Follow AI Builders — Not Influencers

I want this.

I can finally ship the text layout of my dreams

I also wanna write an essay on how I'm short fashion AI. There's no aura, there's no personalization, there's no distance. And most important of all, most people want to fit in, and those who don't do not want advice

For most of last year I lived on a farm with my ex-girlfriend, had two people hired to do the operations of my business I feel like I already experienced what post-AGI life could look like for me I really want to write about this more Because it turns out AGI just feels like what I imagine it's like to have rich parents

Every time I see chat at house party, we're just talking about JavaScript I don't know any JavaScript.

The n̶e̶t̶w̶o̶r̶k̶ agent is the computer

40% of my twitter is translated japanese fashion content, i love twitter

great video on aura and scent https://www.youtube.com/watch?v=_-MthsvOB1c&t=1228s

RT Iheanyi Ekechukwu I gave Codex and Claude Code at work the same prompt for planning a big E2E feature and then had them analyze which plan was better. Both chose Codex, but also identified which gaps from Claude Code's response would help strengthen it, which was my analysis as well. Original tweet: https://x.com/kwuchu/status/2037928225955823776

RT Viv “we’re doing constrained optimization” —> a loose mental model on evals as an input to hill climbing and self-improving agents going back to classical CS class - the Knapsack problem! we’re constrained on the weight allowance of our bag and each item has a weight and value, we want to maximize value the weight allowance is a hard constraint, we can’t submit if the total weight exceeds that very smart computer scientists found algorithms to optimize that objective for any arbitrary task we want to do in the world, there is no algorithm instead we use an agent to do “intelligent search” over the option space via agent actions evals are like a multi-objective constraint on the problem the agent is trying to solve, some evals are “blocking”, we won’t proceed if these fail and some are “soft” but important in the overall hill climbing these hard and soft determinations are often human derived as we move towards self-building agents and harnesses, evals are a great way to constrain the solution space by guiding the agent towards solutions that pass those evals this is why it’s so important that evals capture the behavior we care about! Agents directly or indirectly will use them to define themselves autonomous agents are often search machines over the space of all programs. that’s a massive space and only a subset of it corresponds to being “good at our task” evals are the way we climb the gradient to fin that space which is for another time but… evals are like training data! many interesting analogs between evals, classical optimization theory, ML/RL, agent engineering very exciting time to be in the intersection of any of these :) Original tweet: https://x.com/Vtrivedy10/status/2037924832453530004