The Artificial Investor - Issue 37: Is this the end of the current AI wave?
My name is Aris Xenofontos and I am an investor at Seaya Ventures. This is the weekly version of the Artificial Investor that covers the top AI developments of the last seven days.
This Week’s Story: OpenAI and Google realise that LLMs don’t improve further with more scale.
The most interesting story of last week was leaked news about the fact that OpenAI and Google researchers working on the next version of their respective LLMs realised that the new models are not better than the latest versions. These can be the first signs that the “AI Scaling Law” is not working anymore.
What does this mean about the latest AI wave? Are we entering another AI winter? Is this the end of the road in our pursuit to super intelligence?
If you prefer to listen to this issue, click here for the audio version.
Powered by ElevenLabs
🧱 A dead end?
So, what’s exactly happening? Through various statements and press leaks by employees from OpenAI, Google and Anthropic it appears that the top AI companies are struggling to improve the performance of their current flagship LLMs by simply scaling, i.e. increasing the volume of data and model complexity (number of parameters). Ilya Sutskever, co-founder of OpenAI, made a statement that ”results from scaling up pre-training have plateaued”
There are two graphs that illustrate this very well:
Timeline of LLMs with the number of model parameters. Here we see how researchers and AI companies have been increasing the model size from 100 million in the first GPT to the hundreds of billions in the last cutting-edge LLMs.
LLM performance (MMLU benchmark score) vs. model size (number of parameters). Here we see how scaling different models has resulted in diminishing returns (credit to Chuhan Wu’s Performance Law of Large Language Models). Note: we are not fans of the public model benchmarks, however, it is the main proxy of performance that we currently have.
If we consider the current AI wave as the linear improvement of vanilla Transformer models by increasing their size - which has also translated into larger training datasets (most of the Web’s text has been consumed) and billions of dollars spent in chips and datacentres - then perhaps we are indeed very close to the end of the technical innovation of that wave. Does this mean the current AI wave is finished? Definitely not. However, we do see two serious consequences: i) focus on utility, and ii) the return of the lab/ startup garage.
🔧 Finally some utility
We have written in the past that most of the value in the current AI wave has been concentrated in the infrastructure layer. In last June’s issue 24, the size and vendor landscape of the Gen AI market, we estimated the application layer was worth around 5 billion dollars compared to the infrastructure layer that was worth nearly 70 billion dollars. This ratio is something that we also witnessed in the early phases of the previous Tech wave, the Cloud, which eventually was inverted and the value shifted to the application layer. Sequoia has also written about the 600 billion dollar gap that exists in AI currently, referring to the fact that the amount spent in AI models, which is a cost item in Tech company P&Ls, should correspond to 600 billion dollars of revenues of AI application vendors, which currently doesn’t exist.
This is obvious from a user perspective too. Amazon’s Alexa and Apple’s Siri, two years after the release of Chat GPT-3, still cannot understand when I ask them the question “is it cold out there?”; I need to mention a precise set of words in a certain order to get the weather forecast from them. Microsoft’s much-hyped Office copilots applications have been criticised by users and usage has dropped; as is the case with Salesforce’s recent AI functionality. In many cases, we have even seen user experience worsening in applications that launched Gen AI features, due to the technology’s inherent unreliability (the so-called “hallucinations”) and slow speed.
Where is the killer AI app that I can talk to in the morning on the way to work, get an audio summary of my emails since the previous evening, reply to the most important ones via voice, reschedule the next appointment automatically because I left the house late, cancel my dinner because my wellness app indicated that I slept bad last night and need some rest, get a voice recommendation to skip bread for breakfast because I had enough carbs the night before, etc? The underlying technology exists; will we finally see a focus on building useful AI applications that transform our lives? We hope so.
🔬 Back to the lab (or garage)
Ilya Sutskever also stated that “the 2010s were the age of scaling, now we're back in the age of wonder and discovery once again.” In order to understand better where we are heading, it is important to understand how we got here. If every data scientist could see the diminishing returns of scaling the vanilla Transformer architecture, why have we spent billions of dollars and hundreds of thousands of precious scientist working hours on training bigger LLMs?
Devansh wrote a great piece last week about the incentives behind the model scaling frenzy, explaining how the predictability of AI performance improvement has been very convenient for the various stakeholders.
BigTech company CEOs: The AI race became about chip and datacentre capex, which is great for cash-rich BigTech companies. Plus, there is vendor lock-in in large infrastructure deals.
BigTech’s technology leadership: It’s way easier to throw computers at problems than rebuild everything from scratch.
Data scientist incentives: It has become more likely for a data scientist to produce results they can tangibly demonstrate within their performance review cycle, and in a predictable manner.
Data scientist careers: Scaling reliably improves benchmarks, which makes it a very good match for an academic/ research environment where publications are a must.
Scale-ups raising funds: It’s easy to convince investors that “bigger = better”, which allows people to shift attention away from competitive moats and path to profitability.
So, now that we are in this situation, it is fair to assume that the solution will not come from the hyperscalers, it will rather come from the emerging AI companies. The next big leap in AI will likely take place in a small lab somewhere or even a startup garage. The incentives are simply not there for the hyperscalers or very large AI companies, like OpenAI. And given the plateau that we are hitting on LLM performance, it looks like the playing field is levelling out again.
🔚 Is this the end?
So, does this mean that we have hit a brick wall on the road to super intelligence? We don’t think so. Nevertheless, it is has become obvious for some time now that Transformers (the core technology behind LLMs and underpinning the latest AI wave) have some serious limitations:
They are very expensive (in terms of data and compute) both to train and run. This is because they involve very complex calculations that scale exponentially with the model size, thus precision.
They are unreliable, they hallucinate and come up with some answers, due to being probabilistic, which means that occasionally they can “pick” the alternative answer with 0.5% probability of being correct.
They are black boxes, as we cannot tell how they come up with a response and why.
They cannot address complex problems, such as mathematics, geometry, complex reasoning, planning, etc.
Their skills are largely static after training, which means that they don’t improve substantially with use.
They cannot handle continuous data problems, such as weather forecasting or stock prices (also called time series).
Nevertheless, the end of the Vanilla Transformer S-curve doesn’t mean the end of the entire innovation wave. It may simply mean the creation of a new S-curve; as was the case with other innovation waves. Where could the next AI S-curve come from?
We will look into the different areas of research and try to be as short and simple as possible.
🤖 Path to super intelligence
Cost reduction
One area of research has focused on incremental improvements in Transformers to make them more efficient, by reducing either i) the number of calculations they do, or ii) their size while trading off precision.
Reducing the number of calculations
Transformers have a mechanism called self-attention, which helps the model understand a word by comparing it to every other word in a given text. This means that in a sentence with 20 words, each word is compared with another 19 words. As you can imagine, the more text you feed the model, the more calculations you have (in an exponential way). Imagine being in a room with hundreds of people and trying to listen to everyone at once. It’s overwhelming and inefficient!
There are several ways the number of Transformer calculations are being reduced. One of them is making each word only pay attention to a smaller group of words, such as Sparse Transformers. This is like listening to only a few important conversations instead of everyone in the room, such as the people right next to you or just the key speakers. Other ways to reduce calculations include Linear Transformers and MatMulFree models, where the architecture is simplified. This is short of the equivalent of drawing only straight lines to approximate a trend shown in a graph of dots as opposed to drawing curves, which means the maths are simpler but precision is lower.
Reducing Transformer size
Another field of research addressing the cost of Transformers is related to efforts that shrink a model while trying to maintain its precision. One of the most popular techniques is Quantization, which reduces the precision of numbers used in a model, like changing from 32 bits (high precision) to 8 bits (lower precision), which makes computations faster and uses less memory. This is the equivalent of switching from a heavy box of crayons with 256 colours to a smaller box of 16 colours which is lighter and easier to carry. This still lets you draw good pictures, although not as good as the larger box. Another popular technique is Distilling, in which a large "teacher" AI model teaches a smaller "student" model, helping it learn what’s most important. The smaller model can make predictions more quickly than the large one, while it uses less memory and energy making it easier to deploy on devices like smartphones.
Reasoning and explainability
Another research area is focused on improving explainability and the reasoning capabilities of AI models. One of these approaches is called Chain of Thought (CoT), where instead of producing a direct answer to a question or task, the model generates intermediate steps that lead logically to the final answer.
For example, if you ask a non-CoT model, such as OpenAI’s Chat GPT-4,"What is the sum of the first 10 positive integers?", the answer will be "55". If you ask a COT model, such as OpenAI’s 01, the same question, the answer will be: "The first 10 positive integers are 1, 2, 3, ..., 10."
"To find the sum, we can use the formula for the sum of an arithmetic series: Sum = 𝑛 × ( 𝑛 +1)/2 , where 𝑛 is the number of terms. Here, 𝑛 = 10, so the Sum = 10 x (10+1) / 2 = 10 * 11 / 2 = 55. The sum of the first 10 positive integers is 55."
This makes the model more interpretable and capable of answering more complex questions than a standard Transformer model. Other similar approaches include Self-consistency Decoding and Tree of Thought. In the latter, the model explores several possible reasoning paths in parallel, expanding and cutting branches based on their potential relevance and correctness. This is like a decision tree used by humans to analyse a situation and form a plan of action.
Neuro-symbolic AI is another approach that addresses both explainability and reasoning. The term sounds complex, but it’s actually a simple concept. It’s about combining good old AI that is based on a set of fixed human-written rules (also called symbolic AI) with deep learning, which is probabilistic and recognises patterns. Therefore, this method also addresses the unreliability challenge that Transformers have.
Talking about model unreliability, Knowledge Graphs have also been combined recently with Transformers to improve their factual consistency. A Knowledge Graph is is a structured representation of knowledge where entities (nodes) are connected by relationships (edges). For example, a knowledge graph might contain facts like "Paris is the capital of France."
For example, a Knowledge Graph might contain facts about famous scientists and their field of expertise, such as "Albert Einstein is a physicist". Then, information from a knowledge graph can be embedded into a Transformer model in a way it understands. When the Transformer processes text like "Einstein developed the theory of relativity", it can use the knowledge graph data to understand who Einstein is and why the theory is significant. This would make the Transformer’s answers more accurate and fact-based, while making it easier to trace the “reasoning path” of the model; back to the “black box” challenge!
Going further into the complexity of problems to be solved, we find the combination of World Models and Transformers. World Models are an AI concept that refers to building a model of the environment or "world" in which an AI model operates. For example, in a video game, a World Model would learn how characters, objects, and physics interact over time based on the player’s actions. In such an example, a Transformer model would be capable of imagination-based planning, i.e. simulating different paths or outcomes before choosing an optimal action, leading to more informed decision-making. This addresses the fundamental limitation of Transformers: inability to plan.
Continuous learning
Continual Learning models is another area of research that focuses on tackling the Transformer’s limitation of being static, i.e. not improving post training. Now, one would argue that Chat GPT is a Transformer and it does improve post training, because it collects user feedback through “thumbs up/down” buttons and by presenting users with two alternative answers asking them to choose the one they prefer. Certainly, this is a sample of Reinforcement Learning from Human Feedback (RLHF), however, this is very sporadic (covers a small part of human-model interactions) and allegedly has had little or even adverse effect on Chat GPT’s performance. Beyond RLHF techniques, Continuous Learning includes methods such as Progressive Neural Networks and Memory-Augmented Neural Networks (MANNs). The latter combine traditional neural networks with an external memory that stores results from user interactions and recalls them to improve future performance.
🔮 So, wrapping up, we are probably very close to the end of the a technological S-curve of the plain-vanilla Transformers. However, this doesn’t mean we are at the end of the current AI wave.
First of all, it’s about time that we see companies focusing on the current User Application S-curve, which remains at early stages. Also, new technological S-curves have kicked off during the last couple of years, from Sparse Transformers and Tree of Thought, to Neuro-Symbolic AI and Continuous Learning models. Which one(s) of those will form the basis of the next technological S-curve is hard to predict.
What is perhaps easier to predict is that, taking into account how incentives are currently structured in the AI ecosystem, the next technological innovation will likely come from a small lab or startup garage. Just like it happened when OpenAI had 52 employees and released the very first GPT model in 2018.
See you next week for more AI insights.
So helpful
Excellent!