Google Antigravity Built an OS from a single prompt
antigravity.googleInstalling Antigravity 2.0 breaks the Antigravity IDE due to an amateur Electron deployment mistake. If they can't even get a basic installer right, how are we supposed to trust an OS generated by it? The 2.0 installer lazily drops a new app.asar next to the old app/ folder, creating a loader priority conflict that hijacks the original IDE executable
I've been noticing this type of thing a lot lately, updates that plain don't work because of bugs that would have been caught with any testing, installers that don't work at all..
It's like coding with AI and not testing anything go hand in hand, which I do not understand.
I remain completely unimpressed by marketing announcements that boil down to "the LLM wrote a toy version of something for which there are hundreds of undergraduate GitHub repositories in its training data", even if style transfer was involved, as was the case with the Claude C compiler.
This wasn't possible one year ago. It's not that exciting to recreate something that already exists but being able to reliably write boring code is what most customers need.
So, just a small step up from the same sort of thing I wrote for an AVR once. Just enough for some preemptive multitasking and not much else.
They didn't say anything about what the OS does besides "it can play doom" (which is a single threaded program). For all I know, this could be an even more basic version of dos.
> goes to show just how intrinsically intelligent the Gemini family of models have gotten, even on Flash (Gemini 3.1 Pro was unable to do this)
3.1 Pro couldn't do it.. but why ??
Where is the code? Did I missed the link?
"make && cp vmlinux ailinux"
BTW does the GPL apply to code generated by models trained on GPL code?
lol paid 1k to play doom with worse performance probably