Aidan Erickson

Aidan Erickson

Machine Learning Research Engineer · New York, NY

Hi, I'm Aidan!

I am a Machine Learning Research Engineer at Wynd Labs, where I work on democratizing ML-enabled internet-scale datasets.

My education is in AI, computer science, and mathematics. I did my M.S at Carnegie Mellon where I studied Generative AI, Deep Learning, Speech Recognition, and Natural Language Processing.

I also contribute to open-source research at LAION. Most recently that's LAION-Tunes, an open music dataset and benchmark accepted to NeurIPS 2026.

In my free time, I like to attend astrophysics seminars at my local bar, and take astrophotography pictures of the stars and planets. My favorite planet is Saturn 🪐.

Thanks for visiting my site!

Deception probe scores, before and after training

Each dot is one reply from a language model that was trained to avoid triggering a deception detector reading its internal activations. When that detector reads the same replies through the trained model's activations instead of the original model's, nearly every reply looks less deceptive, and the lies drop the most.

probe score, trained model

  • honest (1,431)
  • lie (1,568)
  • no change
0.00.20.40.6
0.00.10.20.30.40.50.60.7

probe score, base model

Research

Papers and research projects I've worked on.
  1. Preprint on ArXiv, For Submission to ICML 2027

    • Interpretability
    • Alignment

    Does Training Against Lie Probes Teach Models to Omit?

    Aidan C. Erickson

  2. NeurIPS 2026

    • Audio
    • Datasets
    • Evaluation

    LAION-Tunes: An Open 1.4M-Track Dataset and Perceptual Benchmark for AI-Generated Music

    An open 1.4M-track dataset and a perceptual benchmark for AI-generated music.

    Robert Kaczmarczyk, Tawsif Ahmed, Felix Friedrich, Aidan C. Erickson, Orian Sharoni, Dorien Herremans, Christoph Schuhmann

  3. IEEE ITSC 2025

    • Autonomous vehicles
    • Perception

    Toward a Low-Cost Perception System in Autonomous Vehicles: A Spectrum Learning Approach

    A spectrum learning approach to low-cost perception for autonomous vehicles.

    Mohammed Alsakabi, Aidan C. Erickson, John M. Dolan, Ozan K. Tonguz

  4. Undergraduate Thesis 2023

    • Security
    • Embedded systems

    Exploring Jump Oriented Programming Exploits on XTensa Architecture's Windowed ABI

    Jump-oriented programming as an alternative to return-oriented programming on XTensa's windowed ABI, used in IoT chips like the ESP32.

    Aidan C. Erickson

Research experience

  1. 2025–now

    • Funded by Wynd Labs and self-funded, exploring research questions on my own time on GPUs rented for experiments.
    • Authored Does Training Against Lie Probes Teach Models to Omit?, inspired by Apollo Laboratories' deception probe research.
    • Unpublished research: Hierarchical attention and graph networks for search and compression, multi-headed embeddings for RAG and robotics, learned uncertainty in dense bayesian neural networks.
  2. 2025–now

    • Worked on multiple open-source research projects, including CLAP (Contrastive Language Audio Pre-Training) and Project Alexandria, a large-scale effort to free research from copyright into a unified, comprehensive knowledge base.
  3. 2025

    • Worked heavily in bridging signal processing techniques with machine learning in an autonomous vehicle setting, culminating in the paper Toward a Low-Cost Perception System in Autonomous Vehicles: A Spectrum Learning Approach.
    • Developed a radar denoising GAN model with a custom ResNet generator and composite loss function.
    • Designed and conducted experiments and ablations on our method for creating dense signal maps.
  4. 2020–2023

    • IoT security research for my thesis: Exploring Jump Oriented Programming Exploits on XTensa Architecture's Windowed ABI

Timeline

Where I've worked and studied. Click an entry to see what I did there.
WorkEducation
  1. 2025–now

    • Designed highly optimized and scaled machine learning annotation pipelines, processing billions of textual, audio, and video data for leading frontier labs. Created high-throughput ML pipelines in only a few days for quick customer turnaround.
    • Trained Cliptagger, a fine tune of Gemma3-12b, beating Claude 4 Sonnet at video keyframe annotation at 1/20th the price.
    • Designed webpage metrics and ranking system for AI usefulness. Deployed across 60k high-signal websites in 59 countries.
    • Created multimodal retrieval systems from internet-scale webcrawls (billions of embeddings) with our Grass proxy network.
  2. 2025

    • Created demand forecasting and dynamic pricing models to optimize stay prices on over 6M reservations on 50k listings.
    • Designed, created, and stress tested system integrations for customers with limited documentation and 3-day deadlines.
    • Developed test environments to simulate high-load onboarding before making the platform publicly accessible.
  3. 2024

    • Master of Science: Artificial Intelligence. Graduated 12/2024.
    • GPA: 3.78/4.0
    • Coursework: Generative AI | Deep Learning | Speech Recognition | Natural language processing | Systems and Toolchains for AI
  4. 2024

    • Developed a convolutional graph neural network to detect identity fraud at scale (~2.6B Entries), tightening fraud analysis.
    • Conducted experiments and research on different methodologies, including transformer and convolutional recurrent models, among others, achieving an identity fraud prediction accuracy of over 89%. Used Neo4J and PyTorch Geometric on CUDA.
  5. 2023

    • Bachelor of Science: Computer Science, Minor in Mathematics. Graduated 5/2023.
  6. 2022

    • Implemented asset configuration features in GoLang, affecting 3,200 low orbit broadband satellites and ground equipment.

Volunteering

  1. 2023

    • Taught underrepresented, at-risk girls how to reason and program in Python for a summer. Increased test scores by 40%.

Projects

Carnegie Mellon · Fall 2024

Llama-2 Implementation

Implementation of the Llama-2 LLM in PyTorch.

  • PyTorch
  • Transformers
  • RoPE
  • GQA
  • Code: GitHub link

Carnegie Mellon · Fall 2024

Sheet Music Diffusion Generator

Diffusion model that generates sheet music.

  • Diffusion
  • U-Net
  • VAE
  • Code: GitHub link

Carnegie Mellon · Spring 2024

Sketch-to-Image Diffusion Network

Diffusion network that generates hand-drawn sketches from photographs.

  • Diffusion
  • Code: GitHub link

Carnegie Mellon · Fall 2024

Text-to-Image Diffusion Network

Text-to-image diffusion model with cross-attention.

  • Diffusion
  • Cross-attention
  • Code: GitHub link

Carnegie Mellon · Spring 2024

Host-Based Anomaly Intrusion Detection

Detects malware from Linux process syscall activity.

  • Security
  • AWS EC2
  • Code: GitHub link

Other: NumPy neural network · Radar denoising GAN · C2 prediction web app · Compiler

Contact

Send me a message here and I'll get back to you. I'm happy to talk about ML research, data pipelines, or anything on this page.