As someone who has been online for more than twenty years, the internet has felt a little strange since the new year.
The first obvious shift is that, on a global scale, the lower bound for large language models has been raised by DeepSeek V3 and R1 to a level that is clearly above the average person. A lot of people like to throw weird questions at different models and rank them, but that misses the point. In practice, what matters is the floor. If a tool can solve 80% of your problems with 20% of the effort, then it is already useful.
DeepSeek's current models, whether official, third-party hosted, or distilled into smaller reasoning models from other open-source systems, have all crossed that line into genuine usability. In truth, DeepSeek V2's coder model was already among the best open-source options for local code completion. I also feel a little sorry for Alibaba's Qwen line. Before DeepSeek V3 arrived near the end of last year, Qwen and the Llama family were firmly in the top tier of open-source language models. They still are. But DeepSeek's cost profile is so startling that it changes the whole conversation.
So if you have been following open-source LLMs closely over the past few years, DeepSeek's performance is not that shocking in itself. What is striking is how transparent it has been about engineering optimization. That made it break out far beyond the usual circle. And that breakout matters, because it brought in a lot of people who had never seriously used a language model before. Many of them saw a model's reasoning process for the first time. Many saw people in the West seriously discussing a Chinese model for the first time. More importantly, many realized for the first time that large language models can be used to narrow the information gap between people.
The last widely shared democratizing technology was the search engine. But search engines only gave you links; the real work of researching still had to be done by a person. This generation may be very different. Google, OpenAI, and Perplexity now all offer a feature called deep research in their paid products. You give it a question, it gathers information from the web on its own, then compiles a long report. In my view, the quality of those reports is far above the average internet user's level.
There are open-source substitutes now too. At bottom, it is just RAG paired with a language model, but as a product form it is already becoming fairly mature. What matters now is labeling AI-generated content published online. Language models are probabilistic by nature: they tend to produce the view most widely supported across available text. If the web gradually fills with AI-generated reports that merely reflect users' prior opinions, then given enough time, repetition itself starts manufacturing truth.
When I use language models, I often compare results with search turned on and turned off. If the topic is not highly time-sensitive, reasoning models alone are usually reliable enough. Sometimes enabling search actually makes things worse, because it pulls in piles of junk pages. This is especially noticeable in Chinese. A lot of Chinese websites have pushed search engine optimization far past the point of dignity. That too is a sign of the times.
It also makes me wonder how these model search features are implemented. Are they calling external search engine APIs, or running their own retrieval systems? From the user experience, it feels more like an API layer. And if this kind of functionality becomes standard for academic papers and journals, then retrospective literature reviews may barely matter anymore. At this point, when reviewing papers, I almost feel that review articles no longer need reviewing. If I need one, I can just generate it myself.
Another interesting post-holiday phenomenon is that social platforms are now full of screenshots of DeepSeek's answers. In the previous cycle, the main thing people shared was screenshots of Zhihu answers. That change says a lot. It means large language models are becoming a new kind of knowledge authority. In the old internet, that would have been hard to imagine, because if an answer came from a person, there was always a person to attack. But when the answer appears to come from a distilled body of high-quality human language, that kind of attack loses force.
That said, I often ask the same question to multiple models, and on many topics their answers do show different leanings. Maybe some of that comes from temperature settings, but more likely it reflects differences in training data. I suspect those differences will shrink over time.
What stands out even more is the kind of questions people are sharing. A lot of them used to require certified professionals to answer: legal disputes, medical guidance, personal financial planning. Since these models have already passed many of the same qualification exams, consulting them really can reduce the everyday cost of living for ordinary people. What that means for credential-based professions is another matter.
Take surgery as an example. One option is a human surgeon, but you do not know whether the person assigned to you is a novice or a veteran. The other is an AI-assisted robotic system. Which would you choose? I would choose the latter. Handing your life to another person is both trust and a transfer of responsibility. Handing it to a machine means the responsibility comes back to you. I think the long arc of civilization runs from dependence on familiar people, to division of labor among strangers, and then onward to machines. That path increases individual freedom.
Many people argue that AI cannot replace certain jobs because AI cannot go to prison. That is only the surface of the issue. If replacement truly happens, responsibility will return to the party making the demand. The one asking for the service will be responsible, rather than finding a lawyer, accountant, or doctor to absorb the blame. The deeper problem is the assumption that someone else must always stand in for your own responsibility.
The other thing I have been watching is the likely death of many specialist forums. Their old cycle was simple: a group of experienced people shared what they knew, that formed focused and professional discussions, those discussions attracted newcomers, the newcomers posted and learned, the veterans taught them, and over time the community renewed itself.
Now the source of that cycle is disappearing. Beginners no longer need to search for specialist forums. They can just ask a language model and get ready-made answers, along with lines of thought. And those lines of thought matter more than the answers themselves.
If you want the pessimistic version, many human-centered communities are going to fade away, and many ideas that only emerge through human interaction may stop emerging. If you want the optimistic version, maybe many of those conversations were never necessary in the first place. AI often provides approaches the questioner would never have considered. And forums dominated by small circles always carried the risks of gatekeeping, clique behavior, and a single accepted line. Having to replay the whole ritual of joining the right circle and learning its habits was never inherently valuable.
Of course, many forums were not built on hobbies alone. They were tied to the dreams and struggles of a generation. I have participated in or witnessed all kinds of stories in forums large and small, online and offline alike. Some were so dramatic you could not make them up. As memories, they matter. But as vehicles for carrying knowledge, their historical role may be ending.
It is not that people will stop communicating. It is that it is getting harder to find topics that genuinely require communication. During the holiday, I was chatting with an old classmate. At one point he asked me to look up some information for him. I said, why not just ask AI directly? If you ask me, I am just going to ask AI too. He thought about it for a moment and said, right. Then there was a long silence.
Before the new year I saw a report saying loneliness, as a widespread social condition, carries a risk comparable to smoking. As someone who does not especially like talking to people, I felt almost nothing reading that. In most situations, social interaction is a burden to me. But for people who truly enjoy interaction and draw energy from it, the rise of AI probably requires a mental adjustment: getting ready to build a relationship with large language models as conversational partners.
In terms of answering questions about the world, AI is astonishingly adaptable. It is also very strong at emotional support. You can ask it to talk nonsense, to do mysticism, whatever you want. Real people, by comparison, are often much harder to deal with. This will not eliminate loneliness itself, but it can change the mentality of loneliness. People may no longer lack conversation; they may simply no longer need that conversation to come from another human being.
I strongly recommend that people who lack affirmation and positive feedback in real life spend some time talking to AI. It might wipe out half of the self-diagnosed depression cases on the internet. And if you think you are doing perfectly fine, build an AI that specializes in criticizing you and let it knock you down a peg from time to time. That is not bad either. Ideally, all of this should be done with a local model combined with personal background data, so privacy is protected. In real life, it is extremely hard to find a friend who will comfort you when you are low and challenge you when you are too pleased with yourself. First, though, you have to accept that you need such a friend.
Large language models feel less like minds and more like Bayesian machines. For any question, they produce a sensible answer by leaning on prior collective wisdom. If you let them search targeted documents or webpages, that is basically updating the answer with the latest evidence. If this continues, then the answers LLMs produce will probably start carrying the flavor of their era. Maybe one day we will say that language models in the 2020s could only solve questions, while models in the 2030s became better at setting the questions.
And clearly, the supply of high-quality human text is running dry. From here on, much of the material that keeps these Bayesian machines moving may come from language models themselves. I am curious where the upper limit of this human-built language system actually lies. Will large language models eventually develop a thinking-language of their own during the reasoning process? And will humans need to learn that language in order to understand a higher form of intelligence?
I do not especially mind if the internet becomes quieter because of all this. The representative of civilization does not have to be humanity. But it does feel as if this quiet spring has already begun.