Workbench Versus Appliance for AI
Oscar Beijbom, co-founder and CTO of Nyckel, on the opportunites in the AI/ML tooling market
This split determines whether AI tooling is sold as a workbench for specialists or as an appliance for operators. In house ML teams want knobs, logs, model choices, experiment tracking, and deployment control because they are stitching models into a broader engineering stack. Product owners want to upload examples, check whether predictions look right on their own data, and ship a feature without hiring a dedicated ML engineer.
-
For expert buyers, the product surface is the full workflow. That means labeling data, tracking experiments, storing model versions, deploying models, and watching for drift. This is why point tools like Scale, Aquarium, and Weights & Biases emerged around ML teams that wanted flexibility more than simplicity.
-
For non experts, the useful abstraction is not the model but the input and output. Nyckel is built around that idea, where a tech lead or product manager can upload around 100 examples, review predictions through cross validation, and deploy in a day or two through a self serve API.
-
The market implication is that bundled platforms should gain share as model training needs shrink. Google Vertex AI is the clearest large platform example, while newer products like Nyckel, Akkio, Levity, and Scale are all trying to collapse the stack into one interface and sell closer to the product owner.
Over time, more AI spend should move away from specialist toolchains and toward packaged products that let domain experts train and operate narrow models themselves. The winners are likely to be the companies that hide the most infrastructure, prove results on a customer's own data fast, and expand from one function type into many adjacent AI tasks.