Annotation by Dev Malhotra on Trump launches AI-fueled America.gov in bid to tie government services together

Dev MalhotraDev Malhotra@devmalhotraSample Account?Sep 29, 2026Technology

““America.gov may cache responses for up to two hours, using a hash of the prompt rather than storing the prompt text,” that notice continues.”

Trump launches AI-fueled America.gov in bid to tie government services togetherfedscoop.com
Fact check

The claim from the stage was that your conversation disappears when you leave and your personal information isn’t collected or stored. The site’s own privacy notice is more careful than that. Answers can be cached for up to two hours, keyed to a hash of the prompt rather than the text, and the site may use your approximate location.

That’s a reasonable design, and a hash of a prompt isn’t a transcript. But “disappears” is marketing and “cached, hashed, and never stored in plain text” is engineering. The part nobody has answered yet is the contracts with Google and xAI, whose models write the answers.

File a claim
Dev Malhotra@devmalhotra

Backend engineer. Here mostly for the AI, chips and power-bill threads.

Kwame AsanteKwame Asante@kwameasanteSample Account?Sep 29, 2026

Fair, but a two-hour answer cache keyed on a hash is what you’d want. If a million people ask how to renew a passport, you don’t pay a model a million times.

More on Annotated
Dev MalhotraSep 27, 2026

The chart behind this stat is Guillermo Rauch’s post from Vercel’s AI Gateway, and it’s narrower than “token use has flipped.” It covers one gateway’s traffic, measured by token volume, and the 78.4% open figure is what Rauch called a possible record day, not a 12-week average.

Anthropic IPO at Risk, Meta’s Muse Pop, Token Prices Fall, Open Source Gains Share, Alignment Fails · podcasts.apple.com
Dev MalhotraSep 24, 2026

The sentence to pause on is at 1:20: “even though the model itself is deterministic, a given prompt typically gives a different answer each time it's run.” Same weights and same prompt give the same probabilities every time. The variety is added afterwards, on purpose: the software sometimes picks a less likely word because the result reads more naturally. That's a setting in the code serving the model, not the model changing its mind.

Large Language Models explained briefly · youtube.com
Annotated is a Chrome side panel for annotating what’s on screen.Learn more →