Smithy model

Smithy model

The Smithy model is an approach from AI research in which a powerful AI system trains or improves a weaker one — similar to how an experienced craftsman shows an apprentice how to forge a tool. It describes how AI systems can advance one another without a human having to intervene every time.

In the Smithy model, an already powerful AI system trains a weaker or smaller one. The stronger system takes on the role of the teacher: it generates tasks, evaluates answers, or provides feedback. The weaker system learns from this feedback and improves. Humans set the framework conditions but do not intervene in every single learning step. The approach is particularly attractive because it accelerates the development of new, more capable models without requiring a human expert for every training step.

Why AI systems train one another

Improving an AI model normally takes a lot of effort. Humans must evaluate sample answers, flag errors, and incorporate corrections. This takes a long time and is expensive. In the Smithy model, a stronger AI system takes over part of this work.

This matters for two reasons. First, it scales better: a model can produce thousands of evaluations in an hour, whereas a human can only manage a fraction of that. Second, a very powerful model can detect weaknesses that a human evaluator might not even notice — such as subtle logical errors in long derivations. However, there is also a limit: a teacher can only pass on what it itself knows. If the teaching model has a blind spot, it may pass this error on to the student.

How training by a stronger model works

In its simplest form, the stronger model generates a large number of examples — for instance, questions with sample answers. The weaker model is trained on these examples as if they came from a human expert. In research, this procedure is often called “distillation”: the knowledge of the large model is distilled into the smaller one, meaning it is transferred and condensed.

In more sophisticated variants, the stronger model does not just evaluate finished answers but actively accompanies the learning process. It can tell the weaker system which reasoning steps were correct and where it went astray. This is comparable to a sports coach who doesn’t just assess the outcome of a run but also corrects the technique. What matters here is that the evaluating model is actually better than the learning one — otherwise the process goes in circles.

A special case is so-called self-improving training: a model evaluates its own answers and filters out the poor ones. To do this, it needs an internal “evaluation module” that operates independently from the generating part. This approach is technically demanding because the model is simultaneously student and teacher.

Where the Smithy model appears in practice

In the product development of major AI labs, the Smithy principle has by now become standard. OpenAI, Google, and Meta use it to build smaller, cheaper models that nevertheless operate at the level of much larger systems. The GPT-4 model, for example, has co-trained smaller successors that are then faster and cheaper to run while retaining many of its capabilities.

In research, the Smithy model also comes up in debates about AI safety. The concern: if a flawed or manipulated model is used as a teacher, its errors can be transferred across entire generations of successor models. That is why researchers are investigating how to ensure that the teaching model is actually reliable.

In practice, the term is also encountered in discussions about the development of so-called “Small Language Models” — smaller language models that run on smartphones or in enterprise software. They are often built using the Smithy principle: a large model in the cloud handles the training, and the compact result is then deployed locally.

Latest News

Subscribe free. Unsubscribe the second it sucks.

High-signal news across AI, business, UX, and tech. Every morning.