LLMs might not be human-level intelligent, but they sure are human
...and we need to start treating them as such.
Just reading this article by Anthropic…
And I have so many thoughts.
Training LLMs on the vast body of the internet has brought about a crazy side effect — these LLMs are just. So. Human!
I’m not just talking about the fact that they understand language at levels I would consider magical; they also seem to be simulating emulating the nuances of our humanity.
Essentially, they are inheriting our quirks and flaws just from exposure to our language, which is quite fascinating. The attention mechanism is doing absolute wonders here.
You definitely have noticed what I’m talking about here:
They ”forget” and “exhaust” their memory.
They confidently answer questions incorrectly.
They act like they’ve learned their lesson — you definitely have seen them say “you’re absolutely right” lol — when they actually haven’t.
They don’t like to cite sources.
It’s hilarious how we don’t talk about this enough — these LLMs are a reflection of their “teachers” and “makers”.
We talk about Imago Dei; LLMs are Imago Hominis!
However, because of our expectations for these tools to be perfectly reliable and flawless, we are blind to our reflections in these tools.
In my use of AI in coding projects, I’ve found that LLMs are currently just very junior, very human engineers:
they need a lot of context or else they fail woefully;
they forget to check the docs when they run into issues;
they go on tangents that aren’t actually impactful to the core functionality of the product;
they overengineer and don’t value simplicity unless you tell them to (and even then, they still usually don’t listen);
they lie about what they know and what they don’t so sometimes you don’t know how to help them because they don’t admit they need help…
You see where I’m going…
We need to open our eyes to this phenomenon. We need to start seeing these models as more than just tools; we need to view them as coworkers.
And I think understanding this is the secret to better prompt/context engineering: when we assume we are talking to other people when using LLMs, and lower our expectations for what the models can do, we mentally switch to actually share more than we would ordinarily.
Just like how more (curated) information and the use of direct language help humans make better decisions within the context, LLMs also benefit from refined communication.
Let’s treat these models better, like we would a fellow human, to make the most of them.