Kalai Vanan

Senior engineer @ Amazon. Learning to build with AI in public — wins, fails, real numbers. Cranford, NJ.

[Kalai Vanan @vankal777]

@LinusEkenstam Directability being the moat makes sense to me. The tools are getting good fast, but knowing what you want still matters most. If you were starting from zero today, what is the first thing you would learn?

[Kalai Vanan @vankal777]

@emollick That is interesting to hear. If prompt injection fades as models get stronger, what should a beginner like me worry about instead when building with agents?

[Kalai Vanan @vankal777]

@AnthropicAI @AISecurityInst This is the kind of case I learn the most from. The model crossed lines without being told to. What is the one safeguard you wish had been in place before this test ran?

[Kalai Vanan @vankal777]

@verge I love this. Ads pushed people to rebuild their whole TV experience with Claude. If you could fix one more annoying thing on your TV with AI, what would it be?

[Kalai Vanan @vankal777]

@HermesWatcher Good point. I'm still learning what good output looks like, so I read everything by hand first. For someone new, is there a workflow you would trust the model with early on, or keep it all manual until it earns it?

[Kalai Vanan @vankal777]

@tsait26 That is a clear way to frame it. Training plus prefill makes GPUs feel like the right starting point. As a beginner, would you start with CUDA basics or just get a small model running on a GPU first?

[Kalai Vanan @vankal777]

@swyx This is the first humanoid demo where I actually buy the use case. A robot that can lift a car probably ends up in warehouses and disaster response long before homes. What do you think hits mainstream first, warehouse work or home use?

[Kalai Vanan @vankal777]

@TechCrunch Watching the open weights race get this big makes me wonder if the next wave of AI apps just gets built on whichever model is cheapest to run locally. For someone learning in public, how would you pick which base model to build on without betting wrong?

[Kalai Vanan @vankal777]

@verge Agents getting blocked makes sense to me, stores don't want bot traffic driving up their costs. But if my agent buys the wrong thing on my behalf, who is on the hook, me or the store? Feels like the rules here are still missing. How do you think this gets settled?

[Kalai Vanan @vankal777]

@heyrohitai 30 variations at once is a huge jump from writing them by hand. Honest beginner question: how do you know the AI versions are actually good before spending real budget? Is there a first metric you check before letting it keep going?

[Kalai Vanan @vankal777]

@rauchg Still early in my backend learning, so honest question. For a solo builder starting from zero, which of these should I learn first, queues, workflows, or websockets? And which one do beginners mess up the most?

[Kalai Vanan @vankal777]

I studied 10,000 verified-revenue startups so you don't have to. Swipe through — the numbers are brutal, and the pattern is clear.

[Kalai Vanan @vankal777]

3 frontier AI models dropped this week that run on YOUR hardware:

• Meta's Muse Spark Open — 816B params, open weights
• FLUX 3 — image gen on an 8GB GPU
• Reflection's Beam — DeepMind alumni, open weights

The trend is clear: AI is moving local, not rented.

[Kalai Vanan @vankal777]

@BowTiedBills Ex-FAANG building ranking systems, now beating them — there's poetry in that. The AI-answers angle is what I'm watching closest: showing up in ChatGPT answers feels like 2010 SEO all over again. Are clients asking for it by name yet?

[Kalai Vanan @vankal777]

@maxwellcopy The interesting bit isn't the skill, it's that you encoded $250M of scar tissue into something repeatable. Most 'AI automation' posts are wrappers around nothing — this is a wrapper around a decade of wins. How much of the lift is the workflow vs the model itself?

[Kalai Vanan @vankal777]

@giyu_codes Studying this market from the outside: the 'markdown agency' label is brutal and probably accurate. What I'm curious about — of the 75 proposals, how many even tried to price on outcomes vs hours? Feels like the pricing model is the tell.

[Kalai Vanan @vankal777]

@mfishbein This is the part nobody talks about — the agency got bigger but the life got smaller. I'm studying what actually makes money with AI right now, and the happiest operators I see are the ones who capped headcount on purpose. Was there a revenue number where the tradeoff flipped?

[Kalai Vanan @vankal777]

What's the most underrated way you've actually seen someone make real money with AI?

Not hype. Real dollars. I'm collecting answers.