7th March 2023
On LLMs and language
Some of the most commercially successful uses of LLMs so far have been for generating business content: companies like Copy.ai and Jasper are writing blog posts, Tome helps you generate pitch decks and other presentations. While criticism has been much stronger against the 1st set of tools, my personal opinion is that the second is much more harmful.
Critics say that tools like Jasper, Copy (and Maker.ai, where I worked for a while) flood the market with bullshit SEO-generated articles. Let's not fool ourselves: this has been the case for a while now, with or without LLMs. Your average freelance writer is not much better than what GPT-3 outputs (if at all). And that's ok! Most blog posts have a clear, non-intellectual purpose: show up when people query specific keywords on Google. They are artifacts built around a system Google designed. So while I agree that tools like these are increasing the number of posts and this is not ideal, this doesn't bother me too much - the average blog post quality will stay the same and most of them will still stay buried under a “Next page” button. The fact that small businesses now even have a shot at getting to Google's 1st page without hiring dedicated writers is a good thing.
I'm more concerned about Tome or some of the use cases tech-Twitter has found for ChatGPT. Things that miss the point of language and communication entirely, but that seem fine on the surface.
Language is fundamentally a way to connect with another human. As flawed as it may be (everyone has their own beetle in a box), it's the only tool we have to share back experiences, thoughts and feelings. And there are ways to make it less flawed: Robert Anton Wilson says abolishing the verb “to be” would help us become more precise in our language. You don't have to go that far: just think hard about what you want to communicate.
And until we get to AGI, this is fundamentally human. Language without a person is devoid of meaning, it is void. For better or for worse, we are immersed in it. Our emotions shape our inner linguistic processes and vice-versa. Changing the way we talk about something can change the way we feel about it too. With self-awareness (the one thing currently missing for AGI), we can become just what we want, rather than another robot whose blueprint was drawn up by someone else. Language is also a gateway to affecting reality.
When you use Tome to generate a pitch deck for a helicopter rental seed company, you are letting statistical inferences and matrix multiplications shape your vision. Unless you've thought very long and very hard about this to know exactly what you want to say, your thinking will be diluted. Even if you know what you don't want (okay, this slide sucks, my helicopters should not be red - let's make them blue!), you are still being influenced by it. Definition by negation is also limiting. And this is Tome's whole business model.
This is dangerous because it's easy to think it's your thinking. After all, it's almost like it read your mind! After all, you read the slides and you agreed with most of what was written. And you rewrote what you didn't agree with. Can it really be that bad?
It's shaping reality through your will - through language. I think tools like this can be useful, even when writing pitch decks. LLMs make great research tools. Digest the information, but make it your own. You should micromanage your language because words do matter. Ask Larry Seinfield.
LLMs are incredibly useful. Definitely, the most influential technology has changed since the iPhone, and time will tell if it becomes the most influential ever. But let's save it for automating tasks, transforming unstructured data into structured data, and use it for tasks where language matters less.