Sarah Guo

model optimization techniques commoditize, infrastructure compounds

Let's go @figma! (and 2025 swagger IPOs, volatility be damned) https://t.co/nChSWKABEA

why isn’t it more celebrated that many think an AGI “recipe” is known now? large scale pre-training to get to a powerful base model (that can reason), give the model necessary tools, RL (wherever outcome is measurable) against set of real world uses, generalization (!) is it because it’s just hard work/data intensive? that not that many know?

Winners in our hearts 💜 7th Annual @Forbes AI 50! @baseten @Figure_robot @harvey__ai @langchain @MistralAI @pika_labs @SierraPlatform @stackblitz @thinkymachines https://t.co/sLnEE83390

Today, AI model improvements depend a great deal on expert human data. Fun @NoPriorsPod with @BrendanFoody founder @mercor on - how to identify outlier talent - the data we need - the future of the labor markets - verifiability, evals - if we can capture “taste” https://t.co/hiFLp0UwHf full episode everywhere you get you podcasts, every Thursday. YouTube: https://t.co/dwpuXR03Jf

.@TheChainsmokers 😱 are kicking off the @boltdotnew hackathon with a private concert in SF what a vibey timeline we are in

inspired by what @matthuang @FEhrsam @danrobinson have built at @paradigm. the leading investing firm in crypto, a force advancing the ecosystem, a set of independent thinkers (and moneymakers)! Huang is an investor I go to for advice personally https://t.co/KS8WJVuZkz https://t.co/CHTtWu8gns

impressed by the Google model ship. Seeing multiple portfolio companies do things they couldn’t before because of, and only because of Gemini. Hats off @JeffDean @demishassabis @OriolVinyalsML and team

ILYA, book lovers. I’m with you. Context was I was pitched writing a book by a publisher but sensei @karpathy got me kinda sad about the death of the medium and also, I hate writing. almost as hard as thinking

two views of our codegen present https://t.co/Dmy1ajLnlR (this is going to get better, not worse, with scale + RL)

Code AI is working! As they rapidly increase in capability, we’ll start to work with them in AI-native interfaces. I expect “teaming up” with agents like this will bring enable human developers to do step function more. 🙌@ScottWu46 @russelljkaplan @walden_yan + @cognition_labs one of the underexplored capabilities of ai is its scalability. as agentic starts working more and more you should treat it more as a team of engineers (that work for you, human!) than a solo contributor-system. also cc @stevenkplus1 @silasalberti

only @jaltma could get me to 1) be a guest, not a host 2) be candid about how I think venture works 3) do pull-ups in the office

not that many investors I admire, but Neil Mehta is one of them. king-level risk taking https://t.co/F5DewaMnJh

full interview on YT or anywhere you get your podcasts: https://t.co/o9aAz0doo0

Correct

I’m hungry @garrytan

Decided to go full eat pray love with @eladgil @jaltma https://t.co/Q37JPRazOM

AI therapy seems to work https://t.co/rCUT29mnkr https://t.co/EftpYAdtSY

I think we are going to use all the GPUs (and TPUs)

We put a 90” tv on the wall of @conviction’s office. We should background stream: cc @mvernal @igarciacamargo @pranavreddy *I don’t watch tv so have never seen severance, hope this is SFW

I’ve been singularly overweight @gradypb since ~2012! Everyone is about to find out @andrew__reed is in the class of greats too.

the world is going to have more beauty

guys stop worrying the models are good https://t.co/O2FQEWBNAk and the little one https://t.co/G2yzi5Jglq

hell is a Tesla rideshare driver with thrashy start/stop style. sir you can just press the button

I think the personal agent that wins is the one that can handle the most unstructured data messiness. Not just the “read tables in pdfs” problem that many have tackling for a while, but — can you make sense of my Apple notes? My whole life is in there. Apple sure won’t.

is there a book / essay about environments that best support scientific discovery? Not about specifically bell labs, cavendish, xerox parc, los alamos - but why these places were so productive?

vibe code is great but we need vibe debug / refactor

Pace

The most undervalued trait in startup hiring is pace. This is especially true today, when the floor is lava. Models that were state-of-the-art six months ago are now commodities. Code generation companies are fighting weekly battles for performance gains and feature advantages that can determine market leadership. I've watched applied AI teams with modest funding outpace 100X better resourced competitors simply because they could complete thought-to-deployment cycles in days rather than weeks. When hiring, identifying these high-velocity people is worth more than almost any other qualification. How to Recognize It Pace isn't about long hou…

.@CorridorSecure is hiring founding engineers. cracked, mission-driven team with leading expertise, great use case unlocked by AI, smartest security folks in the business around them

hardware folks have such disdain for the kludge, lack of planning, and generally low standards of software folks. didn’t fully appreciate