Thuwarakesh Murallie

Data Scientist @ Stax and 2X top writer on Medium.
I share what I explore on #Python, #ML, #AI, #Datascienc

[Thuwarakesh Murallie @Thuwarakesh]

I was juggling between proprietary models, third-party inference providers, and self-hosting LLMs.

But I’ve learned that self-hosting rarely benefits me—and I think that’s true for most people. Instead, focus on these optimizations.

And be clear on when your GPU infrastructure actually starts to pay off.

Read me recent article on @towards_AI

https://t.co/HKGCmobUiv

[Thuwarakesh Murallie @Thuwarakesh]

If GPT-5.6-style workplace agents become reliable enough for daily use, where would you want them to create value first?

[Thuwarakesh Murallie @Thuwarakesh]

AI saves me time, yet I am not producing twice as much.

That contradiction may explain why AI adoption is soaring while enterprise transformation... not so much!
The work has moved—from creating to verifying, from execution to orchestration, and from using tools to defining processes clearly for machines to follow.

The results of companies building agents before understanding their workflows are often dozens of experiments, a few pilots, and very little measurable impact (if at all).

https://t.co/gpNmyfCqDa

[Thuwarakesh Murallie @Thuwarakesh]

Your agent skills may be working perfectly, yet still wasting money.

Most agent skills do not fail because they lack instructions, but become expensive because they accumulate old rules, duplicated constraints, vague advice, abandoned examples, and prompt debris nobody wants to delete.

Because these instructions can be loaded repeatedly, every unnecessary line consumes tokens and weakens the signal the agent actually needs.

An instruction earns its place only if removing it makes the output worse.

https://t.co/3awq5snCGb

[Thuwarakesh Murallie @Thuwarakesh]

I've used them both. In my opinion Antigravity works well for design the first draft.

But I'd prefer Claude code for serious agentic engineering approach. Especially the nestable CLAUDE.md files are super useful.

I'd rather pick based on the app's complexity.

[Google Gemma @googlegemma]

Unplugging completely! No WiFi and zero notifications. A great way to get deep focus on a project.

Here is a walkthrough showing how to run Gemma 4 (26B A4B) fully offline with LM Studio & OpenCode to parse PDFs, ask questions, and build sites 100% locally.

[Thuwarakesh Murallie @Thuwarakesh]

@prasenx My morning routine is tea☕ → check what model dropped overnight → rewrite my benchmarks → question my life choices. The meme hits different when your roadmap gets outdated every 48 hours 😭

[Thuwarakesh Murallie @Thuwarakesh]

@rohanpaul_ai Schmidt is right about the opportunity, but the winners won't be the ones who build agents first — they'll be the ones who solve the orchestration problem.

Getting multiple agents to collaborate reliably in production is the unsexy, hard problem worth solving.

[Thuwarakesh Murallie @Thuwarakesh]

I was juggling between proprietary models, third-party inference providers, and self-hosting LLMs.

But I’ve learned that self-hosting rarely benefits me—and I think that’s true for most people. Instead, focus on these optimizations.

And be clear on when your GPU infrastructure actually starts to pay off.

Read me recent article on @towards_AI

https://t.co/HKGCmobUiv

[Thuwarakesh Murallie @Thuwarakesh]

@Thishyaketh @Alibaba_Qwen Would it be fair to say MoE models are efficient while Dense models are performant, in general?

The question then boils down to 'would you utilize your GPU in full for a dense specialized model, or a run a much larger MoE which fits in your GPU only during inference time?'

[Thuwarakesh Murallie @Thuwarakesh]

@ClaudeDevs As someone who's broken prod more than once with a seemingly innocent ALTER TABLE, a fleet of agents stress-testing my migration before merge sound like a safety net I needed for years.

I'm curious how it handles multi-step migration chains with rollback dependencies

[Thuwarakesh Murallie @Thuwarakesh]

@TheGeorgePu It's weird you're seeing this differently. I see Claude Code and Coworker are available on all paid plans.

I believe it's an A/B test run on a small population.

[Thuwarakesh Murallie @Thuwarakesh]

https://t.co/501A8JMHlY

[Saeed Anwar @saen_dev]

Long context windows didn't kill RAG. They just made people think they could skip the hard part.
Your LLM knows the future and your private data are two different problems. Context length solves neither.

[Thuwarakesh Murallie @Thuwarakesh]

LLMs have two blind spots: the future (cutoff dates) and your front door (private organizational data).

But long context windows aren't sufficient to kill RAGs.

https://t.co/OOBHVjoDTS

[Thuwarakesh Murallie @Thuwarakesh]

@konstructivizm @grok , is this real?

[Thuwarakesh Murallie @Thuwarakesh]

RAG is the most convenient way to provide private and updated information to an LLM. However, the context size becomes a limiting factor as you scale up.

Context compression keeps costs and latency low while preserving response quality.

https://t.co/0f6zlNWdE0

[Thuwarakesh Murallie @Thuwarakesh]

It may sound counter-intuitive, but we use AI to evaluate AI. But the challenge is writing an evaluation prompt. You get built-in prompts with frameworks. But, they are too generic for niche cases. So you should write the prompt. Here's how.

https://t.co/a13MiTkyrh

[Rohan Paul @rohanpaul_ai]

Anthropic just dropped the beautiful explaination of how they built a multi-agent research system using multiple Claude AI agents.

A MUST read for anyone building multi-agent system.

A lead agent plans research steps, spawns specialized subagents to search in parallel, and then gathers and cites results. It covers architecture, prompt design, tool selection, evaluation methods, and production challenges to make AI research reliable and efficient.

Single-agent research assistants stall when queries branch into many directions. Anthropic links one lead Claude with parallel subagents to chase each thread at once, then fuses their findings.

⚙️ The Core Concepts

Research questions rarely follow a straight path, so a fixed pipeline leaves gaps. One lead agent plans the investigation, spawns subagents that roam in parallel, and later condenses their notes into a coherent answer.

🧠 Why Multi-Agent Architecture Helps

Each subagent brings its own context window, so the system can pour in many more tokens than a single model would hold. Anthropic measured that token volume alone explained 80% of success on BrowseComp, and adding subagents pushed performance 90.2% past a lone Claude Opus 4 on internal tasks.

Running agents in parallel also cuts wall-clock time because searches, tool calls, and reasoning steps happen side by side rather than one after another.

@AnthropicAI