How to view AI at the end of 2025.
Does a bubble exist? Will it burst?
At this point in time, I still firmly believe that an investment bubble is significantly present, and valuations are markedly high.
But will it burst? This is a question I find difficult to answer. If "bursting" means the AI narrative is proven false, then everything I've experienced over the past three-plus years tells me there is a 99% probability that this will not happen.
However, if it means a significant correction in expectations and valuations—for example, tech companies' CAPEX becoming marginally more cautious, Nvidia failing to sell as many GPUs as expected, data center expansion speed being unable to keep up, major adjustments to financial return calculations due to series of delays, terminal revenue growth from AI not being fast enough, or ultimately manifesting as liquidity issues.
I believe the probability of these "examples" manifesting in 2026 is quite high, potentially exceeding one sigma.
What about the models?
Models will certainly continue to improve, but the greater likelihood is that within the "human world," the marginal efficiency gains brought by generative AI are rapidly declining. Fundamentally, this is due to a contradiction: a mismatch between cost and revenue. With the empowerment of Agents, investing more compute can achieve better results and be applied to more scenarios, but it also means higher compute costs. However, because models are essentially data, their generated results are actually a "synthetic" imitation of human digital output. Their actual quality reflects how much data the model has "compressed."
This wasn't very obvious during the text-output era. It only became clearer recently when the "nano banana pro" model began outputting stunning infographics. It turns out Gemini-3's real lead is data: every detail, shape, color, and logo in those infographics corresponds precisely to the "real world."
When we say the Scaling Law is still working, it naturally refers to the fact that as long as high-quality data continues to increase, model performance will continue to improve.
Of course, the know-how here is "data magic": which data to use, how to generate data at scale, and how to "compress" more data through model architecture adjustments...
Compute --> Data --> Ecosystem: clearly, different players are competing across different dimensions.
However, there are still two misconceptions: one regarding Anthropic's Claude, and the other regarding synthetic data.
Claude remains the best programming model for now. It doesn't appear to have a data advantage, nor does it have a compute advantage in the competition, yet its coding ability remains the strongest.
In fact, perhaps counter-intuitively, compared to the vast expanse of "world knowledge," programming data is a highly standardized "small data" set. While Anthropic certainly has its own unique "magic," it likely isn't a secret to other major competitors. While Claude possesses stronger programming capabilities, it does "introduce" more hallucinations, whereas Gemini and GPT models have more baggage and must make certain trade-offs.
As for synthetic data, the explanation is relatively simple: its primary purpose is to reinforce specific data capabilities, but it cannot create something from nothing. For the foreseeable future, this is unlikely to change. This means the limit of human data is the ceiling for models.