Beyond The Model
Throughout history, technological innovations have improved existing systems, but only a small subset can be described as general-purpose technologies; technologies that do not merely optimize a single task but reshape entire economic and social systems over time.
These technologies tend to follow a distinct pattern. They reduce the cost of key activities, expand what is computationally or physically possible, and enable new layers of systems that were previously infeasible. Electricity, the internet, and computing itself are often cited as examples of this category.
Artificial intelligence is increasingly being discussed in the same context.
What makes AI particularly important is the underlying structure that determines how it scales. Across nearly all modern AI systems, from language models to recommendation engines, progress is shaped by three interacting components:
Algorithms,
Data, and
Computing power.
What makes the best of models perform exceptionally is simply the interaction between these components. Remove any one of them, and progress slows down significantly.
How AI Learns
The first part of the triad is algorithms. These the learning methods AI uses to improve its performance. Some learn from labelled examples where the correct answers are already provided. Others find patterns without being explicitly told what to look for, while some others do so through trial and error by receiving rewards for good actions. This same approach has helped AI systems defeat human champions in games like Chess and even strategy games such as StarCraft.
A good example of this is what happens on social media platforms. Every time you use X, TikTok, Instagram, or YouTube, the platform is constantly learning from your behaviour. It pays attention to the posts you engage with, the videos you watch for longer, and even the content you scroll past. Over time, it begins to identify patterns and gets better at predicting, with a keen level of precision, what you are most likely to interact with. The more you use these platforms, the more their algorithms learn about your preferences.
The second part is data. Most people refer to data as the new oil, and for a long time, I thought about it that way too.
But hey, one of the most practical lessons to note is that having more data can often beat having a better algorithm. A simple model trained on a large amount of quality data will frequently outperform a more sophisticated model trained on limited data. Larger datasets also reduce the influence of unusual examples and help models generalize better.
To get the most out of it, the data has to be representative. If important groups or situations are missing from the training data, the model might not be effective when it comes to producing reliable outcomes in real-world scenarios.
Even more concerning is biased data.
Something happened at Amazon some years ago. The company built an AI system to help with hiring and trained it using past hiring data. The problem was that much of the data came from hiring decisions that favoured male candidates. When the system was eventually used, it showed bias against women because that was the pattern it had learned from the data.
If the correct answers in the training data are already biased, the model doesn’t exactly remove those biases. It simply learns them, iterates on them, and can even amplify them over time.
The final piece of the triad is computing power, or compute. This is the raw processing power that allows modern AI systems to train and operate, usually through high-end GPUs and specialized AI chips.
A study showed that the amount of compute used by leading AI systems increased by roughly 300,000 times between 2012 and 2018, and 3-5 times per year, over the last few years, depending on the definition of compute. Many of the breakthroughs we associate with modern AI were not simply the result of better ideas or more data, but also the result of having enough computational power to make those ideas work at scale.
What I also found particularly interesting was the evaluation of the importance of each component. Data may be somewhat overhyped because newer techniques, including synthetic data and methods that require fewer labelled examples, are reducing its relative importance. Algorithms remain extremely valuable, but attracting and retaining top AI talent remains a challenge for many countries and organizations.
The strongest argument, however, is that compute may be the most undervalued part of the entire discussion.
Scaling Intelligence
When we’re talking about advancing or scaling AI, most people’s minds don’t exactly go to compute as the first factor that can influence it considerably. We can see the applications, the products, and the results, but we rarely think about the infrastructure behind them.
Yet computing is increasingly becoming one of the biggest bottlenecks and competitive advantages in AI.
A good example is Anthropic. Over the last year, the company has experienced significant growth in revenue, and even within recent months, the pace of growth has accelerated dramatically. As companies like this grow, more resources go into hiring talent, training larger models, and acquiring more compute. If you’ve been paying attention to the industry, you’ve probably noticed the pace of model releases increasing as well. That acceleration isn’t driven by algorithms alone. Compute plays a major role in making it possible.
This is why conversations around AI today extend far beyond models and applications. They increasingly involve talent, data access, semiconductor supply chains, GPU availability, and national infrastructure.
The AI race is no longer just about building smarter systems. It’s also about who has access to the resources needed to train and run them.
So if you really want to understand what drives the progress and scaling of AI today, start with algorithms, then data, and finally compute.