let users give more access to gemini on mobile. i want the same experience as linux. it should manage my files, optimize or change settings, navigate the phone hands free for me, check my emails, find an old photo, get my contact's address and set it as a destination, find the best cheeseburger on the way, play the last audiobook i was listening to, track my health metrics, and then use all of that to be a good life assistant.
several of those are real failures ive seen on things i know it can do...sometimes. when it wants to? i know some of it is technical misses, some of it is the model not even realizing it can do things eg carbon told me last night it couldn't create a google doc. for the love of god please rl or sft or make a little lora hat for self capabilities. its sometimes surprising in a bad way how bad these frontier models are at using themselves.
gemini in android auto was a big win, and if anybody is gonna make an "everything" app work its the company who can integrate it into the most popular mobile OS, in an unintrusive and helpful way. because an everything app kinda ends up looking like an OS as you brainstorm features and use cases, and you already have that. no new app necessary.
and uhh i dunno something about glasses?
@tn_sanatani@grok Jensen out. Xi in. China gets the entire world supply of chips and the guy who turns them into world class GPUs. Is that what's happening here?
I wrote about similar findings in https://t.co/FEYfT26EeB but flipped - i focused on unembedding voronoi tessellations! In my tests margin max ended up beating Fisher for improving accuracy, and as bonus it's much cheaper in compute.
Of course I separated top k which is the opposite of what you'd want but you could flip it and boost the top advantage key by penalizing the raw score gap, with a gate to protect your heads.
Good luck!
I don't know if people have thought this through but...if weight space programming gets solved, it's basically AGI all at once.
Imagine Astra wiring Astra. No more weeks to months of training and RL. You set up a validation loop, iterate with 1000 instances at a time, take the best one, and it's done.
You want safeguards? Wire them in, only takes a few extra minutes.
Native video? Computer use? Chip design? No problem.
Run out of free dims? Unfreeze a layer or just add some new parameters!
Not only that, but any model of any size will be able to do whatever you need it to do in a very short amount of time, without the mountains of GPUs or training data. Qwen releases will get a trained programmer published a few days later that can immediately find and fix the problems you have, tailored to your exact needs.
This is huge.
@PeterDiamandis I was just thinking today, giving residents of a town a small stake in the earnings of each data center might be a nice baby step towards UHI
@Sauers_ Yep I did it first thing after the announcement too. When i heard "Fable reset" I thought it meant here's a reset because Fable 5.2 was released
@morganlinton@VulcanBench the first thing i noticed was the speed as well. simple tasks get quick answers, it doesn't overthink things and seems very well calibrated. so far using fable for every lane i was running opus, usage looks similar and there's just a bit more common sense in its choices.