In Amazon
9 Best Articles in 2021
Benedict Evans
Why Amazon Has No Profits (And Why It Works) — Benedict Evans
Benedict Evans
10 min read · 57 saves · From 2014 · (I also did a podcast with Ben Horowitz discussing these themes: see here)
Amazon has a tendency to polarize people. On one hand, there is the
ruthless, relentless, ferociously efficient company that’s building the
Sears Roebuck of the 21st Century. But on the other, there is the fact that
almost 20 years after it was launched, it has yet to report a meaningful
profit. This chart captures the contradiction pretty well - massive revenue
growth, no profits, or so it would seem. But actually,
Benedict Evans
Platforms, distribution and audience — Benedict Evans
Benedict Evans
6 min read · 73 saves · From 2015 · It's never been easier to write or harder to get read. Thoughts on
blogging, distribution, platforms and ad-tech
In Bots
Benedict Evans
Chat bots, conversation and AI as an interface — Benedict Evans
Benedict Evans
8 min read · 308 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?
Benedict Evans
Ways to think about cars — Benedict Evans
Benedict Evans
9 min read · 60 saves · From 2015 · Cars are going to change a lot in the next few decades. Electricity on one
hand and software on the other change what a car is, how it gets made and
who might own one. Bu that changes everything else.
Benedict Evans
The VR idea maze — Benedict Evans
Benedict Evans
6 min read · 34 saves · From 2016 · Someone said recently, only slightly flippantly, that you can divide the
world into people who think VR is part of the future and people who haven't
had the demo yet. Unlike 3D, and unlike the primitive VR of the early
1990s, it's extremely hard to have a demo of Oculus (or indeed HTC Vive)
without seeing it playing some major role in the future of tech. After
that, though, the clarity ends - the state of VR now can be seen more than
anything else as a list of questions - an 'idea maze' of
Benedict Evans
TV, mobile and the living room — Benedict Evans
Benedict Evans
5 min read · 36 saves · From 2015 · The tech industry spent decades trying to get to the TV, but now that it's
finally happening, it turns out not to matter much.
In Marketing
Benedict Evans
Search, discovery and marketing — Benedict Evans
Benedict Evans
9 min read · 69 saves · From 2015 ·
"However vast any person’s basic reading may be, there still remain an
enormous number of fundamental works that he has not read"
Italo Calvino
Erasmus is said to have been the last person in Europe to have read
everything. He lived just at the point that printing took off, and so it
was in his lifetime that it ceased actually to be possible to have read
everything. Some time later there was presumably also a last person to have
heard of every book, or every book worth readi
Benedict Evans
Mobile, ecosystems and the death of PCs — Benedict Evans
Benedict Evans
9 min read · 83 saves · From 2015 · When we say 'mobile' we don't mean mobile just as when we said 'PCs' we
didn't mean PC. This isn't about the screen size or keyboard or location or
use. Rather, the ecosystem of ARM, iOS and Android, with 10x the scale of
Wintel, will become the new centre of gravity throughout computing.
More like this
Benedict Evans
The end of a mobile wave — Benedict Evans
Benedict Evans
7 min read · 60 saves · From 2016 · The mobile phone industry has had two waves - first voice and SMS and then
the smartphone. The voice wave has taken it from zero to 5 billion people
on earth with a mobile phone, and now close to 2 billion mobile phones are
sold every year. In parallel, starting 9 years ago, the smartphone wave
converted a larger and larger percentage of those phone sales to
smartphones.
And since smartphones could be sold for higher average prices than feature
phones, revenue grew even faster than unit
In Google
Benedict Evans
What does Google need on mobile? — Benedict Evans
Benedict Evans
6 min read · 30 saves · From 2015 · I generally look at Google as a vast machine learning engine that’s been
stuffed with data for a decade and a half. Everything that Google does is
about reach for that underlying engine - reach to get data in and reach to
surface it out. The legacy web search is just one expression of that, and
so is the search advertising, and so are Gmail and Maps - they’re all
built onto that underlying asset.
Hence, most of the experiments that Google has launched over the years are
best seen as tes
Benedict Evans
Forget about the mobile internet — Benedict Evans
Benedict Evans
4 min read · 125 saves · From 2015 · For as long as the idea of the 'mobile internet' has been around, we've
thought of it as a cut-down subset of the 'real' Internet. It's time to
invert that - to think about mobile as the real internet and the desktop as
the limited, cut-down version.
Benedict Evans
Voice and the uncanny valley of AI
Benedict Evans
7 min read · 104 saves · From 2017 · WHAT DO WE WANT?
Natural language processing!
WHEN DO WE WANT IT?
Sorry, when do we want what?
— Benedict Evans (@BenedictEvans) January 22, 2017
Voice is a Big Deal in tech this year. Amazon has probably sold 10m Echos,
you couldn't move for Alexa partnerships at CES,
In Technology
Benedict Evans
Living in different worlds — Benedict Evans
Benedict Evans
4 min read · 53 saves · From 2015 · What does it mean when you don't understand technology?
In Startups
Benedict Evans
Ways to think about market size. — Benedict Evans
Benedict Evans
7 min read · 51 saves · From 2015 · When you try to work out the market potential for something fundamentally
new, you’re actually trying to resolve two, linked problems.
* First, you have to look past what it is now, and see how much better
and cheaper it might become
* Second, you need to think about who would buy it now, and who else
would buy it once it is better and cheaper, and how it might be
used.
The second problem is actually the hard one. Anyone with a sense of history
ought to have been a
Benedict Evans
Lists are the new search
Benedict Evans
4 min read · 187 saves · From 2016 · All curation grows until it requires search, and all search grows until it
requires curation. So how does curation scale to the internet? Does it need
to?
In Productivity
Benedict Evans
Presentation: Mobile is eating the world
Benedict Evans
35 saves · From 2016 · Updated for spring 2016: why mobile matters, where it is and how it changes everything from cars to productivity to search. For more detailed discussion of the issues this touches on, see my post '16…
Benedict Evans
Learning machine learning — Benedict Evans
Benedict Evans
2 min read · 20 saves · From 2016 · In August 2001, I was a telecoms analyst visiting investors in Tokyo. In
one of these meetings, a portfolio manager at a Very Large Fund asked me
what would happen now that GPRS meant that all mobile voice calls would be
packet-switched and that therefore mobile operators’ voice revenue would
disappear within the next 18 months or so.
This was a surprisingly hard question to answer well. It was nonsense, but
to explain why it was nonsense you had to work out quite which things the
person
