These three things are the same story. US20230076095A1 (published March 9, 2023; a Microsoft research team) generates synthetic data to train language-understanding models. We covered NVIDIA's 2021 synthetic-data patent for vision; this is the same idea pointed at language two years later, evidence the technique migrated up the modality stack.
Connect the dots to the data-wall anxiety. The fear that high-quality human text is finite, and that frontier models will exhaust it, has a standing answer: generate more. By 2023 the major labs were patenting concrete machinery to do this for language tasks, not just acknowledging the idea. The data wall has a workaround, and the workaround is itself IP.
Follow both the money and the IP. Microsoft, deeply invested in language AI across its products and its OpenAI partnership, has direct commercial reason to own methods that keep models well-fed without endless human labeling. Synthetic data lowers the marginal cost of improving a model, which is exactly the lever that decides what's profitable to build.
There's a sharper edge worth naming: synthetic data risks a feedback loop where models train on their own kind of output and drift. The good versions of these methods guard against that, grounding the synthetic data, filtering it, mixing it with real data. The patent is the machinery; the discipline of using it well is the open question.
House caveat: a publication is a method claim, not proof of effect, and synthetic-data quality is everything. But as a dated marker it connects cleanly to the broader arc, from NVIDIA's 2021 vision version to Microsoft's 2023 language version, manufacturing training data went from useful trick to core, patented strategy for staying ahead of the data wall.