Sahaj Garg

if you look at the market right now, it's clear the perfect personal agent experience is yet to be built it's also obvious that it's going to be the one with the deepest personalization stack that learns how you prioritize things/what you care about there's a world where the end goal is an agent that is so personalized it can look through your email, slack, etc. then literally say "nothing pressing right now" and actually be right

nvm lol

companies live and die by: >the speed at which they move >their ability to build useful things for people if spacex moved slowly, they could still be on falcon 1 if iphones only sent texts and played music, 1B+ people wouldn't use them every exceptional company combines both

the biggest pain point that's come as a result of better and better AI is that people's minds feel even more all over the place social media was the first wave where something like this happened, now AI has done this even for people who have resisted the urge to be on social media it'll be interesting to see how companies approach solving for this scattered feeling most people are experiencing

software engineering is interesting right now because everything has changed, but also in many ways nothing has changed fundamentally, you still have to: >deeply understand the user’s problem >know how systems should scale to build a solution >verify that what you built actually solves the user's problem you can build much faster now, but the fundamentals haven't changed

your goal as an interviewer is to get someone to go on a constructive rant about something they're excited about. they get to feel seen. you get to see how deeply they think. the best hires are all obsessed with something. the fun part is finding out what.

i want to share a rule that we have at @WisprFlow, as i believe its existential to most companies: you can give the execution of a task to AI, but you must do the work to keep your understanding of the task. our goal with this rule is that the person responsible for the work should be able to explain the details behind how they solved a problem, even when AI helps with the legwork. in this context, here's how it applies to different teams within the company: in engineering, an agent can implement a feature and complete a pr. however, you need to understand how it fits into the system and how you checked that it behaves as intended. when something breaks, you need enough understanding to investigate it. in marketing, AI can make product animations, but they need human opinion on whether it's a piece of media someone wants to watch. that requires knowledge for pacing, music, and entertainment value. in gtm, mass outreach can be executed by ai. but for large enterprise clients, you almost certainly want to take a personal approach and present the problem your tools solve. to build a successful automated system for this, it first requires an understanding of what great outreach looks like. AI can now help you build the understanding required in these mediums, too. the method by which you acquire the understanding is far less important than having it. as AI takes on more of the execution layer, companies should guide their teams toward keeping their understanding of what the models are doing. doing this will be the only way to direct them effectively.

in your early 20s, one of the best things you can do for yourself professionally is put yourself in extreme situations that challenge you in some way for me, that meant: >doing 40 or 50 hours a week of research during undergrad >working at a startup building optical computing hardware as employee #5 >co-founding @WisprFlow basically, you should try to identify the craziest problem someone is working on, and figure out how you might be able to help shape a solution to that problem

top 1% ML talent can choose to work at any lab in the world by the time you meet them for an interview, their track record and reputation have already done much of the screening your actual goal is sharing how your lab is working on the problems that closely align with their vision, and that it will be a place where their research can come to life you can never convince somebody to fall in love with a problem that they're not interested in

internal memo: If you have not yet, please use ChatGPT Astra 6 with computer use on ultra mode if you are doing knowledge work tasks. This is the first time since Claude Code in Dec 2026 that I've felt a step function change in capabilities - ability to use AI just operates entirely differently. These are expensive background tasks (many run for 1hr+), but completely fundamentally change ability to use models. Warning: never ever ever give up your understanding to these models -- if you don't understand what it's doing, or why the outcome is the right one, you are setting yourself up for long term failure.

openai has dots spacex has grok bot meta has muse very interesting signal that anthropic hasn't opted to go this route thus far

if you want to succeed in a technical interview all you need is the ability to talk for 15 minutes in depth about a problem you solved the catch is that it can take years to be able to do this

"help me send a wire transfer more easily with AI so I no longer have to approve a big money transfer" - said no consumer ever

people have busy lives. if you want to build a complex product, you should start with the simplest version

i've found that a true 10/10 hire doesn't really exist. however, a 9/10 hire does, and they often change the trajectory of your company i realized this a few months ago when we were looking to hire someone on our ml team and i had a call with their reference. when i asked the reference how they'd rate this person as an employee, they gave the candidate a 10/10. normally this would be a great sign, but i've realized a 10/10 is often indicative of one of two things: 1/ the candidate has outgrown their role and should have been promoted. sometimes a 10/10 is given simply because they've outgrown the expectations of the role they're in. once you promote them to a role that reflects their expertise level, you find out what their true rating as an employee is. 2/ their reference has some level of bias. a 10/10 is unlikely to be an objective score, even for a close friend you work with. to better understand the rationale behind the score, i asked: "why wasn't this person promoted if they were a 10/10 employee?" the reference shared that they were indeed promoted, to which i asked "and how would you rate them after the promotion?" they said a 7/10. that gave me a much clearer picture. now, the same person was being evaluated against a different set of expectations. expectations that align more closely with the role we were hiring for. however, if we had gone through the process of evaluating expectations, the person had been promoted, and still received a 9/10 rating—that would be a very strong signal. a true 9/10 rating can be the type of person that changes the trajectory of a company.

it's very unlikely that booking a flight or securing a restaurant reservation will be the killer use case for these ever-present agents that everyone wants to build those aren't points of friction most people care to solve we're still very early in all of this