Here is a bullet point summary of Mark's answers at NIPS 2013:
- wants to publish.
- separate from the ads team (ads team is already a ML powerhouse) and the newsfeed or photo teams.
- wants to use the dataset of FB that he "can't" let out for privacy concerns.
- staff: Marc'Aurelio Ranzato / Yann Lecun (director) / Rob Fergus, a lot of other rumors going around at NIPS.
- acquisition (acquihired?) of a speech-related startup (Mobile Technologies) is in this direction too, using FB messenger's video calls as data.
- will invest in clusters (mentioned GPUs) / large scale (obvious)
- wants to take advantage of multimodal input (ImageNet-like problems with tagged photos and instagram)
- compared the long-time vision he has for the AI/ML lab(s) as an exo-cortex which is the next step after cortex -> neo-cortex.
- http://imgur.com/QDDDMGm Mark Zuckerberg / Yoshua Bengio
- there should be the video on the NIPS website at some point.
To sum up, pretty good PR, with seemingly honests short-time goals (solve ImageNet-like problems) and long-time vision ("strong" AI).
In the book a punk kid steals one of the protagonist's AR glasses and computing devices. When the kid puts these devices on he, in effect, becomes almost like the original owner. He is now "tapped in" to the original owners exo-cortex.