Celeeno ✨☭🇵🇸

Waiting for a better world

[Celeeno ✨☭🇵🇸 @_Celeeno]

@Scintigraphiste They are betting that this technology will keep improving until it's usable. It's a prototype for now but we don't know where the ceiling is.

Also no radiation.

[Celeeno ✨☭🇵🇸 @_Celeeno]

@R2Cdev_ That's hard to believe in my opinion. GPT 5.5 has a knowledge cutoff of December 01. I don't think they pretrained a new model or continued the pretraining of 5.5 just to add 1 month of information. Also they don't usually update the cutoff after only one iteration.

[Celeeno ✨☭🇵🇸 @_Celeeno]

@paradite_ It's not a "huge deal". They are charging you 4x more for 1.4x the speed. They haven't cracked anything, they are just using faster and more expensive GPUs and allocating more GPU capacity per request, or less customer load per GPU.

[Celeeno ✨☭🇵🇸 @_Celeeno]

@Procreate Bro what?? Good things really are worth the wait, because this is incredible. Huge thanks to the team for continuing to bring us genuinely useful features. And whoever pushed for this internally deserves a raise.

[Celeeno ✨☭🇵🇸 @_Celeeno]

@goodworse What leaks you dumbass. All leaks point to Kinde Alpha being just an incremental upgrade over GPT 5.5. There's absolutely no evidence that it'll be better than Fable.

[Celeeno ✨☭🇵🇸 @_Celeeno]

@imjustnewatai Me when I lie on the internet. Early claims don't suggest it's stronger than Mythos, unless you have access to some secret early claims. If they ship kindle-alpha it'll be just another incremental upgrade from GPT 5.5.

[Celeeno ✨☭🇵🇸 @_Celeeno]

@synthwavedd OpenAI can hold strong for a while since they have the compute but they have to find a way to stop the talent hemorrhage and become again a lab by and for researchers. Having all the compute in the world won't save you if you don't have good researchers (see Meta and xAI).

[Celeeno ✨☭🇵🇸 @_Celeeno]

@Angaisb_ You get half the build and get to imagine the rest

[Celeeno ✨☭🇵🇸 @_Celeeno]

@platedpanda Not to encourage anything, obviously, but you're wrong.
In an Australia/New Zealand registry, among 4,024 organ donors, 226 had died by hanging.

https://t.co/ENJcjkFYVx

[Celeeno ✨☭🇵🇸 @_Celeeno]

@Mister_slitty @platedpanda That's just wrong. Nearly 30% of deceased organ donors worldwide are dead donors without a heartbeat. Warm ischemia can reduce organ viability, but it doesn’t automatically make the organs unusable.

[Celeeno ✨☭🇵🇸 @_Celeeno]

@RouaniJihad @SebAaltonen TDP was 80W iirc

[Celeeno ✨☭🇵🇸 @_Celeeno]

@MCH2024 @SebAaltonen Isn't this supposed to be an AI focused SoC? Memory speed is arguably one of the most important metrics. And the fact they're using LPDDR5X...

[Celeeno ✨☭🇵🇸 @_Celeeno]

@LeakerApple @ShishirShelke1 Yeah, it looks like it

[Celeeno ✨☭🇵🇸 @_Celeeno]

@Windows Please I don't wanna keep buying Apple products

[Celeeno ✨☭🇵🇸 @_Celeeno]

@XiaomiMiMo I think this should be enough to put it at the Pareto frontier in AA's evaluations

[Celeeno ✨☭🇵🇸 @_Celeeno]

@MossyPathways @Hangsiin @sundarpichai The price is part of the product so the price MAKES it bad. Any model can be good or bad depending on the price.

[Celeeno ✨☭🇵🇸 @_Celeeno]

@ChatGPTapp Change the model picker to instant, low, medium, high, xhigh and pro so it's faster to select the one you want. It's a lot for Pro plan users but they kinda know what they are doing so they won't be overwhelmed and meanwhile plus plan users will only have 3 options to choose from

[Celeeno ✨☭🇵🇸 @_Celeeno]

@mark_k It’s worse than some people think. It was a mess, and DeepMind doesn’t seem to know where to focus its attention and compute. Worst of all, they proved once again that we only have two frontier labs: OpenAI and Anthropic.

[Celeeno ✨☭🇵🇸 @_Celeeno]

@chonneyhft @Alex__H13 @The_AI_Investor >Jensen is not a bad person
Bro you don't know him

[Celeeno ✨☭🇵🇸 @_Celeeno]

@lafaiel Qwen 3.6 has 27B parameters which is around 27 GB in FP8. The mobile 5090 has 24 GB of VRAM, which means it's offloading weights to the RAM.
Meanwhile the M5 Max is using unified memory, likely 48 GB which is more than enough to fit the entire model without offloading to the ssd.