Executive Summary

A recent remark by Groq founder Jonathan Ross raises an important question. If models begin to improve the quality of their own learning signals, then the AI growth logic we have become familiar with may no longer follow the same path of diminishing returns.

This article does not ask whether Ross’s claim should be accepted at face value. It asks whether the idea behind it is already supported by a set of meaningful weak signals. From Google DeepMind’s continued push into reasoning, to OpenAI’s gradual formalization of high quality feedback, to NVIDIA’s inclusion of post-training, test-time scaling, and agentic scaling in its next generation platform narrative, these developments suggest that AI progress may no longer be only a story of static pretraining and scale expansion.

Still, weak signals do not mean a trend has already been established. At this stage, we still lack cross-task and repeatable proof. We also lack public evidence that a long-cycle flywheel has truly formed. At the same time, issues such as reward hacking, misalignment risk, and the real world challenge of deployment and demand absorption remain unresolved.

The core judgment of this article is therefore a cautious one. Ross’s argument is not an unfounded exaggeration, but neither is it enough to support the view that AI has already entered a new phase of self-accelerating growth. What matters now is whether these weak signals will gradually connect into a structure that can be repeatedly observed and validated. If that happens, then AI’s next growth phase may indeed turn out to be faster than many currently expect.

Introduction

A recent remark by Groq founder Jonathan Ross points to a question worth examining. If models begin to improve the quality of their own learning signals, then the AI growth logic we have become familiar with may no longer be enough.

What he is suggesting, broadly speaking, is that many people look at scaling laws and come away with the impression that AI progress is beginning to slow. Put simply, scaling laws tell us that models generally become stronger with more data, more compute, and larger scale, but each additional improvement often requires disproportionately greater investment. As a result, when much larger amounts of resources produce only relatively limited gains, the sense of diminishing returns becomes more pronounced. Some then go a step further and conclude that AI may not be far from some kind of limit.

What Ross is trying to point out, however, is that this way of thinking may overlook an important assumption. If models begin to improve the quality of their own learning signals, then AI progress may no longer move only along the same curve of diminishing returns. It may begin to develop a different source of growth momentum.

That is clearly a very optimistic framing. But what makes it worth paying attention to is not simply how optimistic it sounds. It is that it points to a deeper question. The AI growth logic we have become familiar with may no longer be sufficient to explain the changes that could come next.

What the Existing Logic May Be Overlooking

What Ross is really challenging is one of the assumptions behind this familiar growth logic. Training data is usually treated as a fixed external condition. The model is expected to learn more effectively from the material it is given, not to improve the quality of that material itself. If that assumption begins to loosen, then the way we have understood AI progress may also need to change.

If stronger models can in turn help generate better learning signals, then AI progress may no longer depend only on more data, more compute, and larger scale. It may begin to take the form of a process in which model capability and data quality reinforce each other. That is also why some recent reinforcement learning approaches have drawn particular attention.

In other words, the real question is not whether AI can continue to improve. It is whether the pace of that improvement could turn out to be faster than we originally expected.

Meaningful Weak Signals

At this stage, I do not think this direction is purely speculative. Some meaningful weak signals have already begun to emerge.

Progress in reasoning is no longer just a concept

Advances in model reasoning are no longer only an abstract research idea. They are beginning to produce clearer results. Google DeepMind offers one example. Across the Gemini series, reasoning has been pushed into a more central role, from explicitly framing the model as a thinking model to introducing Deep Think, which places greater emphasis on the reasoning process itself. This suggests that model improvement is no longer driven only by larger scale pretraining, but is beginning to extend into more complex reasoning and feedback mechanisms. At a minimum, this indicates that AI capability gains are no longer only a story of static pretraining. When models can continue improving on certain tasks through more complex reasoning and feedback, the older idea of growth driven only by more data and more compute begins to loosen. Cases such as DeepSeek-R1 and Kimi k1.5, which place even more explicit emphasis on reinforcement learning, make this shift easier to see.

High quality feedback is becoming formalized

High quality feedback is moving from a research concept toward a formalized process. OpenAI has incorporated reinforcement fine-tuning into its official documentation, explicitly requiring training tasks to be evaluated through graders and emphasizing that this approach is especially suited to clear and verifiable tasks. The importance of this shift is that the key variable in model improvement is no longer simply the amount of data. It is increasingly about whether models can continue to receive effective and evaluable learning signals. In other words, data quality is no longer just a fixed condition outside the model. It is beginning to become part of training system design itself.

The industry’s language is changing as well

The language the industry uses to describe AI growth is also beginning to change. NVIDIA is one example. In its official framing of the Vera Rubin platform, the focus is no longer limited to pretraining. It explicitly incorporates pretraining, post-training, test-time scaling, and agentic scaling into the design logic. This at least suggests that key companies across the industry now recognize that AI growth is no longer a single-path story. It is moving toward a more multi-stage and more complex form of expansion.

These signals did not suddenly appear only recently. On the contrary, they had already begun to emerge in 2025, and by March 2026 some of them were still being extended, formalized, and written into the narratives surrounding new platforms and models. Ross’s recent remarks are therefore less a brand new discovery than an effort to pull together signals that had previously appeared in a more scattered form, and to frame them in a sharper and more optimistic narrative.

Weak Signals Do Not Mean a Trend Has Already Formed

Still, this is precisely where the main uncertainty remains. These weak signals are worth paying attention to, but at this stage they remain only early indications. They are not enough to tell us that a new growth trend has already taken hold.

First, we still lack cross-task and repeatable proof. The strongest results so far remain concentrated in mathematics, coding, and other verifiable reasoning tasks. These tasks are naturally easier to score and easier to evaluate for answer quality. As of March 2026, OpenAI still lists reinforcement fine-tuning as a supported fine-tuning method, which shows that this path has not disappeared. At the same time, it also preserves an important limitation, namely that this method is especially well suited to tasks that are clear and verifiable. That suggests the new path is real, but it still cannot be naturally generalized across the full space of AI capability.

Second, we still lack proof that a long-cycle flywheel has actually formed. For Ross’s argument to truly hold, it is not enough to show that reinforcement learning works. It would require showing that this loop can remain stable across many rounds. Better models would need to produce better data, which would then train even better models. Cases such as Kimi k1.5 and DeepSeek-R1 are closer to showing that this path is possible than to showing that it has already become a stable self-accelerating mechanism.

Third, we still lack proof that misalignment risk has been meaningfully resolved. Reinforcement learning has always carried a basic problem. A model may become better at getting high scores without becoming better at understanding. Anthropic’s 2025 research on faithful reasoning in reasoning models already pointed to this risk, and by 2026 reward hacking was still an issue being actively evaluated in its system card. That suggests this layer of risk has not disappeared simply because models have become more capable. If this problem is not addressed, then what appears to be better data may simply be data that is better at satisfying the scoring system, not data that actually carries more valuable learning signals.

Finally, and most importantly, even if the way capability improves really is beginning to change, that does not mean industrial reality will automatically follow. Stronger models do not automatically create demand. A partial shift in technical constraints does not mean that deployment costs, workflow reorganization, verification burdens, and business model problems will disappear as well. In light of how NVIDIA describes Vera Rubin, a more reasonable interpretation may not be that constraints are vanishing, but that AI expansion is now simultaneously pulling on more layers of demand across pretraining, post-training, inference-time scaling, and agent systems. In many cases, the hardest part is not the model itself, but how technical capability is absorbed into the real world.

For weak signals to become a structural shift, we still need more evidence.

Conclusion

Jonathan Ross’s argument is not an unfounded exaggeration. It does point to an important variable that deserves attention, namely the possibility that model capability, data quality, and feedback mechanisms are beginning to interact in ways that were previously underappreciated.

If these weak signals gradually connect into a real trend, the most immediate implication may be that AI could enter a phase of growth that is faster than previously expected. The point is not that constraints suddenly disappear. It is that the source of model improvement may no longer come only from pretraining scale expansion, but also from high quality feedback accumulated after training, during deployment, and throughout real world use.

If that happens, then the first thing that would change is how we understand AI growth itself. For a long time, AI progress has largely been associated with more data, more compute, and larger scale. But if models begin to improve the quality of their own learning signals, then growth would no longer be only a matter of scaling on top of existing data. It could become a process in which feedback, revision, and reinforcement continue during operation, deployment, and use. In that case, AI growth may not only become larger in scale, but also faster and more self-reinforcing.

For now, however, these signals remain closer to meaningful possibilities than to confirmed reality. What matters most is whether these weak signals will gradually connect into a structure that can be repeatedly observed and validated, and whether that structure will eventually become a clear trend.

Note: AI tools were used both to refine clarity and flow in writing, and as part of the research methodology (semantic analysis). All interpretations and perspectives expressed are entirely my own.