Muhammad Usman

Due with Data | Data Science | Machine Learnin | Artificial Intelligence

[Muhammad Usman @md_usman73]

If you're enthusiastic about learning Machine Learning and Reinforcement Learning, you should definitely check out @rasbt YouTube playlists.

The teaching style is exceptional, and the explanations are backed by practical, easy to follow examples. The quality of the content level

[Muhammad Usman @md_usman73]

Earth men have a talent for ruining big and beautiful things

[Muhammad Usman @md_usman73]

Which one Library you choose for Deep Learning!

Tensorflow or PyTorch
Google Product or Meta Product

[Muhammad Usman @md_usman73]

Vibe testing; means casually trying an LLM app with a few prompts and judging it by feel. Basically I asked it 6 to 8 questions, answers looked good, so I think it works..

[Muhammad Usman @md_usman73]

Ever thought about what “Large” means in Large Language Models?

As @rasbt defines it in his book: “Large” refers to two things;

1: The model’s size in parameters (billions of weights)

2: The massive dataset it’s trained on (trillions of words)

[Muhammad Usman @md_usman73]

@py_poojaa RAG is not dead lol, basic RAG is just becoming outdated.
The focus is shifting toward agentic RAG, better retrieval, and smarter context management.
Fine tuning and long context models have their place, but they don’t eliminate RAG.
RAG is not dead, it’s evolving…

[Muhammad Usman @md_usman73]

Tesla just Launched Roadster: The Flying Car

#roadster

[Muhammad Usman @md_usman73]

Idhar Pakistan mn Alto car abhi aayi hy or udhar Elon Musk ny flying car test kardi

[Muhammad Usman @md_usman73]

Extreme level of Muse Meta AI agent

Relax, because your agent is working..

#meta #MuseAI

[Muhammad Usman @md_usman73]

Docker container
Docker files

[Muhammad Usman @md_usman73]

Why Persistence is important for AI application?

Without persistence, LLMs would lose all session history after every restart. By persisting thread data and user context in external databases, modern ai applications deliver seamless long term continuity.

#LLMs #ai

[Muhammad Usman @md_usman73]

What is Persistence in LLM application?

We all know the base LLMs are stateless, they forgot everything once a request ends. Persistence is the technique of savings chat sessions or threads, user states or context into external databases..

#persistence #llms

[Muhammad Usman @md_usman73]

4_Context Drift / Hallucination: Pushing too much irrelevant background text into the context window creates noise, increasing the chance of hallucinations or off-topic responses.

[Muhammad Usman @md_usman73]

Why does Short Term Memory fail?

While short term memory for short back and forth conversations, it fails as chats get longer due to four main reasons:

Here are the four main reasons discussed in comment section:

#stm #ai #llms

[Muhammad Usman @md_usman73]

High Compute & KV Cache Costs: Attention processing scales quadratically (O(N2)). Re-reading thousands of words of chat history for every single message makes inference slow and expensive.

[Muhammad Usman @md_usman73]

2_Hard Limits & Truncation: Context windows have token limits (e.g., 8k, 32k, 128k tokens). Once a conversation gets too long, older messages are dropped or cut off, causing the AI to "forget" earlier instructions.

[Muhammad Usman @md_usman73]

1_"Lost in the Middle" Effect: Self attention mechanisms focus heavily on the beginning (system instructions) and the end (your latest question) of the context window. Critical details hidden in the middle of a long chat often get overlooked.

[Muhammad Usman @md_usman73]

@rasbt Thank you for sharing your perspective sir, Always great learning from your insights on AI architecture.

[Muhammad Usman @md_usman73]

@rasbt sir mentioned in the book “ Build a Large Language Model from scratch” LLMs can generate surprisingly coherent, human like text. But “understanding” in AI isn’t consciousness or true comprehension. It’s advanced pattern recognition and contextual processing..

#llms

[Muhammad Usman @md_usman73]

@YashHustle_22 Advance AI

[Muhammad Usman @md_usman73]

@YashHustle_22 Hardest question 🙋