Sam Karu (@SamKaru_) header
Sam Karu (@SamKaru_) profile picture

CEO @trylogical | IPhO medalist | Ex @nvidia, @ycombinator

San Francisco, CA
Joined October 2025
Photo by Sam Karu (@SamKaru_)Photo by Sam Karu (@SamKaru_)Photo by Sam Karu (@SamKaru_)Photo by Sam Karu (@SamKaru_)Photo by Sam Karu (@SamKaru_)Photo by Sam Karu (@SamKaru_)Photo by Sam Karu (@SamKaru_)Photo by Sam Karu (@SamKaru_)Photo by Sam Karu (@SamKaru_)Photo by Sam Karu (@SamKaru_)
Your AI notetaker is blind. My meetings are full of screen shares: charts, decks, dashboards. Every notetaker I tried gave me text notes or a pile of blurry screenshots. So we built Nutmeg, the notetaker purpose-built for meetings with screen shares. It rebuilds what was shown into your notes as real charts, tables and numbers - then checks against what was said to fix mistakes and fill in details. No bots. Your notes can finally see what you saw... and what you didn't! nutmeg.so
40
29
30
144
57,482
Your AI agents are not locked out. With Nutmeg they get access to the visual context of your meetings too. Nutmeg builds a visual index of every meeting, so agents find what was on screen fast, without combing through every frame.
1
1
8
263
Nutmeg runs on your Mac while you use Zoom or Google Meet as usual, and it only captures the screen share being shown. Nutmeg is live today! We’re rolling out access in waves, and the first 100 signups get 1 year of Pro on us. Sign-up: nutmeg.so Questions, feature requests, or just here for the spice? Join our Discord: discord.gg/2yH28JCXC
1
10
261
Sam Karu retweeted
Introducing A5S Async V1: the world's most accurate speech recognition model for speaker-attributed transcription. #1 on VoiceArena, ahead of ElevenLabs, AssemblyAI, Deepgram, NVIDIA, Meta & more. The next-best model makes 41% more errors across 5 public benchmarks. And yes, every lab says they're #1. So we open-sourced our benchmark. Try it now👇
57
61
36
208
361,386
how is it possible claude’s search cannot find a chat that contains the word i typed in it 15 times
5
334
Sam Karu retweeted
Give your agents their own iPhone and Mac. Al agents are smart enough to do almost anything, but they've been missing the one tool every human uses: A device. Introducing Glasshand. Glasshand lets your agent tap, type, and swipe on real iPhones and Macs, exactly like a human. That means it can: → Post TikToks → Book Uber rides → Send iMessages → Download any app it needs If a human can do it on a device, your agent can too. Rent an iPhone and a Mac for your agent at glasshand.com
139
104
136
604
444,814
We brought Touch to the Mac! From the all-new iPad mini to the 14-inch MacBook Pro and iMac... and now MacBook Pro with PROLED XDR Display and Touch, creators have new tools to do their best work yet.
6
442
we’re hiring a growth intern at Logical (YC F25). shoot, edit, design, run experiments, manage creator campaigns, and ship under our accounts from week one. no 12 person marketing team to hide behind. $40–60/hr, leading to full time. bonus if you can work with us in SF. if this sounds cool, DM or apply below. hiring fast. ycombinator.com/companies/lo…
4
5
12
1,057
Sam Karu retweeted
12 years ago, twitch played pokemon. today, x plays iphone. click ▶️ to use the phone. xphone.solid.site/
7
5
13
32
21,696
there hasn't been a single reason to think about building a website in php in like a decade. and now there's not even a reason to think about starting to think about building one.
The PHP security team is giving advance notice: Upcoming security release Wed, Oct 21st 2026, 14:00UTC, two critical vulnerabilities, both rated CVSSv4 9.2. PHP users are encouraged to prepare to update. Earlier release if details become public or active exploit evidence.
1
6
459
cursor is cooked. it was good while it lasted.
4
324
in the age of claude, i'm way more impressed by the eng who spends an afternoon optimizing a feature than the one who's always itching to build something new
6
313
you know this is BS when swift is above javascript and python
What programming language does AI know worst? According to AA Omniscience: Java
1
8
356
great, but this graphic is irresponsible at best. the long standing norm in ML research is to give the highlight to the best performer in a row. everyone knows this. say what you want about @ylecun but he would've never allowed this.
Introducing Muse Wonder, our new family of frontier models with state of the art performance for personal and professional agentic work.
3
246
nytimes absolutely cooked here
3
257
i've probably run 50+ interviews across Nvidia and Logical. here's why i think a unfamiliar codebase is good, 60 mins is bad and how leetcode can still work. - anything a human could do with AI in 60 mins, AI can do better alone. at Logical, we measure something called HAV (human-added value) aka the "meat proxy avoidance metric". even after 24 hour take-homes, we realized HAV was difficult to measure and you're basically operating in a noise band. - leetcode can still work, but only if you can get the candidate onsite. online is too easy to cheat these days. what we do - give a reasonably mature codebase. cold start, 7 days. you pick what to build. improvements, bug fixes, new features are all fair game. my experience - bug fixes tests grit. everyone likes adding new features these days, but the difference between vibe software and real software is robustness, speed, security. you have to do the unglamorous work. - improvements and new features test creativity. some smart people are too afraid to push the envelope. unfamiliar codebases show how people get up to speed. in a startup you can't spend weeks reading the codebase like a book. you learn by doing and breaking. you have to adapt to a system. tools, ci/cd, protocol. work trials are the only reliable indicator left.
Goodbye to LeetCode? Google just updated its interview process. AI is now a part of it. They have introduced a new round called Code Comprehension. The idea behind this round is to drop the candidates into a 200-500-line codebase with Gemini available for use. This round evaluates: - Can you read someone else's code - Are you able to debug independently before utilizing AI - Can you catch when Gemini is wrong
2
1
10
583