Bharat Singh

Principal Scientist at Adobe Firefly | Ex Cruise/Amazon, PhD, Computer Vision @UMDCS

[Bharat Singh @bharat_thoughts]

@sonyatweetybird throwing like 16-32 B200s with ring attention, sequence parallel and parallel vae decode , it’s possible with any model, would be interesting if the cost is significantly less. Hosting costs increase when you use GPUs as utilization drops so people don’t do it

[Bharat Singh @bharat_thoughts]

@CalvinMccarter in large scale training for image/video generation you barely see an epoch, so what is better, seeing the whole dataset or seeing half of it ? Might be good to test this on a larger dataset where you have like billion + examples

[Bharat Singh @bharat_thoughts]

@paulg @ycombinator now it will take 101 more months to get its FDA approval

[Bharat Singh @bharat_thoughts]

@elonmusk I used it all day yesterday instead of opus4.7 and it’s very good, didn’t miss the bigger slower model. The speed at which it runs is amazing.

[Bharat Singh @bharat_thoughts]

@om_patel5 This is true everywhere , everyone is an agent manager

[Bharat Singh @bharat_thoughts]

@mark_k @midjourney doing VLM conditioning doesn’t make it AR

[Bharat Singh @bharat_thoughts]

@mark_k @midjourney nope 😂

[Bharat Singh @bharat_thoughts]

@paraschopra make it write flash attention which supports masks

[Bharat Singh @bharat_thoughts]

@cryptopunk7213 as if anthropic checks the license of each github repo it uses for pre-training, many of those could be written by chatgpt even if the license was fine

[Bharat Singh @bharat_thoughts]

@aakashgupta The throughput is the same, so hosting costs don’t change. Also, how do you fit a 500B parameter model on 800 mm of silicon ? That would need 10x scaling of moore’s law

[Bharat Singh @bharat_thoughts]

@bindureddy I want to cancel YouTube and Netflix , that’s what I pay for, any way to do that?

[Bharat Singh @bharat_thoughts]

@karpathy Knowing an AI programming workflow today is >> valuable than how good of a programmer you are

[Bharat Singh @bharat_thoughts]

@mark_k What makes you think Gemini 3 pro image generation is not a diffusion model ?

[Bharat Singh @bharat_thoughts]

@Voltaje9 @ElonClipsX If you do end to end fusion with transformers it’s same as using multiple cameras , if you have parallel models then you will have issues

[Bharat Singh @bharat_thoughts]

@Voltaje9 @ElonClipsX Then LiDAR fills out areas in fog where cameras don’t see

[Bharat Singh @bharat_thoughts]

@jasondeanlee If you use the thinking version and are paying $20, that $ is still better spent on chatGPT

[Bharat Singh @bharat_thoughts]

@jasondeanlee I think this is true in general because Gemini3’s thinking budget is very low compared to GPT5.1. They nerfed it because probably it was too expensive . OpenAI is still ahead in the game.

[Bharat Singh @bharat_thoughts]

@ClementDelangue @Tinder @RobinhoodApp @NotionHQ @figma @Airbnb @DoorDash @Snapchat @stripe Also, no one ever releases pre-training, infra or data recipe in both cases because it’s all built on data which neither have rights to train on. If they did, they would get sued next day…

[Bharat Singh @bharat_thoughts]

@ClementDelangue @Tinder @RobinhoodApp @NotionHQ @figma @Airbnb @DoorDash @Snapchat @stripe If you are not sota, doesn’t hurt to release. Atleast that way you get some visibility. US models today are still sota, so makes no sense to release. If they lag behind, they might do so.