Harikesh Singh

🤝 VISITING MY PROFILE? LET'S CONNECT! If you're building something in tech, AI, or SaaS, I'd love to connect with you. 🚀 Founders 🤖 AI & LLM Engineers 💻 SaaS Builders 📈 Marketers & GTM Engineers ⚡ Automation Builders 🌐 Indie Hackers 👇 COMMENT BELOW — SAY HI & LET'S CONNECT! I'll check out your profile, engage with your content, and look forward to learning from each other. Let's grow together on X! 🚀

Quantization reduces precision. Distillation transfers capabilities. Speculative decoding accelerates generation. Three different techniques, three different trade-offs. Choosing one without understanding the bottleneck is optimization by guesswork.

Your LLM generates tokens one by one. What if a smaller model could draft several tokens and a larger model verify them together? That's the idea behind speculative decoding. Faster generation without simply replacing your main model.

A smaller model can be cheaper and faster, but can it preserve the capabilities your application actually needs? Distillation isn't just model compression. It's deciding which capabilities matter enough to transfer.

Can You Justify Your Model Optimization Choice? 1-Did you compare cost per successful response? 2-Did you measure latency under load? 3-Did you evaluate output quality? 4-Did you compare GPU memory requirements? 5-Did you test failure cases? The cheapest inference setup isn't always the most economical one.

Are You Measuring the Real Impact of Quantization? 1-How much GPU memory did you save? 2-Did tokens-per-second improve? 3-Did response quality change? 4-Are reasoning tasks affected? 5-Have you tested production workloads? Smaller models are useful only when the quality trade-off is acceptable.

If you’re in tech, let’s connect. Developers, engineers, AI builders: say hi 👇

Do You Know Which Optimization Your LLM Needs? 1-Is inference latency your biggest problem? 2-Is GPU memory limiting concurrency? 3-Is model size driving infrastructure costs? 4-Can you afford a trade-off in output quality? 5-Have you benchmarked alternatives? Speculative decoding, quantization, and distillation solve different problems. Don't treat them as interchangeable.

Software engineering is 10% writing code, 90% reading someone else’s code and wondering “who wrote this?” (It was you.)

Everyone’s learning to build AI. Nobody’s learning to maintain the code AI wrote at 2 AM.

Mumbai local trains are the best distributed system I’ve seen. Zero downtime, insane load, no documentation.

Cricket taught me more about debugging than any course: Watch the ball, ignore the noise, and don’t swing at the first thing that moves.

Prompt engineering is just “explaining things clearly to someone who takes you literally.” Welcome to every code review ever.

Every “quick 5-minute fix” in production has ended my evening, my chai, and my faith in humanity.

UPI processes billions of transactions a month without drama. Meanwhile, my to-do app crashes when I add the 11th task.

2 AM thought: AI isn’t taking jobs yet, but it’s definitely taking the “I’ll Google it” excuse away.

Junior dev: “It works on my machine.” Senior dev: “Then we’re shipping your machine.”

I've seen people spend more time managing perceptions than actually solving problems. Have corporate workplaces accidentally made looking busy more rewarding than being effective?

Companies expect employees to give 100% loyalty, but won't think twice before replacing them during restructuring. So why is staying loyal to one company still considered a virtue?

Sometimes the person doing the most work isn't the person getting promoted. In corporate life, does performance matter more, or visibility and relationships?

AI agents can now interact with tools, access systems, and execute workflows. But when an autonomous agent makes a costly mistake, who owns the decision: the developer, the company, or the person who approved it?