Nokia's applied research center recently open-sourced AnyJev to prove one thing: chat AI can also serve as a fast classifier, delivering stable decisions in a single inference pass.
It sounds like hard technical detail, but our editorial judgment is simple—this represents another push along the path of "LLMs are not chatbots." For people doing classification, routing, and content moderation inside enterprises, it's a direction worth watching.
What this is
Today, when we use LLMs for classification tasks (identifying ticket types, judging whether an email is spam), the typical workflow is: let the model generate a piece of text, then parse the label out of that text. The process is slow, expensive, and the results aren't stable—the same input can yield slightly different textual descriptions.
AnyJev takes a different path: skip the generation stage entirely and directly extract signals from the neural network's middle layers (the "internal states") as classifier output. The researchers claim this is more stable, faster, and cheaper, and requires no fine-tuning of the model.
Because it operates on the model's internal layers, it works on locally deployed open-source models too—which is why it has caught attention in the local open-source LLM community.
Industry view
Supporters read this as a signal of enterprise AI use-case differentiation: conversation belongs to conversation, decision-making belongs to decision-making. Extracting classifiers directly from off-the-shelf LLMs is good news for companies without an algorithm team—in theory, it could drop backend classification costs by an order of magnitude.
But cooler voices on Reddit and GitHub push back: first, "trustworthy" is the original author's own claim, and there are currently no third-party independent reproduction reports; second, the method has requirements on model architecture and won't necessarily apply to every open-source model; third, Nokia itself getting into foundational AI research is unusual—a traditional telecom company investing at the foundation-model layer may have genuinely spotted an opportunity, or may simply be R&D budget looking for an outlet.
Impact on regular people
For enterprise IT: backend tasks like classification, routing, ticket triage, and content moderation may in the future directly reuse off-the-shelf LLMs, with no need to train separate classification models. In theory, costs could collapse.
For individual professionals: people doing data analysis, customer service, and content operations will see AI tools respond faster and cheaper—but only if their companies are willing to rework the technical architecture, and in the short term most firms have no motivation to change.
For consumer markets: consumers will barely notice anything. This is an optimization of backend efficiency tooling, not a user-facing new product.