Cyan Ding

new @claudeai diagrams look great!

vercel so slick with the new model training terms :joy:

"I want employees to ask themselves whether they are willing to have any contemplated act appear the next day on the front page of their local paper, to be read by their spouses, children, and friends, with the reporting done by an informed and critical reporter" -- Warren Buffet's "front page test", overhauling Salomon Brothers culture

w @cursor_ai for not draining 99% of my battery with a new update ;)

Keith Rabois on how to identify great talent “What you want to do with every single employee every single day is expand the scope of their responsibilities until it breaks… and that’s the role they should stay in.” Keith tells the story of giving an intern the task of getting smoothies to arrive at the office at 9pm to reward Square’s engineering team. And once the intern proved they could solve this task that had stumped multiple people at the company, he gave them something more important and consequential to do. Everybody has some level of complexity that they can handle, and you want to keep expanding their responsibility until you see where it breaks—as Keith points out, some people will surprise you: “There will be some people who you don’t expect—with different backgrounds, without a lot of experience—who can just handle enormously complicated tasks. So keep testing that and pushing the envelope.” Keith also argues that you should monitor who is going up to other people's desks. If you see people frequently going up to a person's desk, it's a sign that that person can help them. Promote these people and give them more responsibility as fast as you can. Video source: @ycombinator (2014)

you realize you need teammates when you run into a problem you can't even solve with AI

Composer 2 marks the one-year anniversary of our large model training efforts. Since then, we've built an exceptionally talent-dense team of ~40 people with some of the best researchers and engineers from the labs, academia, industry, and more heterogeneous backgrounds. And we are exclusively focused on coding. We don't care about models that can respond to emails, do your tax returns, or be your friend. Every FLOP, token, parameter, and researcher is entirely dedicated to software engineering.

"of all resources on which he could draw, this one was the most precious, the huge pool of crystal essence that he was most reluctant to tap. Buffet would undertake almost any item from his short list of most loathed tasks--get into an angry, critical confrontation; fire someone; give away a vast sum of money; almost anything--rather than make a withdrawal from the Bank of Reputation." -- Snowball

@bradcarryvc bros money is gone

Invested in a founder last year Just asked to see updated financials “Sorry, we aren’t making those yet since we aren’t a public company,” he replied back to me I think my money is gone

The Cursor vs. Claude/Codex feels very flawed and missing the bigger picture The labs have ~infinite money and specialized talent, and are going to win on coding models – that's a runaway train. Composer is impressive, but ultimately more for margin protection / defense than playing to win that game. But to quote Stringer Bell, "there are games beyond the game" and I believe Cursor's destiny is different: becoming the new Github – the place where the whole engineering process lives. Their real competition is with them and Linear, not the labs. Bugbot is a great start. We find it super valuable, no matter what coding agent is used, and is a nice wedge into Cursor getting beyond the coding itself. And of course acquiring @graphite. PR review is the single most essential workflow in Github and very ripe for disruption – Cursor is in an amazing position for this. More on the horizon. The cloud sandbox thing is going to be huge. The new Automations thing aims at GH Actions. And it wouldn't surprise me to see them start getting more into security, observability, etc. Could Anthropic/OpenAI try to compete here? Sure. But I don't think customers want them to. I want my coding agent to be a coding agent and would be happy to pay for another model-agnostic system that sits across the whole menagerie and help me manage it. My prediction is that in a year we'll look back at the "Claude Code is great, therefore Cursor is cooked" discourse as misguided, and understand Cursor as playing a different game entirely.

great reads!

great blog from kimi! 5wvb5ya5wncq4.ok.kimi.link YyWang talks about the process of designing a new architectural component. 1. algorithm proposal (full math) 2. solving memory + memory access problems 3. training overhead as well as the difference between latency and cost optimization, kernel fusion, and team collaboration.

"Hardware upgrades don't necessarily bring benefits: New generation hardware like Blackwell may double compute and bandwidth, but this doesn't mean the latency of these operators will decrease synchronously. Often their bottleneck is not in theoretical compute power, but in launch, synchronization, or memory access. So we even see some topk implementations running slower on newer GPUs." - kimi blog

for those who wear ortho k contact lenses, I found the ultimate cheat code for not getting blurry vision in one eye in the morning: maintain sleep on your back for as long as possible. life changing shit.

@cursor_ai nice blog!

We trained Composer to self-summarize through RL instead of a prompt. This reduces the error from compaction by 50% and allows Composer to succeed on challenging coding tasks requiring hundreds of actions.

"unlike people on Saloman's infamous mortgage desk, [Mozer] did not abuse trainees by hurling food at them or sending them racing out the door to buy twelve pizzas at a time" -- The Snowball

@AnnieLiao_2000 interesting post!

calling it now: AI girlfriends will be normalized by 2027 not just for incels for everyone here's why i think this: current state: replika has 10M+ users character. ai has millions chatting with AI people already forming attachments my friend uses AI: - vents to it daily - gets emotional support - prefers it to real therapy - "doesn't judge me" been doing this for 8 months the progression: 2023: "AI companions are weird" 2024: "some people use them, whatever" 2025: "my friend uses one, seems happy" 2026: "tried it once, actually helpful" 2027: "yeah i have an AI companion" same path as: - online dating (weird → normal) - therapy (weird → normal) - remote work (weird → normal) the use case: not replacing real relationships supplementing them example: person in relationship still uses AI for: - venting about work (partner tired of hearing it) - late night anxiety (partner sleeping) - processing emotions (partner not good at this) AI doesn't replace partner fills gaps partner can't/won't fill the uncomfortable part: this is probably healthy better than: - bottling emotions - trauma-dumping on friends - expecting partner to be everything AI companion as emotional buffer the market: lonely people: obvious market coupled people: hidden market (bigger) everyone has emotional needs partners can't fully meet AI fills gap without judgment, availability issues, or fatigue the objections: "but it's not real connection" neither is therapy, still helps "people will become more isolated" or: people will have better real relationships because AI handles overflow "it's dystopian" so is doom-scrolling, we adapted my prediction: by 2027: - 30% of people have AI companion - 10% admit it publicly - companies offer it as mental health benefit - relationships improve because people stop expecting partners to be therapists the test: if i'm wrong: i'll look dumb if i'm right: i'll look prescient either way: the trend is clear AI companions are growing stigma is decreasing adoption is accelerating whether you think it's good or bad: it's happening prepare accordingly

"If we all grind, we all shine, fk a part time"

me when i hear 150k arr -> 5 customers, 2500 each for one month, project to 12 months. so in reality it may not be so impressive