The Artificial Investor - Issue 30: September 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.

One of the worst months of the year in terms of funds raised by AI startups. Nvidia share’s rollercoaster ride. AI product launches…without AI. The big US AI regulation that never happened. 30 billion dollars for AI infrastructure and the race for the largest AI server cluster. The launch of the best reasoning AI model to date. All this and much more in today’s issue about September 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 120 funding rounds worth an aggregate 2.6 billion dollars during September 2024, approximately half of the amount of July 2024. We know October will help get the monthly average back to 5+ billion dollars, as OpenAI closed a 6 billion-dollar funding round as we are writing this issue.?

In terms of exits, we saw mainly small acquisitions of AI players by large software and other Tech companies, such as Atlassian’s acquisition of Rewatch (meeting recording), Salesforce’s acquisition of Tenyx (AI voice agent) and Instacart’s acquisition of Eversight (pricing and promotions platform).?

Recent mega rounds indicate continuous investor appetite to back foundational model developers (Mistral, Sakana) despite the capital intensiveness of their business model and commoditisation risk. Enterprise productivity (Glean, Codeium) and AI project scaling (Safe Superintelligence on the software side, Applied Digital on the hardware side) also had their fair share of attention.?

?? On pink paper

Exponential. Stripe published a jaw-dropping article stating that AI startups need 11 months to reach 1 million dollar revenues, compared to 15 months for SaaS companies during the Cloud boom. At the same time, the hyper growth of consumer LLM usage seems to continue. OpenAI is leading the march with more than 200 million weekly active users, while 92% of Fortune 500 are using its products. In parallel, it was leaked that Meta AI had 400 million monthly active users and 40 million daily active users in August 2024. The Big Tech also confirmed that its Llama models have been downloaded 350 million times, with 20 million of them taking place just in July 2024.?

Are we in a bubble? Nvidia’s stock fell by more than 10% in a week during September (only to bounce back by the month’s end). We covered in one of our September weekly issues about Nvidia’s share price movements, and our opinion on whether the stock is currently overvalued or not (spoiler alert: if you own the stock you are probably fine for 12-18 months). At the end of the month investors took a liking to Oracle. Investment analysts upgraded their expectations for future growth of the company, as the cloud infrastructure provider confirmed a 10% annual growth for 2024. Overall, the share has risen by 60% in 2024.

New, but not. September is typically a month of new product launches and announcements, and this year could not be different. Apple launched the much-anticipated Apple Intelligence, its privacy-first AI infrastructure, which for now is…invisible. This is because the new iPhone 16 and iOS 18, also launched last month, include some interesting new features, but the main AI functionality will need to wait for a new iOS upgrade expected this October. Amazon announced a revamp of Alexa (we need to see this to believe it) powered by Anthropic’s AI models, as well as the launch of a selling assistant for consumers called Amelia and an ads video generator application for merchants, which for now doesn't have a name (are we running out of female names starting and ending with an “a”?).

AI vs. software 1-0. A piece of news from last August that got lost in the summer media downtime is a tweet on X by Klarna’s CEO, in which he stated that the company is not renewing its Salesforce and Workday licences, because they are building the necessary applications in-house using AI. This is part of the great debate started since Chamath Palihapitiya announced the launch of an incubator called 8090 that plans to use AI to develop 80% of the functionality of common software, such as a CRM, with 90% less effort than traditional SaaS companies. Will AI kill all software? We love this debate, so stay tuned as we pick up the topic in one of October’s weekly issues.

Budget inputs. September is the month that many corporations start their budget cycle for the next year. Consultants know this very well and release market studies to help out. S&P Global published their second annual Global Trends in AI survey. The highlight was the statistic about Enterprise AI adoption: 24% of corporates see Gen AI as an integrated capability deployed across their organisation, with 37% in production but not scaled; only 11% of them are not investing in Gen AI at all. Deloitte also published a survey about enterprise AI adoption, where nearly 75% of respondents ranked data privacy among their top three concerns. Another report claims that 87% of AI projects fail, with the main reason being the lack of Intelligence Engineers, an emerging interdisciplinary job that is missing from most companies.?

?? A double-edged sword

All eyes on the US. California’s governor vetoed the potential AI Safety bill on its deadline date, stating that he would prefer a law that differentiates between the risk of the various AI use cases and treats open source differently. This may not be that bad, as these were the two main differences between that bill and the EU AI Act. Could we finally get some alignment between the two sides of the pond in a Tech-regulation, for a change? We shall see. In the meantime, you can read a piece on AI regulation that we wrote as part of one of September’s weekly issues. While California does not have a broader state law for AI, it does have one (as of recently) about the use of audio and video of deceased people, which basically asks for the obvious: prior family consent. An FTC study recommended a federal data privacy law for the US and it remains to be seen if the policymakers do something about it.?

Together we are stronger. While AI regulation has been fragmented worldwide, there are indeed some efforts for international collaboration and coordination. In September, the EU, UK and US signed the first AI treaty, the AI Convention. The treaty sets out high-level requirements on transparency, oversight, accountability and responsibility, equality and non-discrimination. Signatory countries have some freedom on its interpretation and must ratify the AI Convention for it to have effect in their jurisdiction.

The Cold War continues. The US Commerce Department proposed a ban on all China-originated software and hardware related to autonomous vehicles, driven by fears of misuse of the data collected. Nevertheless, it seems like China is not doing that bad in mitigating the West’s efforts to slow down its technological innovation pace. Instead of disappearing, Nvidia chips are available in China and are actually cheaper to rent than in the US. In addition, we have been following the international success of some Chinese firms, such as Tencent Cloud’s partnership with Etihad Etisalat in Saudi Arabia, and Huawei’s launch of Galaxy AI in North Africa as part of a 430 million dollar investment plan.?

From source to socket. By now everyone has realised that one of the drawbacks of AI is its environmental impact. Public awareness grew in various ways in September. University of California researchers published a study showing that ChatGPT consumes a small bottle of water in an average conversation (water is used directly to cool data centres and indirectly to cool the power plants that supply electricity to them). The White House invited the CEO’s of Nvidia, OpenAI and Anthropic, and Google’s president, to discuss AI infrastructure and energy challenges. Microsoft executives were missing from the meeting; perhaps they were busy closing the deal with Constellation Energy to reopen its Pennsylvania nuclear plant to provide power for its data centres as part of a 20-year deal!

Perfect storm. It looks like a perfect storm has been created for the Tech job market. The 2020-21 Tech boom and bust has certainly been a driver of declined demand for Tech jobs. On top of that, you have the AI-driven productivity gains for software developers. In what is probably the most reliable study to date, researchers carried out randomised controlled trials on three experiments with about 5,000 software developers at Microsoft, Accenture and an undisclosed non-Tech corporation. The study indicated that there was a 26% increase in the number of completed tasks among developers using AI coding tools vs. the ones not using any. The WSJ published an article about how experienced Tech workers are struggling to find new roles and many are resorting to unusual tactics, such as putting up fliers or enrolling in college courses.

?? Laying the groundwork

?? Models

September 2024 was an important milestone for many text-based foundational models with releases by most LLM development incumbents. OpenAI launched its first reasoning model, o1. We covered the capabilities of the model and what it means about how far we are from super intelligence in one of September’s weekly issues. Another model launch for Meta, another model commoditisation for the others. Meta launched its new open-source model, Lama 3.1, which brought it to the top of the benchmarks. Small models also had their lion share, with Google (new small Gemini models), Nvidia (Minitron 8B model) and Mistral (Mistral Small 22B) all advancing further in this space. Finally, Google was responsible for another interesting release in the text-based model area. DataGemma is a pair of models that collaborate with each other to reduce hallucinations.?

On the vision side, we had two impressive model launches coming from the East. Alibaba launched Qwen2-VL (among others, within its new v2.5 model series), a model with state-of-the-art understanding of images of various resolutions and ratios, the ability to understand videos over 20 minutes and the capacity to operate robots. Minimax, another Chinese company, released its video generator model that competes with OpenAI’s Sora and supports 1280 x 720 resolution videos at 25 frames per second Check out some of the demo videos, they are impressive.?

Agents are becoming hot on both sides of the world. China's Ant Group announced an AI 'life assistant' mobile app called Zhixiaobao, which will offer a range of services including ordering meals, hailing taxis, booking tickets and discovering local dining and entertainment options. On the opposite side of the map, Salesforce announced Agentforce, a Gen AI agent platform that will offer pre-built agents for sales reps, service agents, personal shoppers and sales coaches - starting at $2 per conversation. Paradigm launched a new spreadsheet agent powered by Gen AI that automatically fills in spreadsheet cells at 500 cells per minute!?

Other September model launches include AlphaProteo, Google DeepMind's first AI system for designing novel, high-strength protein binders, the “Reflection 70B saga”, a model that was launched and supposedly beat every other model in the main benchmarks, but was then exposed as a fraud, and 1X’s new model that is used by the company to help train robots.

?? Hardware

The North. The South. The Desert. We came across an interesting study that maps the global geography of public cloud GPU compute, based on the countries’ AI compute capacity and relevance. The US (GPU quality) and China (GPU quantity) lead the Compute North group, which has capacity appropriate for AI development. This group also includes another 15 countries, which are all in the Northern Hemisphere (including the EU), with the exception of Australia. Another group, the Compute South, consists of countries whose capacity is more appropriate for AI deployment (as opposed to development) and consists of 13 countries that are in the Southern Hemisphere (including LatAm), with the exception of Switzerland. The rest of the world is classified as a Compute Desert. Adopting this nomenclature then, last month saw Amazon announcing a further investment in the Compute North (8 billion dollar data centre in the UK) and Groq in the Compute South (a big partnership with Saudi Aramco).?

Size matters. The race to build the largest AI server cluster is heating up. Oracle released a huge AI cluster consisting of more than 130,000 of Nvidia’s cutting-edge chips (Blackwell). At the same time, It was leaked that Meta is preparing a cluster to train its next LLM, Llama 4, which will consist of 100,000 tier-2 Nvidia chips (H100) costing 2 billion dollars. AI clusters are not just about chips, they also need land, buildings, power, water, etc. Where is all this money coming from? Blackrock and Microsoft announced the launch of a 30 billion dollar fund to invest in AI infrastructure projects.?

DIY. We have talked in the past about one of the main AI infrastructure-layer trends: hyperscalers developing their own chips and reducing their dependence on Nvidia. Last month it was confirmed that OpenAI has partnered with TSMC to build its own chips, while Microsoft revealed Maia 100, its first in-house AI chip. In an attempt to jump on the chip foundry (the business of manufacturing custom chips for 3rd parties) bandwagon, Intel signed a deal to make custom AI chips for Amazon.

I travel at the speed of light. Groq, a specialist AI chip designer, had made a big impression when they showcased the speed of their infrastructure, which reached 250 tokens per second running Llama 3, which is 5x faster than AWS. This time around, Cerebras Inference, another AI chip designer, unveiled its technology capabilities and claimed it can reach 2x the speed of Groq and 10x the speed of AWS. The chip resolves the memory bandwidth issue, which has been a significant bottleneck in traditional GPU setups. This is achieved by integrating 44GB of SRAM, thus eliminating the need for external memory. Also on the chip innovation side, Infineon announced a breakthrough in using a special compound to manufacture chip wafers, gallium nitride. The compound has lower production costs and it’s very energy-efficient in computations.?

?? Scientific Breakthroughs

Last month saw a number of advancements in AI research, including:

• xLam, a series of LLMs optimised for function calling and AI Agents. Function-calling is a sought-after application for agents, which involves agents completing user commands through a series of function calls.
• GameNGen, a game engine powered by a neural model that allows real-time interaction with complex environments at high quality. The demo video shows how it can recreate the classic game DOOM on the spot; pretty impressive.
• LTM-2-mini, the first LLM with a 100-million-token context window.
• A model distilling methodology that achieves performance comparable to the original model at a fourth of the memory requirements, by converting large transformers into linear RNNs, incorporating a quarter of the attention layers.
• DisTrO, a distributed optimizer that reduces massively the required bandwidth during pre-training, by reducing inter-GPU communication requirements by four to five orders of magnitude without relying on amortised analysis.?
• A comprehensive web-based Federated Learning (FL) solution, which simplifies orchestration of FL tasks while integrating intent-based automation.
• GenRM by Google DeepMind, which improves LLM accuracy by getting them to verify their outputs. The methodology involves majority voting, making effective use of test-time computation.
• OP-RAG (Order-preserve retrieval-augmented generation), a mechanism that achieves higher answer quality with much fewer tokens than long-context-window LLMs.
• Agent Q, a framework for AI agents that combines guided Monte Carlo Tree Search with self-critique mechanism and iterative fine-tuning on agent interactions.?
• MambaByte, a new model architecture that is competitive with state-of-the-art transformers on language modelling tasks, while maintaining the benefits of token-free language models, such as robustness to noise.
• UI-JEPA by Apple, an architecture designed to understand user intentions based on user interface interactions. The design has reduced computational requirements while maintaining high performance and its potential applications include the creation of automated feedback loops for AI agents.
• Cobra, a multi-modal large language model aimed at improving efficiency by employing linear computational complexity.
• We also came across a research paper that compares the three most common AI agent frameworks - LangGraph, Autogen, and Crew AI - evaluating aspects such as ease of usage, tool coverage, memory support and scalability.

??? Fun things to impress at the dinner table

Talk to me. AI restores the ability to speak for an ALS patient using brain implants and machine learning models.

Safety first. Amazon introduced Astro robot, a robot security guard for businesses. No, it doesn’t come armed.

Dr. Chatbot will see you now. About one in four American adults under 30 years old use chatbots at least once a month to find medical advice and information.

GOTCHA. AI is now able to defeat CAPTCHA puzzles with a perfect success rate. Oh boy, I have been finding CAPTCHAs so tough recently, I don’t wanna see the next iteration!?

A(I)rt. Media artist Refik Anadol plans to open the world's first AI arts museum, Dataland, in downtown L.A.

My favourite holding period is forever. This startup is launching an ETF fund that uses AI to mimic Warren Buffett’s investment strategies.

Show me love. 225 million people have downloaded AI companion applications in Google Play to form romantic relationships with AI lovers.

Conspiracy theories. We have all been there: at a dinner party, talking to someone who comes up with extreme conspiracy theories, and we can’t convince them otherwise. This chatbot is the answer.?

??Show me the money

We have tracked nearly 120 funding rounds worth an aggregate 2.6 billion dollars during September 2024, approximately half of the amount of July 2024. Below is a selection of them.

• Safe Superintelligence raising $1B for its AI safety research solution
• Mistral raising $645M for its multimodal AI model solution
• Glean raising $260M for its generative AI for enterprise productivity solution
• Applied Digital raising $160M for its AI cloud-computing and data-centre solution
• Codeium raising $150M for its AI coding platform solution
• Sakana AI raising $100M for its collective intelligence AI models
• Forterra raising $75M for its autonomous systems for defense and industrial sectors
• You.com raising $50M for its AI-powered search and productivity engine
• Pyka raising $40M for its autonomous electric aircraft for agriculture and logistics
• Smartcat raising $43M for its AI-powered translation platform
• Hippocratic AI raising $17M for its safety-focused AI solution for healthcare
• HavocAI raising $11M for its uncrewed surface vessel technology solution
• Blitzy raising $4.4M for its custom software automation solution
• Revrag.ai raising $600K for its AI sales agent solution
• AceUp raising $22.5M for its AI-powered team transformation platform
• Reshape Automation raising $5M for its AI-driven automation solutions for manufacturing
• Champion raising $3.3M for its AI-powered customer advocacy platform
• Campfire raising $3.95M for its AI-native gaming platform
• Root Signals raising $2.8M for its generative AI measurement platform
• Operant AI raising $10M for its runtime application protection platform
• Bounti raising $16M for its AI teammate platform for sales teams
• Martian raising $9M for its large language model router solution
• Track3D raising $4.3M for its AI reality intelligence for construction
• Revrag.ai raising $600K for its AI sales agent solution
• Cerrion raising $5M for its video AI for industrial automation platform
• Vind AI raising €3M for its AI-driven wind park design solution
• IdentifAI raising €2.2M for its AI-generated content identification solution
• Findable raising €9M for its AI-powered solution to automate building documentation
• Cybord raising $8.7M for its AI-based electronic component quality solution
• Protege raising $10M for its AI training data platform
• Cashmere raising $3.6M for its AI platform for wealth management client acquisition
• Convergence raising $12M for its AI agents for personalized productivity
• Fathom raising $17M for its AI-powered meeting intelligence platform
• Prepared raising $27M for its real-time AI-powered 911 dispatcher platform
• Ensemble raising $3.3M for its AI-driven data quality enhancement platform
• Pieces Technologies raising $25M for its AI-driven healthcare solution
• Atomicwork raising $14M for its AI-driven employee experience improvement tools
• Bot Auto raising $20M for its AI-based autonomous trucking solution
• SplxAI raising $2M for its AI cybersecurity platform
• Supermaven raising $12M for its AI-powered coding assistant platform
• Steerlab raising $1.9M for its AI solution for automating RFP responses
• Senvo raising €2.5M for its logistics cost visibility and reconciliation platform
• Spotter raising $7.4M for its AI-powered tools for YouTubers
• Lightium raising $7M for its AI-driven photonic chip technology
• Zenlytic raising $9M for its AI-powered business intelligence platform
• Rep.ai raising $7.5M for its AI digital twin solution for real-time engagement
• Illuma Labs raising $9M for its voice authentication technology for banking
• Osavul raising €2.78M for its AI disinformation-fighting platform
• Mbue raising $1.8M for its AI construction tech platform
• MyndYou raising strategic investment for its healthcare AI solution
• Quin AI raising ￡1.5M for its behavioral generative AI platform
• Vsim raising $24M for its robotics simulation technology
• Fastn raising $2.6M for its composable architecture platform
• ZipBio raising $4M for its AI-driven biotech platform
• Cerrion raising $5M for its video AI for industrial automation
• Reshape Automation raising $5M for its automation solutions platform
• Anytime AI raising $4M for its AI legal assistant for plaintiff lawyers
• Cerrion raising $5M for its industrial automation video AI solution
• Steerlab raising $1.9M for its AI platform for RFPs automation
• Campfire raising $3.95M for its AI-powered gaming engine
• Acuvity raising $9M for its AI solution for GenAI governance
• Martian raising $9M for its AI-powered LLM router
• Cybord raising $8.7M for its AI-driven electronic component quality solution
• Mbue raising $1.8M for its AI platform for architecture
• Vind AI raising €3M for its wind park design optimization
• Fastn raising $2.6M for its AI composable architecture platform
• Operant AI raising $10M for its AI cybersecurity platform
• PharmaForceIQ receiving a growth investment for its AI marketing solution for life sciences

Thanks for reading The Artificial Investor! If you liked it, please share it.