In 2026, AI technology has entered a maturity phase where 'fundamentals' determine performance, beyond explosive innovation. With reinforcement learning (RL) having become the answer for everything, what matters now is not flashy secrets but the product-minded sensibility for handling data and steady refinement through environment scaling.
1. A Shift in Technical Paradigm: From Secrets to Fundamentals
In March 2026, the AI frontier landscape looks different from last year. In the past, people believed that 'secret recipes' known only to certain companies determined model performance, but that illusion is now fading. Looking at models like DeepSeek V4 and GLM 5, the direction of innovation has converged on how to perform RL (reinforcement learning) more efficiently, stably, and diversely.
Kim Sunghyun diagnoses that "the answer to the method has already been found." Whereas before the core was finding new methodologies to solve benchmarks, now it's a battle of how deeply to solve problems with the already-proven answer of RL.
"We found the answer—the answer sheet—for the method, so all we need to do is apply this answer to various problems."
Ultimately, what creates today's performance differences is not magical technology but 'fundamentals'—creating good data, having stable infrastructure, and refining models. More than research innovation, the attitude toward the product—the process of incrementally refining the model through interaction with users—has become more important than ever.
2. The Fog of Progress and the Limits of Prediction
As AI's impact expands into national and social issues, many try to predict the future. Concerns about whether developers will disappear or whether to learn AI usage now all stem from assumptions about 'what future AI will look like.' But as Professor Geoffrey Hinton's 'Fog of Progress' metaphor suggests, we can see the road right ahead but can't predict the distant future.
"The future is hard to predict. You can see very nearby points like fog on a road, but a little further ahead, it becomes hard to know."
In this fog, frontier companies advance slightly ahead with exponential visibility, while the public follows behind, experiencing the changes. What matters is not falling into the error of believing current limitations will last forever. Just as image generation AI once couldn't draw fingers properly but eventually solved it, parts of today's fog will surely clear with time.
3. Environment Scaling: The Decisive Bottleneck of AI Progress
In an era where RL has become mainstream, the only remaining technical bottleneck is 'Environment Scaling.' How complex and diverse the virtual environments in which agents can learn and operate determines model capabilities. Starting from simple function creation, we now need environments for building entire programs and even entire services.
Kim Sunghyun presents three future trajectories depending on how environment scaling unfolds:
- Pessimistic stagnation: Progress halts as the cost of building environments becomes unmanageable as task complexity grows.
- Linear progress: Steady upward progress through technical innovations that gradually reduce complexity.
- Exponential leap: When Continual Learning is solved, AI discovers and learns environments on its own—a bootstrapping scenario.
"If Continual Learning is solved, technology solves technology's problems. When that happens, we'll experience exponential progress."
Particularly in 2026, many researchers are focusing on this continual learning, suggesting a potential paradigm shift beyond simple performance improvements.
4. The Fusion of Models and Harnesses: The Product Is the Model
A characteristic of recent AI products is that the boundary between the model and its surrounding interface—the 'harness'—is collapsing. For example, tools like Anthropic's 'Claude Code' are not mere shells; they are the result of the model understanding that environment and learning tool usage from the training stage.
Context management has also evolved beyond the harness forcibly splitting and summarizing content—the model now reaches the level of judging on its own: "The context is too long now, let me summarize before moving to the next step." Through RL, the model itself has learned the ability to manage its own workflow.
Amid these changes, Kim Sunghyun's proposed strategy is 'the joy of waiting for technology.' Rather than struggling to build complex harnesses to compensate for current model limitations, the attitude is needed to believe that a more advanced model arriving in six months will naturally solve the problem, and to draw a bigger picture.
5. In Closing: A Sense of Balance Amid Uncertainty
AI in 2026 has demonstrated impact that shakes the foundations of society beyond being a simple tool. But the pace of change is so fast that no one can assert the future with certainty. Kim Sunghyun, embarking on a new journey to London, emphasizes an attitude of acknowledging uncertainty and finding balance within it rather than leaning toward pessimism or optimism.
"Rather than being frustrated, enjoying it is our answer. I think the future is too hard to predict for us to be overly optimistic or anxious."
Ultimately, we're walking through fog, but we live in an era that requires the openness to embrace technological progress, fidelity to fundamentals, and the patience to wait for the next innovation (Continual Learning, etc.).