As someone who is building a “Yelp clone” of sorts, this amounts to about 1/1,000,000,000 of the amount of work to actually build a Yelp clone.
I get that this could be helpful for absolute beginners who just want guidance on the very first steps to building any basic CRUD app.
But there are very prominent people in ML who are actively claiming that programming itself will be taken over or whatever by ML soon (Sam Altman, Gwern), and who I think for some odd reason don’t realize that what GPT-3 is doing here would take any average junior software developer about 30 minutes to do tops. Meanwhile no matter how many junior engineers you had they could never clone Yelp (maybe after many years and a few re-writes, ie, they become senior) - it would take more like a team of 10 very senior software engineers multiple years to deliver anything close to a Yelp clone. And it’s not just scaling, the first steps are the easiest parts and the best documented.
It’s akin to an ML art program drawing a single wobbly paint stroke and then writing a blog post about it “drawing all the strokes to make a Picasso clone”.
Except it’s also even harder than that. Because art can be sloppier and generally deals with a simple set of skills put together in forms that aren’t hard to study. Meanwhile software one tiny mistake in one area could break the whole system, and you need deep logical reasoning and tons of iterative problem solving with long term memory to debug it.
If by clone you mean clone of Yelp the day it launched, with basically no data, maybe.
Wrong twice - Product is half the battle, and it’s not easy.
The bare minimum to compete would be native and web apps with good UX, competitive reviews and search, and better social features. Then you can fight the other half.
> The bare minimum to compete would be native and web apps with good UX, competitive reviews and search, and better social features. Then you can fight the other half.
That does not seem like a winning recipe for competition. The bare minimum for competition would be some type of product differentiation that substantially differentiates the experience. Trying to compete on things like data, UX, or search doesn't seem like a winning formula to me.
I totally agree you have to differentiate and avoid replication as much as you can as a startup in general, but this conversation is mixing awkwardly two discussions now.
In context I was pushing back on the idea that cloning it to any competitive degree isn’t a mountain of a task.
But out of context of this thread, then yes as a startup your strategy should involve simpler, novel features that let you avoid some work. Of course if you're aiming to replace Yelp eventually, you do have to do that hard work and I doubt you'll really get far without doing it ultimately. Ratings/search is the product. You can simplify it in other ways though.
I mean just not being a corrupt mob like entity that exists to harass restaurant owners would be a plus to me. Solving the chicken and egg problem of actually getting momentum would be much harder.
Then it is not only competing with yelp but also with Google Reviews which has the advantage of being in the search page itself.
Funny seeing this on HN next to a post on artists worried about Stable Diffusion. Everybody was expecting AI to come for taxi drivers' jobs first, and here we are.
Total clickbait. Why waste our time with code you aren't even competent to evaluate. The author writes:
FYI - I haven’t run this. Never been strong in terminal and I have a tendency to break things in there already.
Running GPT-3 written code in the terminal seems like something best left to the experts…
Or maybe I’m just a chicken.
As someone who is building a “Yelp clone” of sorts, this amounts to about 1/1,000,000,000 of the amount of work to actually build a Yelp clone.
I get that this could be helpful for absolute beginners who just want guidance on the very first steps to building any basic CRUD app.
But there are very prominent people in ML who are actively claiming that programming itself will be taken over or whatever by ML soon (Sam Altman, Gwern), and who I think for some odd reason don’t realize that what GPT-3 is doing here would take any average junior software developer about 30 minutes to do tops. Meanwhile no matter how many junior engineers you had they could never clone Yelp (maybe after many years and a few re-writes, ie, they become senior) - it would take more like a team of 10 very senior software engineers multiple years to deliver anything close to a Yelp clone. And it’s not just scaling, the first steps are the easiest parts and the best documented.
It’s akin to an ML art program drawing a single wobbly paint stroke and then writing a blog post about it “drawing all the strokes to make a Picasso clone”.
Except it’s also even harder than that. Because art can be sloppier and generally deals with a simple set of skills put together in forms that aren’t hard to study. Meanwhile software one tiny mistake in one area could break the whole system, and you need deep logical reasoning and tons of iterative problem solving with long term memory to debug it.