Currently @GoogleDeepMind. // Prev: @Cruise, MEng @MIT
San Francisco, CA
Joined January 2011
- Tweets361
- Following80
- Followers138
- Likes953
1/ I wanted to understand how I was using my AI coding tools and where I could improve my workflow. Since that record was scattered across multiple Claude Code and Codex sessions, I built github.com/vtjeng/session-at… to show everything in one timeline.
4/ Just for fun, Session Atlas also extracts token usage from each turn and estimates what the same usage would have cost at API list prices. (Thankfully, I didn’t have to pay $388 to build this project.)
5/ Finally, the home page puts all your projects on a shared timeline, with a summary for each.
For me, Session Atlas is a fun way to look back at what I worked on and when. Everything runs locally, without uploading your transcripts. Code and setup: github.com/vtjeng/session-at…
Vincent Tjeng retweeted
I can’t believe it - after years of advocacy, exclusionary zoning has ended in Cambridge.
We just passed the single most comprehensive rezoning in the US—legalizing multifamily housing up to 6 stories citywide in a Paris style
Here’s the details 🧵
Vincent Tjeng retweeted
Sure matplotlib is cool, but what if I want to load my loss curves into the 2006 hit Flash game LineRider?
People are getting all excited about artifical intelligence these days, but the real money is in cameras: metacareers.com/v2/jobs/4330…
Love the illustrations in this explainer for diffusion models (and the fun demo of how such models work)
I got a (very tiny) diffusion model running in a webpage madebyoll.in/posts/dino_diff…
I really enjoyed collaborating with @LaurentLessard to study #Wordle - check out his write-up of our results! Thanks, too, to @evanrsparks for running our more computationally intensive code - we'd have been waiting till summer w/o the parallelization provided by Evan's cluster.
My contribution to the #Wordle discourse, with help from @VincentTjeng and @evanrsparks.
A definitive look at the game: insights, heuristics, optimal strategies, and the limits of what is possible.
laurentlessard.com/solving-w…
Vincent Tjeng retweeted
I'm starting an "intro to compilers" blog series, where we build a compiler from scratch, one step at a time. Each episode explains and implements a new component.
First up: eliminating instructions that have no effect, like dividing by 1.
medium.com/@predrag.gruevski…
Vincent Tjeng retweeted
Together with @VincentTjeng, we showed that for #Wordle in Hard Mode,
1. No strategy can guarantee a 4-move win.
2. There is a strategy that always wins in at most 5 moves. First move is "SCAMP". The strategy (in full detail) is here:
github.com/LaurentLessard/wo…
Building on code in @LaurentLessard's github.com/LaurentLessard/wo…, I was able to show that, for Wordle in "hard mode", no strategy guaranteeing a solution in 4 turns exists - but a strategy guaranteeing 5 turns does! Try it out for yourself at github.com/LaurentLessard/wo… (1/5)
Finding a strategy guaranteeing 5 turns takes around 4 minutes; proving that no solution guaranteeing 4 turns exists takes several hours! (4/5)
Just for fun, here's our solve-hard-mode-in-5-steps playing against @qntm's Absurdle (qntm.org/files/wordle/index.…).
Absurdle 5/∞
⬜⬜⬜⬜⬜
⬜⬜⬜⬜⬜
⬜🟩⬜⬜🟨
🟨🟩🟨🟨⬜
🟩🟩🟩🟩🟩
@wordletips @xaqwg
(5/5)
Excited to be featured in @MIT News for research I presented at #ICLR2019: news.mit.edu/2019/how-tell-w…! Thanks to Rob Matheson for a great article😀
I'm presenting work on evaluating robustness of neural networks (arxiv.org/abs/1711.07356) at the #ICLR2019 afternoon poster session today! We scale complete verification to larger networks, and our code is open source: github.com/vtjeng/MIPVerify.… (joint w @KaiKaiXiao, Russ Tedrake)
Really cool work on learning descriptors for robotic manipulation by @peteflorence, Lucas Manuelli and Russ Tedrake featured on MIT news today! (Also, my shoe is in the training set somewhere... )
news.mit.edu/2018/mit-csail-…