Co-founder, Acyclic Labs (YC F26) building faster, cheaper and bigger agent swarms Cambridge CS 22-25: research on parallel PL
London
Joined October 2021
- Tweets74
- Following97
- Followers97
- Likes166
Pinned Tweet
Hello SF, I'm Varun!
@RamV2003 and I just moved here for YC F26 with the goal of getting all agent work to log time execution via swarms.
Would love to connect with those building super ambitious infrastructure and crazy agent setups. If you have any wild projects, please send them my way!
Varun Latthe retweeted
This is our YC launch.
I talk about armies of tiny servers.
you should watch it.
Varun Latthe retweeted
We built a lie detector for AI.
Today we're launching Latent (YC F26). It reads a model's mind while it's answering and stops the wrong answers before anyone sees them.
Works with any model, at 500x less than paying another AI to check the work.
Years of interpretability research taught me one thing: the model often knows it's about to be wrong. It just doesn't tell you. Now it does.
Try it free at runlatent.ai
We gave blender to an agent swarm over just 15 minutes to create a complete model of an airship and this is what it achieved:
- 3 floors, 30 rooms and an exterior shell
- 4313 distinct parts and a 271 piece furniture library consisting of over 338k total triangles
- A rover hangar, drones, turbine hall, hydroponic garden, 3 distinct laboratories, an infirmary, radio room, a flight bridge
- Sleeping quarters, laundromat, communal mess, sky lounge (incl. musical instruments), library and much more
Resource usage:
- 230 agents
- 228 peak, 93 average CPU cores
- 2998 tool calls
Later this week i'll give it a shot at something photorealistic
Varun Latthe retweeted
🚀 Introducing DeepInfer, our in-house inference framework, now fully open source!
Powering DeepSeek-V4.1-Flash at scale with high throughput and low latency. We’re sharing our production stack to help the community lower compute costs, reduce energy use, and make AI more affordable.
🔗 github.com/deepseek-ai/DeepI…
1/4
there's so much you can do with a primitive like this, these projects will be 🔥
@IncTerse is throwing the first (to our knowledge) Actor Hackathon at @ycombinator!
Join us October 31st and build same scary good agents, drink some coffee made by me, and build alongside the top builders in SF.
Prizes include:
- Guaranteed YC Interview
- iPhone duo
- Mac mini
- more to be announced soon!
events.ycombinator.com/haunt…
Varun Latthe retweeted
Demo 3: acyclic in 37 seconds.
npm install → sign in → describe the task. The swarm splits the work across agents in parallel, and one command applies the patch. Or let Claude Code drive it.
acyclic.dev/guides/setup
Recorded this at SFO after 113 agents researched 153 companies and drafted tailored outreach for each.
Not mass spam. Parallel research, personalised drafts, human review.
This is what agent swarms unlock.
Varun Latthe retweeted
agent swarm demo 1: medicare glp-1 eligibility check for a clinic
what happened:
21 agents, 64 web lookups, every quote checked against its source
it caught a prompt injection hidden in white text on a flyer and didn't follow it
logs below
starting a miniseries on how we're rebuilding the cloud for swarms.
first up is the low-latency forkable durable stream that provides the foundation for everything else we make.
went to a @ycombinator event with @paulg @jesslivingston @garrytan and the stories he had were wild!!
unfortunately didn't manage to fight through the crowd tho 😕
Varun Latthe retweeted
YC just sent us a merch pack with a bodybag coat to show us what might happen if we don't get to 1M ARR by demo day
YC just sent us a merch pack with a bodybag coat to show us what might happen if we don't get to 1M ARR by demo day
I'm so tired of people anthropomorphising swarms
here's a short list of operations that you can do in a swarm that you can't (shouldn't) do to people:
1) clone them
2) kill them
3) edit their memory
Read @hallerite's essay here: primeintellect.ai/blog/on-th…
overheard in Ghirardelli Chocolate
guy 1: I'm also open to surrogacy
guy 2: holy shit what a skill issue
big if true
Backprop has been the only credit assignment algorithm capable of training large neural nets.
Introducing Dust, the first zeroth-order method to pretrain transformers to approach and sometimes even *exceed* backprop with large amounts of computation (bitter lesson!)
- Dust is on the order of 1,000 to 10,000x more compute efficient than EGGROLL, the state-of-the-art ES method, for training transformers.
- We chose pretraining transformers because it's arguably the hardest task possible for zeroth-order optimization.
- Search in high dimensions is really misunderstood: zeroth-order usually gets better with overparameterization, not worse, i.e. at a fixed population size, larger models (up to 120x) reach a lower loss than smaller ones.
- Backprop-like gradients emerge from a large population of activation perturbations, and the alignment holds up at every scale we tested, up to 1B tokens, which is critical for scaling.
The core idea is a *virtual population*, which helps scale to large populations in transformers by perturbing activations in parallel.
w/ @bishmdl76, @cs_serdar, @akshayvegesna