Post

More from @akshay_pachaar

Akshay 🚀

@akshay_pachaar

Nov 23

You’re in an ML Engineer interview at Google.

Interviewer: We need to train an LLM across 1,000 GPUs. How would you make sure all GPUs share what they learn?

You: Use a central parameter server to aggregate and redistribute the weights.

Interview over.

Here’s what you missed:

One major run-time bottleneck in multi-GPU training happens during GPU synchronization.

For instance, in multi-GPU training via data parallelism:

- The same model is distributed to different GPUs.
- Each GPU processes a different subset of the whole dataset.

Check this 👇

This leads to different gradients across different devices.

So, before updating the model parameters on each GPU device, we must communicate the gradients to all other devices to sync them.

Let’s understand 2 common strategies next!

Read 14 tweets

Akshay 🚀

@akshay_pachaar

Nov 21

NOBODY wants to send their data to Google or OpenAI.

Yet here we are, shipping proprietary code, customer information, and sensitive business logic to closed-source APIs we don't control.

While everyone's chasing the latest closed-source releases, open-source models are quietly becoming the practical choice for many production systems.

Here's what everyone is missing:

Open-source models are catching up fast, and they bring something the big labs can't: privacy, speed, and control.

I built a playground to test this myself. Used CometML's Opik to evaluate models on real code generation tasks - testing correctness, readability, and best practices against actual GitHub repos.

Here's what surprised me:

OSS models like MiniMax-M2, Kimi k2 performed on par with the likes of Gemini 3 and Claude Sonnet 4.5 on most tasks.

But practically MiniMax-M2 turns out to be a winner as it's twice as fast and 12x cheaper when you compare it to models like Sonnet 4.5.

Well, this isn't just about saving money.

When your model is smaller and faster, you can deploy it in places closed-source APIs can't reach:

↳ Real-time applications that need sub-second responses
↳ Edge devices where latency kills user experience
↳ On-premise systems where data never leaves your infrastructure

MiniMax-M2 runs with only 10B activated parameters. That efficiency means lower latency, higher throughput, and the ability to handle interactive agents without breaking the bank.

The intelligence-to-cost ratio here changes what's possible.

You're not choosing between quality and affordability anymore. You're not sacrificing privacy for performance. The gap is closing, and in many cases, it's already closed.

If you're building anything that needs to be fast, private, or deployed at scale, it's worth taking a look at what's now available.

MiniMax-M2 is 100% open-source, free for developers right now. I have shared the link to their GitHub repo in the next tweet.

You will also find the code for the playground and evaluations I've done.

@MiniMax__AI GitHub repo for M2:

(don't forget to star 🌟)
github.com/MiniMax-AI/Min…

@MiniMax__AI Find the code for the playground and the evaluation done using @Cometml Opik: github.com/patchy631/ai-e…

Read 4 tweets

Akshay 🚀

@akshay_pachaar

Oct 27

Claude Skills might be the biggest upgrade to AI agents so far!

Some say it's even bigger than MCP.

I've been testing skills for the past 3-4 days, and they're solving a problem most people don't talk about: agents just keep forgetting everything.

In this video, I'll share everything I've learned so far.

It covers:

> The core idea (skills as SOPs for agents)
> Anatomy of a skill
> Skills vs. MCP vs. Projects vs. Subagents
> Building your own skill
> Hands-on example

Skills are the early signs of continual learning, and they can change how we work with agents forever!

Here's everything you need to know:

Skills vs. Projects vs. Subagents:

https://x.com/akshay_pachaar/status/1982817709323612628

If you found it insightful, reshare with your network.

Find me → @akshay_pachaar ✔️
For more insights and tutorials on LLMs, AI Agents, and Machine Learning!

https://x.com/akshay_pachaar/status/1982817709323612628

Read 4 tweets

Akshay 🚀

@akshay_pachaar

Oct 25

I've been coding in Python for 9 years now.

If I were to start over today, here's a complete roadmap:

While everyone's vibecoding, a few truly understand what's actually happening.

This roadmap matters more now than ever.

So, let's dive in! 🚀

1️⃣ Python bootcamp by @freeCodeCamp

4 hours Python bootcamp with over 46M views!! It covers:

- Installing Python
- Setting up an IDE
- Basic Syntax
- Variables & Datatypes
- Looping in Python
- Exception handling
- Modules & pip
- Mini hands-on projects

Check this out👇

Read 9 tweets

Akshay 🚀

@akshay_pachaar

Oct 20

You're in an ML Engineer interview at OpenAI.

The interviewer asks:

"Our GPT model generates 100 tokens in 42 seconds. How do you make it 5x faster?"

You: "I'll optimize the model architecture and use a better GPU."

Interview over.

Here's what you missed:

The real bottleneck isn't compute—it's redundant computation.

Without KV caching, your model recalculates keys and values for each token, repeating work.

- with KV caching → 9 seconds
- without KV caching → 42 seconds (~5x slower)

Check this out👇

So let's dive in and understand how KV caching works...👇

Read 11 tweets

Akshay 🚀

@akshay_pachaar

Oct 6

You're in a Research Scientist interview at OpenAI.

The interviewer asks:

"How would you expand the context length of an LLM from 2K to 128K tokens?"

You: "I will fine-tune the model on longer docs with 128K context"

Interview over.

Here's what you missed:

Extending the context window isn't just about larger matrices.

In a traditional transformer, expanding tokens by 8x increases memory needs by 64x due to the quadratic complexity of attention. Refer to the image below!

So, how do we manage it?

continue...👇

1) Sparse Attention

It limits the attention computation to a subset of tokens by:

- Using local attention (tokens attend only to their neighbors).
- Letting the model learn which tokens to focus on.

But this has a trade-off between computational complexity and performance.

Read 10 tweets

Share this page!

Enter URL or ID to Unroll

Akshay 🚀

Try unrolling a thread yourself!

More from @akshay_pachaar

Akshay 🚀

Akshay 🚀

Akshay 🚀

Akshay 🚀

Akshay 🚀

Akshay 🚀

Did Thread Reader help you today?

Don't want to be a Premium member but still want to support us?

Send Email!