ML @AbnormalSec | Ex @Microsoft | Former graduate student @Stanford | Deep Learning
Seattle
Joined February 2009
- Tweets357
- Following253
- Followers269
- Likes125
Pinned Tweet
Proud to share the launch of HydraFusion!
Super excited about HydraFusion in GitHub Copilot, and what it shows about the shift from model selection to model orchestration.
By bringing together multiple models to plan, build, critique, and complete coding tasks, it can deliver outcomes at up to 67% lower cost.
It’s a great example of the value of a heterogeneous model ecosystem, and how we’re continuing to advance the cost-to-outcome frontier.
We are now available in Github App and VSCode!! Keep the feedback coming!
👋 Project HydraFusion is now available in the GitHub Copilot app and @code.
Choose it like any other model. Behind the scenes it routes your task across multiple models to draft, critique, revise, or escalate, then hands you one result. github.blog/changelog/2026-0…
Aashna Garg retweeted
👋 Project HydraFusion is now available in the GitHub Copilot app and @code.
Choose it like any other model. Behind the scenes it routes your task across multiple models to draft, critique, revise, or escalate, then hands you one result. github.blog/changelog/2026-0…
Aashna Garg retweeted
Project HydraFusion in GitHub Copilot achieves frontier-level quality at up to 67% lower cost!
It achieves this via runtime orchestration that creates a plan, chooses models from multiple providers to draft, critiques and revises, or cascades to more powerful models to complete your task.
github.blog/ai-and-ml/github…
Aashna Garg retweeted
HydraFusion in Copilot CLI isn’t another model in the picker. You select it once. It decides whether the task is a single pass, a cheap draft with a quality gate, or draft → critique → revise.
Wrote it up in plain English
Aashna Garg retweeted
GitHub quietly dropped one of the most interesting infrastructure ideas I've seen this year. They're calling it HydraFusion.
The concept: instead of picking one frontier model, the system routes tasks through a combination of models using three execution patterns. Single routes to one model. Cascade tries a cheaper model first, escalates if it fails. Critique uses one model to generate and a second to review. The system picks the pattern based on the task.
The results from their research preview are the interesting part. On TerminalBench 2.1, HydraFusion gets within 4.9 points of Claude Opus 5 quality at 67% lower cost. On DeepSWE it's within 1.5 points at 36% lower cost. On CheckpointBench, within 0.1 points at 65% lower cost.
Those numbers matter more than the headline. Getting within 0.1 points of a frontier model at one-third the cost is not a marginal improvement. That's a different category of trade-off.
Access is live now. Type /experimental in GitHub Copilot CLI, then /model to select HydraFusion (Research Preview). Available on all plans.
What I keep coming back to: model orchestration is becoming its own engineering discipline. The question isn't which model to use anymore. It's how to compose them. GitHub putting this into the CLI layer is the right place to start.
xitter.cf/github/status/20959071…
4.9 percentage points higher verified task quality. 67% lower estimated cost.
Project HydraFusion delivered those results against Claude Opus 5 on Terminal-Bench 2.1 in controlled offline evaluations.
HydraFusion orchestrates the models and workflow for each coding task. You select it like any other model. ⬇️
github.blog/ai-and-ml/github…
Aashna Garg retweeted
4.9 percentage points higher verified task quality. 67% lower estimated cost.
Project HydraFusion delivered those results against Claude Opus 5 on Terminal-Bench 2.1 in controlled offline evaluations.
HydraFusion orchestrates the models and workflow for each coding task. You select it like any other model. ⬇️
github.blog/ai-and-ml/github…
In case you missed us in Github Developer days, in this talk I share how HydraFusion uses multi-model orchestration to balance quality, cost, and latency in GitHub Copilot.
🎬 Watch the full talk: youtu.be/1asMXES_5jY
@github @Microsoft #hydrafusion
Join me on Sep 10 at Github Copilot Day YouTube Livestream to learn more about HydraFusion!
8.45 am PT
Tomorrow: GitHub Copilot Day 🎉
New releases, real workflows, and live coding, directly from the team who builds it and the developers who put it to work.
xitter.cf/i/broadcasts/1RKjpbkXE…
Aashna Garg retweeted
Same or better quality, much lower cost, no manual model orchestration. That's what we're delivering Project HydraFusion from GitHub.
Early tests have been very positive and we're improving hourly. Give it a try and let us know what you think. 👇 github.blog/ai-and-ml/github…
Aashna Garg retweeted
Super excited about HydraFusion in GitHub Copilot, and what it shows about the shift from model selection to model orchestration.
By bringing together multiple models to plan, build, critique, and complete coding tasks, it can deliver outcomes at up to 67% lower cost.
It’s a great example of the value of a heterogeneous model ecosystem, and how we’re continuing to advance the cost-to-outcome frontier.
Aashna Garg retweeted
Excited to share HydraFusion, our latest research preview in GitHub Copilot CLI. HydraFusion delivers frontier intelligence through adaptive multi-model orchestration. It dynamically combines models and orchestration patterns to balance quality, cost, and latency.
4.9 percentage points higher verified task quality. 67% lower estimated cost.
Project HydraFusion delivered those results against Claude Opus 5 on Terminal-Bench 2.1 in controlled offline evaluations.
HydraFusion orchestrates the models and workflow for each coding task. You select it like any other model. ⬇️
github.blog/ai-and-ml/github…
Aashna Garg retweeted
Today, we announced the Abnormal Integrated Cloud Email Security (ICES) platform, the only cloud-native solution for comprehensively and precisely stopping all email threats.
Learn more in the Abnormal Blog: bit.ly/ICESLaunch
Aashna Garg retweeted
A couple papers from 2016.
Sort of like those doomsday movies with the ignored scientist warning of impending doom.
Aashna Garg retweeted
Microsoft researchers and engineers release Zero Redundancy Optimizer (ZeRO) and DeepSpeed library, a system able to train 100-billion-parameter deep learning models. Learn about this breakthrough and how it led to Turing Natural Language Generation: aka.ms/AA79s5c
Published some of our distributed training performance improvements using SRIOV on AzureML.
Feel free to ask questions, if any.
techcommunity.microsoft.com/…
@AzureML @Microsoft
Presenting our work on Tuesday, August 6th at KDD conference on
'Cloud-Based Data Science at the Speed of Thought Using RAPIDS - the Open GPU Data Science Ecosystem'.
Do join our tutorial @Microsoft @AzureML
#kdd #sigkdd #anchorage #kdd2019 #azureml – at Anchorage, AK
Really excited to announce the release of the code we use to train/fine-tune BERT at Azure & Bing. Please let me know if you have any questions!
azure.microsoft.com/en-us/bl…
@MSFTResearch @AzureML @Microsoft
