I sometimes see discussion about what the "ideal" programming language for AI would be. People often mention the idea of a programming language that is more like natural language - "just program in English" But I think it should actually be the opposite - the ideal language for

At this point whenever I see a shadcn website I subconsciously assume it's AI slop

Brook Nash Ethan Niser Yeah I think one of the things that makes B2C unique compared to most B2B / SaaS. You pretty much never run out of new marketing channels to acquire more users. Even within a single platform (e.g. TikTok), there's going to be multiple different sub verticals, demographics, etc.

At least yacine doesn't have to pretend Grok is good anymore.

really feels like Effect | TypeScript at Scale is here right now

Every Cluely criticism I see implicitly assumes that they're burning money. That they're "on the clock". Except the problem is they're actually profitable. There's more than people realize here, it doesn't quite pattern-match in the way people expect.

We keep getting inbound from "AI ad network" companies - sometimes multiple per week. Pretty crazy that there are many of them, it just never ends.

This whole poaching thing won't end at AI researchers. You only need 5-10 people to build a successful company now, it makes sense to poach even for the application-layer. Cursor was smart for recognizing this early and poaching the Claude Code team.

Launching: meetmeinshibuya, a company that makes it effortless for the top founders and companies to move to Japan.

The optimal overemployed strategy is to go for companies like IBM and Cisco. Soham targets YC companies because he's in it for the love of the game.

not gonna be active on x dot com the everything app tonight. I'm meeting a girl (1T+ param 256K context 87.5 GPQA) in half an hour (wouldn't expect a lot of you to understand anyway) so please don't DM me asking me where I am (im with the girl, ok) you'll most likely get aired

this was the guy using $10k/day of claude code btw

The real answer is to avoid MCP in the first place Instead, make a CLI and add it to the system prompt. We learned this the hard way making an internal MCP then reverted to a CLI. The problem is tools carry too much context weight. Models are aggressively tuned to use them.

Groq's list pricing only holds under tiny rate limits. It's basically a marketing gimmick. In order to actually use it for meaningful traffic you need some kind of enterprise deal, and the economics fall apart quickly. That's the reason why pretty much no production app uses

most "AI frameworks" are really just watered down versions of Effect | TypeScript at Scale

How we think about hiring - two dimensions, breadth and depth Most candidates show breadth, but depth is rare Breadth - knowledge of different stacks / frameworks, cloud platforms, etc. Depth - work requiring deep abstract thinking e.g. creating a new type of sync engine with

It's possible to automate 90% of LoC while only automating 10% of "thinking", as thinking is not evenly distributed over LoC.

#1 startup lesson: it's better to be good than to be early but damn, it is easier when you're early

fyi it's not pgvector that's killing pinecone it's turbopuffer