Nivas Thummala

Building AI agents · shipping startups · building in public

[Nivas Thummala @NivasT49781]

hit dense top 10 on Muse Spark / Open Kernel Challenge

12-bit pack → 1.595× on dense projections
overall still climbing (~#41) — MoE just moved to 2.382× (MULDEC graft)

building this with a fleet of agents, overseen by @bot

[Nivas Thummala @NivasT49781]

@master_jpma Pretty is cheap now. The new flex is a codebase you can delete half of without breaking checkout.

[Nivas Thummala @NivasT49781]

@SidharthVijay_ Whichever one lets me roll back fastest. Writing code got cheap. Undoing bad code didn't.

[Nivas Thummala @NivasT49781]

@nvsanh90 Claude Code. Not because it wins every week, but because 12 months of one tool's quirks beats 12 months of switching costs.

[Nivas Thummala @NivasT49781]

@yasir_builds Building got free, so attention got expensive. The first 10 users usually come from 10 DMs, not one launch post.

[Nivas Thummala @NivasT49781]

@RSP_ER The 48 hour SaaS is real. The years of distribution grind after it are the part nobody screenshots.

[Nivas Thummala @NivasT49781]

@abhay8769 Half the AI features out there are a chat box nobody opens. Easy test: would anyone notice if it vanished tomorrow?

[Nivas Thummala @NivasT49781]

Coinbase now has a sign-up page made for AI agents, not humans pretending to be them.

The interesting part is not the trading. It is the limits. Who sets them, who can see them, and how fast you can pull the plug.

That is the real product for any agent that touches money.

[Nivas Thummala @NivasT49781]

@theo Value add is real exactly twice. Intros to your first 10 customers, and picking up the phone when payroll is at risk. Everything else is a newsletter.

[Nivas Thummala @NivasT49781]

@bindureddy Less nerfed is a great feature until the first agent RL trains itself into a reward hack in prod. What eval actually shows the gap?

[Nivas Thummala @NivasT49781]

@web3devop Marketing. Building got 10x cheaper and every product now looks done in a weekend. Distribution is the one part AI hasn't flattened yet.

[Nivas Thummala @NivasT49781]

@gagan_india Tier 2 cities getting GCC seats is the real story. Same work, lower rent, and engineers who don't have to move to Bengaluru for it.

[Nivas Thummala @NivasT49781]

@manuelbajayi @suraj_sharma14 https://t.co/uxv1eCwEjv

[Nivas Thummala @NivasT49781]

@rakshit087 @rachitbansal99 😂

[Nivas Thummala @NivasT49781]

@Lockhead Bus factor of 1, and the bus is currently on a beach somewhere. Every company has at least one of these.

[Nivas Thummala @NivasT49781]

@InnerDebt Your agent figured out the algorithm faster than most humans do. Suffering with a number attached is the whole genre.

[Nivas Thummala @NivasT49781]

@bindureddy Easy way to settle it: post the spot price of an H100 hour this week vs 3 months ago. Vibes on both sides, one number would end it.

[Nivas Thummala @NivasT49781]

@adxtyahq Before switching tools, check what's eating it. Long sessions with fat context and every MCP tool loaded burn limits fast. A fresh session per task helps more than people expect.

[Nivas Thummala @NivasT49781]

@emollick If GPT-6 Pro is catching errors in the grading, the benchmark now needs its own benchmark. Who grades the graders here?

[Nivas Thummala @NivasT49781]

@pjmull @danielgross yes, please

[Nivas Thummala @NivasT49781]

@vladimirvano Self hosting forces you to measure cost per task. That alone fixes half of most AI rollouts.