---
title: "Ankita Maji (@ankitamaji) | Hashnode"
description: "From silicon to softmax — ML systems, LLM internals, and research at the edge. Read the latest articles by Ankita Maji on Hashnode."
url: "https://hashnode.com/@ankitamaji"
---

@ankitamaji

# Ankita Maji

@ankitamaji · Chandrapur, India · Joined August 2026

From silicon to softmax — ML systems, LLM internals, and research at the edge.

### [AI Is Now Training Other AI. Here’s Why That’s Both Brilliant and Terrifying.](https://silicontosoftmax.hashnode.dev/ai-is-now-training-other-ai-here-s-why-that-s-both-brilliant-and-terrifying?utm_source=hashnode&utm_medium=feed)

How Constitutional AI swaps human raters for the model itself, what RLAIF means for alignment, and what's still unsolved at DeepMind and Anthropic.

7 min read · [silicontosoftmax.hashnode.dev](https://silicontosoftmax.hashnode.dev/ai-is-now-training-other-ai-here-s-why-that-s-both-brilliant-and-terrifying?utm_source=hashnode&utm_medium=feed)

[1](https://hashnode.com/posts/ai-is-now-training-other-ai-here-s-why-that-s-both-brilliant-and-terrifying/6aa90d3e6c3789f27339a01c) [Sep 15](https://hashnode.com/posts/ai-is-now-training-other-ai-here-s-why-that-s-both-brilliant-and-terrifying/6aa90d3e6c3789f27339a01c)

### [Why AI Benchmarks Are Broken — MMLU, HumanEval, and Chatbot Arena](https://silicontosoftmax.hashnode.dev/why-ai-benchmarks-are-broken-mmlu-humaneval-and-chatbot-arena?utm_source=hashnode&utm_medium=feed)

MMLU is contaminated. HumanEval is saturated. Chatbot Arena has selection bias. how AI benchmarks fail and what actually matters instead

7 min read · [silicontosoftmax.hashnode.dev](https://silicontosoftmax.hashnode.dev/why-ai-benchmarks-are-broken-mmlu-humaneval-and-chatbot-arena?utm_source=hashnode&utm_medium=feed)

[Sep 4](https://hashnode.com/posts/why-ai-benchmarks-are-broken-mmlu-humaneval-and-chatbot-arena/6a9a6298897c5c74cb5e66ef)

### [Google’s Gemini Handles a Million Tokens. Here’s the Architecture Trick That Makes It Possible](https://silicontosoftmax.hashnode.dev/google-s-gemini-handles-a-million-tokens-here-s-the-architecture-trick-that-makes-it-possible?utm_source=hashnode&utm_medium=feed)

MoE lets Gemini handle 1M tokens efficiently by activating just 2 of 8 experts per token—cutting inference costs while preserving massive model capacity.

7 min read · [silicontosoftmax.hashnode.dev](https://silicontosoftmax.hashnode.dev/google-s-gemini-handles-a-million-tokens-here-s-the-architecture-trick-that-makes-it-possible?utm_source=hashnode&utm_medium=feed)

[Aug 24](https://hashnode.com/posts/google-s-gemini-handles-a-million-tokens-here-s-the-architecture-trick-that-makes-it-possible/6a8c0780bfb07440a68a0cca)
