LLM influenced workflow changes
Google search is Javascript only
This happened a little bit ago (2025-01?). Google search requires Chrome, Firefox, Edge, Safari, or Opera (by user agent) and javascript to be enabled.
w3m 'www.google.com/search?q=plaintext'Update your browser Your browser isn't supported anymore. To continue your search, upgrade to a recent version. Learn more
curl -sL "www.google.com$(curl -sL www.google.com/search?q=date | grep -Po '/httpservice/retry/[^"]+'|sed 1q)"| pandoc -f html -t plain |head -n3[Google] Turn on JavaScript to keep searching
Charitable interpretations of this change include combating (LLM training set related increases in) bot traffic or aligning browser requirements across all google web services. To the first, Javascript might be essential to the fingerprinting needed to more correctly track ad views.
But I suspect LLM inference outputs in search results require javascript and this change at least helps push LLM results forward. Then again, is the number of actual humans without javascript even noticeable to google? Bot traffic must be larger than organic requests from non-js browsers.
Alternatives
My use case was to search from within emacs using w3m.el. At the time, https://news.ycombinator.com/item?id=42747092 reported lynx (user agent Lynx/2.8.6rel.5 libwww-FM/2.14) would still work. But that no longer seems to be the case.
seirdy.one has an excellent write-up of the search engine landscape and is a good place to mine for alternatives. mwmbl github readme has a smaller list.
- duckduckgo is bing and/or yandex but doesn't require JS, has Tor hidden service
- https://www.mojeek.com/
- https://mwmbl.org/ (open source with foundation hosted central server)
- https://stract.com (seridy favorite "smaller")
- https://search.marginalia.nu/ (seridy favorite "non-generalist")
Firefox built in inference
Firefox is running its own local LLM inference for link summary and tab grouping. I only noticed monitoring about:performance. Inference was using ~250Mb or RAM. But it's also been a reported cause of high CPU usage.
For me, while 250Mb/16Gb isn't large, it's still unnecessary. I toggled the about:config option browser.ml.enable to disable.
Yelling at clouds
Things change. Previous workflows break. Tools update with new features/bloat. It's the price of being connected, of benefiting from specialists moving their piece of the world forward.
Whereas google's incentives align with user-hostile practices where "forward" progress is regressive from my perspective, I think the new firefox bloat could at least be benign if not enabling.
Language model drive features could be a useful part of the browser experience. Firefox getting out in front with local models that don't phone home to a central server puts FF in a place to better respect users privacy and choice. And even better that it's easy to disable. Too bad there's no such toggle on google search.