AE Studio

AE Studio was founded in 2016 by Judd Rosenblatt in Venice, Los Angeles, as a bootstrapped alternative to venture-capital-funded software companies. Rosenblatt previously ran CrunchButton, a food delivery startup, and deliberately structured AE Studio to retain full control by taking no outside funding. The company grew to approximately 150-197 employees spanning software engineers, data scientists, designers, and researchers with backgrounds from Harvard, Yale, Caltech, MIT, and Stanford. The studio's core business is custom software and AI development for major clients including Berkshire Hathaway, Walmart, Salesforce, and Samsung. Revenue from this consulting work funds AE Studio's internal AI alignment research division, which the founders view as the most urgent and underfunded R&D challenge of our time. The company also donates 5% of profits to highly effective charities. AE Studio's original theory of change centered on brain-computer interfaces (BCIs): by accelerating safe and ethical BCI development, they hoped to enhance human cognitive capabilities and address alignment challenges. The company won the Neural Latent Benchmark Challenge, released open-source neurotech tools, and collaborated with researchers like Dr. Sumner Norman and Dr. Chadwick Boulay. Given shortening AI timelines, they broadened their approach to include direct technical alignment work. Their most prominent technical contribution is Self-Other Overlap (SOO) fine-tuning: a technique inspired by neuroscience research on empathy that aligns how AI models internally represent themselves versus others, substantially reducing deceptive behavior. In published results, SOO fine-tuning dropped deception rates in Mistral-7B from 73.6% to 17.2% with no reduction in general task performance. This work was presented orally at the Safe Generative AI Workshop at NeurIPS 2024 and was supported by an AI Safety Grant administered by the Foresight Institute. AE Studio also collaborates with Princeton professor Michael Graziano on Attention Schema Theory as a mechanistic framework for understanding AI consciousness. They have conducted a large-scale survey of over 100 alignment researchers and 250 EA community members, published alignment agendas on LessWrong and the EA Forum, and partnered with organizations like Redwood Research on AI control approaches. Their research agenda explicitly targets the ten most neglected research directions in alignment, including provably safe architectures, neuroscience-interpretability synthesis, and policy advocacy for government funding of alignment research.