2024 04 01 experiments

Experiments with Agents

The ideas that you can now experiment and build with Open weight models like mistral and gemma on single card GPU is mindblowing. Current goal is to build small applications for everyday which are run locally. Keep building everyday and making the Agents smarter.

Hyperparameter tuning for Prompts is the current bottleneck to solve, How does one make the best of current resources. The 7B models are quite good for simple tasks, utilise them consciously. Constraints bring out the best application.

When your access to GPU cluster is cut-off over the weekend. The wait becomes long to experiment new ideas/prompts into making the agents complete the task on first attempt. The road to Mars, is now possible. Its only a matter of when and not IF. keep crushing ahead