Yash Dugriyal

some weeks back I made a thing github.com/0xV0YD/Mnemosy… because I kept losing the thread between agent chats it stores the conversations along the way, organizes them into useful memory, and uses retrieval algorithms to remember decisions we made earlier when a similar topic comes up again. not limited to Codex either: plug it into multiple IDEs/editors through MCP and the memory layer can travel with the work. local-first, searchable, source-cited memory for coding agents. u can try it out

Explore enough to learn Exploit enough to earn

@goyalayus the milk finally arrived

@0xVoyd proud of you son

so the difference is that DP already knows what moves it can make then it figures out corret possible ways while RL also have to fiugure out the possible moves along with the path as the possible moves will be the ones which would lead to the correct goal

the value of now depends on reward now plus value of later

lets see our codex lifetime usage till now

5 sept milenge bol aaiga

so gradient tells us whether loss will increase or decrease if we move that particular parameter in a specific direction and thats how backprop. uses the gradient to decide the updation direction of the parameters

how beautiful the way `fork`, `exec`, and `wait` work together is.

x→y^​→L L→dL/dy^​→dL/dw​,dL/db

Can I ask a dumb question... Why do most vibe coding success stories involve web apps?? You rarely see people vibe coding: • Database engines • Linux kernels • Compilers • Distributed systems What's different about web development that makes it such a good fit for AI-assisted coding?

Let's model this input–output relationship: y = 0 for x ≤ 0 y = x for 0 < x ≤ 2 y = 2 for x > 2 In words: • Before x = 0, output stays at 0. • Between 0 and 2, output increases with x. • After x = 2, output stops increasing and remains 2. we can build this using two ReLU neurons. ** Neuron 1 h₁ = ReLU(x) which means h₁ = 0 for x ≤ 0 h₁ = x for x > 0 Neuron 1 begins increasing after x = 0. ** Neuron 2 h₂ = ReLU(x − 2) which means h₂ = 0 for x ≤ 2 h₂ = x − 2 for x > 2 Neuron 2 remains inactive until x passes 2. Combining the neurons y = h₁ − h₂ Therefore, y = ReLU(x) − ReLU(x − 2) Now each region Region 1: x ≤ 0 Both neurons are inactive: h₁ = 0 h₂ = 0 Therefore, y = 0 − 0 = 0 Region 2: 0 < x ≤ 2 Neuron 1 is active, but neuron 2 is inactive: h₁ = x h₂ = 0 Therefore, y = x − 0 = x Region 3: x > 2 Both neurons are active: h₁ = x h₂ = x − 2 Therefore, y = x − (x − 2) = x − x + 2 = 2 The second neuron cancels the further increase created by the first.

so basically activation functions are used to show the different relationships between input and output in a single input range

just realised how RAG is based on embeddings

goal is one of the craziest feature I explored recently which enables my eating and sleeping time also to be fruitful

@thedankoe strongly agreed

If you don't have a goal so meaningful it makes other people's opinions irrelevant, you will lose control of your life. You will adopt the goals assigned to you by your parents, peers, or society.

dL/dX , dL/dY , dL/dW -> crux of understanding NNs and BackPropgtn

gm “ If it makes you happy - take the risks… “