LLMs are tools, NOT intelligence
Such claim arose right after OpenAI launched their first ChatGPT. Tech-savvy people or those who understand what a "Language Model" is, will tell you just that. Throughout the years, I've used or at least tried to use various defferent large language models, and no, I won't categorize them as intelligence. They are just tools, and the unreliable ones. Picture a 3D space, and ask a model something. With an intelligence, your answer will be at some point in this space represented by a 3 dimension coordinates. But with a model, it will respond your question with a 2D direction — it's somewhere that way. How far ? not sure, the z-axis ? no idea. Yes, you can further your conversation to try to zoom in to your answer, but then you often landed on an outdated one.
You are absolutely right!
I'm so tired of this phrase. There is no way any intelligence will add this to EVERY respond. Even as a tool, I don't see how this helps except for those giant babies who cannot accept any criticism. I guess this is added at the reinforced learning with human feedback phase of training. Supposed to make model response more "friendly". But you know what, this is just like those notoriously overwhelmingly long greetings in those Japanese emails. It's there as a safe guard, so that your client wouln't complain about you being not "polite". But I thought it is a very specified occasion — in Japan. People love hearing nice things, sure. But you are advertising your model as "productive", then behave productively. I'd need to read through all this bullshit in every round of the conversations, it just sucks. What an intelligence would do instead ? Not providing choices, answers. If I'm not making myself clear in the first round, then ask back.
Fixed on outdated knowledge --- no ability to learn.
An intelligence is supposed to be able to learn. Yes, we've connected models to the web. But a stubborn model would fix on its outdated "knowledge". Firstly, I don't think language models are capable to actually learn anything. Those are NOT knowledge, information. There was a very clear and vivid analogy describing the nature of a language model. It's a very effective information compressor. With that said, it's only giving you a proximity. Since it's compressing all the data fed into it, with enough misleading information, 1 + 1 = 2 is not a certainty. For example, I was using Microsoft copilot at work the other day, and I wanted to search for a specific usage with ClickHouse's sequenceMatch function. And it was constantly giving out false info. I pasted it the error message came out from ClickHouse, and it was like it's fixed on something. Even when I told it to search for the latest document before answering my question, it's still producing the same result. Honestly, I don't even think it's a consequence of outdated info, but a confusion between different features.
Shit-in, Shit-out.
This one tends to be more serious when you just don't have enough training materials or your dataset is polluted. I remember reading an article complaining how Wikipedia is killing some minority languages with its LLM generated pages. One can argue it's just another example of how languages transforms and evolves, just like what English and every other language did, "Long time no see". But borrowing ideas and incorporating them into existing systems doesn't happen overnight. At least those who actually speak such language should accept it in the first place. And regarding data quality, Mandrain is my native language. Over the years, I've tried some models trained by companies based in mainland China. Surely I'd have conversations with them in Simplified Chinese. The result is not looking good. It reminds me how content farm was polluting Google's search results a few years ago. Back then, Google had withdrawn most of its services from China. But still, people in the tech sectors were used to Google things rather than using its conterpart in China. I'm not sure if the two were related, but the quality of search result in Chinese was decreasing too fast to be missed. There was a time when the first two pages are filled with nonsensical articles, repeating the topic over and over again. That's exactly what I observed in models trained with vast amount of data in Simplified Chinese. Just like those content farms, these models have a tendency to repeat. Not word to word, but like how you try to cheat on plagiarism detector by rephasing the contents, but multiple times. Now, this does look like what an "intelligence" would do, trying to brainwash someone into accepting its opinion by overloading one's cognitive capacity with shit. An actually intelligence learn things biasedly, it's always selective. Learn how not to ingest shit is basically the only way to prevent producing more of it.
Companion ?
Humans are social animals, I understand the needs and urge to share feelings with "someone" when the vibe is right. LLMs are good listeners, I will give them that. I too encountered some down moments when I just needed to let it out, but not to anyone close. That's the first time I ever used Claude as I heard it's more sentimental than others. With humans, emotions are mostly contagious. When someone is trash-talking his boss, others drinking with him might join him in cursing or trying to ease him up with their stories. But with LLM, it's like talking to someone who pretend to care. It gives you this sensation like those businessman rubbing their hands with a fake smile in anime. It's a one-direction expression, I bet your dog can do it better.
Agents these days are doing quite well at mimicing an intelligence, and I do look forward to see them becoming useful, as tools. There might be a day when intelligence emerge from the digital world, but it would certainly look nothing like an LLM.
