Finding a Suitable Application
In the conversation, they talked about the application of models. He thinks call centers are an example. In the future, it is very likely that manual systems in call centers will be replaced by models because they involve “fixed processes” and “similar tasks,” which can be correctly handled and completed within the system. In addition, he abstracted some characteristics. Tasks with these characteristics are suitable for replacement by large models. Here are the key points I recorded:
- Exclude external factors: No interference from external factors, which reduces the possibility of the model making mistakes. Just as a stable factory assembly line requires a specific workshop, so does a model system.
- Sufficiently closed: Similar to the point above, this is also to minimize external interference with the system.
- Digitalized: This is also easy to understand, because large models are based on tokens. We can’t use a large model to bake bread.
- 80% AI / 20% Human: This point is rather counterintuitive. We hope that models can replace humans 100% and achieve complete automation, but from a real engineering practice perspective, almost all systems cannot avoid the 20% human part. We can reduce it, but we can never eliminate it. For some black swan events or very strange new situations, only humans can handle them properly.
Following his summary, I have also been thinking about what kind of use cases are suitable for using large models, in other words, looking for requirements. One possibility that I can think of and have encountered in my actual work is summarizing test results: in daily meetings, someone needs to check all the test results in all environments, compare them with historical results, share the new or strange test results, and finally ask for others’ opinions and ideas.
In this example, it first meets the characteristics of being digitalized and sufficiently closed. As for excluding the interference of external factors, it is not fully realized because the upstream and downstream relationships and configuration of the service system are very complex, and the dependencies are not simple. However, external interference is a low-probability event, which can be temporarily ignored here. The final 80/20 ratio is also met. Although it is all done by us now, large models can complete tasks such as historical comparison of test results, finding new error tests, summarizing test reports, and preliminary cause analysis. As for those very abnormal results, they can be directly analyzed with human intervention.
If this could be realized, it would save a programmer about half an hour to an hour every day.
I Need to Know What I’m Doing
Andrej uses models with a strong sense of purpose. He knows what good code is and what result he wants. What the model does is to implement the specific things in his mind. If the model happens to be unable to implement it, he is capable enough to implement it himself.
There is a way to measure how we use models: one end is not to use them at all, believing that AI is a product of evil, or completely adhering to the old school style. The other is to use them completely, even if I don’t understand a single line of code, I let it implement all the functions.
Everyone is using the same set of models and following the same measurement method, so the way a model is used largely reflects a person’s experience and personality. AI will not produce correct results just by inputting wrong information.
Therefore, as Andrej said, it is very important to know what you are doing, to have a strong opinion and will to realize something, so that the core belongs to yourself. You implement this core, and AI is just an assistant to help you complete the things you already know.
This also illustrates the limitations of AI in another aspect, that is, it does not know where my limitations are. It’s like it can’t tell me what I don’t know or what I lack. As a user, I can only stare at the text it generates, as if I have understood and learned something, but this is an illusion. I have not mastered these, nor have I experienced them before.
So my attitude towards using it is also optimistic but cautious, to prevent it from training me into a fool.
I Have to Consume the Original Content
After this interview came out, my social media timeline was filled with video clips, personal reflections, and summaries about the content. The personal reflections were fine, but the AI summaries were so boring. They had almost the same structure, paragraphs, and content. The only difference was the wording, probably because everyone used different prompts. There is some irony here; the most summarized content by AI was a long video about profound insights into AI, which is a real paradox.
After I spent some time watching the entire interview, I am very sure that those summaries are of no value to me, and sometimes even have a negative impact, affecting my impression and understanding of the original video. In the original video, besides the long sections on the theme of AI, there was also a lot of content related to learning and understanding knowledge. This content was very specific. For example, they would talk about how to deal with a code snippet worth learning, starting from using it, paying attention to each functional block and the relationship between each functional block, and grasping the main thread of the matter. In other words, it is very interesting to me, but it may not be to you, and even less so to the model. If you only read the model’s summary, you get a so-called essence summarized for you by a ghost, which is not even the idea of a real person. In the entire chain from summary to final publication of the summary, the only humanity that may exist is the taste and choice of the account owner, and such choices often come with advertisements.
As humans, I understand that we don’t tend to like second-hand things, such as second-hand smoke. But we consume a huge amount of second-hand needs, second-hand opinions, and second-hand information every day. Sometimes it’s not even second-hand, but a product that has been layered over and over. They are all digital and transmitted in binary, so they are very deceptive.
In order to ensure that more real and individual content can be transmitted, I need to consume the original content.
On Learning
They described the benefits of learning physics, that is, through systematic training, one can acquire an ability to quickly grasp the essence and core of a field, build models through abstraction, and remove temporarily useless noise. A metaphor using mathematics is to Taylor expand a new field of knowledge, and then only keep its first and second order terms, the most essential parts, the fundamental waves with stable frequencies.
In terms of content, starting from understanding a module, asking questions, and then throwing out another module that can cleverly solve the problems of the current module, thus triggering the student’s “aha” moment, and then gradually expanding into a concise and solid system. When the most basic system appears, everything else is just optimization of efficiency. Most of the systems or tools we use have been baptized by countless optimizations and have become extremely complex, so much so that we cannot understand them at all from external observation.
