Co.Design
AI Is Inventing Languages Humans Can’t Understand. Should We Worry?
Researchers at Facebook realized their bots were chattering in a new language. Then they stopped it.
Recently shared by
Take a deep dive
More like this
TNW
Facebook's AI accidentally created its own language
TNW
3 min read · 160 saves · From 2017 · Researchers from the Facebook Artificial Intelligence Research lab (FAIR) recently made an unexpected discovery while trying to improve chatbots. The bots — known as “dialog agents” — were creating…
Benedict Evans
Chat bots, conversation and AI as an interface — Benedict Evans
Benedict Evans
8 min read · 312 saves · From 2016 · Chat bots tap into two very current preoccupations. On one hand, the hope
that they can actually work is a reflection of the ongoing explosion of AI,
and on the other, they offer a way to reach users without having to get
them to install an app.
One way to look at any given AI problem, in the crudest terms, is to ask
whether solving this needs 'general AI' or whether your domain is narrow
enough and your solution broad and flexible enough that you can deal with a
wide enough percentage of potential scenarios without having HAL 9000. That
is, we're clearly pretty close to making a car that can handle highway
driving with no human input, since the range of possible events is pretty
small, but is driving in central Rome or Moscow is an entirely different
matter that needs an entirely different level of decision-making? What does
that hand gesture mean?
Hence, the challenge in plugging an AI into a 'conversational' chat bot
interface is that you don't have HAL 9000 but are in some sense pretending
to the user that you do. You speak to it, it speaks back, and it uses
natural language (either in voice or text), but it isn't general AI at all.
So, how likely is it that the users can ask something that breaks it? How
badly? How narrow is the domain of questions that they might ask, and how
can you define the users' expectations such that they understand the
domain? Perhaps more importantly, can you get people to accept the domain?
A good way to see this problem in action is to compare Siri and Google Now,
both of which are of course bots avant la lettre. Google Now is push-based
- it only says anything if it thinks it has something for you. In contrast,
Siri has to cope with being asked anything, and of course it can't always
understand. Google Now covers the gaps by keeping quiet, whereas Siri
covers them with canned jokes, or by giving you lists of what you can ask.
The actual intelligence might (for the sake of argument) be identical, but
you see Siri failing.
To invert this, the Siri 'this is what you can ask' screen is essentially a
command line help prompt, whereas the whole point of a GUI was that you
didn't have to know what you could type anymore. Though natural language
processing means you don't need to know a specific syntax for Siri or a
bot, there's still a basic discovery issue - what can I ask, given I can’t
ask anything?
All of this means that for now, it seems that a bot or conversational UI
might work best for something very specific - where the user knows what
they can ask, and where those are the only things that they will ask.
However, when it does work, it becomes very interesting indeed,
particularly now because it happens to align pretty well with the second
preoccupation - getting around the app-installation problem.
Instead of going through all the pain of persuading someone to install your
app, what if you could have an immediate, fluid and easy way to engage
through a conversational interface? Could you make chat bots a third
runtime, after the web and native apps, that gives you a new interaction
model? And, of course, could that have a new way to acquire users that
works better than the App Store? (In other words, ‘never mind the dream of
artificial intelligence - will this let me stop giving so much money to
Facebook?’)
That is, the immediate promise of a conversational UI is less something
that you do within your own app than that it might make it much easier to
interact with your users without having to get an app installed in the
first place.
The practical question here, of course, is, well, what runtime? We had the
smartphone OS with its APIs, and we have the web browser, but what's the
local target for a bot? If a bot is a command line UI, what's the terminal?
How does it get onto the phone? And does it have its own meta-mechanisms
for user acquisition and discovery of services as well as the runtime
itself? Where are those?
One pretty obvious smartphone trend has been for interaction models not
just to proliferate but to move up and down the stack - up the stack into
new apps that themselves act as platforms (Facebook, WeChat, maps etc) and
down into the OS (Siri, Now, deep links etc). So too with bots: Apple has
not opened up Siri to developers (yet), but Google is selectively opening
Now, and it would certainly seem to be strategically desirable for the OS
provider to be the one creating any new runtimes.
However, the challenge for both Apple and Google here is the other end.
They can put a bot terminal on the phone, but for anything in the cloud to
talk to you on your phone, that service has to know where to look. Both
companies have some form of cloud identity platform, but Apple's privacy
strategy precludes many uses of this, and neither Apple nor Google have a
platform today that Gap or Instacart might ask you to log into when you
place your order, such that they can talk to you about it afterwards. Of
course, this could change - in particular, the rumour that Apple will
extend Apple Pay (and by implication Touch ID) to the web opens up lots of
possibilities in this direction.
Moving up the stack, Facebook has both ends - the terminal, in Messenger
(which now has 1bn mobile MAUs), and the web identity platform. Facebook’s
challenge is in a different place: would publishers want to give it so much
control, and would they trust it to continue the platform into the future,
given its track record?
Meanwhile, I found Microsoft's heavy focus on bots and associated tools at
'Build' pretty interesting, because of course Microsoft has neither end -
it doesn't have a mass-market smartphone platform or a web identity
platform that could be the connective glue. (It does have Skype, with 300m
MAUs, but I'm not at all clear that that has the right market positioning
to be a viable terminal in this sense.) But a lot of Microsoft's narrative
is about building cloud platforms to enable other people to create bots
that could then run between any two other end-points - between FB Messenger
or Slack at one end and Gap at the other, logged in through a Facebook
identity, but with the actual capabilities coming from Microsoft. That is,
Microsoft as, well, a development platform. Naturally, this is Google's
self-image too - what is Google if not a platform for understanding and
communicating data at massive scale?
I’ve deliberately been using the rather retro term ‘terminal’ to describe
the place a bot UI lives in on the phone because that seems to be the way
many people think about it - there’s an AI in the cloud and you talk to it
through a dumb text or voice UI on the phone or PC. That also reflects the
way Now works - the device is mostly dumb glass. This is an extension of
the shifting of interaction layers up and down the stack I mentioned
earlier - here you’re in some senses shifting the interaction up off the
device entirely. Coincidentally, ‘terminal’ is the word telcos used to use
to refer to mobile phones before smart happened at all, and indeed many of
the sample use cases I’ve seen for chat bots could be done with SMS, or
perhaps USSD.
But that’s not the only potential model. WeChat uses messaging as a
platform but doesn’t push interaction down to the level of ‘dumb text’ - it
provides its own social and acquisition model but much of the actual
interaction with third party services happens in rich UIs (mostly built as
web views, as it happens). And after all, if you’re offering a user a
question to which there are only two options, should you tell them ‘you can
reply ‘red’ or ‘green’’, or should you give them two buttons within the
chat? What if there are five options? Should you perhaps construct some
sort of on-screen interface for your users that lays out, graphically, the
options? You could call, it, perhaps, a ‘GUI’. You could have ‘links’ that
you tap on, that load new ‘pages’… And indeed, if you’ve got your chat bot
working, does that need to be in Facebook, or could it be on your own
website too? It depends what kind of interactions you’re looking for, and
maybe whether you’re solving your own problems or your users’.
There’s an old computer science idea that a computer should never ask a
question that it should be able to work out the answer to. One promise of
AI is that you can get, not so much a computer that you can talk to like a
person, but one that you don’t need to talk to at all, or much less - that
you can remove the mental load and friction and maintenance of engaging
with a computer or online service. The computer can ask fewer questions and
work out more by itself. (This is somewhat analogous to the reduction in
mental load involved in moving to iOS or Android from Windows or Mac).
Instead of going back to the website and logging into your order and
editing it, you can just send a message: ‘cancel the shirt and get me the
green bag instead of the grey one’. A computer ought to be able to work
that out, and will probably be able to. You ought to be able to use AI, and
chat, in some way, to achieve the same aims with much less work. But it
doesn’t follow that that’s a great way to do everything. If it takes you
more time to work out what you can ask the AI assistant than to drag the
meeting to a new slot on your calendar, you’re doing it wrong. An AI
shouldn't be more mental load than just tapping the damn button - an IVR
with better buzzwords. That might be a better user acquisition model for
you, but that’s tough. And so this comes back to my opening point - is this
a domain where you can just tell it what to do and move on, or one where
the ‘general intelligence’ still needs to be between the chair and the
screen?
Facebook AI
The first AI model that translates 100 languages without relying on English data
Facebook AI
8 min read · 141 saves · 2020-10-19 · Facebook AI is introducing M2M-100, the first multilingual machine translation model that can translate between any pair of 100 languages without relying on English data.
VentureBeat
Facebook’s TransCoder AI converts code from one programming language into another
VentureBeat
3 min read · 104 saves · 2020-06-08 · Facebook's TransCoder AI model can translate among programming languages like Java, C++, and Python with high computational accuracy.
WIRED
Building AI Is Hard—So Facebook Is Building AI That Builds AI
WIRED
6 min read · 110 saves · From 2016 · By forcing computers to do more of the grunt work, the world's biggest tech companies are accelerating how quickly AI enters the everyday world.
VentureBeat
Language AI is really heating up
VentureBeat
5 min read · 132 saves · Jan 17th · Any procedure that still requires a human to understand language is now at the verge of being disrupted or automated.
digitaljournal.com
Researchers shut down AI that invented its own language
digitaljournal.com
2 min read · 113 saves · From 2017 · An artificial intelligence system being developed at Facebook has created its own language. It developed a system of code words to make communication more efficient. The researchers shut the system…
VentureBeat
Facebook’s AI speeds up natural language processing without additional training
VentureBeat
2 min read · 84 saves · 2020-02-19 · In a preprint paper, Facebook researchers propose a new approach that speeds up natural language processing without requiring additional training.
