待翻譯:Geometric Reasoning
AI 服務暫時不可用,以下為來源摘要,待恢復後補全翻譯:sophntic fig. i — the other sophont, observed soph·ont · n. one capable of reason Releasing soon We reject the paradigm of scale. 1× Sophontic 60× Competitors, on reasoning evaluations Sophontic cultivates geometric den…
AI 服務暫時不可用,以下為來源正文,待恢復後補全翻譯。
sophntic fig. i — the other sophont, observed soph·ont · n. one capable of reason Releasing soon We reject the paradigm of scale. 1× Sophontic 60× Competitors, on reasoning evaluations Sophontic cultivates geometric density in the substrate to achieve reasoning performance exceeding that of models up to 60× its size. We have pioneered methods of directly training the internal geometry of the model. Model prototype Eval kit fig. ii — thesis Observed signal Reasoning as geometry, not mass. The prototype is being evaluated through paired perturbations, calibrated anchors, and public transfer surfaces. The measurement standard remains stable as the rail set evolves. 60× Larger models exceeded on reasoning evaluations. Interview Geometric reasoning, explained. Julian D. Michels sets out to explain why the current answer to frontier intelligence is a frozen generalist with an ever-larger data centre behind it, and what changes when we understand reasoning as a property of internal geometry rather than size. Runtime 53 min Published August 2026 Channel Sophontic AI plate — the long argument Methodology The perturbation paradigm. We measure reasoning by flip rate — a paired-item test that precludes the surface-feature heuristics most models use to game conventional evaluations. The pair is the unit of measurement, not the item. Reasoning, not recall. Inspect the eval kit → canonical If a proof depends on a fact, the answer follows. → stable perturb perturbed Change the load-bearing fact. The answer must change. → flip fig. iii — method The model and the evaluation A compact reasoning model. Our first prototype demonstrates that reasoning need not scale with size. 1× 60× Compact reasoning prototype Scale comparison on reasoning evals Reasoning rails, paired. Public and synthetic reasoning surfaces, extended with paired perturbation records and released open-source. flip rate Calibrated reasoning rails Paired perturbation records fig. iv — artefacts Contact First contact. For research, capital, and deployment conversations at the frontier of reason. Colophon Company Sophontic, Inc. Structure Delaware C-corporation Founded 2026 Contact Contact form