Sushma Rao

🔍 Breaking down AI nuances and complex topics to simple daily insights on X.
⚡️ Engineering Agentic AI systems.
🤖 AI-native •SaaS
📩 Open for Collabs

[Stijn Noorman @stijnnoorman]

Say no to bad clients.

The money isn't worth the headache.

[Sushma Rao @magicwithAI]

https://t.co/ZrV9YabTfP

[mal @mal_shaik]

https://t.co/3Rzi5WKGV2

[Sushma Rao @magicwithAI]

Great post.
Highlights that we can use, to implement in the code we design.

[rvivek @rvivek]

I asked Claude to analyze the leaked Claude Code source code.

512,000 lines across 1,900 files.

5 things about how it actually works that most people don't understand 🧵

[Nikhil Kamath @nikhilkamathcio]

If AI is going to change the world, at this minute, (for however long) Dario is on the absolute top of this new world.

One thing this has left me asking is, for all the AI companies which aren't really profitable yet and have cukoo valuations;
if they do make most jobs redundant, including high paying ones from consultants to programmers, to doctors and engineers, where does the revenue growth come from & who is the new consumer in the capitalistic world we live in?

Capital can no longer be a Moat if increased productivity, inevitably leads to deflation..

[Paweł Huryn @PawelHuryn]

https://t.co/rnkK6kImP0

[Sushma Rao @magicwithAI]

Input & Embedding: Turning Words into Numbers https://t.co/C0NCOw1FjY via @telestreak

[Sushma Rao @magicwithAI]

https://t.co/4VmUeF86Wg

[Sushma Rao @magicwithAI]

Zero-Shot Learning: Model performs tasks it wasn't explicitly trained for! Uses general knowledge to handle new scenarios. LLMs excel at this. No examples needed! 🎯 #LLM #MachineLearning

[Sushma Rao @magicwithAI]

Vanishing Gradient: Gradients become extremely small in deep networks, stopping learning. Major problem in early deep learning. Solutions: ReLU, LSTM, residual connections. 🌊 #DeepLearning #Training

[Sushma Rao @magicwithAI]

Unsupervised Learning: Learning from unlabeled data! Finds hidden patterns without ground truth. Clustering & dimensionality reduction are common. Exploratory & discovery-focused! 🔍 #MachineLearning

[Sushma Rao @magicwithAI]

Underfitting: The model is too simple to capture data patterns. High error on both training & test data. Opposite of overfitting. Solution: more complex model or better features! 📉 #MachineLearning

[Sushma Rao @magicwithAI]

Transfer Learning: Using knowledge from one task for another! Pre-train on a large dataset, fine-tune for a specific task. Saves time & data. Foundation of modern NLP & computer vision! ♻️ #DeepLearning

[Sushma Rao @magicwithAI]

https://t.co/Ltx5iZCc8g

[Sushma Rao @magicwithAI]

Supervised Learning

Learning from labeled data! Model learns mapping from inputs to known outputs. Classification & regression are supervised. Most common ML paradigm. Needs ground truth! 📚 #MachineLearning

[Sushma Rao @magicwithAI]

Softmax

Converts raw scores into probabilities that sum to 1. Used in a multi-class classification output layer. Higher score = higher probability. Final decision maker! 🎲 #DeepLearning

[Sushma Rao @magicwithAI]

Stop Words

Common words removed during text preprocessing (the, is, at, and). They carry little meaning. Reduces noise & dimensionality. But context matters - sometimes they're important! 🚫 #NLP

[Sushma Rao @magicwithAI]

Sequence-to-Sequence

Maps the input sequence to the output sequence. Translation, summarization, speech recognition. Encoder-decoder architecture. "Hello" → "Bonjour". Powers many NLP tasks! 🔄 #NLP #DeepLearning

[Sushma Rao @magicwithAI]

Semi-Supervised Learning

Uses both labeled & unlabeled data for training. Best of both worlds! Leverages abundant unlabeled data with limited labels. Practical for real-world scenarios. 🎯 #MachineLearning

[Sushma Rao @magicwithAI]

@rvivek @albinder @letsblinkit Amazing tech. Reactive vs proactive, AI predictions on sentiment, news, and other related events, Kafka streaming, microservices, Redis, and locking, Weather configured, and much more.