https://wiki.alt-text.eu/index.php?action=history&feed=atom&title=Jascha_Sohl-DicksteinJascha Sohl-Dickstein - Revision history2026-09-15T06:24:05ZRevision history for this page on the wikiMediaWiki 1.45.3https://wiki.alt-text.eu/index.php?title=Jascha_Sohl-Dickstein&diff=165&oldid=previmported>ALT-TEXT: Import: 52 additional AI people glossary entries2026-09-07T20:10:00Z<p>Import: 52 additional AI people glossary entries</p>
<p><b>New page</b></p><div>== Jascha Sohl-Dickstein ==<br />
An American researcher with a background in physics who introduced diffusion models to machine learning in 2015, borrowing from non-equilibrium thermodynamics: gradually destroy the structure in an image by adding noise, then train a network to reverse each step, so that running the reverse process from pure noise generates a new sample. The idea was largely overlooked for five years before becoming the basis of modern image generation. He has worked at Google Brain and later Anthropic, and has written on scaling laws and the pathologies of optimisation. (See also: [[Jonathan Ho]], [[Generative AI]], [[Synthetic media]], [[Scaling laws]])<br />
<br />
[[Category:People]]<br />
[[Category:Artificial Intelligence]]<br />
[[Category:Deep Learning Pioneers]]<br />
</div>imported>ALT-TEXT