Sebastian Raschka starts with a useful correction to the Jev hype: classification isn’t new. The interesting engineering question is whether one model can handle many different bounded decisions without retraining for each one. His long technical essay puts Jev beside bag-of-words, recurrent networks, ModernBERT, and generative models, then reports his own Jev tests.

An actual task, not a launch chart

  • Raschka reports testing Jev on the IMDb test set of 25,000 movie reviews: 96.47% accuracy using choice, and 96.20% using binary noul. Those are his measurements, not tests reproduced here.
  • He reports about 22–23 minutes and roughly $0.65 for each complete API run. Those timings include his run conditions and should not be mistaken for a controlled local-versus-cloud latency comparison.
  • His task-trained ModernBERT was competitive on movie reviews; Jev’s potential advantage is answering a different decision task without another fine-tune. An IMDb benchmark alone cannot establish that broader advantage.
  • He notes that the public IMDb test set could have been included in Jev’s undisclosed training data, and a repeat Jev run gave slightly different answers.

What an open alternative would need

  • Raschka sketches a reusable scoring head: encode the state, instruction, and each candidate description, score each with the same learned head, then normalize across the candidates. The head need not be resized when the option set changes.
  • Building that interface on ModernBERT is straightforward; teaching it to perform reliably across unfamiliar tasks is the hard part. He built a prototype but deliberately did not publish it as a Jev replacement.
  • His own trials of Contrastive Language Models and Laya reported lower IMDb accuracy and failed his Tetris demonstration. Those are his particular tests, not a comprehensive or independently replicated ranking of open models.

Probabilities need their own test

  • Calibration asks whether predictions assigned, say, 70% probability are correct about 70% of the time on comparable held-out examples. Producing a number—or an answer in the required schema—does not establish either calibration or correctness.
  • Raschka explains temperature scaling and a calibration penalty he tried on a ModernBERT classifier; he says the extra penalty offered only a modest benefit in that setup. The associated training experiment is not a demonstration of Jev’s proprietary method.
  • TypeSafe calls its training method RLCD but hasn’t published the recipe. Raschka explicitly labels his comparison with a separate, published calibration-reward method as speculation, not a description of Jev’s internals.

His conclusion is narrower than “classifiers are obsolete”: use simple task-specific classifiers when labels and a stable task justify them; use a general decision interface when tasks keep changing; evaluate the probabilities, errors, and fallbacks before putting either in charge of a consequential branch.