Varun Joshi

ML @ Apple | UT Masters | UTD Alum | All opinions are my own!

[kjcodes_ @kjcodes_]

tried using my own harness novus over claude code to create a portfolio and in just one prompt the result was substantially better

[kjcodes_ @kjcodes_]

would anyone be down to try out my new product its in early beta

[kjcodes_ @kjcodes_]

Reinforcement learning has libraries.
What it’s missing is workflow discipline.

Shipping RLCLI today.
1. Local-first
2. Reproducible
3. Built for real experiment iteration

GitHub: https://t.co/63AJqOPOhG
Install: https://t.co/OPFFBGSDD1

[Varun Joshi @domxjoshi]

@kjcodes_ runpod is pretty good idk if it’s the cheapest. Thunder compute markets itself as the cheapest

[kjcodes_ @kjcodes_]

you can try playing as the player yourself against any of these levels. been trying to beat level 2 for the past hour buts its too fast. would love to see someone try beating it.

source code - https://t.co/Xi4hpUZRaR

command - python play_human.py --level 2 --seed 0 --fps 60

[Varun Joshi @domxjoshi]

@kjcodes_ Amazing stuff

[Varun Joshi @domxjoshi]

@kjcodes_ more chance of overfitting when changing billions of parameters in full ft, might shift the models pretrained structure, can run into catastrophic forgetting and might have waited hours for it. Full ft better for high quality data and if target task far from pre trained dist.

[Varun Joshi @domxjoshi]

@ananayarora Yeah that’s why using pickle files in enterprise isn’t recommended

[Varun Joshi @domxjoshi]

Balancing autonomy with reliability in AI agents is important. MCP really is a step in the process.
Complete game changer! #AIAgents #MCP #Anthropic

[Varun Joshi @domxjoshi]

@PicturesFoIder And somehow we are responsible!

[Varun Joshi @domxjoshi]

@sama Steve Jobs moment right here