On the Relationship Between AI and Knowledge

Everyone, I was really happy to see you yesterday, but due to some other reasons, my enthusiasm for in-depth discussion wasn’t very high at the time. Reflecting on it today, everything else is fine, but there is one key issue I still want to bring up. I spent some time this afternoon writing some rambling thoughts; it’s a bit long, but I’m sharing them with you.

I remember A once asked a question: Are there any apps you can recommend now? Off the top of my head, I said Reddit, but we didn’t go into it deeply. For this “answer” I gave, I expect you all might have subconsciously understood it as a specific software, but that’s not what I meant at all — what I wanted to say was a type of knowledge carrier like Reddit, specifically the open knowledge community type.

Why do I say that?

In the past, in our process of acquiring knowledge and expanding our thinking frameworks, the sources and driving forces came more from highly complete objects and traditional channels like works, papers, and monographs. But we face two key problems.

The first problem: These contents need to be seen by you (discovery barrier). This depends on too many contingent factors — timing, channels, the completeness of the content (completeness, not value itself), and even the prevailing academic research trends and cultural preferences at the time, i.e., the conditions for knowledge to reach you. Only when all these conditions align will valuable content be placed before you. Otherwise, no matter how good something is, it cannot become present value and is more likely to become a discovery for later generations, like Van Gogh or Bach.

The second problem: Even if presented to you through traditional channels, individual energy, free time, and the limitations of cognitive breadth will impose another filter (absorption bottleneck). As long as you are human, this ceiling is unchangeable — it’s a hard constraint at the physiological level. There is an insurmountable upper limit in the allocation of time and energy and the ability for cross-disciplinary learning. In short, “deep specialization” and “broad expansion” are fundamentally mutually exclusive.

So, what exactly do large language models bring us now?

Most people will subconsciously regard large language models as all-knowing, lofty knowledge authorities. (This is also why I said yesterday that most people treat LLMs as a database — it’s the same judgment.) But the truth is completely different. When the data fed to LLMs during training (which is indeed traditional knowledge content) undergoes embedding, reranking, and vectorization, what the model acquires is the ubiquitous relationality inside knowledge — that is, a knowledge relational network. And this relationality is precisely the most basic and core thinking paradigm and “insight” needed in the processes of summarizing, inducing, classifying, categorizing, deducing, analogizing, reasoning, and judging that we rely on when studying problems. In one sentence: What LLMs provide is not knowledge, but logical reasoning ability.

On the way back to school in the car, we talked about matters related to B’s thesis. C, your judgment and line of thinking are already well on the path of in-depth research. This topic can certainly be written about, and it can be written well. The logic and path needed are not complicated: treat the work as the endpoint of various influences from political, economic, cultural, and other fields, rather than searching within the work itself for its own inherent laws and evolutionary paths.

Why do I say that? Because there is a logic I cannot deny: artistic works, as the final manifestation of the mutual influence and interaction effects of factors like society, politics, economy, and culture, will inevitably contain evidence of all preceding causes, whether subtle or obvious. Doing good research of this kind requires only one thing: breadth of knowledge.

As a side note. I consider myself relatively fast at learning — not just among the older generation my age, but even compared to people one or two generations younger, I often run ahead. Many times, when I and others (including you) simultaneously approach a new problem or field, I might also take a temporary lead (though the final result might be different, but this is likely because I lack the patience to persist — a character trait brought about by genes that is hard to overcome). I’m not speaking casually; I have seriously made phased comparisons on many things. The reason I can do this is actually just one: after my thinking ability matured, I did my best to constantly try hopping between different knowledge domains. You might also notice this from my educational background. So, in my view, cross-domain transfer learning significantly enhances cognitive efficiency and depth, meaning: The breadth of knowledge is the sole factor determining the height of knowledge.

Back to the main topic, why did I mention Reddit?

Before D’s thesis defense, I was forced to once again tackle semiotics. In this process, the greatest inspiration didn’t come from any academic paper or monograph — those were all familiar — but from a Chinese blog post. It pushed my understanding and thinking about film semiotics forward by a huge leap. The result was that I fundamentally negated the theoretical and methodological applicability of her thesis (she eventually passed her defense, of course; I won’t comment on that).

Before E’s thesis defense, a similar situation occurred. I was forced to delve deeply into Bourdieu’s field, habitus, and capital, which at the time was a missing piece in my knowledge structure. The biggest gain in this process also didn’t come from papers or monographs but from a set of discussions on Reddit. It was precisely thanks to this discussion that I quickly found some fundamental logical errors in the thesis.

This kind of good fortune isn’t the norm. Me being able to see these things was a matter of chance, not necessity. From another perspective, I have no physiological ability to exploit such channels for this kind of content on a regular basis.

But in today’s information ecology, the value of such content might far exceed what we are familiar with from conventional channels and paths. To make an inappropriate analogy: the information or knowledge mentioned above is still in its nascent stage, with a long way to go in terms of influence, reach, and the journey from personal viewpoints to knowledge consensus. Before completing the familiar knowledge growth cycle, such content will always be drowned in the background noise of information, unknown and unseen.

From my personal experience: if I hadn’t stumbled upon them, my loss would be immeasurable. Of course, another possibility is that if I had delved deeper into the content conveyed through conventional knowledge pipelines, I might have drawn similar conclusions or judgments — but what about efficiency?

I’m quite scared of the results deduced from this. Even setting aside utilitarian goals (like being the first to propose newer, deeper insights or publish papers), purely from the perspective of knowledge learning, this feeling of FOMO (fear of missing out) is terrible — it truly feels like a fear of being abandoned on some “track.”

The logical ability of LLMs mentioned earlier might fundamentally change this situation.

Since the late 1990s, one result of the internet has been the democratization of knowledge — but it hasn’t changed the basic model of knowledge iteration and renewal because it still relies on the “academic system” built over centuries.

But now, the situation is different. If we leverage the logical capabilities of LLMs, the content previously buried beneath the noise floor of information, beyond human reach, can be easily extracted and solidified, becoming a brick in one’s own knowledge structure, height, and breadth almost immediately. Ultimately, this could build, with high efficiency and quality, a magnificent skyscraper entirely one’s own, akin to those in Minecraft. Those with a habit of broad knowledge who are forced by physiological limits to compromise on depth, or those with a habit of deep knowledge who are similarly forced to compromise on breadth, can both leverage LLMs to perfectly compensate for their weaknesses.

The feeling AI has given me over these past six months is like what I joked about yesterday: I feel like I “can’t keep up.” This isn’t about the “big words” thrown around in terms of content, but about quantity. The temptation of its logical ability and output quality has led me to spend at least 10 times more time on it than before, unable to stop. In other words, if I could ignore basic needs like eating, sleeping, and work commitments, it could completely seduce and consume all my time — and I still feel far from its ceiling of output.

Where is that ceiling? My personal judgment is: we haven’t yet seen signs of the scaling law plateauing. If that day ever comes — let me throw out a conclusion off the top of my head — if we imagine the future logical ability of LLMs as being as vast as the Earth, then in that coordinate system, the physiological limit of a perfect, ideal human might be less than a speck of dust on Earth.

I strongly agree with one viewpoint: LLMs will exacerbate the stratification of people. Those who can use them and those who can’t, those who use APIs to purposefully achieve their goals and those who only ask a chatbot like Doubao or ChatGPT in a chat window, will become two different species in the future society. People who leverage the logical ability of LLMs are like driving a combine harvester, harvesting hundreds of acres by noon, while the other group is bent over, sweating profusely, wielding a sickle on three acres of land, toiling endlessly. Thinking about this kind of dimensionality reduction scenario is both thrilling and spine-chilling.

Returning to the starting point — what software to recommend?

A year ago, there was a viewpoint: Model is software; all software will disappear, and models will replace everything. I was skeptical of this judgment at the time. The reason is simple: models provide logical capabilities beyond human physiological limits, but they won’t become a combination of “knowledge base + logical ability” — this goes against the basic architecture of LLM training. Whether it’s the Transformer or the later-introduced Mamba, or the not-yet-mainstream, unclear-future RWKV, BASED, xLSTM, etc., none supports that conclusion.

But another concept proposed in the first half of this year — Agent is software — although derived from a commercial logic standpoint, I believe it points to the right future direction. Establishing an “execution” middle layer between the model’s capabilities and human needs, this architecture aligns with all my beautiful imaginings of the AI paradigm.

So, are there any traditional software/apps to recommend now? My answer is: No.

But, I suggest you start leveraging LLMs to understand and initially use Agents.

What you need to do is not to go to a platform and pick up a ready-made Agent made by someone else — like Workbuddy on the Tencent platform, for example. Doing that is essentially still using “traditional software,” because its internal process is “black-boxed,” which doesn’t help much in understanding new methods and paths.

What I suggest you do is:

1. Purchase and subscribe to some model APIs — such as Anthropic, ChatGPT, or open-source DeepSeek, Zhipu, Qwen, etc. If you’re in the initial familiarization phase, you can buy APIs from aggregation platforms to easily switch between different models;

2. Try using some basic open front-end tools — such as Open WebUI, Lobe-Hub, or more pure Agents (like the well-known Open-Claw, Hermes, etc.), or even true open frameworks like LangChain, LlamaIndex, AutoGen, etc. In these tools and frameworks, try building a local knowledge base and, in a connected state, extracting and processing knowledge from vast amounts of information.

Honestly, what you build yourself will definitely be far inferior in quality and user experience to products developed by professional software maintenance teams in the future. But the value and significance of personally going through this process lie in: you can deeply understand your own needs, see the directions and paths of different ideas and thoughts taking shape, sense the sweet spots and pain points of your personalized needs… All of this will become targeted judgment for choosing better Agents in the future (or more abstractly, called the “middle layer between models and needs”), as well as important basis for personalized tuning and improving accuracy and efficiency of your chosen tools.

I’ve rambled on enough. Let’s stop here for today. I hope we have the chance to sit down and talk again next year — a lot can happen in a year.