Comment Can we kill blockchain hype with blockchain? (Score 1) 164
Can blockchain be applied to stop blockchain hype?
Comment Re:!cloud == privacy (Score 2) 50
Our current launch plan is browser addon or plugin. That got lost in the news articles. I'm the project coordinator.
Comment Re:Sounds like a great extension (Score 1) 50
We've been debating a few integration options including an extension, a plugin with an install on first usage, or a captive process that Firefox launches and communicates with. Part of this depends on how much performance we can get out of WebAssembly vs native C++ code with CPU-specific routines. I'm the project coordinator.
Comment Re:Deep Leaning seems so slanted (Score 1) 50
Yes, we are worried about bias in translation. Here's a specific example from people we collaborate with: https://ancillary-proxy.atarimworker.io?url=https%3A%2F%2Fgenderbiasnlp.talp.cat...
Comment Re:The German one is CATASTROPHIC. (Score 1) 50
I can't speak to the spellchecker. But the translation system is able to handle German compounds via Byte Pair Encoding: https://ancillary-proxy.atarimworker.io?url=https%3A%2F%2Fgithub.com%2Frsennrich%2Fs... .
Comment Re:!cloud == privacy (Score 2) 50
Would you be interested in joining a user study? The difficulty with addons is that WebAssembly doesn't support vector instructions over 128 bits wide, which make neural machine translation's matrix operations faster.
Comment Re:!cloud == privacy (Score 4, Informative) 50
That's a good suggestion! The main difficulty with PDF files is you have to reassemble soft-returns into sentences (but not hard returns), possibly OCR, avoid page numbers in the middle of things, etc. It's doable with heuristics but a lot more software is involved.
Comment Re:!cloud == privacy (Score 5, Informative) 50
Your desktop does the translation. The models will be trained on a cluster at Mozilla. I run the project.
Comment Re:Does this actually exist yet? (Score 2) 50
Not in a user-friendly way, but we're working on it! Follow our code at https://ancillary-proxy.atarimworker.io?url=https%3A%2F%2Fgithub.com%2Fbrowsermt%2F and we'll announce features on https://ancillary-proxy.atarimworker.io?url=https%3A%2F%2Ftwitter.com%2Fbergamotpr... .
Comment Re:Since I doubt it will be client-side only (Score 4, Informative) 50
It will be offline. Here's an earlier paper of ours on efficiency: https://ancillary-proxy.atarimworker.io?url=https%3A%2F%2Fneural.mt%2Fpapers%2Fedinb... . Since then we've improved to 982 words per second per core on a Skylake with ~38 MB RAM consumption.
I https://ancillary-proxy.atarimworker.io?url=https%3A%2F%2Fneural.mt%2F coordinate the Bergamot project (https://ancillary-proxy.atarimworker.io?url=https%3A%2F%2Fbrowser.mt%2F) which is a consortium of the University of Edinburgh (lead), University of Tartu, University of Sheffield, Charles University in Prague, and Mozilla.
Comment Re:Useful (Score 4, Informative) 50
It is and will be FOSS with an API that allows integration outside Firefox. The code is a combination of Mozilla Public License and Marian https://ancillary-proxy.atarimworker.io?url=https%3A%2F%2Fmarian-nmt.github.io%2F under MIT license, as well as software to be written. Follow our code on https://ancillary-proxy.atarimworker.io?url=https%3A%2F%2Fgithub.com%2Fbrowsermt%2F and https://ancillary-proxy.atarimworker.io?url=https%3A%2F%2Fgithub.com%2Fmarian-nmt%2F... .
I coordinate the Bergamot project.
Comment Duh (Score 4, Interesting) 61
That's how mobile roaming works. Data is carried back to the mobile provider who then connects it to the Internet locally. US SIM cards block sites that are censored in the US. Film at 11.
Comment Re:AP Stats is the easy AP math (Score 1) 155
Statistics isn't easy. AP Statistics is easy. It's basically how to turn a story problem into numbers you plug into your calculator.
Comment Not what they said (Score 2) 144
They said:
Effective immediately, if we receive information indicating that someone is using our services to traffic in stolen intellectual or physical property from Microsoft, we will not inspect a customer’s private content ourselves. Instead, we will refer the matter to law enforcement if further action is required.
One narrow circumstance that probably won't happen again. In all other circumstances they can read the customer's private content?