VP engineering Mercor
San Francisco
Joined January 2009
- Tweets3.7K
- Following3.2K
- Followers2.3K
- Likes17K
Congrats @cyril_mathew and the Latitude team!
We raised a $35M Series A led by @oakhcft . That brings our total capital raised to $43M.
We spent the last decade watching money get stuck at borders across roles at Uber, Coinbase, Stripe, and Zero Hash. Latitude (@rtp) is what we built with those lessons.
3
353
Time to update the @JeffDean facts
Capex is negative since compute pays to work with Jeff and Sanjay :)
We created a pitch deck to tell a handful of VC firms about us and what we were up to (a fun experience!). Here’s a few slides about our background and some of the things we’ve worked on from the pitch deck (it was fun putting together the list of people in our teams who have gone on to found a whole range of exciting companies). We are delighted to have selected @radicalvcfund
and @khoslaventures to lead our initial funding round, along with participation from @lightspeedvp, @kleinerperkins, Doerr Capital (@johndoerr), and Alphabet (@Google). We’ll be working with them to close our seed round over the next few weeks.
2
526
Interesting work
The HF science team just made async RL weight sync ~100x cheaper on bandwidth, and you don't need a shared cluster anymore.
The problem: every RL step, the trainer typically has to sync fresh weights to the inference engine. for a 7B in bf16 that's ~14GB. for a frontier 1T fp8 checkpoint, that's ~1TB; in bf16 it would be ~2TB. per sync.
The insight: between two RL steps, ~99% of bf16 weights are bit-identical. at RL learning rates, the optimizer is whispering and bf16 literally cannot hear most of it. the stored bf16 bits don't change.
What they shipped in TRL: only the changed elements get encoded as a sparse safetensors file, dropped into a Hugging Face Bucket, and fetched by vLLM. on Qwen3-0.6B, per-step payload goes from 1.2 GB to 20 to 35 MB. This is exactly what we built Buckets for: S3-like object storage on the Hub, Xet-backed (so even full snapshots only transfer the changed chunks).
The cherry on top: we ran a FULL disaggregated training where:
- the trainer lived on one box
- vLLM ran inside a Hugging Face Space
- the Wordle environment ran in another Space
- weights flowed through one Hub bucket
no shared cluster. no RDMA. no VPN. no NCCL across clouds. just HTTPS and a bucket.
one GPU + a Hugging Face account is now enough to do real disaggregated RL. multi-replica inference fleets across regions become a small devops exercise, not a research project.
Full write-up: huggingface.co/blog/delta-we…
Open source RL keeps eating the moat!
2
421
I would be surprised if othe hyperscalers AWS, Microsoft don’t do an equity raise similar to Google before the OpenAI and Anthropic fund raises to have enough dry powder
1
3
1,489
Srini Rajagopal retweeted
Recently met @srush_nlp and he started giving me an impromptu lecture on how targeted on-policy self-distillation works.
I asked him if I could record it on my iPhone.
The basic idea is this: if the model made a mistake at some point in the rollout (for example, calling a tool that doesn't exist), we want to discourage this specific error, but we don't want to just learn from the final reward, because it's a very noisy signal spread out over the whole trajectory.
So we have another model read this trajectory and figure where the error was made. It simply inserts some hint tokens to the part of the trajectory right above where the mistake was made.
Now with these injected hint tokens, have the model run a forward pass. You're not having to regenerate a new rollout - aka no new decode required.
The hint causes the model to assign lower probabilities to the error tokens. You then trains the original model to match these new probabilities, teaching it to downweight that specific mistake.
Excited for what @cyril_mathew and team are building!
Happy to have a played a small part in it as an advisor
Over a decade ago at Uber, we were paying drivers across 40 countries. Their top ask: "Pay me faster." That stuck with me through Coinbase and Stripe, where I kept running into the same wall from different angles.
Today, Latitude (@rtp) is launching out of stealth. We raised an $8M seed round in Jan 2025 and have been building quietly since then.
4
363
Srini Rajagopal retweeted
Over a decade ago at Uber, we were paying drivers across 40 countries. Their top ask: "Pay me faster." That stuck with me through Coinbase and Stripe, where I kept running into the same wall from different angles.
Today, Latitude (@rtp) is launching out of stealth. We raised an $8M seed round in Jan 2025 and have been building quietly since then.
Srini Rajagopal retweeted
We knew the Ivy League was racist against east asian and white men, but little did we know they were EXTRA racist against south asians.
The famous SFFA case treated Indians and East Asians as a single group. This masked significant heterogeneity: It's way harder to get in if you're Indian!
In Columbia's internal admissions database (h/t @cremieuxrecueil), East Asian applicants had a 41% lower odds of admission than equally qualified White applicants, whereas South Asian applicants had 63% lower odds.
Srini Rajagopal retweeted
Performance Hints
Over the years, my colleague Sanjay Ghemawat and I have done a fair bit of diving into performance tuning of various pieces of code. We wrote an internal Performance Hints document a couple of years ago as a way of identifying some general principles and we've recently published a version of it externally.
We'd love any feedback you might have!
Read the full doc at: abseil.io/fast/hints.html
Hope for a speedy recovery, one of the good guys. Great sense of clarity in all conversations.
On Monday, I will ask President Trump, via X, to help save my life. He offered to help me if I needed it.
I need it.
As many of you know, I have metastasized prostate cancer.
My healthcare provider, Kaiser of Northern California, has approved my application to receive a newly FDA-approved drug called Pluvicto. But they have dropped the ball in scheduling the brief IV to administer it and I can’t seem to fix that.
I am declining fast. I will ask President Trump if he can get Kaiser of Northern California to respond and schedule it for Monday. That will give me a fighting chance to stick around on this planet a little bit longer.
It is not a cure, but it does give good results to many people.
2
928
Interesting
RL learning is sparse
On policy distillation combines student trajectories with token level teacher feedback so both relevant and dense feedback
Our latest post explores on-policy distillation, a training approach that unites the error-correcting relevance of RL with the reward density of SFT. When training it for math reasoning and as an internal chat assistant, we find that on-policy distillation can outperform other approaches for a fraction of the cost.
thinkingmachines.ai/blog/on-…
4
1,267
Revenue growth is vertical, many novel and interesting engineering problems to solve.
A year ago, our entire company fit into one conference room.
Now, we’ve moved offices twice to keep up with our growth. We’re a team of more than 300 people globally, across engineering, product, and operations. I’m proud of how far we’ve come, and even more energized about where we’re going.
This raise wasn’t out of necessity. Mercor is profitable and operating from strength. The round was driven by inbound investor interest, and we closed it in 5 days.
We’re hiring across nearly every function. Join us: mercor.com/careers
1
12
1,372