iiinova — where your digital self keeps evolving

Rashmi Banthia @rashmib.bsky.social · 25/04/2025

+1 O3 speaks authoritatively (occasionally wrong too)

Ines Montani 〰️ @inesmontani.bsky.social · 31/03/2025

Reposted by Rashmi Banthia

I don't normally write advice columns but this question I got deserved its own post: With all the hype around Generative AI, how can you, as a developer, communicate the nuances and advocate for new and modular solutions that are better, easier and cheaper? explosion.ai/blog/modular...

3258

Rashmi Banthia @rashmib.bsky.social · 16/04/2025

Does anyone know why are the meta-llama/Llama-3.1-8B-Instruct requests on @hf.co are getting denied ? I see lot of posts on the community discussion tab as well ?

000

Rashmi Banthia @rashmib.bsky.social · 26/03/2025

Watch this video on AI Agents - youtu.be/tx5OapbK-8A?... This is actually without the hype ! (Hear what he says about Devin 😆 at 17:30

youtu.be

How to Build Effective AI Agents (without the hype)

YouTube video by Dave Ebbelaar

Rashmi Banthia @rashmib.bsky.social · 25/03/2025

ChatGPT Desktop App , projects feature always uses 4o model and there's no way to change it for the project 🤷‍♀️

000

Rashmi Banthia @rashmib.bsky.social · 20/03/2025

cuDF-pandas (%load_ext cudf.pandas) with Rapids ... work similarly and super cool to see we will be able to speed up scikit-learn

111

Rashmi Banthia @rashmib.bsky.social · 18/03/2025

Yes and less work than creating a notebook 😅

Rashmi Banthia @rashmib.bsky.social · 18/03/2025

I recently shared a Claude Artifact on Kaggle www.kaggle.com/competitions... I really like Claude Artifacts - it uses React, so viz is better and so easy to create (oh and available in free version too) support.anthropic.com/en/articles/...

kaggle.com

BirdCLEF+ 2025

Species identification from audio, focused on birds, amphibians, mammals and insects from the Middle Magdalena Valley of Colombia.

120

Rashmi Banthia @rashmib.bsky.social · 06/03/2025

I was using aistudio.google.com with Gemini 2.0 Flash and default settings on the right. I am not able to upload .ipynb file - I keep getting "The current model doesn't support files of this type." Which model works for you and should I change any settings ?

100

Rashmi Banthia @rashmib.bsky.social · 02/03/2025

I wanted the model to explain code in a .ipynb file (over 21,000 tokens, according to Gemini). It refused the .ipynb file, but I changed the extension to .txt, and Gemini 2.0 Flash explained it and answered all my questions well.

210

Rashmi Banthia @rashmib.bsky.social · 01/03/2025

Impressed with Gemini 2.0 Flash, I gave the same prompt to 4o, O3-Mini, Deepseek R1, and Claude 3.7 - all either denied or hallucinated. Gemini 2.0 Flash provided the answers.

Rashmi Banthia @rashmib.bsky.social · 25/02/2025

Claude 3.7 Sonnet - Did "Ah, I see the issue" become "I see the issue" ? 🤣

Rashmi Banthia @rashmib.bsky.social · 24/02/2025

Congratulations 🎉 I tried Marino notebooks yesterday first time. It looks really nice - dependency graphs, variables list super cool. How do people share these notebooks, without having another person to install Marino ?

100

Rashmi Banthia @rashmib.bsky.social · 21/02/2025

Model size 0.18 MB - Tensorflow GPU memory usage 22GB 🤬🤬🤬

Rashmi Banthia @rashmib.bsky.social · 16/02/2025

+1 for the “from scratch “ , I don’t understand it either otherwise !

Rashmi Banthia @rashmib.bsky.social · 16/02/2025

I haven't tried Huggingface GRPO trainer, They have so many different trainers huggingface.co/docs/trl/grp... Anyways, Thank you for this !

Rashmi Banthia @rashmib.bsky.social · 16/02/2025

I tried it , however I had even less memory (24GB) than L4 so I couldn't do higher batch size. I think this intuition by Claude seems good - share.specstory.com/stories/e32f... Model gets loaded also twice, so its kinda slow.

Rashmi Banthia @rashmib.bsky.social · 15/02/2025

Thanks, I'll change the clip ratio.. I plan to try on my local 4090 - was yours on L4 ?

100

Rashmi Banthia @rashmib.bsky.social · 15/02/2025

Thank you for sharing this, I'll try it !

Rashmi Banthia @rashmib.bsky.social · 12/02/2025

Cursor Agent running commands in terminal is 🔥🔥🔥

Rashmi Banthia @rashmib.bsky.social · 11/02/2025

Reddit’s Top 25 Artificial Intelligence and Machine Learning communities

Rashmi Banthia @rashmib.bsky.social · 30/01/2025

Wow 🤯 my local store has only 5 5090s for sale when they open … and there are probably 100 people lined up

030

Rashmi Banthia @rashmib.bsky.social · 28/01/2025

its under Cursor Settings -> Models

Cursor Settings -> Models (List of available models)

120

Rashmi Banthia @rashmib.bsky.social · 28/01/2025

Gemini is there, last time I checked it didn’t work with composer. Deepseek is also available in the latest version

210

Rashmi Banthia @rashmib.bsky.social · 22/01/2025

marketplace.visualstudio.com/items?itemNa... Really cool extension for cursor.ai - saves all your chat history in .md files 🔥🔥🔥 PS: I have 359 questions in 100+ files 😆

marketplace.visualstudio.com

SpecStory (Cursor Extension) - Visual Studio Marketplace

Extension for Visual Studio Code - (Cursor Extension) Capture, search and learn from every AI coding journey

000

Reposted by Rashmi Banthia

JF Puget @jfpuget.bsky.social · 13/01/2025

NVIDIA’s Academic Grant Program is accepting proposals to accelerate data processing, graph analytics, graph neural networks, operational research, route optimization, and predictive modeling for scientific research using NVIDIA technology. Deadline to apply is March 31: nvda.ws/3ZNxzuW 1/2

184

Rashmi Banthia @rashmib.bsky.social · 10/01/2025

Snowing in Atlanta ❄️❄️❄️

Rashmi Banthia @rashmib.bsky.social · 08/01/2025

Cursor with Claude Sonnet 3.5, I ask it to provide me suggestions/strategies without any code and then I pick one to implement. I find model is too eager to provide me code for every question I ask.

000

Rashmi Banthia @rashmib.bsky.social · 06/01/2025

Anthropic has meta prompt generator - docs.anthropic.com/en/docs/buil... I have used it - works really well - writes structured instructive prompts

docs.anthropic.com

Automatically generate first draft prompt templates - Anthropic

042

Rashmi Banthia @rashmib.bsky.social · 04/01/2025

OpenAI allows you to request "Do not train on my content" privacy.openai.com/policies (Make a privacy request)

privacy.openai.com

Rashmi Banthia @rashmib.bsky.social · 27/12/2024

it could be peft/lora fine tuning also ... may be a separate card for different fine tuning strategies

000

Rashmi Banthia @rashmib.bsky.social · 26/12/2024

😂 And you were able to recover?