Shubham Kejriwal

camera calibration coming in strong 💪

@animesh_garg Sure, will take a look.

@skpro19 great work. check out, cobalt-teleop.github.io (2025) roboturk.stanford.edu (2018) This will soon be opensourcd, and feel free to contribute.

ARKit is cool

Our internal chat playground is finally open to everyone! It’s still early, so expect some rate limits and slower speeds. No fluff, just the models, the best way to see what the tech can really do. Try it here: chat.maruthlabs.com

We’ve been working on a small 150M parameter model called Madhuram-v0.5. It’s a model designed to be finetuned for specific tasks. We're quite happy with how it holds up. Here is a comparison of our Madhuram-v0.5-Base with other models of similar sizes.

Boy, those are a lot of tokens! @cursor_ai

@leandroOnX Where is this clip from?

Placed my first Zepto order using the ChatGPT `Agent`. Took me 10 minutes 😅

@ericzakariasson I would rather use `Gemini 2.5 Pro` given how expensive Sonnet is. Also, `o3` seems to have a propensity to use up a lot of tokens.

o3 for research and planning and sonnet 4 for implementation is the best combo right now. it’s not even close

@ericzakariasson @xai Because of the ElMo `tweet`?

@nabilajamal_ Aren't dividers supposed to be 'in the middle of the road'?

Who approved a divider that suddenly starts in the middle of the road? Total setup for disaster!! Engineer or authority responsible needs to be held accountable. This is criminal road design *Speeding car slammed into a divider and overturned on #Bhopal's Subhashnagar flyover. Both front tyres ripped off, two young passengers injured, hospitalised

i don’t think you understand you train a small model on scikit-learn it gets 80% accuracy you fine-tune a huggingface transformer it gets 90% accuracy you build a custom data pipeline it gets 93% accuracy you ensemble three models with voting it gets 95% accuracy you throw in feature engineering and SHAP it gets 96% accuracy you replace it with a gradient boosting tree it gets 97% accuracy you build a custom loss function for your niche problem it gets 98% accuracy you realize the labels were wrong fix them, retrain it gets 99% accuracy you make a tiny architecture change it beats human baseline you remove half the parameters it runs in real-time you quantize the model it runs on a potato you distill it it fits in a JavaScript bundle you deploy it to edge it autocompletes before users even type you realize you haven’t written a paper but you built god i don’t think you understand

@maharshii @birdabo Why not a Macbook pro?

@birdabo lenovo legion but with RTX 4060 with 8GB VRAM, local models go brr

hi from the new laptop

Can we please add a `branch` feature in ChatGPT web app already? @ChatGPTapp

@arindam___paul Hey, any plans for a Reddit AMA?

Will be launching our 3rd major category after fans and mixer grinders later this month ( smart locks are a small category) Perhaps the strongest use case of IOT seen in appliances/electronics In a category where most consumers “feel cheated “ Any guesses

@Rainmaker1973 What is that transparent blob?

When chemistry paints its masterpiece

@ericzakariasson @cursor_ai No honourable mentions for `o4-mini` ?

we get a lot of questions about which model to use in @cursor_ai, so we put together a guide on how you can think about selecting models based on what we've seen work well

@AravSrinivas How about chip design / fabrication?

At this point, you can use any AIs (perplexity, chatgpt, etc) as a PhD advisor on whatever topic you're getting deep into. It's pretty good. I have my preference. The core point: Research advice is no longer an elite academic university thing.

@ericzakariasson @cursor_ai Difference between auto-attached and always?

last week i posted about how to work with large codebases in @cursor_ai which gauged a lot of interest. we've now turned this into a long form post that goes more in depth see link below

@ryolu_ @ItHowandwas Makes sense. What about 'Auto Attached'?

@skpro19 @ItHowandwas agent will use it at will vs agent will always use it (almost like a system prompt)

Using Cursor well = fast, clean code. Using it wrong = AI spaghetti you’ll be cleaning up all week. Here’s how to actually use it right: 1. Set 5-10 clear project rules upfront so Cursor knows your structure and constraints. Try /generate rules for existing codebases. 2. Be specific in prompts. Spell out tech stack, behavior, and constraints like a mini spec. 3. Work file by file; generate, test, and review in small, focused chunks. 4. Write tests first, lock them, and generate code until all tests pass. 5. Always review AI output and hard‑fix anything that breaks, then tell Cursor to use them as examples. 6. Use @ file, @ folders, @ git to scope Cursor’s attention to the right parts of your codebase. 7. Keep design docs and checklists in .cursor/ so the agent has full context on what to do next. 8. If code is wrong, just write it yourself. Cursor learns faster from edits than explanations. 9. Use chat history to iterate on old prompts without starting over. 10. Choose models intentionally. Gemini for precision, Claude for breadth. 11. In new or unfamiliar stacks, paste in link to documentation. Make Cursor explain all errors and fixes line by line. 12.Let big projects index overnight and limit context scope to keep performance snappy. Structure and control wins (for now) Treat Cursor agent like a powerful junior — it can go far, fast, if you show it the way.