AI Tooling Shift To Product Platforms
Oscar Beijbom, co-founder and CTO of Nyckel, on the opportunites in the AI/ML tooling market
The big shift is that AI tooling is moving up the stack, from helping ML engineers build pipelines to helping product teams ship features with a small batch of labeled examples. In that world, the fastest growing product is not experiment tracking or annotation workflow software, it is the simple API or UI where a product manager uploads 50 to 100 examples, checks predictions on their own data, and puts a classifier into production the same day.
-
The interview lays out two different markets. One is expert MLOps, where teams stitch together labeling, experiment tracking, model registries, deployment, and monitoring. The other is the non expert workflow, where the user only specifies inputs and outputs and the platform handles training, deployment, and monitoring behind the scenes.
-
That shift matters because smaller training sets remove the need for a lot of outsourced labeling and pipeline glue. Nyckel describes customers annotating about 100 examples themselves in roughly 20 minutes, then training and deploying immediately. That is a very different workflow from the old Scale style model built around large annotation volumes.
-
The implication is not that labeling disappears, but that demand gets narrower and more specialized. The strongest remaining cases are open ended environments like autonomous driving, where edge cases explode, while more bounded document and business workflow tasks can increasingly be handled with pre trained models plus light fine tuning.
Going forward, AI infrastructure should consolidate around products that hide the model and expose a clean business workflow. The winners are likely to look less like tool vendors for ML teams and more like application platforms for developers and product owners, with human labeling shifting from mass annotation toward targeted review, evaluation, and hard edge cases.