Pinned Tweet
We won 1st place at the @JaneStreetGroup x @GPU_MODE Hackathon in NYC this weekend! 🚀⚡
Our challenge was to a real-time inference server for an ensemble of sequential stateful models (Mamba2, xLSTM, etc.) processing streaming market data. The goal was to maximize the PnL of our trading system, which demanded high model accuracy while decreasing latency, maximizing throughput, and maintaining live uptime in production. We worked across the full inference stack, from high-level batching algorithms to GPU utilization optimizations.
Our production engine ultimately processed ~400 requests per second with ~30ms of latency, achieving the highest PnL of the competition! Our key techniques:
💡 Dynamic batching and state management to maximize throughput while preserving sequential inference accuracy
💡 Profiled and eliminated CPU <-> GPU communication overhead, removing synchronization points and bottlenecks
💡 Reduced kernel launch overhead with PyTorch optimizations like torch.compile
💡 Fast state expansion/reduction strategies to minimize batch latency
💡 Explored model quantization and custom Triton kernels to fuse operations and improve GPU utilization
Huge thanks to @marksaroufim and @GuggerSylvain for designing a deeply interesting open-ended technical challenge and a smooth contest experience, and the real-world insights throughout!
Rathul Anand retweeted
Introducing TimesFM-3, a state-of-the-art time series foundation model that enables accurate multivariate time series forecasting in a single forward pass, significantly outperforming other forecasting models across major benchmarks.
More on the blog →goo.gle/4x5WGpD
ALT TimesFM-3 architecture.
Rathul Anand retweeted
The frontier has increasingly shifted to hybrid models - from Qwen to Kimi-Linear and now with NVIDIA's Nemotron-3 Super - that rely on a strong linear sequence model. Today we release Mamba-3, the most powerful linear model to date.
xitter.cf/_albertgu/status/20339…
The newest model in the Mamba series is finally here 🐍
Hybrid models have become increasingly popular, raising the importance of designing the next generation of linear models.
We've introduced several SSM-centric ideas to significantly increase Mamba-2's modeling capabilities without compromising on speed. The resulting Mamba-3 model has noticeable performance gains over the most popular previous linear models (such as Mamba-2 and Gated DeltaNet) at all sizes.
This is the first Mamba that was student led: all credit to @aakash_lahoti @kevinyli_ @_berlinchen @caitWW9, and of course @tri_dao!
Check out our implementation at github.com/rathull/jane-stre… :)
This wouldn’t be possible without my incredible teammates Kyle Yu (linkedin.com/in/kylecyu/) and Aswinkumar (linkedin.com/in/aswinkumar99…) 🙌
Especially grateful to Jane Street team and the broader GPU MODE community for giving us a taste of the demands of ML infra for low-latency trading, to Tri Dao and the PyTorch team for sharing their insights on the future of GPU programming models, and to CoreWeave and Northflank for the H100s and support! 💸
Rathul Anand retweeted
Final moments at the @JaneStreetGroup hackathon today, everyone monitoring the leaderboard before the winners are declared
Rathul Anand retweeted
We just wanted to cowork past 5pm. Turns out the entire city did too :)
@samanthaaouyang and I first crossed paths in Turkey earlier this year. In April, we reconnected at a women founders’ event and nerded out for hours about café culture, cities that sleep too early, and what it would take to build something different this summer.
Two days later, we had a full Notion doc and a dream for a late-night popup called Elsewhere. A third space that would be a little less lonely than working from home, but a little more magical than your usual café.
When @elsewhere_today launched on July 16th, we got way more inbound than expected: 100K+ views in 24 hours and 700+ new followers from almost nothing. Still, the real test: would SF actually show up?
400+ people bought in. During our first pop-up last Thursday, we had a line out of the door before 7pm. We served hojicha lattes, blue Thai tea, crepe cakes, fruit tarts, coconut macarons, and, of course, endless matcha. People brought their laptops and stayed there working until 11pm on passion projects. Someone even paid in USDC on Solana! (other chains tap in 👀)
At first, some skeptics on X were quick to dismiss “just another late-night café attempt.” But honestly, if we were able to make SF even 1% more alive that night, I think that’s beautiful :)
Huge thanks to the best team in the world: Sam, Akshaya (@akshayadinesh19), and Daniel (@Bluezmango123). Thank you to Fayeeza and the Entrepreneurs First team (@join_ef) for an incredible venue. And much love to Jade and Homeroom for an amazing collab!
Let us know what other cities to stop by! Stay tuned for more from Elsewhere ❤️
Rathul Anand retweeted
🔥 Semgrep is officially live on Cursor!
You can now harness the power of @semgrep directly in your AI coding assistant, combining fast, accurate static analysis with LLMs to help developers ship code that’s secure from the start, fast.
From securing code at leading AI companies to joining the @cursor_ai tools ecosystem, Semgrep is becoming essential for dev-first security in the modern stack.
Shoutout to our team for making this integration happen, and to our customers, partners, and community for pushing us forward 🚀
docs.cursor.com/tools/mcp
#Cursor #AppSec #DeveloperTools #SecureCoding #LLM #StaticAnalysis
