Two items posted on LinkedIn in July & September 2017 about machines arriving at solutions – together & without direction – and what this might mean for our development and use of, and indeed learning from, their capabilities.
“Intelligence artificielle : les machines préfèrent parler dans des langues que nous ne comprenons pas,” Romain Bonnemaison, PaparGeek.fr, 24 July 2017 (posted July 2017)
Nice short summary by Romain Pomian-Bonnemaison (in French) of GNMT’s & FAIR’s experiences with AI programs that developed their own languages to communicate with each other.
My thoughts: Would be interested to see how sophisticated those languages might become. How for example might they communicate the contents of what we now know as a resume, or a project bid – things that are complex conglomerations of smaller bits of information that might also be edited on the fly to make the final whole more optimal for the intended outcome?
Wrt the concern about machines communicating in ways we don’t understand, maybe one answer might be an AI linguist to analyze AI languages for the benefit of humans? In any event, it makes no sense to force AI programs to speak English (or any other human language) in their inter-communication. What might we learn from the ways to communicate that they develop?
“What intelligent machines can learn from a school of fish,” Radhika Nagpal, TED Talk, April 2017 (posted September 2017)
Listening to Radhika Nagpal’s interesting TED talk on artificial collective power in robots (modeling swarm behavior where intelligence is distributed), wondering again about applications with bots each seeking to exchange information and optimize relationships based on individually unique (though comparable) criteria. Here not talking about physical robots but rather “intelligent agents” operating in a virtual space.
Although different than a swarm having a common more -or-less tangible objective, this would be similar in that the exact patterns and connections of the whole – the interacting mass of bots – or of any subset would not be designed by any central intelligence. Working on the premise that discrete subsets of information and goals could be programmed into individual bots then sent into an interactive “marketspace” at the behest of individuals or organizations, where they would interact with other bots. Anyway, one Saturday morning’s reflection inspired by this talk.