Own the ML workflow not models

The durable company in AI tooling owns the workflow, not the model. Nyckel is built so customers describe the job in plain terms, classify this text, flag this image, detect this scam, and then upload a small set of their own examples. Behind the scenes, Nyckel can swap among in house models, open source models, or foundation models, because the product it sells is reliable prediction on customer data, not access to one architecture.

• This is why Nyckel hides model choice from the user. Customers upload data, label roughly 10 to 100 examples, see cross validated predictions on their own records, then tweak labels or add more samples. The value is the fast loop from data to working output, not the specific engine underneath.
  1 sacra
• Foundation models push the market toward this abstraction. As pre trained models reduce the amount of labeled data needed, point tools built around one training method get weaker, while products that combine labeling, training, deployment, and monitoring into one flow get stronger.
  1 sacra 2 sacra
• The contrast with Scale shows the strategic difference. Scale started as a data labeling business tied to huge edge case heavy datasets, then had to expand into the rest of the stack as foundation models reduced dependence on massive manual labeling. That shift shows why anchoring a company to one technical layer is fragile.
  2 sacra 3 sacra

The market is heading toward ML products that feel like Twilio or Stripe for prediction APIs. Winners will keep changing the underlying models as costs fall and capabilities improve, while keeping the user experience fixed around inputs, outputs, evaluation, and deployment on real customer data.