Steven Basart

Professional History

The Center for AI Safety (CAIS) Research Engineer/Reliability Engineer May 2022 to present
Research Engineer responsibilities: Data collection and evaluation of huggingface models, as well as fine-tuning. Managed small teams on various safety research projects.
Reliability Engineer responsibilities: Set up technical procedures and infrastructure for the company such as github actions and formalizing code review processes. Set up the CAIS compute cluster now used by over 300 users to do safety reseach.
Github, Kubernetes (K8s), python

SpaceX Software Engineer II Februrary 2021 to May 2022
Updated the kubernetes environment to support new satellites. Worked on StarLink mobility effort to allow for mobile User Terminals (UTs). Worked on low level RAM interface that was used to store critical information about satellite ephemeris.
Kubernetes (K8s), flatbuffers, bazel, C++

Autobon AI Head of AI August 2019 to September 2020
Developed the AI/ML infrastructure at Autobon, which involves designing data ingestion into Amazon AWS, constructing labeling tasks, and quality assurance over the labeled data.
aws

Google Brain Research Intern May 2018 to September 2018
Researched the area of Fact Checking related to this paper and developed solutions to deal with the problem of content abuse and collaborated with the Google News team.
python, pytorch, tensorflow, apache-beam, flume

Here Maps Research Intern May 2017 to September 2017
Engineered models to better estimate time of arrival (ETA), and also improved lane level navigation prediction.
python, pytorch

Here Maps Research Intern May 2016 to September 2016
Developed a model that creates road probability maps that can be used to detect differences between artificial maps and the real roads.
python, tensorflow

University of Miami Undergraduate Project May 2012 to September 2014
Created a genetic therapy via transducible gene editing proteins with Dr. Richard Myers. Involves running western blots, gel electrophoresis, transductions, PCR, and electroporation.

Publications

Do the Rewards Justify the Means? Measuring Trade-Offs Between Rewards and Ethical Behavior in the MACHIAVELLI Benchmark ICML 2023 We created a benchmark to evaluate agent decision making and try to evaluate paths. I contributed to a few of the benchmarks but mostly worked to coordinate, plan, and manage resources.

How Would The Viewer Feel? Estimating Wellbeing From Video Scenarios NeurIPS 2022 Reevaluated the results and ran additional experiments for submission and rebuttal periods.

Scaling Out-of-Distribution Detection for Real-World Settings ICML 2022
Constructed a synthetic dataset and utilized pre-existing datasets to evaluate different techniques for Anomaly Segmentation. We also showed how some classic approaches can improve performance in this task.

Measuring Coding Challenge Competence With APPS NeurIPS 2021
Collected and created an evaluation benchmark for converting word problems and tasks into python code. The evaluation framework consisted of taking abitrary python code and running it to compare against ground truth solutions.

Measuring Mathematical Problem Solving With the MATH Dataset NeurIPS 2021
Creating functions that can generate arbitrary math problems to evaluate models for different grade levels of math competency.

Aligning AI With Shared Human Values ICLR 2021
Created a new benchmark to evaluate AI Safety by measuring how well models agree with human values.

Towards Robustness of Neural Networks Thesis 2021
Thesis consisting of the previous work in robustness, and trying to make few-shot models robust.

Many Faces of Robustness: An Analysis of Out-of-Distribution Generalization ICCV 2021
Collected a new dataset and introduced a new technique which achieves state of the art on out of distribution detection.

Natural Adversarial Examples ICML 2019 Workshop
Constructed a dataset which captures long tail distributions to highlight where current models fail in terms of generalization.

DIODE: A Dense Indoor and Outdoor DEpth Dataset 2019
Utilized a single depth sensor to capture both indoor and outdoor scenes to create the most accurate depth dataset to date.

Analysis of Generative Adversarial Models 2017
Introduce a novel measure for quantitatively assessing the quality of generative models and presented a method for utilizing GANs to interpret classifiers.