Generative AI is eating the world in the quest for AGI
A primer to ChatGPT and its AI competitors (October 2024) | Quantum Edge
Oct 04, 2024
From buzz word to key word
A few short years ago, AI was mostly a buzzword among the technically minded. So how did the words "LLM" and "generative AI" become permanent fixtures in our business lexicon?
Enter ChatGPT (Generative Pre-trained Transformer).
OpenAI released a better Google that allowed users to engage, tweak, and even break it to get tailored responses through the use of prompts. The viral interactive chatbot contained all these features within a visually appealing and intuitive interface to captivate audiences globally. This killer use case spawned a Cambrian explosion of generative AI and nudged the general public to contemplate the benefits (and consequences) of AI-everything. Soon after, creative AI bots like Midjourney and Dall-e elevated user imagination by turning anyone into a digital creative with its generative image capabilities - all made possible with a few keystrokes.?
Faster everything
ChatGPT became one of the fastest apps to achieve both 1 million users (5 days) and 100 million users (2 months) since its public launch in November 2022. This technological leap ushered a new era of AI that catalyzed a gold rush for GPUs (see NVIDIA stock price), which provide computing power for the vast amounts of data needed to train and operate large language models (LLMs). The data is organized through neural networks, or architecture loosely designed on the pathways of the human brain, and comes with layers of unprecedented complexity. This shift in the zeitgeist catalyzed an arms race between companies and startups toward the quest to achieve Artificial General Intelligence (AGI), which is the ability to match and surpass human cognitive abilities across a range of tasks. Such a daunting endeavor towards the next phase of the industrial revolution broadly requires high CapEx for data center infrastructure, unwarranted demand for computing power, quantum leaps in clean energy generation, fierce competition for niche talent in a nascent space, and visionary leaders to navigate the potential for AI enablement while upending the status quo. Simply put, generative AI will leave an indelible mark on humanity and the way we work.?
Unlocking AI potential
Is AI simply a hammer looking for a nail? What problems can generative AI truly solve? Will the world end up like a sci-fi movie? The possibilities are nearly unimaginable at this stage.
"AI will be the most important technology of the 21st century, just as electricity was in the 20th century."
- Demis Hassabis (co-founder of DeepMind)
"I think the impact of AI will be larger than the impact of the personal computer, maybe even larger than the impact of the internet. I think it's like the impact of electricity or the steam engine—a real general-purpose technology that's going to affect almost everything."
- Sam Altman (CEO of OpenAI)
"We should be thinking hard about how to create AIs that are not just powerful, but also aligned with human values and interests."
- Dario Amodei (Co-founder and Chief Research Scientist of Anthropic)
"Any sufficiently advanced technology is indistinguishable from magic."
- Arthur C Clarke (British science fiction author)
Perhaps we truly are at that internet or social media moment, where our lives are about to be inextricably integrated with new tech in ways that we couldn't imagine. For now, with a short prompt and perhaps a few iterations, a user can essentially get any answer to almost anything across the public domain - almost like instant access to a digital consciousness. Soon we may all have our own personalized interactive virtual assistants at our beck and call (akin to Jarvis for Tony Stark in Ironman). Meanwhile, vast resources in capital, infrastructure, and computing power are being funneled towards this leap in technological capability.
Bubble hype or New Paradigm?
Investors allocated the majority of investment dollars across the AI ecosystem in 2021, with deal flow declining quarterly while unicorn rounds drove spikes in funding value. This trend indicates consolidation among investors towards high-potential unicorns, with concentrated bets channeled towards those perceived winners. In 2024, major funding included $1B+ rounds to AI infrastructure plays like OpenAI (October), Anthropic (March), and Moonshot China (February), which provide the horsepower for generative AI chatbots that users have grown to know and love.
As we can see from this Q1 CB Insights report, the AI landscape is fairly distributed and does not cover the base layer of compute infrastructure that powers this generation of technology. This article will focus on generative AI (housed within AI infrastructure), and we will evaluate key players, valuation & funding, focus & capabilities, success factors, and other intangible considerations specific to the space.
Big Numbers for Big Ideas
OpenAI leads the pack with a whopping $157b valuation from their latest round closed in early October (superseding the previous $80b valuation), followed by a host of unicorns worth multiple billions, including Databricks ($43b) and Anthropic ($18b) who is rumored to be raising for their next round at around $50b.
In venture capital circles, AI went from an enhancement feature to a necessary integration across most tech startups. This is in large part thanks to OpenAI, which democratized generative AI for the non-technical user and paved the path for commercial opportunities across industries.?
Valuations are often derived by Revenue multiples for private companies. Generative AI maintains a wide range, spanning from 8-150x Valuation/Revenue. OpenAI and Anthropic both come in at approximately 40x and 21x revenue multiples respectively, which is not unreasonable in the context of their comparables.? In a competitive arena of "growth at all costs", billions are being raised for billions in returns, as generative AI companies trailblaze speedy runways to $1B ARR (annual recurring revenue) propelling them into the league of tech startup heavyweights in under 3 years. For comparison, one of the most successful GPU chipmakers (NVIDIA) trades at a 31.6x revenue multiple as an established publicly traded firm, thus indicating relative fair value for the eye-watering valuations generated by fundraising rounds for OpenAI and Anthropic among others. OpenAI would be ranked #64 with $157b market capitalization in the S&P 500 if it were a publicly traded company as of this writing
Are chatbots really worth more than most established industrial Fortune 500 companies? What's driving these supersize valuations after effectively two years since inception? Why would some of the world's largest venture capital funds and tech companies invest at such high valuations - where is the upside?
The simple answer - industry pioneering. Imagine being Elon Musk pre-Tesla, empowered with a vision to electrify a segment of the world's fleet of vehicles. The sheer cost and scale of the battery supply chain and charging infrastructure are enough to dissuade the most ambitious of founders and investors. However, once these Herculean tasks were overcome, Tesla enjoyed the first-mover advantage for many years before meaningful competition was introduced into the marketplace. At one point in December 2020, Tesla traded at more than 21x revenue and now trades at approximately 8x EV/Rev - a healthy premium to other car manufacturers at 1-2x EV/Rev in 2024. The value creation from innovation often requires high capital expenditures, complex supply chains and infrastructure, new technology, specialized engineering, killer use cases for customers, tectonic narrative shifts in ways of working, regulatory challenges, and a host of other considerations. The top AI players have embarked on such a journey, and they are putting their vision into motion.
Since a new industry is not always a "winner-takes-all" game, there is enough opportunity for the major players to explore their thesis while making significant advancements to the space. In fact, many tech companies and VC investors have taken stakes in multiple AI firms within the same competitive space. Major tech firms such as Microsoft, Amazon, and Google have symbiotic investment structures in play, such that the receiving startup uses those funds for host native infrastructure or compute services from the parent company to effectively round-trip funds in a circular ecosystem. In the latest funding round for OpenAI, the company requested investors to limit cross-pollination of funding across competitors. Time will tell if investors follow suit.
How various Generative AI models stack up
LLM Performance is industrially measured by “GPQA” (general-purpose question answering) - it is considered an IQ approximation for LLM effectiveness to evaluate model comprehension, reasoning, and ability to generalize. The earliest publicly available model (OpenAI ChatGPT 3.5) scored 28%, which is higher than the average human, yet lower than a general PhD. Subsequent ChatGPT versions improved that measure to 78%, leading the pack in model performance. Anthropic's Claude LLM follows with an impressive 59% - 67% for their Claude 3 models. Various other LLMs from Google (Gemini), Meta (LLaMa), Pi (Inflection), and 01.AI (Yi-Large) scored in the 30-50% range for their latest models (LifeArchitect.ai ). The top hemisphere represents more user-centric generalized use cases for LLMs, and the bottom hemisphere characterizes more specialized or multi-modal offerings. Use cases may vary for company offerings thus focus may not be mutually exclusive.
The AI Playbook
In addition to model performance and user segmentation, let us evaluate critical success factors for generative AI companies from a strategic perspective. Competitors could vary on technical factors (model performance, open/closed network), target audience, use cases, monetization strategies, and many other dimensions. In almost all cases, however, they all share the need for computing power, training model efficiency, talent pools, regulatory factors, and ethical considerations.
The three core pillars for generative AI companies are Compute, Infrastructure, and Go-To-Market (GTM).
Other critical success factors include the following:
Unbundling Unicorns
Let's review the financial metrics of the two largest generative AI startups, OpenAI and Anthropic:
OPENAI
OpenAI is currently in its hypergrowth stage, with 2023 ARR of $1.6b, and ~$3.5 - 4.5b expected revenue in 2024 ($11b+ expected in 2025). The lion's share of revenue comes from their flagship chatbot ChatGPT Plus ($20/month), ChatGPT Enterprise for large companies, and developer access for product APIs in the form of subscriptions. Licensing deals provide integration of AI models with software that enterprises use to build on OpenAI models. Other revenue streams include Partnerships, Grants, and custom solutions or consulting offerings for specialized AI deployments.
领英推荐
Computing and infrastructure costs dominate the P&L, which include GPUs, cloud services, and the cost of running inference at scale for its cloud services, subscriptions, and enterprise offerings. Furthermore, direct costs include the processing of tokens for user requests through API services. Separately R&D model development involves model training and iterative development, which is capital intensive and competitive depending on the datasets and variables used to train LLMs. Their latest model "o1" leads the market across many dimensions. Personnel expenses represent employee costs for staff engineers, researchers, and other key roles, which are offered at a premium on market rates to attract and retain niche talent. SG&A and other operational expenditures include sales, marketing, administrative, legal, and other miscellaneous operational costs.
Significant funding rounds help fund their development through loss-earning periods, as OpenAI expects to spend $8.5b in total direct and operating costs this year to recover revenue in future years and establish a market-leading position through novel innovation and new product monetization. Investors are not as concerned with operating losses in the initial years, as revenue growth and expansion take center stage. In later years when revenue plateaus or the company files an S-1 for an IPO on public markets, cost efficiency and profitability become more important priorities.
OpenAI recently announced a successful funding round in early October, which brings their total funds raised to a whopping $17.9b. "The new funding will allow us to double down on our leadership in frontier AI research, increase compute capacity, and continue building tools that help people solve hard problems,"… OpenAI mentioned in a recent blog post. The new tranche of funding will significantly ramp up the OpenAI war chest to pursue capital-intensive long-term strategies such as vertically integrating AI chip production and data centers while reducing reliance on GPU and infrastructure providers like Nvidia.
As a market leader in the space, OpenAI's capitalization could further attract top talent and enable productization of AI capabilities, while giving them the license to explore innovative strategic plays like the recently announced partnership with Jony Ive (Apple's former Chief Design Officer) to build the "iPhone of AI" - such a play can bolster user adoption and retain users within their ecosystem of software offerings (similar to Apple and its App Store's closed-loop ecosystem which creates a defensive moat of user loyalty).
OpenAI manages a $175m internal fund to partner with early-stage startups where "AI tools can empower people by helping them be more productive". The Fund has taken stakes in generative AI applications across various domains, thus expanding its footprint in the space and exploring ideas, stickiness, and talent by proxy.?
ANTHROPIC
Anthropic's financials share a similar revenue and cost structure, reflecting the stratospheric scale of market leaders in the generative AI segment. Anthropic saw 2023 ARR of $100m and anticipates 2024 revenue of $850m - $1b. Revenue from its chatbot (Claude Pro) accounted for a significantly smaller portion of its revenue than ChatGPT, as the majority of its revenue comes from business and developer API access into their impressively performing Claude models. Pricing is based on a token model and varies based on task complexity and usage. Licensing Enterprise features also added a revenue stream, which offers increased security and integration opportunities.
Similar to OpenAI, computing and infrastructure costs reflect the majority of expenses. Cloud services, GPUs, API token-retrieval, and other resources are needed to operate AI models and provide functionality to customers, representing a significant cost component that currently overshadows revenue. Language and model training costs, which are likely absorbed in earlier iterations of the market-leading Claude LLM, are estimated to cost $100m for further R&D and model development. Personnel expenses and SG&A are significantly lower than OpenAI proportionally, likely reflecting efficiencies from not being the first mover.
Per TechCrunch, Anthropic intends to raise another $5b over the next two years to build a frontier model (10x more capable than today's models) and enter over a dozen industries. Their frontier model, dubbed Claude Next, intends to serve as a "next-gen algorithm for self-teaching" - a reference to an AI training technique under their internal umbrella of "Constitutional AI", which seeks to provide a way to align AI with human intentions in an ethical manner by letting systems respond to questions and perform tasks using a simple set of guiding principles. Anthropic aims to automate large portions of the economy per their latest pitch deck, and intends to include domain-specific applications while making their model “much less likely to produce harmful outputs,” “easier to converse with” and “more steerable.”
According to a General Partner at Spark Capital (current investor), "Anthropic has assembled a world-class technical team that is dedicated to building safe and capable AI systems". The company intentionally prioritizes safety over rapid advancement. This approach includes measures such as withholding certain versions of their AI system, Claude, until they have sufficient confidence in the robustness of their safety features. Despite being cautious, Anthropic has managed to develop one of the most advanced versions of AI technology, demonstrating that safety and high performance can co-exist. Anthropic's CEO, Dario Amodei, believes responsible AI must be developed with sufficient advancement within the guardrails of safety, however, it is upon governments and binding industry regulations to guide safe development of AI.
Structurally, Anthropic operates as a public benefit corporation (PBC), a type of for-profit corporation that legally prioritizes public benefit alongside profit-making. This structure gives its Board the flexibility to prioritize broader social welfare over shareholder profits, allowing it to focus on long-term safety objectives. This corporate structure is also intended to mitigate some of the structural challenges that OpenAI faces in its current conversion from a non-profit to a capped-profit entity governed by a non-profit Board - such a distinction may seem trivial to employees, however, this guides incentives and decisions, thus influencing leadership retention and company morale. Equipped with talented staff, a sufficiently advanced LLM model, and a clear thesis on building responsible AI with profit and performance in mind, Anthropic is positioned to capture a meaningful niche in the vastly expanding generative AI ecosystem.?
Similar to OpenAI, Anthropic manages a $100m Anthology Fund in partnership with key investor Menlo Ventures to enable the burgeoning ecosystem of AI-enabled development across 5 key areas. Anthropic's incentives to build on its ecosystem will expand the Claude models' potential use cases while offering the company an opportunity to "try before you buy" in order to effectively scale out its long-term strategy for novel use cases within the context of responsible AI development.
Pay to Play
As we can see, the scale to train LLMs requires tremendous computing power, which translates into a significant portion of the cost structure of these companies. In the discovery phase of technological potential, subsequent models increase in data capacity and training cost, thus no plateaus of efficiency appear to be realized as of yet. This trajectory will continue and thus requires additional infrastructure, more efficient training, higher computing power, and lower cost. No small feat!
Tokens are the building blocks for training LLMs. The model breaks text into small pieces called tokens, which are assigned numbers that computers can process. These tokens help the model learn language patterns, and they are converted into embeddings—complex numbers that capture meaning and context. This helps LLMs understand relationships between words and generate accurate responses. Training LLMs is costly because it requires powerful hardware, large datasets, and a lot of time. Methods like supervised learning (learning from labeled data), unsupervised learning (learning from raw text), and reinforcement learning (adjusting based on rewards) are used. The high cost is due to the large scale of computations and energy required for training.
Compute and infrastructure costs are significant components of AI model training and management. The vast computational needs require ample compute resources, cloud service providers, training optimization techniques, efficient architecture, and data efficiency. In addition to the incumbent big tech providers, the cloud GPU space is expanding to cater to this demand. There is an opportunity for partnerships, as is the case with strategic investments by NVIDIA (OpenAI) and Amazon AWS (Anthropic) and vertical integration through build or buy strategies, and well-funded companies appear to be sufficiently capitalized to pursue every avenue for competitive advantage.
Forecasting the Future
During 2021-2022, the majority of generative AI funding was dispersed across various applications. Multi-modal, media, interfaces, content curation, and APIs constituted the majority of deals. New applications are curated as ideation increases around the potential for generative AI. The space will likely evolve from generalized applications (chatbots as a better Google) to novel use cases across industries, functions, and cognitive capabilities. Per McKinsey's research report on the Economic Potential of Generative AI, below are a few trends identified in their study:
Per McKinsey, Marketing & Sales were outlined to have the greatest impact from AI. This will likely present itself in the form of automated or enhanced sales/marketing task support and execution throughout the entirety of the sales cycle. Currently, chatbots for customer service dominate the post-sale phase, and this could expand into the earlier phases of the sales cycle. Other functional opportunities include technical development (engineering) and customer operations. Considering the highly technical nature of engineering, this function will likely reduce reliance on humans (smaller yet more productive teams), but not entirely eliminate them due to the higher level thinking and output refinement involved in the field. Manual digital activities within supply chain, manufacturing, procurement, and compliance offer moderate opportunity and impact to reduce task workload and increase productivity. Highly subjective tasks that require creativity and reflexivity such as strategy, pricing, and HR are less likely to be made redundant, but may be enhanced through generative AI.
According to the research report, technology and banking industries will be affected the most by generative AI, and subsequently have the greatest impact on their annual revenue. Highly specialized industries or those with high SKU complexity like advanced manufacturing, electronics, and CPG appear to anticipate the lowest impact. Other highly impacted industries include pharmaceuticals, education, and telecom. Industries with a high level of creativity (media/entertainment) and specialization (insurance) appear to have a lower anticipation of impact on revenue.?
Generative AI's flagship product - the chatbot - is primarily trained in modalities encompassing logic, reasoning, linguistic understanding & generation. Media generators exploring multi-modal features such as digital art and video expand on creative abilities. As models establish core competencies in generative abilities, the next phase will involve exploration and training of social & emotional factors to mirror human empathy by modeling human consciousness. McKinsey estimates developments for most of these capabilities to initiate this decade, and mature across the next 5-15 years (vs. previous estimates of 10-50 years pre Gen-AI).
What’s Next?
Are our jobs safe? Will we get Universal Basic Income (UBI)? Will human interaction be slowly replaced with AI?
There are mixed views on the use of AI. The classic dystopian view highlights job redundancies. LinkedIn co-founder Reid Hoffman has suggested that the traditional 9-to-5 job will be obsolete by 2034, largely due to advancements in AI. Others see AI as a tremendous productivity measure and technological advancement to simplify the way we work and live.
Despite speculation (no one truly knows), one can embrace the technology, experiment with it, and see how it can work for us. I personally use both ChatGPT and Claude as my virtual staff - personal trainer, nutritionist, research analyst, strategist, life coach, therapist, and idea generator. I also enjoy iterating threads so the algorithms are tailored to my needs with context, and this has enabled a measured improvement in life satisfaction (for the low price of $20/month each!).
Like time, money, or access, Artificial Intelligence is a resource whose value depends not on its inherent morality, but rather on its effective utility. With near-infinite access to a collective superintelligence, we have a tremendous resource at our fingertips, and soon, we won't be able to imagine life without it.
"AI is not just a technology. It’s a way of amplifying our human intelligence and creativity."
– Satya Nadella (CEO of Microsoft)
"Artificial intelligence, deep learning, machine learning - whatever you’re doing if you don’t understand it - learn it. Because otherwise, you’re going to be a dinosaur within three years."
– Mark Cuban (Dallas Mavericks owner, technology investor)
"We must look at AI not as a replacement, but as an enhancement that can help us achieve far beyond our limitations."
– Ginni Rometty (former CEO of IBM)
~Nikhil Jacob
Strategy Consultant (Tech, Healthcare, Finance)
ex-BCG, ex-VC, INSEAD MBA alum
Research Sources | CB Insights - Q1 2024 AI report , CB Insights - AI 100 (2024) , Aventis Advisors - AI Valuation Multiples 2024 , AI Multiple Research , Stanford AI Index Report 2024 , McKinsey research - Generative AI , Sam Altman blog , Lifearchitect.ai , mindsDB , OpenAI blog , Anthropic AI blog , TechCrunch, Crunchbase, The Information, NY Times, Time, Wired, Financial Times, decoder, Visual Capitalist, Chartr, Sacra, pymts.com, OpenAI ChatGPT, Anthropic Claude AI
Disclaimer: This article was written by a human (myself), not by generative AI. The latter was used purely for research purposes and concept synthesis. All sources referenced.
This article was originally published on Substack
Follow my substack [Quantum Edge ] for more trends, research, and perspectives on technology and business!