AI in Automotive Podcast · Jayesh Jagasia

AI in Automotive - #504 - Dr Jason Corso, Co-founder and Chief Scientist, Voxel51

·50 min

The episode starts with a broad setup about the convergence between automotive and energy. Utilities are moving toward EV leasing, and OEMs are moving toward energy services. AI is presented as a big accelerator for that shift. From there, the conversation narrows to the challenge of video data. The host points to how much of global internet traffic is video and how hard it is to pull value from it. The guest breaks down the practical questions teams face when building vision AI systems. The discussion centers on how to find the right subset of video to train models effectively. It also focuses on how to rethink annotation so development moves faster. A major thread is the gap between expectation and reality after the first model training pass. The guest describes the familiar situation where a team has already invested in data and still needs more performance before deployment. The conversation notes that there has often been very little good science around what comes next. Instead, teams have leaned on folklore and perceived best practices. One common habit is to keep looping through data, annotation, and training without a clear selection strategy. The episode argues that this can waste money without guaranteeing better results. The guest uses road signage as a concrete example. Most roads have limited signage, and the structure of roads creates uneven patterns that random sampling can miss. Uniform sampling tends to overrepresent ordinary scenes and underrepresent the rare but important ones. That creates a grossly imbalanced dataset. The conversation connects that imbalance to why tools like Voxel51 are useful. The value lies in helping teams select data samples more intelligently. It also lies in helping them manage annotation and keep the model-development loop moving with less waste. By the end, the episode places Voxel51 inside a larger shift. Data is becoming the central asset in visual machine learning. The long-term opportunity is not just general annotation, but a fuller development platform for the way data and training work together.

As heard by us

A sober case for better data workflows over random iteration.

This episode keeps its attention on the practical middle of vision AI: picking useful video samples, labeling them, and avoiding the costly cycle of random collection followed by tiny gains.

Read the full review in PlayNext →

Why you'd press play

Press play if you want a grounded look at how automotive and energy are converging, with AI framed as the accelerator.

Read the full recommendation in PlayNext →

Listen to the show on

Apple Podcasts Spotify