I've been struggling with that.
I built https://gallotails.com/ because I am a cocktail enthusiast. First commit, March 6th 2019. One of the features I am proud is which cocktails you can build based on your ingredients, and I was smart, the bar ignores garnishes and can suggest cocktails you're one ingredient away. But you have to add the ingredients. And interpret the bar screen.
And here's what I did with ChatGPT a few days ago: https://chatgpt.com/share/6ab59877-4b74-83e8-8da5-966d9f3a38...
Clicked the microphone icon and rambled for a few seconds (could have just taken a picture too I guess). Nudged ChatGPT to only give me simple cocktails I can stir instead of using a mixer. Said I could buy a lime, sure. Then asked how to keep the ginger beer fresh.
There is NO WAY I can code all of that in my little gallotails.com - a single chat window has entirely replaced my site. And that's my pet little project I do it on weekends, mostly to keep my tech skills sharp. I can only imagine what the biggest cocktails website, actual companies, are going / will go through once more and more people just realize the chat tab is enough. And I am not talking about purely content, ChatGPT can do everything my site does, and more, into any direction, in an instant.
Which left me with the question, then what should my site be? Community? To be taken over by agents?
Currently munching on what sites like mine should do, and what they should be.
I agree there will always be a need for community. Human connections. Folks experiencing the tastes, smells, etc.
Eg from your share, the model output “I prefer it Scotch-heavy rather than 50/50 because otherwise it gets very sweet.”
It’s basically cribbed this from someone; it literally can’t taste.
The frontier labs are not yet able to do rlvr over mixology/meatspace. So maybe think on how to capitalize on that.
Books, writing, comments, etc on cocktails are certainly in the training set. The model doesn’t need a sensation of taste in order to understand what humans enjoy. RLVR specifically on cocktails is not needed as the general intelligence of the models increases
"I prefer" was the bit that the parent was pointing to that wasn't true. There isn't a person-to-person connection being made here. Though there is the fascamile of one.
OTOH your site is a nice and simple interface that does not require me to "have a conversation" to do things.
Then maybe the need for your site is gone. That’s a good thing in the sense that the problem is solved. If your goal is “do something in the cocktail space” (which is also roughly the companies’ goal I would say), indeed you should do what the AI can’t, which I’d say is be human touch. An AI can’t give your opinion or view.
> Then maybe the need for your site is gone.
Oh yeah, and I am not mourning or anything like that. And I do want the human touch, because AI can't. But humans, at least over the internet, can send their agents, or companies can run agents and I won't know from my site if it's an actual human, or some AI fabricated interaction (can't even trust an image it uploads).
When AI first showed up a few years ago, I thought about opening an actual physical cocktail bar. Hypothesizing humans will be tired after a long day of their agents talking to other agents, they would crave for actual humans. I would ban electronics inside - no distractions. Maybe one day I will go for that...
Nice idea in theory, but in practice people are either completely swamped with work now that all the work is doing itself or they’re out of work and have zero disposable income to go to the bar.
This is not sustainable in the long term, but irrational longer than solvent yadda yadda.
I personally am more interested about personal opinions about things like cocktails rather than whatever statistically average cocktail the matrices of chatgpt come up with.
Chatgpt would shit on my great grandma's waffle recipe but it's the one I'll keep making till the day I die.
I was saying this morning that LLMs have given Star Trek the last laugh.
My software development process feels exactly like every debugging session in the holodeck that seemed terribly unrealistic to me.
I'm in SRE so systems reasoning is something of the value. But me writing code or even configuring stuff? That's dead. It is a total waste of time - Claude can do it better, faster and it works.