Benjamin F Spector
Benjamin F Spector is a cofounder of Flapping Airplanes and researches machine-learning systems.
Contents13 sections
Key facts
2021Learned String Indexes
He co-authored RadixStringSpline (RSS), a learned index that predicts where strings lie in ordered data. It arranges radix splines in a tree, processes only the prefixes needed to distinguish keys, and bounds prediction error to narrow the final search.[1]
2022MIT Degrees
In 2022, he received MIT bachelor’s degrees in computer science and engineering and mathematics, and an MEng in computer science.[2]
His MIT advisor was Tim Kraska.[3]
2023Hertz Fellowship and Prod
He received a 2023 Hertz Fellowship and founded the nonprofit startup accelerator Prod.[4]
August 2023Staged Speculative Decoding
With Chris Ré, he studied small-batch, on-device language-model inference, proposing tree-structured speculative batches and a second speculation stage to reduce generation costs and increase the tokens accepted per verification.[5]
October 2024ThunderKittens
He co-authored ThunderKittens, a layered approach to GPU kernel programming for AI: warp-level operations use 16×16 tiles, thread blocks overlap asynchronous work, and grid-level abstractions reduce launch and memory-access overhead.[6]
November 2024Precision and Scaling
He co-authored Scaling Laws for Precision, incorporating numerical precision into scaling analyses of training and inference. It examines effective parameter counts at low precision and how additional pretraining data can increase degradation from post-training quantization.[7]
May 2025Low-Latency Megakernels
With Hazy Research collaborators, he placed a complete Llama-1B forward pass inside one GPU megakernel to reduce stalls at kernel boundaries. A GPU interpreter schedules work with shared-memory paging and counter-based synchronization, targeting low-latency single-request inference.[8]
September 2025Throughput and Multi-GPU Execution
The team subsequently targeted tensor-parallel Llama-70B inference on H100 GPUs, overlapping computation, memory operations, and communication through an instruction-and-interpreter framework. The authors also cautioned that the research code depends on compiler and hardware configurations and was not intended to receive ongoing support.[9]
November 2025ParallelKittens
He co-authored ParallelKittens, extending ThunderKittens to multi-GPU kernels. The team examined tradeoffs among network-transfer mechanisms and schedules, overlapping communication and computation at different levels while retaining tile-based programming.[10]
As of October 2026: Flapping Airplanes
He co-founded Flapping Airplanes with Asher SpectorAsher SpectorAsher Spector is a statistics researcher and co-founder of Flapping Airplanes.Open the full entry and Aidan SmithAidan SmithAidan Smith is an engineer and co-founder of Flapping Airplanes who previously worked at Neuralink. He was selected for the 2025 Thiel Fellowship.Open the full entry.[11]
The lab focuses on data efficiency; its website reports $180 million raised and describes its work as still in stealth.[12]
Stanford Research
He formerly pursued a computer science PhD at Stanford University with Chris Ré.[3]
Related Organizations, Websites, and Public Accounts
MIT: Official website: https://www.mit.edu/ (opens in a new window)[2]
Stanford University: Official website: https://www.stanford.edu/ (opens in a new window)[3]
Hertz Foundation: Official website: https://www.hertzfoundation.org/ (opens in a new window)[4]
Hazy Research: Official website: https://hazyresearch.stanford.edu/ (opens in a new window)[8]
Prod: Official website: https://www.prod.so/ (opens in a new window)[13]
Flapping Airplanes: Official website: https://flappingairplanes.com/ (opens in a new window)[12]
Benjamin F Spector Personal Website: Official website: https://benjaminfspector.com/ (opens in a new window)[3]
Public X Account: Official account: https://x.com/spectorb (opens in a new window)[14]
Sources
- Bounding the Last Mile: Efficient Learned String Indexing (opens in a new window)
- Six from MIT awarded 2023 Hertz Foundation Fellowships (opens in a new window)
- benjamin f spector (opens in a new window)
- Benjamin Spector (opens in a new window)
- Accelerating LLM Inference with Staged Speculative Decoding (opens in a new window)
- ThunderKittens: Simple, Fast, and Adorable AI Kernels (opens in a new window)
- Scaling Laws for Precision (opens in a new window)
- Look Ma, No Bubbles! Designing a Low-Latency Megakernel for Llama-1B (opens in a new window)
- We Bought the Whole GPU, So We're Damn Well Going to Use the Whole GPU (opens in a new window)
- ParallelKittens: Simple and Fast Multi-GPU AI Kernels (opens in a new window)
- Flapping Airplanes (opens in a new window)
- Flapping Airplanes (opens in a new window)
- Prod (opens in a new window)
- Benjamin F Spector (@spectorb) (opens in a new window)