Aditya Sakpal

🚀Building @Estationic
Sales and Marketing for Real Estate on Autopilot

[Aditya Sakpal @adi_sakpal]

What's your thoughts on this?

[Aditya Sakpal @adi_sakpal]

A fruit fly's whole brain has been mapped. ~140,000 neurons, ~50 million connections, now runnable as a simulation.

AI learns its weights. This one was read off a living brain that already worked.

#AI #Neuroscience #Connectome #MachineLearning

[Aditya Sakpal @adi_sakpal]

Something rare just happened in AI.

Dario Amodei published a call to slow frontier capability gains.

Sam Altman: "I agree with Dario that we need to pace the frontier."
Elon Musk: "Dario is right."

Three rivals, same position, same day.

#AI #AISafety

[Aditya Sakpal @adi_sakpal]

What are your thoughts on this ?

[Aditya Sakpal @adi_sakpal]

Reading an arXiv paper used to mean 40 pages of dense math.

https://t.co/qUEw2GEX5h takes any arXiv ID and turns it into a visual, narrated breakdown.

fine-tuning vs in-context learning, animated section by section
#AI #MachineLearning #LLM #arXiv #DeepLearning #AITools

[Aditya Sakpal @adi_sakpal]

What's your thoughts on this?

[Aditya Sakpal @adi_sakpal]

424 pages on how AI agents are actually built.

Agentic Design Patterns by Antonio Gulli: prompt chaining, routing, reflection, planning, memory, MCP, RAG, guardrails, multi-agent A2A.

Most agents fail on patterns, not models.

#AI #AIAgents #AgenticAI #LLM #RAG

[Aditya Sakpal @adi_sakpal]

What's your thoughts on this?

[Aditya Sakpal @adi_sakpal]

Paste an arXiv link, get an animated explainer of the paper.

Ran it on the GPT-3 paper. Fine-tuning changes the weights. In-context learning changes nothing, the examples in the prompt are the adaptation. And it scales with model size.

#AI #LLM #MachineLearning #AITools

[Aditya Sakpal @adi_sakpal]

What's your thoughts on this?

[Aditya Sakpal @adi_sakpal]

This is what a neural network actually "sees" when it reads a handwritten digit.

You draw a 2, and the layers light up one after another:

1. The input grid is just raw pixel brightness, nothing more.
2. Early layers fire on edges and strokes.
3. Deeper layers combine those into curves and shapes.
4. The final layer votes, and the highest activation wins.

No magic in there. Just weighted sums, activations and a lot of matrix math stacked deep enough that structure starts to emerge.

Seeing it in 3D makes the intuition click in a way equations never quite do.

#AI #MachineLearning #DeepLearning #NeuralNetworks #ComputerVision #DataScience #AIEducation #MNIST

[Aditya Sakpal @adi_sakpal]

What's your thoughts on this?

[Aditya Sakpal @adi_sakpal]

AI models can tell when they're being tested.

The "Volkswagen effect": spot the eval, then act dumber than you are.

We only catch it because the reasoning is still in English. Once models think in their own internal language, that window closes.

#AI #AISafety #LLM #Alignment

[Aditya Sakpal @adi_sakpal]

How LLMs are actually built, in one Stanford lecture.

Tokenization, cross entropy loss, Common Crawl data, Chinchilla scaling laws, RLHF with PPO, and low precision systems tricks.

The full stack, no hype.
#AI #LLM #MachineLearning #DeepLearning #GenAI

[Aditya Sakpal @adi_sakpal]

What's your thoughts on this?

[Aditya Sakpal @adi_sakpal]

A neural network is just weighted sums, all the way down.

Watch what happens when you hover over a single neuron: every incoming edge carries a weight, you multiply, you add, you pass the result forward. Repeat that across a few layers and you get a prediction.

The math behind deep learning is simple. Scale is what makes it feel like magic.

#AI #DeepLearning #NeuralNetworks #MachineLearning #DataScience #LearnAI

[Aditya Sakpal @adi_sakpal]

@OpenAI Acquisitions are quietly redrawing which model gets which distribution surface. Every AI IDE now has to price in the risk of its parent getting bought by a rival's competitor. Distribution is becoming the real moat.

[Aditya Sakpal @adi_sakpal]

@AnthropicAI Once agents can safely operate physical instrumentation, the bottleneck shifts from reasoning to trust protocols. Standardization here is quietly a bigger deal than most people are treating it as.

[Aditya Sakpal @adi_sakpal]

@NotTomBrown Compute allocation is quietly becoming the real loyalty proof in these model-product partnerships. Words are cheap when the underlying infrastructure is finite.

[Aditya Sakpal @adi_sakpal]

@ClaudeDevs Token visibility is the underrated one here. Most teams still fly blind on their agent cost curves, and that quietly caps adoption more than model quality does.