Choice / Score speeds up classification, but it is not AGI

What Choice / Score actually does

I would remove the AGI talk from this discussion right away: this is not general intelligence, but a sensible restriction of the answer space. Choice / Score does not generate text freely. The model receives predefined options and returns a score or probability for each one.

Google’s guide to text classification describes the same basic format: the output is a set of class scores that become a probability distribution after normalization. Instead of generating tokens one after another, the system encodes the input and performs a small label-scoring step. That makes the speedup a straightforward architectural consequence, not magic.

The claimed range is 20–200× compared with standard text-generation models. Such a gap is possible because a generator repeatedly runs its decoder to select the next token, while a classifier usually needs one pass and a score calculation for a short list of options. The longer the free-form answer would have been, the more visible the difference becomes.

• Classification and intent detection fit naturally into a closed set of labels.
• Routing and moderation get lower latency without unnecessary text output.
• Probabilities can support thresholds and handling for uncertain cases.
• The model cannot create a new answer that is absent from the options.

The comparison with Google TabFM only works at the level of principle. TabFM handles tabular data, while Choice / Score takes text, but both replace general-purpose generation with structured prediction. They are related by engineering logic rather than being the same architecture.

Where the savings are real and where marketing begins

For closed tasks, this can genuinely be a better replacement for a large LLM, but it is not a new class of intelligence. A compact model can be trained on soft labels, distributions, or rankings produced by a larger language model, then run inference more cheaply and quickly.

The first things I would check are not the attractive speed multiplier, but quality on rare classes, probability calibration, and behavior on inputs outside the training distribution. Without an open model, a comparison methodology, and identical test conditions, the 20–200× range should not be treated as a universal result.

The main limitation is built into the idea itself: the system is only as good as the options defined in advance. Choice / Score removes expensive generation where it was never needed, but loud positioning can turn a careful classifier into a promise of something it is not technically designed to be.

We previously covered how to measure LLM-as-a-Judge reliability with IRT metrics. That evaluation framework complements choice-score models by making model comparisons more consistent and measurable.