Editing done π
Adding something new in the packet.
And yes uploaded, obviously schedule it at 8 am :)
1
49
Going to try to make videos on YouTube!!!!
Sounds weird for me personally but inner me force me to go try.
I personally feels I should go and make,
Btw,
yes I have made it but not posted yet will post on here also but the best one only which I found ok...ok type, oo byeπ
2
51
Happy Sunday so far :)
- Started learning Kubernetes β completed the first lecture and implementation
- Fixed a major project bug
- Went deep into fine-tuning
One task left before sleep DSA which i'm doing.
@rishabh10x @kirat_tw
Fine-tuning, and what actually happens inside the model
Prompting and RAG donβt change the model. Fine-tuning does. RAG adds knowledge, fine-tuning changes behavior and style.
LoRA freezes the base model and trains tiny adapters....more
Did deep dive into Recursive Language Models RLMs today. Fascinating how they turn context into code variables, helping reduce context rot and enabling work with much larger contexts.
Started working on a project "Code Reviewer" where Iβll be using RLMs.
Thanks to @rishabh10x π
Back on track after getting through the worst phase. We all know what that phase is πΆ
Anyway:
Naive RAG breaks in production because of context rot and distractor noise. To scale AI systems, context must be treated as a scarce resource, not something to fill with tokens.
3
152
LLMs are frozen in time + know nothing about your docs + hallucinate confidently when they don't know. RAG fixes that: Retrieve β Augment β Generate.
Naive chunking alone can tank faithfulness from 0.80 β 0.47. The pipeline is only as strong as its weakest link.
Agents sound complicated. They're not.
Just built one from scratch and it's an LLM, a few tools, and a while loop.
The model writes text, my code parses it, runs the tool, and feeds the result back. Repeat until done.
Demystified.
#AIAgents #Python #LearningInPublic
Built a mini-LLM from scratch on Shakespeare
Used:
β RMSNorm + Pre-Norm
β RoPE
β GQA (8 Q / 2 KV heads)
β SwiGLU
β 4 Transformer blocks + weight tying
β AdamW + gradient clipping
Loss: ~4.17 β Shakespeare-ish text.
Great lecture bhaiya @rishabh10x
LLM architecture since the original 2017 Transformer:
Turns out it's 4 upgrades:
β LayerNorm β RMSNorm (simpler, faster)
β ReLU β SwiGLU (learned gating)
β Sinusoidal β RoPE (relative position via rotation)
β Multi-Head β Grouped Query Attention (4-times less KV cache)
Every AI model β from a simple neural net to GPT-4 β is secretly doing ONE thing: matrix multiplication.
Scalars β Vectors β Matrices β Tensors (literally just stacking dimensions)
And the bonus part it feels too easy to understand when @rishabh10x is your mentor :)
My audio worked perfectly locally but stopped working after deployment.
The issue?
macOS is usually case-insensitive, while Linux servers are case-sensitive.
BTW you can review my portfolio though , it is not fully completed yet :) - ipranav.in
1
65
#5
It was the fun class many things revised first and then learned more concept:
~ Understand Self-Attention
~ Know Q, K, V
~ Know the Attention Formula
~ Understand Context-Aware Meaning
~ Understand KV Cache
~ Understand Multi-Head Attention and more....
#4
Initially it's little bit tough to understand but gradually all are digestible
Just because of @rishabh10x bhaiyaπ
: Text Tokenization Approaches & Their Trade-offs
: Practical Implications of Tokenization
: The Concept of Embeddings
: Semantic Vector Arithmetic... and more
#3
Great class, Learned a lot:
> Forward Pass: Data flows through layers to make a prediction.
> Calculate Loss: Measure exactly how wrong the prediction is.
> Backpropagation: Trace the error backward to assign "blame" to specific weights. And much more....
@rishabh10x π«‘



































