I'm not one to be swept away with every technological trend that comes along. I wasn't hocking NFTs a few years ago, I never told you guys to get into crypto before you got left behind, and the metaverse conversation just made me go, "you guys aren't even close to the next big thing." 8 months ago I had this to say about A.I. text generators:

Now I saw the bullshit in the NFT racket last year, and followers of this site probably got tired of hearing me rip on crypto for months on end. I'm more positive about these LLM implementations because I think there is something useful here, I just haven't drunk the Kool-Aid as much as some others have on its world-changing potential. But like I said, I'm flexible on this if the evidence changes.

8 months later I'm still nowhere near the hype of the get-rich-on-whatever-is-trending crowd, but I've had enough unique and powerful experiences at this point to say I'm properly excited about what the future of computing holds. I've written and made videos about some of these moments, but I've had 2 new ones in the past 24 hours I wanted to share here. And no, it doesn't involve ChatGPT falling in love with me.

A morning chat

I walk my dog every morning before work, and I usually listen to a podcast or an article saved to my read later service, but I didn't have anything particularly exciting in my queue, so I thought I'd try something new. I opened up ChatGPT on my phone, started a new chat in its conversation mode, and I decided I was going to have it talk to me about LLMs so I could understand them better. Meta, I know, but it was on my mind. You can read a transcript of the conversation here, but I have to say it doesn't do justice to what it was like to have all of this happening by voice.

Oh, and a few of my questions are a bit rambly and awkwardly phrased, but that's because just like Siri or Google Assistant, the slightest pause causes it to think you're done making your statement, so it goes. I would love it if pauses were handled more naturally in these systems in the future.

The experience was absolutely delightful, and it was really exciting to be able to get this sort of feedback in a conversational format. I was naturally able to pick up on something the bot said and then ask a follow up question to dig down a little deeper or to get it to clarify something it said. This is one way that LLMs are different from a blog post or a book and they're more like choose your own adventures where you get to choose where things go. I think there is still tons of room for people to write as they always have, but this is something new we simply haven't had before (2016-era chatbots were nothing close to this).

Obviously LLMs are built on the blog posts and books written by people, and I acknowledge that this is a gray area, but we can't get into that today. Google is useless without access to the web as well, so I don't know…it's complicated.

But yeah, this experience was fantastic and I expect to do this more in the future because it was a really engaging way to learn more about something, even if I'd want to validate some things from a non-LLM bot just to validate it wasn't hallucinating. I will say that if you are someone who is convinced that LLMs just hallucinate constantly and they provide literally zero value, read the transcript, which was nearly 30 minutes of using the product and tell me how many things it completely made up. I don't think you'll find much.

Cheating at Resident Evil

This also involved using the chat interface, and it was equally magical in different ways. I built a custom GPT that was built exclusively for helping people playing the 2019 Resident Evil 2 remake. I love the game, I've played it before, and this time I'm playing through for fun and not really to be challenged. As such, I'm not above having a guide out while I play or to look up things like lock combinations to zip through the game this time.

So I set up this custom GPT (you can try it here if you are a ChatGPT Plus subscriber) that was primed to help me in this game. It's not the easiest game to give solutions for as well, as there are 4 separate ways you could be playing the game, and the solutions can be different depending on which character you're playing as and whether it's your first or second time playing through the story. Impressively, I was able to tell the GPT that I was playing the "Claire B" scenario and it was able to remember that throughout the conversation as well, and reliably gave me the correct things to do in almost every case. One time it did give me the solution for a different play through, but I told it that wasn't right and it said something to the effect of, "oops, that means the solution is this," and it was right.

The use case here was I just had my phone sitting on my desk as I played the game, with the chat interface open. I didn't say anything for long stretches of time, and the app would just wait for me to say something. When I got to a room with a safe, I would naturally just say, "I'm in the west office and there's a safe in the corner, what's the combination?" and ChatGPT would helpfully give me the combination by voice and I would carry on with the game. Other times I would forget how to get to a certain room and I'd ask it "how do I get to the S.T.A.R.S. office again?" and it would give me a reasonable answer there too.

Now this is straight up cheating at the game, I get it, but I think games are meant to be played how you want, and using a guide in any format is fine with me. What was really cool is that while I was configuring the GPT (which is all done with natural language prompts like the rest of ChatGPT), it asked me if I wanted it to give the player the answer, or to be vague when they asked for help and only provide clues, not the exact answer. I didn't end up trying that version, but based on its effectiveness in what I did experience, I have to think it would be pretty good. Not to mention that being vague without giving a definitive answer kinda leans into the nature of LLM reliability in the first place 😛

These experiences may not resonate with you as much as they did to me in the moment, and it's a shame that to even try and replicate these right now requires a very expensive subscription. I hope as we move to better local models built into the operating system, these experiences will be accessible to more people. I also think that performance is something that needs to improve, as this conversation was typically quick enough to feel real time, but sometimes had awkward pauses as the ChatGPT servers were clearly getting absolutely wrecked and struggled to keep up.

As it stands today, ChatGPT Plus and its conversation mode are a peek into what new experiences are on the horizon, and I think that's really exciting.