Gursimran Bhullar

[Gursimran Bhullar @isimmabhullar]

@EricFriedman I think it also because of free usage. None of the other provide free VMs for free users. Even the limits for the $20 sub ones are limited so no one wants to pay $100 to do same thing that Muse can do effortlessly for them. Also less setup required.

[Gursimran Bhullar @isimmabhullar]

@quxiaoyin Apple and the others are so stuck on selling hardware to push services, but AI agents are about to kill that whole vibe. Once stuff like @AIatMeta smart glasses really takes off and we stop needing screens ( 2035), the whole "app store" economy is basically cooked.

[Gursimran Bhullar @isimmabhullar]

@Karmedge I am more bullish on @Meta , than @Apple. Why would you need a phone if AI can do most stuff, except content consumption. Once you use Muse through Meta glasses, why would you use your phone? I think Apple can be much more successful in Enterprise Environments.

[Gursimran Bhullar @isimmabhullar]

@KairosPraxis You cant ask it to browse e-commerce sites specially Walmart and stuff because it will easily give captchas and then their model will refuse to solve them on their own. Muse doesnt get any captchas and their prolly breaking all the ToC, but user doesn't care.

[Gursimran Bhullar @isimmabhullar]

@startupoppa @vidythatte They can't, their ai team, vertex team, gemini team, ai studio team, all are separate teams who seems have no clue what android users or google ecosystem needs from AI perspective. Meta has @finkd who can align all the teams for one common goal. Google missed it, so far.

[Gursimran Bhullar @isimmabhullar]

@fabknowledge Because just like threads the onboarding is super smooth, just press one button to select the meta account you already logged on another apps and you are in the chat with AI. No passwords or questions asked, Also the 1B free token is working like charm.

[Gursimran Bhullar @isimmabhullar]

@tobi Waiting on Android. Seems like launched on App store for now

[Gursimran Bhullar @isimmabhullar]

@da_fant Also it would make safer to direct prompt injection, there naty be new methods invested in future. But so far jev seems not easy to prompt inject.

[Gursimran Bhullar @isimmabhullar]

@kinglycrow LLMs has to reason which would always be latency bound, not sure how reasoning is done on jev side, but it seems to be just answering without any delays, and cost is to cheap to meter.

[Gursimran Bhullar @isimmabhullar]

@NotPro2XL Bro Camera processing is bad, I am real bad. Currently using pixel 10 pro xl. It can't take 5 Full res images in row without disabling the capture button for 3s. The hardware is so mediocre that it can't support the features they promote, same for 8K video.

[Gursimran Bhullar @isimmabhullar]

@haider1 It's never about SOTA or small model. It's about whether you can let model be free and trust it. Right now Google's models are not there. Opus been there since 4.6 and Fable is even better, same for OpenAI and Chinese SOTA models.

[Gursimran Bhullar @isimmabhullar]

@DannyWestside6 Also they don't know how to or doesn't have models to do so yet. For e.g. aistudio let's you create Web apps/Sites with iterative improvements for free but I don't see anybody using it except tech twitter may be. I think AI certainly will put downward pressure on $20 Saas subs.

[Gursimran Bhullar @isimmabhullar]

@DannyWestside6 In theory it is true they can do it, but they won't in near future unless models become deterministic. The cost to let a consumer create their own custom app is high given their distribution size. Until their local models can do it, they won't.

[Gursimran Bhullar @isimmabhullar]

@jun_song Totally Disagree! GLM5.2 (Chinese SOTA) churns through tokens like crazy, it works but only after overthinking same problem that Grok4.5 solves much efficiently. Using SuperGrok,https://t.co/QE0yn5PyzY Pro plan, both are lasting me same amount of time, Grok 3x cheaper tho.

[Gursimran Bhullar @isimmabhullar]

@sudoingX @LarryAGuy1 Currently getting 90-100 Tok/s using unsloth qwen 3.6 35B-A3B IQ4 gguf on my p520 with dual Radeon Vii 16G around 100k context and MTP *2

[Gursimran Bhullar @isimmabhullar]

@bendee983 I think you are reading AA wrong. It is similarly worse on AA as well. AA chart depends upon model selected for comparison. You can select few which will come at 56 and would set Flash to 9-10 place.

[Gursimran Bhullar @isimmabhullar]

@scaling01 I think it's becoming more about data now. Google has good amount of media data from Photos and maps etc. But they don't have anything related to code except from general internet. So their multimodality is good but worse on code as compared to OpenAI and Anthropic

[Gursimran Bhullar @isimmabhullar]

How do say no to this?
@Shopify

[Gursimran Bhullar @isimmabhullar]

@LottoLabs Kimi 2.6 is extremely slow, specially with Hermes as it has big system prompt. I had to custom tweak the timeouts to at least 5 minutes to work with frontier models on Ollama. GLM 5.1 is quick when it works though. But Sub is still a good value considering amount of token you get

[Gursimran Bhullar @isimmabhullar]

@LottoLabs Tried Ollama pro, it's better than OpenAI or Anthropics $20 plan. Only quirks are the latest model will be slow and often timeouts. So if you are okay using GLM 5.1 or deepseek4 flash not pro, you can essentially hook up with Hermes and would never run out of limits.