A week into building my portfolio site and I still don’t feel ready to put it out. There’s always something to change, fix, move around or obsess over that probably nobody else would even notice. At this point, I’m starting to think the hardest part of building something is deciding when to stop 🥲

If you’re in Indiranagar, don’t miss the Chini Malai Toast at Khao Aur Khane Do 🤌🏻

Made with @claudeai Opus 5.5 in 12min. The model is a delight to work with. What a comeback 🤌🏻

most misalignment in small teams doesn't come from disagreement, it comes from assumptions that were never written down whoever wrote the requirement knew exactly what they meant, whoever built from it made their best guess on the gaps. by the time it's in front of you, 3 different interpretations have already quietly baked in write for the person who wasn't in the room when the idea was born this from @thenayazubuko is worth the read, changes how you think about writing requirements

“Handwritten” no longer means pen and paper, and instead means you wrote it using human intelligence (HI) rather than prompting AI.

everyone's focused on which AI model is best this month, that's the wrong game the real position to be in - you own your data, you own your compute, your AI knows things nobody else's does, and nobody can pull the plug on it or reprice you out of using it that's not where most companies are today, but it's where the ones who last will end up

day by day nothing changes, but when you look back everything is different

Bengaluru really said, double it 🌈🌈

Me: I’ll just grab a coffee Also me: walks out with four desserts Glen’s Bakehouse, you’re making this a habit 🤌

As usual, my brother called me at the last min, “Can you get my ITR filed?” I didn’t want to bother my CA at night, so I thought - let me try this with Codex. This was also my first time filing an ITR myself. Codex started by asking a bunch of questions, researching and making a plan. It figured out which ITR was needed and went through everything - Form 16s, HDFC and ICICI bank stmts, Zerodha tax P&L, tradebook and AIS. Then it: - Read and cross-checked all the records - Matched the salary, bank interest and stock trades - Prepared the tax calculation draft, I reviewed and approved it - Navigated a tricky Income tax site and filled like 30–40 forms and countless fields - Stopped for tax payment and challan - Added the challan, validated the return and completed the filing It ran for a couple of hours on 5.6 Sol High, Fast mode. Pinged me only when a human was needed - for confirmations, OTPs, Aadhaar verification and payment. Watching Codex use Chrome and handle everything step by step was just crazy! @ktejesh2203

building a startup in 2026 is a completely different game than 3 yrs ago, this podcast is the clearest breakdown of what actually matters now on ideas - ideas are not valuable. everyone has them. what matters is founder-market fit. did you live this problem so deeply that no one can tell you it's not real? Shopify founder just wanted to sell snowboards. Stripe founders just thought internet payments should work better. the company came from frustration, not a smart idea. on building - Wispr Flow's biggest insight wasn't voice to text, it was that when you dictate, you're running two brain processes at once - one thinking about the idea, one thinking about punctuation and formatting, they killed the second one. That's why users don't say it makes them fast, they say it feels effortless, that's the depth of insight that builds a billion dollar product. on competing with big companies - big companies can only obsess over a few things at once, everything else ships at 60%, that 40% gap is exactly where startups win. Anthropic won enterprise by saying no to everything except code for 3 straight years, Focus is not a strategy - it's the whole game. on distribution - the cost of building software has collapsed, the new bottleneck isn't how hard it is to build, it's how fast you can get it in front of ppl and whether they stay. Great talent is also scarce now coz the best ppl just start their own thing instead of joining yours. Wispr Flow solves this by hiring ex-founders and giving them entire business units to run with agents. on markets - just because the market doesn't exist doesn't mean the idea is wrong. Uber, Airbnb, Stripe - none of those markets existed before they created them. the founder believed it before anyone else did, that belief is the only thing that keeps you going when 20 VCs say no. on iteration - iteration velocity solves every problem, not just engineering ones, every great company looks obvious in hindsight. Google had 10 search competitors, the devil is always in the details nobody wants to talk about. on SF - SF is the only place in the world where a billionaire brags about driving a Chevy Bolt, money is low signal, building is the only status symbol. Everywhere else conversations feel one click behind not coz ppl aren't smart, but the environment around you decides whether your delusional ambition gets encouraged or killed. SF encourages it, most places don't. on finding your ppl - if you feel like no one around you gets it, don't go looking for a community, build something and put it out. the right ppl find you. A Discord, a GitHub repo, a tweet - whatever, every person in this podcast found each other by doing stuff publicly. what all of this adds up to - the game has changed but the filter hasn't, are you building or are you talking about building?

slop isn't a prompt problem, it's a systems problem. a factory shipping defective units doesn't have a worker problem, it has a quality control problem. nobody is checking the output before it leaves the building. the fix isn't to blame the worker, it's to build the checkpoint. same with AI output, the prompt got you 80% there. the missing 20% is the review step most teams skipped when they automated the workflow. someone needs to own the output before it reaches the customer, a human eye, a second pass, a standard for what good looks like. the teams producing slop aren't using worse AI, they just never built the gate.

Most people confuse AI SDK, AI Gateway, LLMs, and Embedding Models. Here's how they actually fit together: 𝗔𝗜 𝗦𝗗𝗞 = the library you install in your codebase It's how you write code that talks to AI models, generateText(), streamText(), useChat(). Runs inside your app. 𝗩𝗲𝗿𝗰𝗲𝗹 𝗔𝗜 𝗚𝗮𝘁𝗲𝘄𝗮𝘆 = cloud infrastructure your requests pass through one API key. Built-in caching, cost tracking, rate limiting across all providers. Not a model. Not required, but useful at scale. They're not alternatives. They work together: Your Code (AI SDK) → Vercel AI Gateway → Claude / GPT-4 / Gemini 𝗟𝗟𝗠𝘀 𝘃𝘀 𝗘𝗺𝗯𝗲𝗱𝗱𝗶𝗻𝗴 𝗠𝗼𝗱𝗲𝗹𝘀: two completely different jobs: LLMs → generate text (Claude, GPT-4, Gemini) Embedding Models → convert text to numbers to find meaning and similarity. They don't generate anything. They're used together in RAG: → Embedding model finds the right documents from your knowledge base → LLM reads them and generates the answer Building AI products means knowing which tool does what. Confusing these leads to wrong architecture decisions.

Vibe coding is not agentic coding. Vibe coding is accepting output because it looks right. Agentic coding is pair programming with a very fast junior engineer who sometimes lies with perfect confidence. Read the diff. Challenge the plan. Use sub-agents for review. The human is still the control plane.

CLIs are becoming the product surface again. Not because terminals got fashionable. Because coding agents read `--help`, try commands, recover from errors, and keep going. A 401 should not just say "unauthorized". It should tell the agent the exact reauth command. Interactive for humans. Non-interactive for agents. Useful errors for both. Anyone shipping a CLI in 2026 is actually shipping 2 products.