Thinking Machines Lab · posted · checked
Software Engineer, Training Infrastructure
C++PythonPytorchRust$300k – $400kOn-site · New York and San FranciscoMid-level
Who can apply
On-site · New York and San Francisco
Location: New York / San Francisco
ABOUT THINKING MACHINES
The mission of Thinking Machines is to build AI that extends human will and judgment. We are training frontier models with Inkling, developing Tinker to let people make models their own, and crafting interfaces that broaden human-AI communication. We believe the future worth building is human, and we're hiring people who want to build it.
ABOUT THE ROLE
We're hiring a Software Engineer, Training Infrastructure to build the systems that train our frontier models and power Tinker, our fine-tuning API. You will work on the distributed training stack that runs supervised and reinforcement learning jobs reliably across large GPU clusters, from scheduling and fault tolerance to throughput and cost.
Tinker exposes a small set of low-level primitives, so researchers and developers can write their own training loops while we manage the infrastructure behind them. Your job is to make that abstraction hold up at scale: as we add new model architectures, more users, and increasingly demanding workloads like long-horizon reinforcement learning, the platform has to stay fast, stable, and easy to reason about.
WHAT YOU'LL DO
Design, build, and operate the distributed training infrastructure behind Tinker and our internal frontier model training, spanning scheduling, orchestration, checkpointing, and recovery
Improve training throughput and cost efficiency through profiling, parallelism strategies, memory optimization, and kernel- and communication-level tuning
Bring up and optimize training support for new open-weights and in-house model architectures, including mixture-of-experts models, and make new models available to users quickly
Build reliability into the platform: automated failure detection, job recovery, multi-tenant isolation, and observability so that long-running jobs survive hardware and network faults
Support efficient multi-tenant LoRA fine-tuning and tightly coupled sampling and training for reinforcement learning workloads
Collaborate with the Tinker API, frameworks, and product teams to keep the infrastructure and the user-facing SDK consistent and well-behaved