In 2022, artificial intelligence felt like it leapt forward overnight. New tools appeared every week. Capabilities that once seemed academic suddenly became accessible to anyone with a browser. Investors rushed in, with global AI investment at about $91.9 billion in 2022. Businesses scrambled to adapt. Commentators predicted everything from instant automation of knowledge work to artificial general intelligence within a few years.
Since then, something curious has happened. Spending on AI has exploded. Infrastructure investment has surged. Yet the improvements that most users see feel incremental rather than revolutionary. New models are better, faster, and more reliable. But the sense of exponential surprise has slowed.
This pattern is not unique to AI. It is the normal rhythm of innovation.
The breakthrough phase
Technological revolutions often begin with a foundational breakthrough. In the case of modern AI, that breakthrough was the transformer architecture, introduced in 2017 and refined in the years that followed. Whereas previous architectures could process only a few sentences, transformers can efficiently process entire books. They unlocked the ability to train large language models that can understand and generate text, code, and images at scale.
When the first widely accessible generative AI tools reached the public in 2022, it felt like the future had arrived all at once. Years of research suddenly translated into products that anyone could use. The jump in capability between 2019 and 2022 was dramatic. For many people, it felt exponential.
This phase tends to create inflated expectations. When cars were invented, people immediately started predicting flying cars. Over a century later, we still drive on roads. Aviation exists, of course, but the seamless fusion of road and sky remains largely conceptual. Early breakthroughs trigger imagination faster than reality can keep up.
AI followed the same path. The first wave of generative models was so impressive that it encouraged predictions of near-term superintelligence and fully autonomous businesses. Those predictions assumed that progress would continue at the same visible pace.
History suggests otherwise.
The optimization phase
After a breakthrough comes a long period of refinement. During this stage, progress continues but becomes less dramatic to the outside world. Engineers focus on efficiency, reliability, safety, and integration. Systems become more usable and more scalable. Costs decrease. Infrastructure expands.
This is where AI sits today. Companies are spending billions on compute, chips, and data centers. Models are becoming more capable. But these gains are often incremental. A model might be 20% more accurate or twice as fast.
Those improvements matter deeply for businesses deploying AI at scale, but they don’t always create the same sense of wonder that accompanied the first wave of generative tools.
Crucially, this phase cannot be rushed simply by increasing budgets. Money can accelerate engineering and infrastructure. It cannot guarantee conceptual breakthroughs. The transformer architecture emerged from years of research and experimentation. The next revolutionary idea will likely emerge the same way.
Innovation at this level depends on new ways of thinking, not just feeding the system more and more training data.
Await the next leap
Every major technological era has its defining breakthroughs separated by long stretches of incremental progress. Electricity, aviation, computing, and the internet all followed this pattern. AI’s next leap will come when researchers discover a new architecture, paradigm, or method that unlocks capabilities beyond current systems.
If history is any guide, that moment will arrive later than most forecasts suggest. Breakthroughs rarely happen on schedule. They emerge from sustained research, unexpected insights, and sometimes pure luck.
This does not mean progress has stalled. It means we are in the middle of the build-out phase that turns a breakthrough into an industry.
What businesses should do now
For companies, the current moment is not a pause. It is an opportunity. Waiting for the next revolutionary advancement is a mistake. The tools available today are already powerful enough to transform workflows, reduce costs, and create new products.
Instead of standing still, organizations should focus on integrating existing AI capabilities into daily operations. For example:
- Automate repetitive tasks and start from low-hanging fruit like customer support (especially email-based) or report generation.
- Develop what we call “evals” or test cases that allow you to quantifiably test different tools and objectively determine which one performs best for each use case.
When the next major leap arrives, the companies that have already adapted will be positioned to capitalize on it immediately. Those who waited for perfection will have to start from scratch.
AI innovation goes fast and then slow. The fast phase captures attention. The slow phase creates lasting value.
Komninos Chatzipapas is the founder of Numus Arbitration Center.