Distillation, GEO, Atmosphere Programming. How many of the "Top 10 AI Buzzwords of 2025" can you understand?

In 2025, the development in the field of AI was dizzying, with a series of new concepts emerging continuously and reshaping the industry landscape. MIT Technology Review selected the top ten annual AI buzzwords. Understanding these words might mean understanding how we were changed by AI in 2025.

1. Vibe Coding

Thirty years ago, Steve Jobs, the co - founder of Apple, proposed that everyone should learn programming. Today, programming is being redefined. "Vibe Coding" proposed by Andrej Karpathy, the co - founder of OpenAI, is not a new language but a new way: people only need to express goals and logic in natural language, and the specific code is automatically completed by AI.

In Vibe Coding, developers only need to tell AI what kind of application they want to create, what functions are needed, and what the overall experience should be like; AI is responsible for generating code, adjusting details, and continuously iterating through repeated dialogues.

2. Reasoning Model

In 2025, "reasoning" was repeatedly mentioned and gradually became a core term in AI discussions. Behind this popularity is the rise of reasoning models: these large language models start to handle more complex problems through multi - step decomposition and continuous deduction.

After OpenAI released the o1 and o3 series of reasoning models, DeepSeek quickly followed up. Now, mainstream chatbots have introduced reasoning technology and reached the level of top human experts in mathematics and programming competitions. However, the question of whether AI truly has the "reasoning" ability has once again triggered people's thinking about the essence of intelligence.

3. World Model

Large language models are good at "talking" but do not really understand the world. They can generate fluent text but often lack basic common sense and even make mistakes in simple physical problems. To make up for this shortcoming, AI research is turning to a key direction - the World Model: enabling AI not only to learn languages but also to understand the causal relationships, physical laws, and time evolution of the real world, so as to judge what is reasonable and predict what will happen next.

Whether it is Genie 3 from Google DeepMind, Marble from Fei - Fei Li's team, or the new research direction that Yann LeCun focused on after leaving Meta, in essence, they are all trying to let AI master the basic laws of the world's operation through simulating and making trial - and - error predictions of video evolution or building virtual environments.

4. Ultra - large - scale Data Center

As the demand for computing power in AI surges, technology giants are building "super data centers" dedicated to AI on an unprecedented scale. For example, the "Stargate" project jointly promoted by OpenAI and the US government plans to invest $500 billion to build the largest - scale data center network in the history of the United States.

However, these behemoths have also raised many social concerns: their astonishing energy consumption may drive up the electricity bills of local residents, they are difficult to rely entirely on clean energy in the short term, and they can create limited long - term jobs for the community. This is becoming an increasingly prominent contradiction between the rapid development of technology and the interests of the people's livelihood.

5. Bubble

Currently, AI is becoming one of the most crowded tracks for capital. The valuations of companies represented by OpenAI and Anthropic continue to rise, but the reality is that most of them are still in a stage of high investment and have not established a stable profit model. Investors are betting that AI will start the next wave of wealth, but the actual value that this technology can ultimately release remains to be verified by time.

However, compared with the Internet bubble back then, today's top AI companies have rapid revenue growth, and they are supported by technology giants like Microsoft and Google with strong financial strength and solid capabilities, which provides them with stable support.

6. Agent

"Agent" might be the hottest and most vaguely - defined concept in the AI circle at present. Although every company is promoting that its AI can complete tasks autonomously like a "smart assistant", there is still no unified standard in the entire industry for what really constitutes agent behavior. Even though current AI still has difficulty working stably and reliably in complex and changeable environments, this does not prevent "agent" from becoming one of the most popular labels in product promotion.

7. Distillation

At the beginning of 2025, the R1 model released by DeepSeek showed us the ingenuity of "distillation" technology. It enables small models to learn the essence of large models and achieve performance close to that of top - level models at a very low cost. This makes the industry realize that building a powerful AI model does not necessarily rely on piling up expensive computing power. Efficient algorithm design can also bring new possibilities.

8. AI Junk

"AI Junk" has become a well - known term among the public, specifically referring to low - quality AI content produced in batches to attract traffic. Now, "junk" has evolved into a suffix used to describe various things that lack substance and are dull, such as "work junk" and "social junk". Behind this reflects people's general reflection on the quality and authenticity of content in the AI era.

9. Physical Intelligence

Compared with language and reasoning abilities, AI's ability to act in the real world is still a major shortcoming. Although robots can learn faster in specific tasks now and the simulation of autonomous driving has become more and more realistic, many products labeled as "smart home assistants" actually still need human remote control.

To improve this ability, some robot companies have started to pay ordinary people to collect videos of doing housework. This shows that there is still a long way to go to enable AI to truly understand and adapt to the physical world we live in.

10. GEO

As AI increasingly provides direct answers, the way people obtain information is changing. Traditional Search Engine Optimization (SEO) is giving way to a new approach - Generation Engine Optimization (GEO).

In the past, brands and content creators competed for the ranking of web pages in search results; now, the question has become: when users directly ask AI a question, will AI mention your brand, your views, or even quote your content in the answer. Under this rule, content providers either learn to be quoted and absorbed by AI or may gradually disappear from people's view.

This article is from "Tencent Technology", author: Jin Lu, published by 36Kr with authorization.