Vijay @EqualsAI

Intelligence in the AI era - Ability to infer new possibilities using existing knowledge and new real time information.

My Minikube cluster is now running a production-style LLM stack on consumer grade NVIDIA GPU. Here is the breakdown of milestones achieved 1. Minikube + NVidia GPU Operator setup for sharing GPU to multiple pods 2. vLLM multi-model hosting (Gemma + LLaMA) 3. Local Docker…

keyboard_arrow_left Previous keyboard_arrow_right Next

When you started thinking, you started growing.

Multi-LLM serving on a single GPU cluster on local. 1. Minikube + NVIDIA GPU Operator 2. vLLM CUDA-enabled pods 3. Local Hugging Face model mounts 4. GPU time-slicing for concurrency 5. Ingress for unified API access Already fluent in Docker/K8s/Ingress,…

keyboard_arrow_left Previous keyboard_arrow_right Next

LLM deployed on the local was trained on the entire internet data. Can you work offline and search what you want through model inferencing?

Failures of AI projects are due to knowledge gaps and inaccurate data.

Full Fine-tuning Gemma-270M-it with Effective Batch Size = 24 across different LR scheduler types. Key takeaways from the experiment: 1. Eval Loss stays ~1.20 across scheduler types 2. Train Loss rises with certain schedulers. 3. GPU VRAM ~13.9 GB consistently Even small…

Experimenting with Gemma-270m-it model and gradient accumulation steps. Carried out the Full PFT runs with higher effective batch size which led to better fine-tuning performance and lower loss. Plotted Batch Size vs Loss & VRAM :- Created custom model trainer python script…

An AI version of interstellars. We don't need more LLMs, we ran out of algorithms. We need that God's equation to achieve ASI.

Just finished an experiment of Full PFT with Gemma 270m instruct model on consumer grade GPU. Purpose of the experiment was to observe the effect of batch sizes on metrics such as eval loss, train loss, VRAM requirement. Plotted Batch Size vs Loss & VRAM → clear scaling…

Even if the AI bubble bursts, the revolutionary AI technology is going to stay in the future.

Without reinforcement there is no improvement in the AI systems.

Most people don't know the details of the Internet and HTTP but they used it to create breakthrough ideas which generated wealth. The same will be with artificial intelligence.

Human minds always want to evolve through reinforcement learning. Exploratory minds are innovators, and Exploitation minds are efficiency seekers. One who managed to both explore and exploit is exceptional.

Reinforcement learning is the closest to the kind of learning that humans and other animals do.

What are the benefits of the image or video generation models?

Future of jobs and the economy will be completely different from current predictions.

Reinforcement learning is essential for both humans and machines to learn and evolve.