Jeremy Nixon

Jeremy Nixon is an AI researcher and entrepreneur working on automated machine-learning research and information retrieval. He is the founder and CEO of Infinity as of September 2026. With Annika Singh, he coauthored OMEGA, a framework that turns algorithmic ideas into executable classifiers and evaluates them. Earlier, his work on Omni explored bringing relevant information and counterarguments directly into a person's working environment.[4][5][2]

Nixon and Singh published OMEGA: Optimizing Machine Learning by Evaluating Generated Algorithms on 29 April 2026. Their framework accepts ideas submitted by humans or generated by language models. For autonomous idea generation, it supplies models and research principles, then asks a language model to propose modifications. Each idea becomes a prompt for generating a classifier.[5]

OMEGA couples generation with execution, repair and evaluation. Generated classifiers must follow scikit-learn conventions, inheriting from BaseEstimator and implementing fit() and predict(), so they can enter existing workflows. Execution errors are fed back into the generation loop for repair; after a fixed number of unsuccessful retries, the framework returns no code. Valid classifiers are evaluated on 20 classification datasets from scikit-learn and OpenML, with accuracy normalized within each dataset before aggregation.[5]

One generated model, MetaSynthesisClassifier, uses stacking: predictions from several base learners become inputs to a second model that learns how to combine them. Its training predictions are produced through cross-validation to avoid exposing the second model to the base learners' training-set bias. Nixon and Singh report that generated classifiers surpassed selected scikit-learn baselines on their benchmark. Those results concern the tested classification datasets; image and video datasets were not tested.[5]

Their comparison of Claude Sonnet 4.5, GPT-4.1 mini, Gemini 2.5 Flash and grok-code-fast-1 found Gemini strongest on average in the presented prompt experiment. They explicitly distinguish generated-classifier performance from algorithmic novelty or long-horizon reasoning. In a separate improvement experiment, revising the generation prompt outperformed revising the code for three of the four models.[5]

Before OMEGA, Nixon described a different application of machine learning: integrating retrieval into creative and analytical work. In his August 2022 essay “Immersive Information Retrieval”, he described building Omni to bring decision-relevant snippets into the interface where work was happening, reducing the need to leave a task, search, read and return.[2]

His proposed mechanisms went beyond finding similar documents. Semantic similarity could surface related ideas and forgotten personal notes; contradiction detection could bring disagreeing research to a scientist writing a paper; and entailment could expose assumptions and implications for criticism. Question answering would supply answers as factual questions arose. These were Nixon's proposed uses, grounded in his broader argument that “Machine learning can replicate the major processes of thought, leading to the decomposition and automation of thought.”[2]

Nixon subsequently presented Omni at the San Francisco AI | UX meetup covered by Latent.Space on 26 April 2023. The publisher listed his demo under syntopical reading, proactive search and omnidirectional linking.[1]

Nixon's earlier research addressed how models' confidence is evaluated and how their training can be improved. He contributed to deep-learning research at Google AI Brain before founding Infinity, and his personal bibliography includes the calibration and learned-optimizer collaborations from 2019.[4][3]

In Measuring Calibration in Deep Learning, Nixon collaborated with Mike Dusenberry, Ghassen Jerfel, Timothy Nguyen, Jeremiah Liu, Linchuan Zhang and Dustin Tran. Calibration measures how closely predicted probabilities match observed accuracy. The work was first submitted in April 2019; its August 2020 revision reports that measurement choices substantially changed the ranking of recalibration methods across MNIST, Fashion MNIST, CIFAR-10/100 and ImageNet. The authors recommended class-conditioned evaluation, L2 comparisons and adaptive binning for the evaluation properties they studied.[6]

Nixon also coauthored Understanding and correcting pathologies in the training of learned optimizers with Luke Metz, Niru Maheswaranathan, Daniel Freeman and Jascha Sohl-Dickstein at ICML 2019. Training an optimizer through short portions of an optimization trajectory produces biased gradients; longer portions can produce exploding gradient norms. Their method dynamically weighted two unbiased gradient estimators to address both problems. By training against validation loss, they also targeted generalization. On the convolutional-network problems tested, the authors reported faster wall-clock training and improved test loss relative to tuned first-order methods.[7]