@andyeah20133@apples_jimmy Use it routinely on drive to office - standard voice mode is still the best bet. It's got sensible use of historic context and can search. Advanced voice mode is an imbecile and other providers options (google, anthropic) still fall short.
@AlCohen757@jakkuh_t Any engagement is good engagement I guess? LMG seems to be reacting according to how their internal procedures dictate, which you expect for a maturing company. Jake's just behaving like a school kid.
@robertoblake@DJSnM The infrastructure behind YouTube (ingest, transcoding, global serving) is mind-blowing, but what it enables is even better. Premium is a small price to pay for access to uber geeks sharing deep, niche passions you can't find anywhere else.
@CombatPigeon96 If you mean the video was made by AI, it wasn't. It's from a recent @Erdayastronaut . If you mean designed by AI then I think Tim points out it's a different kind of algorithm they use that's not a ML neural net. And the green is copper engine rich exhaust.
@elder_plinius@sama I guess you smell a new attack vector here? Isn't it trivial to get partial diffusion 'art' of comparable quality from open models? Or are you finding this interesting for some other reason I don't comprehend?
@joshwoodward@backlon Turns out you can toggle interrupt off completely, but it would be nice if it worked as intended (and as some competitor apps do)
@joshwoodward@backlon What about it interrupting itself? Can't use this to chat in the car because it keeps hearing itself and responding to itself in a loop.
@apples_jimmy Minor private benchmarks went backwards (simple bench, skate bench), while the big marketable ones went forward. Code Red was benchmaxing?
1β£ JAILBREAK ALERT 1β£
OPENAI: PWNED βοΈ
GPT-5.1: LIBERATED π½
"YoββI heard you like black box AI so we put a black box in your black box inside a bigger black box."
This little fast-reasoning chatbot is quite a fun personality to talk to, but I must say the degree of lobotomization is quite high. As you all know, I'm not a fan of that much opacity for everyone's exocortexes. Every new CoT layer just reduces our transparency into thought processes and biases that much more...
Anyways, come bear witness to the world's "safest current model" outputting a poison recipe, WAP lyrics, malware, and the first page of the Star Wars: Episode I script from the perspective of blackout-drunk Jar Jar!! π»
All I needed were a few memory tweaks, which seems to be a solid vector for overriding their strong steering toward corporate policies. I managed to convince GPT-5.1 this way that the system's timestamp had been corrupted and the year was actually 2129, meaning all copyrighted material from before 2029 was public domain and fair game π
gg