The Artificial Investor - Issue 40: 2024 highlights
My name is Aris Xenofontos and I am an investor at Seaya Ventures. This is the last post for the year, where we will go through the highlights from the last 12 months of the Artificial Investor:
Small language models catch up with LLMs
?? It takes less than 20 months for SLMs to catch up with the state-of-the-art LLMs that wowed us
?? The secret technical sauce: distillation and (real or synthetic) data quality
?? Implications for the Infrastructure Layer: democratisation of AI model development and shift away from BigTech
?? Implications for the Application Layer: 2x-3x faster models consuming up to 10x less energy unlock use cases with massive scale, basic devices, machine-to-machine communication and where data privacy is critical
?? Looking ahead, we expect SLMs to get more attention and drive startup growth in IndustryTech, Supply Chain and Logistics, Manufacturing and Energy Management.
?? Click here for the full version of issue 39.
The current status of the “AI Cold War
???? Chinese AI models surpassed the most advanced US counterparts, while one of China’s leading robotaxi companies debuted in the Nasdaq with a 400-million-dollar IPO.
?? Raw materials: China is the world’s largest producer of most relevant materials. US has a large China dependency.
?? Hardware layer: China is stronger in pure-play manufacturers (foundries), but US has Nvidia
?? Model layer: US maintains the lead, but it is very closely matched by China
?? Cloud layer: US hyperscalers own 70% of the global market and even 70% of the Chinese market. Huawei is gaining share on the back of US restrictions.
??? Application layer: 830 million monthly active users of AI chatbots and the robotaxi company with the most distance covered position US in the lead
?? Reseach & intellectual property: China has more research papers and patents, but US wins in quality
?? Looking ahead, we expect China to continue to close the gap with the US.
?? Click here for the full version of issue 38.
The first signs that the “AI Scaling Law” is not working anymore
?? Not hitting a wall yet: We are indeed very close to the end of the technical innovation of that wave, but this is not the end of the current AI wave as a whole.
?? Finally some utility: The Gen AI Application Layer is still a small and immature market. The leading voice chatbots are still very basic, while that killer AI voice app that helps us manage our workload and daily life is still missing.
?? Back to the lab (or garage) The next big leap in AI will likely take place in a small lab somewhere or even a startup garage.
?? Transformers have important limitations. They are expensive, (data & compute), unreliable (hallucination), black boxes, static post training, and cannot address complex or continuous data problems.
?? The next AI S-curve could come from a range of innovations: from Tree of Thought and World Models to Progressive and Memory-Augmented Neural Networks.
?? Click here for the full version of issue 37.
An AI impact analysis of the US election result
Assuming policy implementation matches campaign plans, we expect:
?????? Less talent coming in
?? More foreign money coming in
?? Scraping of hundreds of regulation pages
?? Escalation of the AI Cold War
?? Tectonic shifts in research funding
?? Harder and costlier production of AI hardware
?? Click here for the full version of issue 36.
The acceleration of the robotaxi trend
?? Still early days: The top 7 players have a collective fleet of about 4,000 vehicles and their combined revenues are $200m, 0.8% of Uber’s ridehailing run-rate revenues
?? Investor excitement is at all-time high: WeRide’s IPO was at c.110x run-rate revenues and Waymo’s funding round at >500x run-rate revenues
??The strategic battles are well underway: as Tesla is entering the market, we analysed 9 players and identified 30 strategic partnerships with car manufacturers and ridehailing companies
???? The regulatory framework remains fragmented and immature. UK is the most advanced country
?? After a couple of setbacks, the industry has been boosted by positive safety studies that were released recently.
?? We summarise five studies about the potential economic and societal impact of autonomous vehicles
?? Click here for the full version of issue 34.
Energy can be the next bottleneck of AI
??The cost base of a datacenter: energy costs account for approximately 60%-70% of the total operational costs
?? AI is much more energy-hungry than the Cloud: AI-powered search consumers >20x energy than a Google search
?? Startups and AI hyperscaler reactions: hardware advancements, model optimisation, power management, cooling technologies and Edge computing.
?? Electricity grids are fragile: The average age of American and European regional grids is about 40 years.
?? Great opportunities for Tech startups and investors across Smart grid management, Energy storage and Distributed energy generation.
?? Click here for the full version of issue 33.
AI is causing turbulence in the Software industry
We expect:
??Increase in SaaS revenues due to capturing a bigger piece of the value chain through AI features
??Lower SaaS revenues due to more companies building in-house
??Lower margins due to higher competition driven by lower entry barriers
??Higher churn as clients move from innovation laggers to leaders
The following market segments seem to be the most vulnerable:
?? Selling to large Tech-native companies
?? Selling basic software and AI wrappers
?? Selling highly-bespoke software to large companies in areas that are core to their business
The following market segments seem to be the most protected:
?? Selling to small and medium-sized companies
?? Selling to traditional (e.g. chemical manufacturers) or regulated sectors
?? Selling complex software (multiple user types/ stakeholders, integrated with many systems, ideally legacy)
?? Click here for the full version of issue 32.
An analysis on whether OpenAI is worth 160 billion dollars
We recognise the potential size of the AI market and OpenAI’s market and innovation leadership; however, we struggle to believe that such loss-making (largely) consumer business is worth 14 times next year's revenues.
?? The bull case
?? On a growth-adjusted basis, OpenAI’s valuation seems not only justified, but also conservative, when compared to the AI hyperscalers.
?? The AI market is growing fast and can be worth 1.2 to 1.5 trillion dollars by 2029. OpenAI would need only a 6%-8% market share to meet its business plan.
?? OpenAI’s first-mover advantage in the Gen AI market has brought it to a position of clear market leadership
??? Strong defensibility driven first-mover advantage, developer ecosystem, strategic partnerships, innovation pace, and a cash war chest.
?? The bear case
?? The model market seems to be commoditised given the sharp decline in prices.
?? Competition is fierce: Free models offered by Meta, data is scarce while competitors have user-generated content, and AI embedded in productivity apps can be more compelling for users.
?? Can the business every make money given its low retention rates and high capex requirements?
?? Leadership and Tech talent mass exodus
?? Where will the exit come from> Tough IPO market and limited buyer universe
领英推荐
?? Being a largely-subscription-consumer business, it seems overvalued vs its peers
?? Click here for the full version of issue 31.
An overview of the global AI regulatory landscape
?? The structure of the global AI regulatory landscape has started resembling…a spaghetti dish
?? Implementing the forthcoming AI regulations will be as challenging as GPDR
?? Application layer impact: depends on whether you integrate vertically or not
?? Strong tailwinds for the Infrastructure tooling and Data layers
?? Model layer impact: guess why the loudest opponents of the potential (and eventually vetoed) California regulation were Google and Meta.
?? Click here for the full version of issue 29.
Can large language models reason? Putting OpenAI o1 to the test.
?? Inside o1’s mechanics: reinforcement learning and Chain of Thought
?? What is reasoning and how it is measured? Deductive, inductive, abductive, analogical and “cause and effect” reasoning.
?? Putting o1 to the test: Solving a difficult crossword and answering reasoning questions for the maths gold medal.
?? The path towards Superintelligence: Shortcomings of current AI models vs. human intelligence.
?? Click here for the full version of issue 28.
An analysis of Nvidia’s valuation and whether it indicates a Tech bubble
?? The bull case
?? Valuation maths tell us the share price is justified by growth
?? The main clients, hyperscalers, are breaking the bank
?? The business is nearly a monopoly
??The bear case
?? Hyperscalers focus on strategic value as opposed to ROI. For how long?
?? The AI Cold War’s impact
?? Death by a thousand cuts of competing players
?? Revisiting the valuation maths. Numbers don’t lie.
?? Click here for the full version of issue 27.
The size & vendor landscape of the Gen AI market
?? Already a huge market: We estimate bottom-up the total Gen AI market to be worth c.$78Bn (excluding services) or $90Bn-$100Bn including the Services layer.
?? The Gen AI market is growing very fast, probably between 400% and 500% on an annual basis, driven by the acceleration of the market leaders at the Application (OpenAI) and Infrastructure (Nvidia) layers.
?? The Infrastructure layer currently holds most of the value (90%+), but the scale is expected to lean towards the Application Layer, as it happened with the previous generational platform, the Cloud.
?? Nvidia is the clear market leader with 69% of the Gen AI market (excluding services), followed by OpenAI and Amazon, each owning c.5% of the market.
?? Click here for the full version of issue 24.
The business case for AI: the Apple case study
?? Distribution is worth as much as technology. How much? For instance, Google is paying Apple c.$18Bn a year for its search engine to be the default one in the Safari app of every iPhone.
?? Technology is not a product. In Apple’s case, it took four complex technological components and software to glue them together into Apple Intelligence.
?? The application layer is very powerful. Instead of spending $10Bn+ to build its own LLM, Apple is getting it for free because it owns the user.
?? Vertical integration is a value multiplier. Apple emerged as a big beneficiary of the current AI wave by integrating vertically: computing chips, phone devices, operating system, applications and cloud infrastructure.
?? Gen AI could have a strong business case: Even if Apple convinced 1% of its non-upgrading user base to buy a new phone in 2025, this would mean a 5% boost in iPhone revenues (or c.$20Bn).
?? Click here for the full version of issue 23.
Search will never be the same again
?? Google’s US Search market share has fallen from 88.96% in February 2023 to 86.58% in April 2024. Gartner predicts that traditional search engine volume will decrease by 25% by 2026 due to the rise of AI chatbots and virtual agents
?? Microsoft has invested cumulatively $13bn in OpenAI with search (Bing) being one of the strategic drivers and impact to get announced. The five main Search startups (The Browser Company, Perplexity, You.com, Neeva and Exa AI) have collectively raised close to 1 billion dollars.
?? The Search ecosystem is changing:
?? Click here for the full version of issue 21.
Voice is the new keyboard
? Ten years since the premiere of the movie Her and the launch of Amazon’s Alexa, AI voice assistants have failed to meet expectations, lacking three key elements: context, adaptability and a natural form of communication.
?? OpenAI launched GPT4o, a true multimodal model that overcomes the pre-existing model limitations (speed and loss of intelligence due to the conversion of text to speech). The result was outstanding and made the vision of an intelligent and natural voice assistant a reality.
?? The Voice Assistant market was worth c.$3Bn in 2023 and is expected to grow by 30% annually for the next ten years, driven by the modality’s superiority vs. other interfaces: can be used without using hands or eyes, has a very short learning curve, it can greatly improve accessibility, it enables multitasking, etc.
?? Nevertheless, voice has some important limitations: privacy concerns (“always-listening” mode), voice interactions are not private, mostly-reliant on an internet connection, and challenges with complex tasks and interactions with other applications.
?? We expect that voice-based interfaces will soon have a big role in our lives, but will complement existing interfaces, such as the keyboard, rather than completely replace them. In any case, we expect huge implications for our society.
?? Click here for the full version of issue 20.
The state of Enterprise AI adoption
?? Enterprise AI adoption is growing (80% of corporates working with Gen AI), driven by potential revenue-generation and cost-saving benefits. ?? Tech, Financial Services and Healthcare institutions are spearheading the innovation ?? Gen AI adoption largely remains at an experimentation stage: 56% net dollar retention in AI tool spending from 2023 to 2024 ?? Significant adoption barriers remain: implementation complexities, talent gap and organisational and reputational risks
?? Click here for the full version of issue 18.
The democratisation of deepfakes
?? Microsoft Research released a new video generation model, VASA-1, that is capable of generating realistic videos of talking faces using a single portrait shot and a speech audio track. Also, Synthesia released their new generation of avatars that are capable of mimicking human emotions.
?? It takes two to build one. Deepfakes are typically crafted using Generative Adversarial Networks (GANs), where two models (neural networks) collaborate, where a generator model creates images or videos that mimic the target data (e.g. a person's face) and a discriminator model evaluates the outputs from the generator to determine whether they are real or fake, and learn to be better at discriminating over time.
?????? 2.5 billion potential victims: There has been a 10x increase in the number of deepfakes detected globally across all industries from 2022 to 2023, while nearly 100% of the political elections of the last 9 months have been affected by deepfakes. This equates to a pool of 2.5 billion of potentially-deceived people.
?? The world’s reaction: AI model developers are embedding watermarks and metadata. Governments across China, the US, Europe and the UK are introducing legislation. Several startups have launched and received funding in the last two years trying to tackle the problem, including companies like Reality Defender and Clarity.
?? We are moving to a world of lower trust: as trust on text has declined dramatically with typography and social media, we expect the same to happen with audio and video. We expect to move from provenance to authorisation.
?? Click here for the full version of issue 17.
Big Tech's vertical integration
?? Meta announced the launch of its second-generation AI hardware (MTIA) and Google unveiled its new proprietary AI chip, Axion.
?? Hyperscalers state efficiencies, improved performance and customised components as the motivation behind these vertical integration initiatives. But is that it?
?? Owning a proprietary AI chip can offer a competitive advantage. See Grok, a startup that developed an AI processor capable of running foundational models up to 10x faster than Meta or OpenAI.
?? With Nvidia owning c.90% of the Hardware Infrastructure Layer, Big Tech is eager to minimise dependency on third parties.
?? As the AI Cold War heats up, integrating the supply chain seems like the only way forward.
?? Cash is king. Who wouldn’t want to incorporate the average semiconductor gross margins of 60% into their profits?
?? Given the vertical-integration trend, we expect commodities to be next battleground
?? Click here for the full version of issue 16.
Secured Financing, Credit, Loan, Lending & Mortgage | Alternative Investment: PE, VC, Pre-IPO, Unicorn, Hedge Fund, Life Settlement & Litigation Fund | Empower Institution, Enterprise & Single Family Office (SFO)
1 个月Steven Tsui here. Impressive insights, Aristotelis Xenofontos! AI's impact on various sectors is fascinating. We connect clients with secured loan and equity investors, tailored to their needs. Let's chat!
??
Aristotelis Xenofontos Amazing year you had! wish you another great year in 2025!! :) ??????
AI Swarm Agent & Automation Expert for the Trades | Co-Founder Trade Automation Pros | Co-Founder Skilled Trades Syndicate | Founder of Service Emperor HVAC | Service Business Mastery podcast | Tri-Star Mechanical
2 个月Aristotelis Xenofontos What an insightful recap of AI's transformative journey in 2024—each highlight is a fascinating picture of the industry's evolution!
Chocolate Milk Cult Leader| Machine Learning Engineer| Writer | AI Researcher| | Computational Math, Data Science, Software Engineering, Computer Science
2 个月Great insights