FigureAsia 35 Under 35 · Science
Tri Dao
Age 31 · Hardware-aware machine learning systems · Vietnam / United States
Creator of FlashAttention and a leading architect of hardware-efficient sequence models.
- Approximate age at the edition eligibility date
- 31
- Field
- Computer science
- Country or region
- Vietnam / United States
- FigureAsia U35 Assessment
- 94.6 / 100
Profile
Career and documented record
Tri Dao's central insight is that better machine learning systems do not come from algorithms alone; they come from algorithms designed around the physical movement of data. FlashAttention reorganised the attention operation to reduce expensive memory traffic, delivering exact results with far better practical speed and memory use.
By 2025, the method had become infrastructure across major model developers and software frameworks. Dao's continuing work on faster decoding and state-space architectures kept the programme at the centre of the efficiency debate, where compute, energy and accessibility increasingly determine which research can be attempted.
The distinction is not that Dao made a particular language model more capable. He changed a primitive used across the field. That influence is unusually legible: open implementations, reproducible system benchmarks and wide independent adoption connect the original research to day-to-day scientific computing.
FigureAsia selection
Why Tri Dao is on the list
Dao is one of the rare young scientists whose work became part of the operating substrate of a global technology. His 2025–26 record extends rather than merely trades on FlashAttention, preserving a coherent research programme in which mathematical structure, hardware limits and open implementation reinforce one another.
Verified work
The 2025–26 record
Hardware-efficient decoding
Advanced attention implementations that raised long-context decoding throughput on modern accelerators.
MLSys recognition
Received an Outstanding Paper Honorable Mention for systems research in the same programme.
Infrastructure adoption
FlashAttention remained integrated across major model stacks and research frameworks.
Field context
The work in its field
As AI models scale, memory movement often matters as much as arithmetic. Hardware-aware algorithms can therefore shift the feasible frontier without changing model quality or requiring new chips.
FigureAsia U35 Assessment
Assessment breakdown
94.6out of 100
Substantive 2025–2026 contribution
18.7 / 20
Advanced attention implementations that raised long-context decoding throughput on modern accelerators.
Verified scientific impact
14 / 15
FlashAttention is independently adopted infrastructure, and Dao's later work continues to move the systems frontier rather than resting on the original paper.
Originality and distinction
9.5 / 10
The distinction lies in recasting attention around input-output complexity and accelerator memory hierarchies while preserving exact computation.
Field influence
9.4 / 10
The contribution gives hardware-aware machine learning systems a new method, limit or line of argument with relevance beyond one paper.
Individual agency
9.5 / 10
Dao originated and led the FlashAttention programme and continues it through academic and industrial research roles.
Durability and trajectory
4.8 / 5
The record shows continuity at Princeton University and Together AI: this contribution belongs to a wider, sustained agenda.
Asian significance and global relevance
4.8 / 5
Vietnamese-born computer scientist whose work connects the Southeast Asian diaspora with leading United States research institutions.
Evidential validity and reproducibility
7.6 / 8
Open code and hardware benchmarks permit replication across architectures, with speed claims kept specific to measured configurations.
Advance in scientific knowledge
6.7 / 7
The work made memory traffic a first-class object in the design of modern sequence algorithms.
Translational or methodological utility
4.8 / 5
It lowers memory demand and increases throughput for research and production systems using attention.
Responsible research stewardship
4.8 / 5
Open implementations improve inspectability; the profile does not equate compute efficiency with responsible AI as a whole.