
21 yo building @stanford
Palo Alto, California
Joined June 2016
- Tweets1.4K
- Following821
- Followers773
- Likes2.9K
genuinely don't understand how people formed trust in business during covid, meeting face-to-face is so important
313
Darrow retweeted
AI is running into energy limits, and people are asking whether AI investments will earn back what they cost. That’s exactly why this moment is so exciting. We need to focus on the sectors and workloads that create the most value, and build models and chips that do more with every unit of energy.
Really enjoyed working with @JonSaadFalcon, @Avanika15, and the entire team on Joules-to-GDP! Thanks to everyone who helped along the way, especially @Jon_Hartley_, @Azaliamirh, and @HazyResearch.
AI is using more and more of the world’s energy. But how much economic value are we getting from AI?
We’re introducing Joules-to-GDP (J2G), a framework that measures how well AI turns energy into economic value and shows which tasks, models, and hardware give the best return.
Across 44 occupations in finance, engineering, law, administrative work, and more in @OpenAI's GDPval, we arrive at three findings:
1) The tasks with the highest economic value are not the tasks with the highest economic value per joule. For example, CAD design earns ~10% more per task than building websites, but the latter earns ~27% more value per joule.
2) The most powerful model isn’t always the best use of energy. Gemma 4 31B delivers 56% of GLM 5.2’s net value with 23% of the energy, for 2.5x the value per joule.
3) More chips don’t guarantee proportionate returns. Moving the same model from 1 to 8 NVIDIA H100s uses 3.2x the energy but cuts completion time by only 22%, with about the same value.
With @SnorkelAI, we’re also releasing GDPWorks, a benchmark for economically valuable tasks across 36 additional occupations.
Joint work co-led by @amirzeinaalii and @Avanika15, with @Mersad_Abbasi, @Jon_Hartley_, John Hennessy, @HazyResearch, and @Azaliamirh.
More details below 🧵
Darrow retweeted
How well do you understand training dynamics for deep nets?
Here’s a CS312 (Deep Learning Alchemy) prediction question: midway through training an LM, we temporarily lower the learning rate and then restore it. How does validation loss change relative to the original run? ↓
We train a 35M-parameter LM with a WSD learning-rate schedule (peak LR 0.003, tuned roughly to be optimal for this setting). Consider two different learning rate schedules
A short LR drop: between 40-50% of training, dropping to 0.1x the original LR.
A longer, shallower LR drop: between 35-55%, dropping to 0.55x the original LR.
These LR drops are matched so that the total sum of the removed LRs is the same in both.
How does the shape of the dip affect validation loss during the dip and after the LR returns?
The image shows four qualitative predictions.
A: The deeper LR drop leads to a larger loss drop and a large penalty after
B: A deeper LR drop leads to a large temporary loss drop, but both match soon after.
C: Both LR drops keep the loss curve monotone, but a larger LR drop leads to slower progress
D: The deeper LR drop leads to a larger temporary loss drop, and this gain persists long term
Try the full quiz: deep-learning-alchemy.github…
Darrow retweeted
Can we forecast benchmark accuracy from models that score zero?
We (co-leads @LukeBailey181 @rckpudi) turn a benchmark into a series of easier hinted versions. Jointly modeling scale C and hint level H enables forecasting GSM8K acc (H=0) within 3 pts from 5600x less compute 🧵
Darrow retweeted
This week in CS 312, we explored one of the most unintuitive concepts in pretraining: sharpness
In this post, I'll explain what sharpness is, how sharpness appears in real training, and our unanswered questions about sharpness
This week in CS 312, we explored one of the most unintuitive concepts in pretraining: sharpness
In this post, I'll explain what sharpness is, how sharpness appears in real training, and our unanswered questions about sharpness
Darrow retweeted
People long believed in human-like computer use. But software (even terminals) is built for humans, not AI agents. Why not let agents build their own?
Meet 𝗔𝗴𝗲𝗻𝘁 𝗦𝟰, an agent scientist evolving its own interfaces, letting an open 27B model beats GPT-5.5 & Opus 4.8.
With its evolved interfaces, the same model improves 𝟯.𝟴× and costs <𝟱% as much per task as either frontier model.
Agents today read screenshots and act through click, scroll, type, maybe a shell. So an agent fills a spreadsheet cell by cell when one line of code would do, or prints two thousand lines of a file to find one function.
Every new app or model release means a human reads trajectories, diagnoses failures, and hand-writes new primitives. Skill-distillation methods can reorder what the interface allows, but they can’t extend it. Agents compensate with more test-time compute instead of a better interface.
𝗠𝗼𝗱𝗲𝗹𝘀 𝗶𝗺𝗽𝗿𝗼𝘃𝗲; 𝗶𝗻𝘁𝗲𝗿𝗳𝗮𝗰𝗲𝘀 𝘀𝘁𝗮𝘆 𝗳𝗿𝗼𝘇𝗲𝗻.
“We want AI agents that can discover like we can, not which contain what we have discovered.”
— Rich Sutton, The Bitter Lesson
Hand-designed interfaces are what we have discovered. Agents can read their own trajectories, spot what’s missing, and write and test code. 𝗜𝗻𝘁𝗲𝗿𝗳𝗮𝗰𝗲 𝗱𝗲𝘀𝗶𝗴𝗻 𝘀𝗵𝗼𝘂𝗹𝗱 𝗯𝗲 𝗮𝗻 𝗮𝗴𝗲𝗻𝘁 𝗰𝗮𝗽𝗮𝗯𝗶𝗹𝗶𝘁𝘆.
That’s what Agent S4 does.
Agent S4 makes a 𝗿𝗲𝗽𝗼𝘀𝗶𝘁𝗼𝗿𝘆 𝗼𝗳 𝗶𝗻𝘁𝗲𝗿𝗳𝗮𝗰𝗲𝘀 𝗮𝗻𝗱 𝘂𝘀𝗮𝗴𝗲 𝗸𝗻𝗼𝘄𝗹𝗲𝗱𝗴𝗲 the unit of evolution. A task agent attempts synthetic tasks with no success labels or verifiers. A meta-agent studies the trajectories, forms hypotheses about which actions, observations, or knowledge would help, and builds descendants through mutation or crossover. Validation tasks the meta-agent never sees decide which repositories survive.
The model and harness stay fixed, and the meta-agent 𝗻𝗲𝘃𝗲𝗿 𝘀𝗲𝗲𝘀 𝗮 𝗯𝗲𝗻𝗰𝗵𝗺𝗮𝗿𝗸 𝘁𝗮𝘀𝗸. 𝗢𝗻𝗹𝘆 𝘁𝗵𝗲 𝗶𝗻𝘁𝗲𝗿𝗳𝗮𝗰𝗲 𝗲𝘃𝗼𝗹𝘃𝗲𝘀.
With the evolved repository, Qwen3.8-27B goes from 𝟱.𝟲% to 𝟮𝟭.𝟯% on OSWorld V2, 𝟯.𝟴× the same model with fixed interfaces and a new state of the art for open-source computer-use agents.
It beats GPT-5.5 (13.0%, $25.50/task) and edges past Opus 4.8 (20.6%, $72.40/task) at $𝟭.𝟭𝟬 per task.
On ScienceBoard, the same model goes from 41.4% to 58.6%.
The analysis is the more interesting part. Evolution mostly deepens interfaces the agent already has: 4 new interfaces and 127 extensions. Give the meta-agent a scratch VM to test its code, and it builds 12 new interfaces and 253 extensions, reaching music scores, audio projects, and 3D scenes. The best repository came from crossing two earlier branches.
In one case study, the seed agent hand-builds a browser client mid-task and debugs through failure after failure; the evolved agent calls page_text and reaches the same message at tool call 𝟭𝟭 instead of 𝟯𝟳.
Human-like computer use was a starting point, not the destination. The interface an agent acts through is a second axis of improvement, beyond better models and harnesses, and agents can now design it themselves. 𝗘𝘃𝗼𝗹𝘃𝗶𝗻𝗴 𝗮𝗴𝗲𝗻𝘁 𝗶𝗻𝘁𝗲𝗿𝗳𝗮𝗰𝗲𝘀 𝗶𝘀 𝗮 𝗻𝗲𝘄 𝗱𝗶𝗺𝗲𝗻𝘀𝗶𝗼𝗻 𝗼𝗳 𝗮𝗴𝗲𝗻𝘁 𝘀𝗲𝗹𝗳-𝗶𝗺𝗽𝗿𝗼𝘃𝗲𝗺𝗲𝗻𝘁.
Darrow retweeted
3-step guide to a viral launch from SF founders:
1. jump out of a plane to launch pancake.ai
2. set a realistic goal, like hitting #1 on Product Hunt in 24 hours with 0 prep
3. make your users a promise you’ll keep, like breaking a Guinness World Record for the tallest pancake stack if they help you win
observe the consequences.
we're picking a place to break the record rn btw
Darrow retweeted
Soft launching +EV: plusev.tech/. @EV_Fellowship
I think this is *very* important for the world.. Otherwise I'd retire 🫠
Notes:
> It's NOT a venture fund
> The goal is to unblock top ambition/talent, underserved young people with big dreams. Where do small grants/mentorship make a *real* difference?
> Every company and person helping/contributing time, capital, compute credits is (obviously) doing this pro-bono.
> The goal is ambition with social good in *some* way... Meme coin trading apps, etc aren't the target. Again, this isn't a venture fund.
> Literally launching, so be patient if website issues etc
> If you want to help, please email: [email protected]!
> With the rise of AI, most applications for anything get a lot of noise & bullshit... It's easy to spam them! The vast majority of people will be rejected for obvious reasons. If you don't get an interview, you're rejected. The point of this is to cater to the top 1% underserved, young people who *could* change the world but don't have the connections, capital, mentorship. This does *not* cater to the 99%.
> Decisions are made by an independent committee. Harvard iLab mentors, ex/current founders, VCs, etc.
wifi wasn't working on my computer
super satisfying to realize I can use local qwen 3.8 27b in opencode to debug the issue
8
817
I made an explainer video on our latest paper at Simular!
arxiv.org/abs/2609.36927
Darrow retweeted
Today, we're officially launching Colmez, backed by a16z @speedrun, Andreessen Horowitz's early-stage fund.
The $248bn hazardous waste industry is broken. After Stanford and quant trading at SIG, I left to fix it.
We just launched our private beta two weeks ago and are already working with 5 of the top 7 operators in the industry, including the largest environmental services companies in the world.
Thank you to my incredible team, @brianxmeng1 and @numairarazzak , our speedrun partner, @JoshLu, and the @StartX community. If you want to build the backbone of America’s industrial infrastructure, we’re hiring.
kat's the first agent to ask me questions about myself like this! love it
hello.kat.ai @kdrag0n
1
7
770
didn't plan to launch today but ok 🐱
hello.kat.ai
ranking every text message personal assistant. (worst to best)
4. poke - blocked, spammed too much
3. instinct - blocked, too slow + goes rogue. doing too much
2. lucas - good
1. kat - literally the best by far. super fast. super proactive. and it just doesn’t mess up.
they technically haven’t released it yet but you can get access by texting this number
+1 (425) 537-0753




















