It seems like the biggest hurdles to LLMs currently are lack of imagination and lack of a habit of using them.
It will take a while for memes of how to use LLMs effectively are widely spread, and then another while for people to see everyone else using LLMs and to copy them.
It's all social inertia though. If GPT-4 were the last language model ever trained, we'd still be discovering new ways to use it and transforming society for decades.
What a great article! He seems a bit skeptical which honestly is not my view, but actually overall the article is very balanced.
The article compresses quite a lot of these LLM discussions and balances with point and counterpoint.
To pick up on the "what do you do with them" question, one trend on the horizon might be virtual employees (aka Agents) and virtual teams.
So the starting interface might go from a chat session that opens with some example prompts, to pictures of fake people with their job titles and skills under them.
The agents you select would plug in to a team and iterate on whatever tasks they are assigned in your project, up to your given budget.
So the thing you need to figure out is more like a mission statement for a company with occasionally a few high level goals, rather than a lot of detailed instructions and corrections for every individual task.
Another route along these lines would be an assistant that follows you around, watches what you are doing, and provides suggestions on demand that automatically incorporate context. I think many (if not most) large software providers are building this in.
At dinner last night my son in law went on verbal rampage about LLM “hallucinations” and tangentially on a fear of real AGI.
My point to him was that my own use of ChatGPT, Anthropic, and Hugging Face models always involves my providing text to work with, be that my own ideas that I want feedback or more extended ideation on and CoPilot auto completions in my Emacs buffers. Hallucinations are much less a problem in this mode of using LLMs.
It will take a while for memes of how to use LLMs effectively are widely spread, and then another while for people to see everyone else using LLMs and to copy them.
It's all social inertia though. If GPT-4 were the last language model ever trained, we'd still be discovering new ways to use it and transforming society for decades.