This challenge has also arisen at @OptibleAI. During the final 10 percent of work, we recently added support for Gemini-based translation within Chromium to keep translations inside the browser, since our files are often large; our testing shows it performs well. We will observe how it manages different cases in the coming days.
On-device inference isn't free. It moves the cost into engineering time, edge cases, and your users' laptops.
We replaced 14 in-browser translation models with a cloud model. No more big downloads, failures we can reproduce, and better translations.
Nico explains the tradeoff: https://t.co/l3Xgsbaffl
Our Chrome extension can't read the page it sits next to.
It puts the assessment panel beside a record in whatever system your team already works in, so an assessor scores where their pipeline lives instead of copying between two tabs. The easy way to build that is to let the extension read the page and pull the record out of it, which hands it access to everything on every page of your CRM, for everyone who installs it.
Ours reads the address bar and nothing else. The record is already identified in the web address, so there was nothing on the page worth asking for.
If you've built one of these, where did you land on permissions? I want to know whether anyone needed the page contents.
A programme officer asked me last week where the private notes go in our product, and who could see them. Then answered it for me: the assessors will just use pen and paper.
Our notes field is shared with the rest of the panel once you submit an assessment. Assessors are scoring public money, so they assume anything typed in could be pulled back up later. If that's the deal, paper wins.
So we shipped a second notes surface that stays in the browser. It never reaches our servers, and it's gone when browser data is cleared. We can't read it and neither can the other assessors.
It came up on a Wednesday morning call. It was in the product by lunchtime.
How much of what your users actually do is deliberately off your system?
Our scores cite their evidence. Each one links back to the answer or the page it came from.
We got this wrong at first. Applicants edit answers, documents get replaced during correction rounds, and the link still resolves afterwards. The score keeps looking well supported while the thing it points at has changed.
So we check sources now instead of just storing them. If an answer or a document moved after the assessment ran, the score says so.
Curious if anyone's handled this properly. Seems like it'd apply to anything that cites its inputs.
@VivekRabara @UberINSupport Yes uber people never read message. They automate the message. I tried raising query but got tired of their nonsense replies