollars since mid-2023, which is double the figure disclosed in the last quarterly report in April. About three-fourths of the AI bookings are from consulting, with the rest from software.
Searching for growth. OpenAI officially announced the launch of its search product, SearchGPT. The product allows users to use the company’s latest models, search the internet for relevant location-specific results and get a summary of findings with links to the sources. This is the latest move in the new chapter of the Search market, which comes shortly after Google launched its AI Overviews and announced its exclusivity deal with Reddit for access to the latter’s user-generated content. The battle started with the emergence of AI-powered search startups, such as Perplexity, which has been forced to follow the revenue-share route with publishers, after legal actions taken against it for improper use of Web content by media incumbents, such as Condé Nast (The New Yorker, Wired, etc). OpenAI’s search market entry should boost the company’s consumer revenues, which, according to a leak, account for 55% of the company’s 3.4 billion dollars of annualised revenues. By the way, this means that 45% of OpenAI revenues come from sales to businesses (ChatGPT Enterprise, OpenAI API, ChatGPT Team).?
Put a ring on it. Smart rings are officially becoming a category, as the first large Tech company, Samsung, announced that it is entering the market with Galaxy Ring. Other announcements of the South Korean giant as part of the annual Galaxy Unpacked event include AirPods-like Galaxy Buds 3 and 3 Pro, a beefy Galaxy Watch Ultra, and a redesigned Galaxy Z Fold 6 that comes with a new image-generating AI tool called "sketch to image" that uses generative AI to turn rough sketches into detailed images.
The AI Olympics. A few interesting reports comparing AI adoption across countries were released last month. A survey carried out by SAS, a data analytics company, showed that China is leading in the adoption of generative AI, with 83% of Chinese respondents using the technology, compared to 70% for the UK and 65% for the US. However, the US is ahead in terms of implementation with 24%, compared to China’s 19% and the UK’s 11%. In terms of AI readiness, IMF has issued an AI Preparedness Index, which tracks 174 economies' readiness in digital infrastructure, human capital, labour policies, innovation, integration and regulation. Singapore leads the index with 0.8 points. The US scores 0.77, the UK 0.73 and China 0.64. Overall, there is a correlation between GDP/capita and AI readiness; from a regional perspective, the Nordics score relatively high (unsurprisingly). In terms of AI research, a report of the UN’s World IP Organization showed that China is leading the GenAI patent race, filing more than 38,000 patents between 2014 and 2023 against 6,276 filed by the US in the same period.?
?? A double-edged sword
The countdown. Another US state, this time California, is considering introducing another AI law. Legislators are concerned about potential risks associated with powerful AI technology, including data privacy threats, such as the collection of private data by autonomous car companies like WeRide. On the other side of the pond, the timeline of the EU AI Act was published: i) by 2nd of February 2025 certain AI use cases become prohibited, such as biometric categorisation, untargeted scraping of faces from the internet, emotion-reading systems in workplaces and schools, etc, ii) by 2nd of May 2025 developers are required to comply with legal compliance benchmarks, key performance indicators, and transparency requirements, and iii) by August 2026 all rules of the AI Act will apply generally to companies operating in the EU.
In the end, no one truly wins a war. The US/China AI Cold War continues uninterrupted. Nvidia is developing a Chinese-market version of its new AI chips that will be compatible with U.S. export controls.? Washington tightened its controls on exports of semiconductors to China in 2023, prompting Nvidia to develop chips specifically for the Chinese market. In the meantime, some distributors are being creative in finding ways to smuggle Nvidia chips from places like Singapore into China. It is unclear how effective are America’s measures, as China seems to be catching up with the US in the foundational model battle. Chinese companies have launched text-based open source models that beat their US counterparts (Alibaba’s Qwen 2 beats Meta’s Llama 3 and Mistral’s Large in benchmarks), as well as have released text-to-video models to the public prior to their US equivalents (Kuaishou’s Kling model was released before OpenAI’s Sora model).?
Strength in unity. A number of international collaborations announced in the last 30 days indicate that this AI wave may be more likely to result in strong bonds between traditional competitors than previous Tech waves, such as social media. The US, UK and EU competition authorities have announced an unprecedented collaboration to address international antitrust matters, in particular i) control of key AI inputs, ii) entrenching or extending market power in AI-related markets, and iii) arrangements involving key AI players that could amplify risks. At the same time, in a contrarian move that goes against the unofficial AI Cold War, China and the US supported each other in a U.N. resolution addressing the need for wealthy developed nations to narrow the gap with developing countries in accessing and benefiting from AI. Coincidentally, this is exactly one of the conclusions of the IMF report mentioned earlier (correlation between GDP/capita and AI preparedness). Finally, industry leaders like Google, IBM, Intel, Microsoft, NVIDIA and PayPal have formed the Coalition for Secure AI (CoSAI) focusing on i) software supply chain security for AI systems, ii) preparing defenders for a changing cybersecurity landscape, and iii) AI security governance.?
Fragile. Scraping, the act of collecting data from a website using bots, is governed by a legal and technical regulatory system that seems fragile. Websites typically have terms of service agreements that outline what is and isn’t allowed, including restrictions on scraping. Also, a website developer typically creates and exposes a robots.txt file to indicate which parts of the site can be crawled by web crawlers or bots. It looks like this system is being put to the test due to the recent data collection spree of AI model developers. Some websites are complaining about Anthropic’s scraping practices, which keeps changing the name of its bots in order to avoid website restrictions.
Together we stand. At a federal US level, a bipartisan group of senators has introduced the COPIED Act, aimed at protecting artists, songwriters and journalists from having their content used to train AI models or generate AI content without their consent. The bill would require companies that develop AI tools to allow users to attach content provenance information to their content within two years, and give content owners the ability to protect their work and set terms of use, including compensation. On the other hand, video game performers seem to feel left out so far. They are set to go on strike due to failed negotiations with major game studios over artificial intelligence protections, marking the second strike for this group.
Check, please! Meta Platforms, Facebook’s parent, has agreed to pay 1.4 billion dollars to Texas to resolve a lawsuit about the illegal use of facial-recognition technology, the largest accord ever by any single state. Texas accused Facebook of capturing biometric information "billions of times" from photos and videos as part of a discontinued feature called Tag Suggestions. Meta previously settled a biometric privacy class action in Illinois for 650 million dollars.
There is no Planet B. Microsoft and Amazon are preparing to develop large data centres that use renewable sources of energy. In Microsoft’s case it’s a 1 billion dollar project in Kenya for a geothermal-powered data centre and in Amazon’s case it's a project for a data centre connected to a nuclear plant in the US East Coast. The latter recently also acquired a nuclear-powered data centre in Pennsylvania for 650 million dollars. In any case, it is hard to calculate the net impact of new and complex technologies, such as Cloud and AI. For instance, in the case of the Cloud trend, a recent report claims that it has been overall net positive in terms of energy savings. A study found that Cloud-based IT services significantly improved users’ energy efficiency, particularly after 2006 and even more so after 2010. The Cloud software layer has facilitated energy-efficient production, while the Cloud infrastructure layer has reduced the energy consumption of internal IT equipment and infrastructure. These savings exceed the total energy expenditure in the cloud service vendor industries and are comparable to the total electricity consumption in U.S. data centres.
?? Laying the groundwork
?? Models
Open-source AI models have challenged the dominance of proprietary systems in a watershed moment for the industry, potentially reshaping the AI landscape.?
Meta has released Llama 3.1 405B, the first frontier-level open source AI model, expanding context length to 128K, adding support across eight languages. Performance evaluations on over 150 benchmark datasets suggest that Llama 3.1 405B is competitive with leading closed foundation models, and the model was trained on over 15 trillion tokens using over 16,000 H100 GPUs.?
Mistral, a well-funded French AI startup, has introduced three new large language models (LLMs):
领英推荐
SenseTime released its SenseNova 5.5 LLM with a 30% improved performance and claimed to surpass GPT-4o in five out of eight key metrics. This is the first large AI model in China to realise a new means of human-AI interaction by integrating cross-modal information such as sound, text, images and video in real time.?
On the small model side, we recently saw the launch of various small, specialised Vision-Language Models (VLMs) that democratise VLMs at a lower cost, while still maintaining performance. Examples include LLaVA-Next, PaliGemma, Phi-3 Vision, Florence-2 and InternLM-XComposer 2.5. OpenAI introduced GPT-4o mini, the company’s most cost-efficient small model (priced at 15 cents per million input tokens and 60 cents per million output tokens). The model has a context window of 128K tokens, surpasses GPT-3.5 Turbo and other small models on academic benchmarks, and supports text and vision in its API. Huggingface released SmolLM, a family of small text-based language models with 135M, 360M and 1.7B parameters. Google released the second version of its smaller model series, Gemma, with 27B and 9B parameter sizes. Particular focus has been placed on i) safety (ShieldGemma introduces safety classifiers to detect and mitigate harmful content in AI model inputs and outputs) and ii) explainability (Gemma Scope allows researchers to analyse the internal processes of language models and gain insights into their workings).
On the large multimodal model side, a new study has shown that VLMs do not see in the way humans do. The study tested the biggest VLMs on simple visual tasks, and they all had great difficulty with tasks that even a first-grader could do perfectly, such as defining whether two shapes overlap or counting interlocking circles. The models seem to be matching patterns in their training data rather than truly understanding visual input. Meta has introduced Meta 3D Gen, an AI system that creates high-quality 3D assets from text descriptions in less than a minute, potentially transforming industries from video game development to industrial design and architecture.
Other recent announcements include Baidu’s Ernie 4.0 Turbo model, KyutAI’s Moshi model (world's first open-access voice AI) and OpenAI’s CriticGPT model (identifies errors in ChatGPT's code output).
On the logical reasoning side, Numina’s NuminaMath 7B TIR, a fine-tuned version of DeepSeekMath-Base 7B, won the AI Math Olympiad, and AlphaProof and AlphaGeometry 2 have achieved a silver-medal in the (human) International Mathematical Olympiad for the first time, demonstrating AI’s improving mathematical reasoning capabilities.
Also, OpenAI released a classification system for AI technologies that measures how close we are to Artificial General Intelligence (AGI), ranging from Level 1 that corresponds to a machine that can interact in conversational language with people, to Level 5, which corresponds to a machine that can do the work of an organisation. The company’s developers claim OpenAI’s models are in Level 2.
?? Hardware
Robotics have attracted their fair share of attention last month: Amazon announced the launch of a business-focused version of its Astro home robot, priced at 2,350 dollars and designed to support various tasks, such as assisting with medical care and improving customer service. Tesla showcased the third generation of its humanoid robot, the Optimus, at the World Artificial Intelligence Conference in Shanghai, which is still not at the state of production at scale. A newcomer, Mytra, released its warehouse robots focusing on moving, storing and retrieving materials, as well as automating complex pallet and case handling tasks.?
We have talked in the past about our prediction that all Big Tech companies will soon have their own AI chip. A couple of steps were made towards that direction, as Microsoft unveiled its first homegrown artificial intelligence chip, the Maia 100. Microsoft designed the chip to help AI systems process data required for tasks like speech and image recognition. The hardware was manufactured by TSMC. Also, OpenAI allegedly is in discussions with Broadcom for the manufacturing of its own AI chip.
In other hardware news, apparently the X (aka Twitter) spinoff, X.AI, is developing a new AI cluster, the Memphis Supercluster, which will feature 100,000 H100 servers it has acquired by Nvidia. This cluster will be 6x larger than the one used to train Llama 3.1, making it the most powerful AI training cluster in the world.
?? Scientific Breakthroughs
JARVIS-1 is an AI agent designed to perceive multimodal input, generate plans and perform control within the open-world Minecraft universe. The agent is developed on top of pre-trained multimodal language models, utilising memory augmentation to facilitate planning using both pre-trained knowledge and actual game experiences. The agent is capable of completing over 200 different tasks in Minecraft, ranging from short-horizon tasks to long-horizon tasks, and can self-improve following a life-long learning paradigm thanks to growing multimodal memory, leading to improved autonomy and general intelligence.
Apple’s researchers published a paper about 4M-21, a new any-to-any vision model trained on tens of diverse modalities and tasks, such as images and text along with several semantic and geometric modalities. The model can solve at least 3x more tasks/ modalities than existing models and doing so without a loss in performance. In addition, the research enables more fine-grained and controllable multimodal generation capabilities and allows studying the distillation of models trained on diverse data and objectives into one unified model. Training was scaled to three billion parameters and different datasets.?
Google DeepMind's new JEST training method accelerates training speed and energy efficiency by 13 times and 10 times, respectively. JEST breaks apart from traditional AI model training techniques by training based on entire batches, utilising a smaller AI model to grade data quality and a larger model for training findings. The JEST method relies on the quality of its training data and is more difficult for amateur AI developers to match without expert-level research skills.
A team of Chinese researchers has developed the world's first open-source brain-on-chip intelligent system. It’s a tiny organoid from human stem cells that is integrated into a small robot, creating a unique system that can be controlled by brain tissue. The brain tissue is connected to a neural interface, allowing it to transmit instructions to the robot. The scientists treated the organoids with low-intensity ultrasound, finding that it could support the formation of networks within the host, offering a potential non-invasive method to help patients with brain damage. The use of ultrasound could help bridge the gap between organoids and a computing interface, representing a small step towards a future where lab-grown brain tissue could aid in restoring functions in the human brain.
Baidu has developed a "self-reasoning" framework for AI systems to critically evaluate their own knowledge and decision-making processes. The self-reasoning AI outperformed existing state-of-the-art models in question-answering and fact verification datasets, achieving comparable performance to GPT-4 while using only 2,000 training samples.
Researchers developed a neuromorphic device that can perform on-chip training, eliminating the need to train AI models in a computer environment before transferring them to the chip. They managed to create a two-layer neural network based on electrochemical random-access memory (EC-RAM) components made of organic materials, but more layers are needed for practical applicability.
Other scientific papers that attracted our attention include: RouteLLM, a set of efficient router models that dynamically select between a stronger and a weaker LLM during inference. CoCoNot, a resource for training and evaluating models' noncompliance. A study that showed how to utilise tree search algorithms, such as Monte Carlo Tree Search, to significantly improve LLM performance on complex mathematical reasoning tasks in an efficient manner. A system that enables non-experts to program robots using natural language prompts. A couple of recent papers that addressed LLM vulnerabilities, such as self-evaluation as a defence against adversarial attacks, and an information-theoretic threat model called inferential adversaries who exploit impermissible information leakage from model outputs to achieve malicious goals. Larimar, a novel architecture for the process of updating the knowledge stored in an LLM using a distributed episodic memory. MobileLLM, which optimises sub-billion parameter language models for on-device use cases. FlashAttention-3, which allows for more efficient GPU utilisation, better performance with lower precision and longer context. MA-LMM, which integrates a vision model into a large language model to store past video information in a memory bank. MUSCLE, a set of evaluation metrics for compatibility to prior model versions and a training strategy to minimise inconsistencies in model updates. The NeedleBench framework, which assesses long-context capabilities and logical reasoning of LLMs. SpreadsheetLLM, an efficient encoding method of spreadsheets into LLMs. A paper that analysed model collapse, the phenomenon of AI producing nonsense when trained on AI-generated data. SAM 2, a unified model for real-time object segmentation in images and videos that achieves state-of-the-art performance.?
??? Fun things to impress at the dinner table
Forever young. An impressive and scary AI-generated video of celebrities meeting their younger selves. A must watch.?
A man’s best friend. New York state's Office for the Aging has partnered with Ageless Innovations to provide robotic companion animals for the ageing population, including a retriever-like dog, a cat, and red and blue birds.
Video killed the radio star. SYNDI8 is a new AI-generated pop group that was created by AI voice tech company Supertone.?
Intelligentia Artificialis. An AI-generated video that shows how life was in the Roman Empire.?
Shaping minds, touching hearts. Morehouse College is set to utilise AI teaching assistants, virtual 3D spatial avatars that use OpenAI models to engage in two-way oral conversations with students, leveraging professor-created content and 3D models for lessons.
Your guests have arrived. Bird Buddy is a bird feeder that uses AI to identify individual birds and sends users notifications when they visit.
Starry eyes. Research showed that AI-generated fakes can be detected by analysing human eyes using techniques similar to studying pictures of galaxies.?
Social good. A Virginia Congresswoman used AI to address the House, as she has been diagnosed with progressive supranuclear palsy, making it difficult for her to speak.
Trust no one. Meta is rolling out its AI Studio tools for custom chatbots on Instagram, allowing creators to make AI personas to chat with followers and fans on their behalf.?
??Show me the money
We have tracked nearly 50 funding rounds worth an aggregate 5.1 billion dollars during July 2024, an 19% increase compared to the previous month: