Selective training data, lora fine tuning or MOE are other solutionsZ Sure, creating a model with 100 billion parameters will yield good results, but it’s sort of like employing a million random people to play darts. Or shooting sparrows with A nuclear bomb.
There has been two attacks on the home of Sm Altman in recent days, which he for some odd reason finds perplexing. One could hope that the attack would have a similar effect as Sarah Connor’s visit to the lead engineer at Skynet, but it seems that Altman will just double down. Anyways, I wanted to do a little deep dive into OpenAIs new policy document, which mostly feels like fluff, but had a few things about «responibilty» that caught my eye.
I agree. I remember working as a front end web developer years ago and needing to parse and transform xml/xslt from an api, and spending days and days trying to get it to work. The problem was the version of the IIs web server, version of xslt parser, and version format of the xslt etc. spent days on stack overflow trying to get a working solution. With ai I would have solved that idiotic problem in 1 minute.
Ai is tool, that lets us do what we want to do.
I also remember trying to create my first iOS app in Xcode and thinking «this is simply beyond me». Wouldn’t say app coding is trivial with ai, but its at least feasible now.
The bad part that none of us have a competitive edge anymore, and are close to unemployable. We can’t all be self taught founders who starts our own businesses. It’s going to get weird.
I haven’t made iPhone apps but my guess it that the size comes from bundling libraries and dependencies needed, sort of like how venv includes as snapshot of python and all dependencies into your project folder. Apple core apps probably uses iOS libraries (25GB) already on the device
It’s all fun and games vibecoding until you
A) have customers who depend on your product
B) it breaks or the one person prompting and has access to the servers and api keys gets incapacited (or just bored).
Sure we can vibecode oneoff projects that does something useful (my fav is browser extensions) but as soon as we ask others to use our code on a regular basis the technical debt clock starts running. And we all know how fast dependencies in a project breaks.
Two thoughts:
1 the field of ai research moves so fast that any attempt to make a full documentary would we obsolete long before it was released.
2 all I want ai do to right now is to remove generic «dramatic» music from YouTube clips.