Scattered thoughts about AI

Below are some scattered thoughts about AI, which I hope to update over time.

Last updated: 1 month, 1 week ago

I largely think that LLM-based B2B AI tools are too early to show any real success/takeoff, and consumer AI products are the most likely to find early success. That being said, it does not mean that you shouldn't attempt to build a B2B AI tool right now, because there's value in setting up proprietary data sourcing and collection. However, the fruits of that labor are likely going to take longer.
I've used a number of the AI coding tools, and while they're certainly impressive for simple things N, they can't really produce valuable code for N+1 level of difficulty, even with a very detailed report on how to solve the problem. I've talked to a number of software engineers, and they've come to the same conclusion.
It's interesting to me that, while everyone agrees that data cleaning is the most important part, nobody actually seems to be investing in it? RLHF is like a bandaid more than anything.
Chat interfaces are a really boring UI, and prompts are a subpar UX. Why do users have to constantly find new ways to beg, cry, and threaten a computer to get the answers they want?
I do, however, believe that language models wouldn't have received nearly as much attention as quickly without the chat UI. It was (at the time) a really important design choice/innovation that showed what language models could do and what they were capable of in a really simple way, so much so that I knew 10-year-olds using ChatGPT in January 2023! The chat UI is still useful for some purposes, but I don't think it should be the end-all be-all for all interactions with language models.
I think search engines' decline is going to be far slower than what people imagine. I still find it easier to just search for an answer to query than wait for an LLM to generate a response for me.
Building auxiliary devices and marketing them as a replacement for existing XYZ device is as good as lighting money on fire, and proves that a company doesn't understand their customers (especially if they're selling to the consumer market).
The barrier to entry ($) will drop faster than you can blink. You can already train Llama2-7B-level models for $100k.