←
Alex Wa68 posts
Alex Wa avatar

Alex Wa

@_djdumpling

@yale | @modal, residency @primeintellect

◷ Joined November 2022▧ 17 Photos and videos
Alex Wa avatar
Alex Wa@_djdumpling·Feb 18

new blog! What methodologies do labs use to train frontier models? The blog distills 7 open-weight model reports from frontier labs, covering architecture, stability, optimizers, data curation, pre/mid/post-training + RL, and behaviors/safety djdumpling.github.io/2026/01…

Post attachment
Locamage avatar
⟳ Alex Wa retweeted
Locamage@Locamage·Oct 26

Meet Finch 1, our first step toward one model that works with any sensor, at any resolution, anywhere on Earth. Trained on imagery from 50+ sensors across satellites, aircraft and drones.

0:34
Modal avatar
⟳ Alex Wa retweeted
Modal@modal·Oct 14

Today we're announcing that Modal has crossed $1B in annualized run rate, as well as raised a $750M Series D at a $15.75B valuation, led by @Accel. Since our Series C, usage has tripled and we’ve grown more than 15x year-over-year. We are deeply grateful to the developers and companies building on Modal. There's still so much more to build, and we can't think of a better time to be building software.

1:41

Erik Bernhardsson @bernhardsson Oct 14 Article Modal crosses $1B in Annualized Run Rate and raises $750M Series D Just five months ago, when we announced our Series C, I wrote that demand was accelerating. Since then, usage has tripled and we’ve grown more than 15x year-over-year. We recently crossed $1B in

Changxiu (Leo) ji avatar
⟳ Alex Wa retweeted
Changxiu (Leo) ji@Changxiu_leo_ji·Oct 5

New research 🚀 from the ACE (Agentic Context Engineering) team at Stanford We took the failure lessons out of an LLM agent's memory, and it got better! 📈 Why? We find that failure recovery tips could hurt LLM agent performance when they sit in context all the time, even when nothing is going wrong. So we build Sentry 🛡️: a test-time framework that recovers LLM agents from failures. It keeps failure knowledge outside the agent and only steps in when a failure actually happens. On average: +39% over our own ACE, +37% over the best runtime-intervention baseline. 📄 Paper: arxiv.org/pdf/2610.02994 💻 Code: github.com/nuglifeleoji/Sent… Demo Here!!!!!!! More details below 🧵 (1/N)

2:35
amitsai avatar
⟳ Alex Wa retweeted
amitsai@asaiiie·Sep 8

I just wrapped up my internship @modal. This summer I worked on container filesystems, speeding up agentic workloads by up to 4x and directory traversals by 12x. I wrote about this systems work behind it here: amitsai.com/writing/erofs/?v… Huge thanks to @amitpr, @jonobelotti_IO, the runtime team, and everyone at Modal for an amazing summer!

Ashton Chew avatar
⟳ Alex Wa retweeted
Ashton Chew@ashtonchew·Aug 26

Now open source! Have codex/cc set it up for you and read the code behind the optimizations and results. link below to repo and docs

Post attachment

Ashton Chew @ashtonchew Jul 31 Article How I Got Computer-Use Clicks to under 10 ms on Modal A model that uses a computer needs a computer to use. Not yours, so you rent one from your favorite local sandbox company. A Linux desktop runs in a sandbox and your computer-use agent (CUA) takes the

Melody Yu avatar
⟳ Alex Wa retweeted
Melody Yu@melodyexists·Aug 21

It's my last day as an intern on @modal's training team. With the rise of RL, it's becoming increasingly critical to have robust, reliable observability. I wanted to share some of my work building an observability system for the Modal Training Gym (gym.modal.dev), built on top of @slime_framework & @radixark miles. Thank you @qjoyliu, @peywalt, training team, and Modal for an awesome summer!

0:18
Scott Hao avatar
⟳ Alex Wa retweeted
Scott Hao@shao256·Aug 18

It's my last week @modal. This summer, I had the chance to work alongside @waltergoat, @colin_weld, @paulgb on our Sandbox product. It was my first time working on systems at this scale, and I couldn't be more grateful for their mentorship. Sandboxes are full of gnarly distributed systems problems. I wrote about one of them - distributed locking - on my blog: scotthao.com/writing/distrib… To ship it, I needed to break rules for practicality and performance. This led me to formally verify the design with TLA+ to catch the many race conditions. This is also my first blog post! I don't share my writing often, but I'd like to do it more. Hope you enjoy.

Alex Wa avatar
Alex Wa@_djdumpling·Aug 13

Main takeaway is that when long-horizon self-play rewards didn’t transfer, the rollouts were still useful, as market data. Modeling prices, valuing liquidity, and distilling the resulting decisions worked. More technical details about the implementation here: djdumpling.github.io/2026/08…

Alex Wa avatar
Alex Wa@_djdumpling·Aug 13

Finally, we distilled the analytic bidder back into the 4B model through one round of expert iteration: ~4.7k bid decisions, including ~1.2k expert deviations. With no selector at inference, the weights improved paired margin by +8.31 [3.21, 13.42] over SFT.

Post attachment
Alex Wa avatar
Alex Wa@_djdumpling·Aug 13

The missing variable was the option value of cash. We added a shadow price for liquidity, charging each bid for the future flexibility it destroyed: U(b) = (V̂ − (1+lambda) * b)P̂(win | b) That flipped the result: • first-place rate: 36.0% → 59.3% • paired margin: +12.16 [7.61, 16.71]

Post attachment
Alex Wa avatar
Alex Wa@_djdumpling·Aug 13

But offline improvement didn’t survive live play. This myopic selector changed paired margin by −5.01 points. It bought better gems, but overspent, depleted its liquidity, and lost investment returns. Optimizing one auction at a time had made the full-game policy worse.

Alex Wa avatar
Alex Wa@_djdumpling·Aug 13

Each bid was modeled by U(b) = (V̂ − b)P̂(win | b) Why a distribution? 55% of auctions cleared within 1 coin. Even at 1.10 MAE, point predictions misclassified win/loss 26.1% of the time. Modeling residual uncertainty reduced offline regret by another 0.456 coins/decision.

Post attachment
Alex Wa avatar
Alex Wa@_djdumpling·Aug 13

The failed runs still produced something valuable: thousands of bids. We fit 2 models: • value MLP V̂ predicts each gem's final value and mission value • bid GBDT P̂ predicts other policies' discrete bid distributions