In the last week OpenAI released the latest version of their chatbot based on deep learning, ChatGPT. It's been quietly taking the tech industry by storm ever since and I finally got around to trying it out today. Let me tell you, despite a few shortcomings, it's actually quite mind-blowing. This article won't be about how it works, some Googling will give you the basics there if you're interested, rather here I'll be sharing some of the prompts and responses I've been getting. Let's dive in.
I actually struggled to come up with a sufficiently complex use case to get started with, but some browsing on Twitter gave me inspiration. We traveled to Iceland in April and Brittany did a huge amount of research into what we should see, where we should go and how we should get there. I thought that might be a quick test to see the level of complexity it can handle.
Aside from getting a couple of the locations wrong (Laugavegur trail is in the south, not the North), this was actually very impressive. It covered most of the locations we saw in April, as well as some we didn't. It didn't really do a great job of ensuring each day's itinerary was possible in terms of driving time, but I suspect with a bit of clarification in my prompt I could get that level of detail as well. I'd call this a success, moving on.
Next one up was asking for a recipe. I get so frustrated with the level of SEO bullshit that comes with every recipe on the internet, so I wondered if I could use this to cut through a lot of that. Here's the result.
Again, success! Not sure how it's gonna taste, but I'm definitely gonna try it. And we avoided all the SEO BS. Wild.
Specific Advice for a Video Game
Next up I tried to see if it could tell me how to solve one of the most challenging parts of the Xbox game I'm currently playing, Doom Eternal. Mixed bag here as you'll see.
Most of this is... simply not accurate. There's a very specific strategy for beating a Marauder involving either the super shotgun or the ballista weapons. Technically 5 and 6 are correct but apply generally to all enemy encounters, everything else has clearly been adapted/lifted from documented strategies for beating other enemy types. Again, maybe some additional specificity in the prompt may help here? Closest thing to a fail so far.
Coding and Recruitment Challenges
I'm learning that the potential for this tool to solve coding and technical challenges is incredible, as people have been posting about it all week. I decided to put it to the test using a classic problem I've solved in a variety of languages before.
This impressed me with the structure it created, but the actual implementation of the methods were missing. A simple tweak to the prompt I used, adding "include the code for logic in each method" produced the entire code for the game. I've included a sample of this below for reference. Looks like the code output gets truncated at a certain length, but wow it looks good.
I'd call this one a success, and somewhat terrifying for anyone trying to recruit using coding tests/exams/exercises.
Impressed so far, I wanted to try some short-form creative-writing use cases. First one I tried was the following.
I thought correctly that style mimicry might be a bridge too far here, but as far as content goes it's mostly bang on, if a little dated. References to Canucks now-departed head coach Travis Green and maybe-retired Brandon Sutter give away a little about how far in the past it's looking in terms of training corpus. Ultimately this is amazing, but it is more Thomas Drance than Ryan Lambert. This might be expected given Drance is a Canucks beat writer and Lambert writes about draft prospects.
The next and final (for now) was a similar article about cricket which again, basically nails the generalizable core premise.
Paragraph 2 about the constant state of being always "on" in a game that lasts so long isn't something I'd thought about but absolutely resonates. Paragraph 3 summarizes my thoughts about lapses in concentration almost perfectly. Paragraph 4 is basically a rehash of paragraph 2, but still, makes sense. Paragraph 5 could apply to any sport. The final sentence is perfection. In all, not a bad effort, but kind of vanilla from a stylistic perspective.
That's all for now, but I plan to write more about this as I have more time to experiment with the tool. Stay tuned!