Sounds like this is similar to a decompilation (pay walled article so I haven't read it).
A while back a listened to a podcast where they interviewed people doing this and it was quite interesting. (And it's an engineering podcast so many technical details.)
Given their estimates of a Mac being able to generate $1k (per month?) a 5090 with a lot more power would be able to generate $50k. For a $3k piece of hardware. Which is obviously not realistic. (As in, nobody is paying that much for the images, which seems to match well with no actual requests on the system.)
But with waymos it would be possible. Mark those streets as "extremely slow" and never go there unless you are dropping someone off. (The computer has more patience than human drivers.)
If that's too annoying then bad parking by school areas so the situation doesn't happen.
The field of "AI assisted programming" is progressing extremely rapidly.
Claude code (arguably the most recent large change, even if it wasn't the first of it's type) was released one year ago.
After watching the video I'd say that it is similar to my own reaction when opening my own code that is 2 years old. (To be clear, code I myself wrote 2 years ago, without AI.) Or even more realistically, code I wrote 6 months ago.
Buy it mostly reads like someone making an exaggerating claim to get s boost from a populist narrative.
If you just want to play I believe the Google alternative can even run on the free tokens you get from them. It's not going to do all that much before running out of tokens but you can probably have it make a simple single page web site for a company or something like that.
The typical "best practice" for these tools tend to be to ask it something like
"I want you to do feature X. Analyse the code for me and make suggestions how to implement this feature."
Then it will go off and work for a while and typically come back after a bit with some suggestions. Then iterate on those if needed and end with.
"Ok. Now take these decided upon ideas and create a plan for how to implement. And create new tests where appropriate."
Then it will go off and come back with a plan for what to do. And then you send it off with.
"Ok, start implementing."
So sure. You probably can work on this to make it easier to use than with a CLI chat. It would likely be less like an IDE and more like a planning tool you'd use with human colleagues though.
It is typically neither free nor open to develop on consoles. As in, you pay to access developer tools.
iOS and Android less so (even if there is a one time charge for Android and a yearly charge on Apple). OTOH I have not heard of them usually reaching out to more than a handful of devs for promotion purposes.
"People Make Games" on YouTube made some videos about it. Recommended watching for any parent who is trying to figure out if they should allow their child on Roblox. (Spoilers, the answer is "no".)
There are Python bindings for the framework as well.
Personally I could see Go being quite nice to use if you want to deploy something as eg a compiled serverless function.
I'm assuming the framework behaves the same way regardless of language so you could test using Python first if you want and then move over to eg Go if needed.
I have a M4 Max with 128 GB memory. Even on that machine I would not consider 70B+ models to be useable. Once you go below 20 tokens/s it becomes more like having a pen pal than an AI assistant.
MoE models can still be pretty fast. As are smaller models.
(This is mostly a warning for anyone who is enamored by the idea of running these things locally to make sure to test it before you spend a lot of money.)
Currently I'd probably say the Nvidia RTX pro 6000 is a Challenger if you want local models. It "only" has 96 GB of RAM, but it's very fast (1800 GB/s). If you can fit the model on it and it's good enough for your use case then it's probably worth it even at $10k.
And while it has unified memory the memory is quite slow. 250GB/s compared to 500+ for M4 Max or 1800 GB/s for a 5090. So it's fast for a CPU, but pretty slow for a GPU.
(That said, there are not a lot of cheap options for running large models locally. They all have significant compromises.)
Using chip and pin is what makes the system secure. The PIN is verified by the chip on the card so it can't be skimmed. (You can copy the PIN of course, but you also need the same physical chip to clone a transaction.)
Signing the recipe is a useless security theatre. That's why it's not done in Europe.
A while back a listened to a podcast where they interviewed people doing this and it was quite interesting. (And it's an engineering podcast so many technical details.)
https://softwareengineeringdaily.com/2024/05/24/bonus-episod...