Udita Pal 🧂

Currently traveling and my parents randomly showed up at the station at 9 AM—just for five minutes—to hug me, kiss me, and slip me some cash. Made this 29-year-old woman cry like a baby. 🥹😭❤️ I don’t deserve my parents 😢

@ismaelsevi87 @WilliamsF1 This is huge! Congratulations a lot of folks dream of this 🎀

Llevo toda la vida viendo estos coches de colores pensando 'ojalá formar parte de eso algún día', y ya no es un deseo. Muy contento de poder contaros que he pasado a trabajar para Williams F1 como diseñador gráfico y artista 3D freelance :)

The most honest thing about this piece is what it doesn't say: that no app, no system, no second brain fixes the underlying tension between a biological mind and an information environment that was never built for it. We've been treating the symptom for years. This is about the cause.

there is a world outside f1 and i need to focus on that

@kocharpulkit lmao, you so mean xD

they liked this comment 💀 hope they watch it

Started my TouchDesigner journey w @ThineAI keeping me company and providing inspiration. Extremely satisfying when the animations react to midi + piano.

Her insta is taste.guaranteed 👀

My mom created her cooking channel last year. She finds / invents dishes, records them, edits them, find captions on chatgpt and post. All because she is passionate 🙌😭💗

@harnidhish So happy for you. 💗

Got to HOLD my BOOK for the FIRST TIME in my HANDS and SCREAMED AND ALMOST HAD A PANIC ATTACK!!! Preorder now!!!! amzn.in/d/0ccy42t0

@idleshubh Loved it!

If you don’t believe in superstitions you can pick similar bangles up from peepal tree near your house 👀

mad respect to them for running races they know they've lost :) flying w @thineai

“how can you forget” sire, i forgot my tiffin as a kid and my laptop as an adult. forgetting is literally the one thing i’ve been consistent at my whole life

@KithRags sending you all the love in the world <3

It’s been difficult to share anything about her and to process my grief publicly but today I miss her SO much and I just felt like I needed to continue to share my beautiful girl on my timeline ♥️ She is my whole world.

@manan_0308 kar kya rahe the tum aisa?

What on earth is this now, Claude?

From sending blurry pictures of class notes on WhatsApp in 2013 to screensharing on it in 2026 to help my mom with upi payments. Same app, completely different version of me.

Psa: you do not get bonus points for frying your brain.

Pretty 💗

LangChain shipped Deep Agents last week. @harisonchase is arguing the harness IS the product - built-in task planning, sub-agent spawning, middleware, hooks, the works. Garry Tan dropped "Thin Harness, Fat Skills" the same week. Opposite thesis: the harness should be as thin as possible. I'm firmly in Garry's camp. Here's why. Every piece of logic you put in the harness is reasoning you're taking away from the model. 40 tool definitions. God-tools with 5-second round trips. Brittle prompt chains babysitting the model at every step. You haven't built an intelligent agent. You've built a cage around one. This was defensible two years ago. GPT-3.5 couldn't follow multi-step instructions. Even through Claude 3.5 Sonnet, models genuinely needed guardrails for basic tool use. But with the current generation - Opus 4, GPT-5, Gemini 2.5 - if you're still hardcoding orchestration logic, you're actively capping what the model can do. The real test: what happens on model upgrade day? Fat harness → your product breaks. You spend weeks re-engineering scaffolding to accommodate capabilities the model already has. Thin harness → your product gets better automatically. No ceiling on reasoning. The upgrade flows straight into better outcomes. This is the architecture that separates the 2x engineer from the 100x engineer. Same model. Same API key. Same context window. What thin harness actually looks like in practice: > Fat skills: judgment in markdown. Domain knowledge, heuristics, the fuzzy human stuff. Make these as rich as possible. > Fat code: deterministic operations. Auth, validation, DB writes. Things that must be right every time. > Thin harness: just the loop. Context in, model reasons, tool executes, result back. That's it. @hwchase17 is right that memory matters. But memory is a skill problem and a data problem - not an orchestration problem. You don't need hooks and middleware intercepting every model call. You need well-structured skills that tell the model how to use its memory. The best AI engineering looks almost too simple. Boring infra. Judgment in markdown. A model you actually let think.

Harvard: we want humans to remember everything Me, who actively suppresses 90% of my memories to function normally, please don't