Learning Curve
We’re in the midst of the most important shift in civilization since the invention of the steam engine — the pervasive application of intelligence into every aspect of the world.
My goal today is to equip you with the tools you need for thinking about a world of pervasive intelligence, because that will describe both the sorts of investments you will be presented with, and the overall environment within which you will make those investments.
So without further ado, let’s Go…
PART ONE: GO HARD / GO FACEBOOK
In July, the Chinese government made a public commitment to lead the world in artificial intelligence, sealing the deal with a $200 billion investment fund.
Why did that happen? Well, I have a suspicion…
You might have heard that back in May, a computer program became the best player of Go in history.
Go is a 2500 year-old Chinese board game that is broadly believed to be the most complex and subtle of all board games — so important in ancient China that knowledge of Go was considered one of the four ‘essential arts’ of the cultivated person.
Go is not like that most famous of board games, chess. A computer can simulate all possible moves on a chessboards thousands of times faster than a human.
The number of possible moves in a Go game is 10 trillion trillion trillion trillion trillion. No computer can handle that kind of calculation. Instead, to succeed at Go, you have to learn how to play.
The only way you can learn how to play Go is by playing Go.
So researchers in the UK taught a computer program to play Go, but — and here’s the important bit — they also gave it the capacity to learn from its mistakes.
What this meant in practice is that the computer program studied the game board after every move it made. It learned which moves left it weakened, and which moved it toward victory.
Not that there was a lot of victory at the beginning.
Even a very bad human opponent could beat this computer program — named AlphaGo.
No human could expect to win their Go first games. AlphaGo didn’t either. But it learned from every loss. Every bad move made it better, fed into a continuing stream of data used to improve its performance.
In every game it did just a bit better than the game before it. Several thousand games later, AlphaGo could defeat a novice human player.
At this point, AlphaGo’s creators upped the pressure, matching the program against more expert players. AlphaGo lost more matches — but learned from those losses.
Simultaneously, these researchers did something quite sensible — they set AlphaGo to play matches against itself.
In addition to the thousands of games it played against increasingly proficient Go players, it now played tens of thousands of matches against itself.
That’s when AlphaGo started to get very good.
Early last year it beat a top-level grandmaster of Go.
In May of this year, AlphaGo played against Ke Jie, ranked as the #1 Go player in the world.
AlphaGo wiped him out utterly, in 5 matches out of 5.
At a post-tournament press conference, Ke Jie marveled at an AlphaGo that “played like a god”.
A few weeks later, AlphaGo’s creators announced it would be retiring from Go competitions.
AlphaGo had no one left to beat.
We hear the phrase ‘artificial intelligence’ thrown around all the time — these days. That’s why it was necessary to explain in a bit of detail how AlphaGo came to be.
There is no magic to artificial intelligence. It’s about building systems that learn from their own mistakes — just as we do — using that stream of errors to improve their performance.
There’s a second example I want to share with you, one that you’re almost all familiar with.
How many of you use Facebook at least once a month?
Of course that’s most of you — you’re some of Facebook’s two billion monthly users.
You know how Facebook makes money: advertising.
Facebook needs you to see their ads.
They get eyeballs by making Facebook so interesting, so irresistible, users return time after time.
But how does Facebook make itself irresistible?
That’s something Facebook doesn’t talk about.
Facebook is watching users all the time they’re on the app. Everything a user does is observed and recorded.
Facebook also uses cookies to track users even when they’re not on Facebook — so Facebook has a good idea of what its users are up to when they’re online.
All of that observational data becomes part of a user’s profile.
The feed a user sees when they visit Facebook is a reflection of what Facebook has learned by watching that user.
Facebook selects what goes into a user’s feed. If Facebook gets that wrong — makes a mistake in curation that results in a drop in user engagement — Facebook learns from that mistake.
Facebook feeds its mistakes back into its newsfeed curation, constantly working to make a user’s feed something they’ll want to linger upon.
Over time, a user’s feed becomes more and more interesting. Because Facebook has learned what a keeps a user glued to Facebook, and gives them more of that.
Facebook’s machine learning system has been designed with one goal in mind — increase user engagement.
It works. Facebook is on a trajectory to become one of the first trillion-dollar companies because it works.
It’s probably the most widely deployed example of machine learning. It’s one that everyone is familiar with.
But it’s also invisible, because it operates behind the scenes, making things work better.
This same capacity to learn from mistakes in order to achieve goals — that’s coming to pretty much everything.
PART TWO: START SMART
As you’re all investors, I reckon the best thing I can do for you is show you exactly how machine learning forms the foundation for three extraordinary Australian startups.
Each of them are using machine learning to solve a particular class of problem in a specific way — but with huge implications. Get it right here, and each of these businesses will grow into giants.
The first one we’ll look at is Premonition.
What they do is almost ridiculously simple. They use machine learning to improve the efficiency of vehicle fleets.
If you have a fleet, you distribute the Premonition app to your drivers.
Driver smartphones stream location information into Premonition’s machine learning system — along with current weather and traffic information, plus all of the historic data already fed into Premonition.
Then, Premonition starts making recommendations to the drivers.
At the beginning these recommendations won’t be perfect. They won’t improve efficiency very much.
But Premonition feed the results of those recommendations back in. It learns, and those recommendations improve.
Learning takes time. Premonition reports that it can take as long as two months to reach peak efficiency.
But if the use case looks like something Premonition already learned — and the system figures that out quickly — that learning time can decrease to a fortnight.
On a system that’s already highly optimised for efficiency, Premonition can regularly add 12% to 14% improvement. Some clients have seen 20% increases. In a worst-case a Premonition customer see 6%.
All just by giving drivers an app that gives them route suggestions.
Logistics is a very clear use case for machine learning. Premonition provides a basic software-as-a-service that’s a no-brainer for any business with a vehicle fleet.
It’s the specific implementation of a more general principle: machine learning applied to increase efficiencies at every point in the economy.
That’s going to a hugely fertile ground for startups over the next few years. Perhaps we’ll even see companies pitch themselves as ‘Premonition for X’.
Next, we come to Maxwell MRI, a Brisbane based startup on a path toward constructing an ‘AI genome’.
What’s that?
Imagine every bit of medical data you generate — every test and scan and GP visit, every tablet swallowed and every exercise performed — all going into an evolving, learning model that represents, as completely as possible, your health.
That model becomes an invaluable reference — for you, for your doctors, for anyone you partner with for wellness.
It doesn’t start out accurate, but it learns.
It tries to predict — and where it fails, and learns from that. It tries to recommend — and where it falls short, it learns from that. It never stops learning. It only grows more accurate.
Can we do that today? There’s a long road of development between what we can achieve today and an ‘AI Genome’.
The first step on that road for Maxwell MRI is simple and clear — improving the diagnostic capabilities of MRI scans for cancer.
Starting with prostate cancer, today plagued by false diagnosis and unnecessary surgery.
They’re building a tool to increase the accuracy of those scans, a tool that learns — both from the patient, and from thousands and thousands of scans.
Can they do that today? Yes, because it’s a simple and clear learning task, where the successes and failures get fed back into the diagnostic engine. With every diagnosis, Maxwell MRI improves.
That’s the beachhead. Once they get that bit right, Maxwell MRI moves on to improving diagnostics for breast and lung cancers, learning more about scans, learning about patients, getting smarter.
This is not a quick process — building the ‘AI Genome’ will take years.
But neither is it an all-or-nothing ‘moonshot’. Even the problem they’re working on today will save lives and make for a great business.
It also perfectly describes this moment: at the beginning of learning how to learn.
Putting that learning to work to solve some of the most important problems we have.
Because — and I don’t think this really bears saying, but I’ll say it anyway — the closer Maxwell MRI comes toward their goal, the more indispensable it becomes to every human being on the planet.
They’re creating the universally useful medical aid — fully reflective of the individual, learning, improving, and able.
Finally, we come to a startup that is the least mature — but possibly the most innovative, opening the door into an entirely new way to harness machine learning.
APS is a machine learning startup currently going through the Sydney University INCUBATE program — where I mentor.
They’re working in a space we don’t normally hear too much about in the investment community — process engineering.
Process engineering is the pointy end of chemical engineering — how you get from a range of input compounds to an output compound.
That’s a combination of chemicals and processes, in a very specific order, with very specific durations. Change anything and you wreck that process.
It’s painstaking work.
Fortunately, simulation tools have been developed to aid process engineers.
They can develop a process, test it in the simulator, and adjust or adapt the process as required.
Process engineers use the simulator to identify their mistakes, feeding those errors back into their process engineering.
That’s the essence of a machine learning process, and that’s the core innovation APS is working to bring to market.
APS has a machine learning system that develops processes — almost at random, to start with — feeding these into the simulator, looking at the mistakes, feeding those back into the machine learning system, iterating, learning, and improving, as it closes in on the goal.
This is process engineering that’s entirely machine-driven — and more than occasionally it develops processes no human would have thought to create, processes never seen before — truly original processes.
That’s interesting in itself.
The global process engineering market is worth at least a few billion dollars a year, and APS could take a big chunk of that — but what’s more interesting is the fundamental innovation — the coupling of machine learning with simulation.
This coupling between simulation and learning is how AlphaGo got so smart — it played games against a simulated opponent — itself.
This coupling is going to become a generalised technique in machine learning.
Learning from your mistakes is good. Learning from your mistakes via simulation is even better — because it’s faster.
Any process we can simulate is now a process we can build a learning model for.
Suddenly, this isn’t just about the revolution in machine learning. It’s also about the revolution in simulation — something much bigger than virtual and mixed reality.
The marriage of learning and simulation amplifies the value of both tremendously.
That’s an important thing to note. Perhaps the most important thing to note in this talk.
To recap, here are three Australian startups:
Premonition implements a specific instance of a generalised machine learning technique to increase the efficiency of almost any organisation.
How many other ways can that be used in organisations? How many clones and spin-offs in different areas will we see over the next few years?
Maxwell MRI shows how to evolve from MVP into breakthrough innovation.
Each step along the path represents an improvement in the ability to know and treat a patient.
It starts small and grows to encompass the entire medical history and well-being of — well, everyone.
APS couples machine learning with simulation to produce an automated solution for process engineering.
Yet they’ve also shown how simulations of all sorts will now become intimately linked with machine learning systems, accelerating learning and capacity.
That’s not bad for three startups here in Australia.
And of course that’s also happening everywhere else in the world.
And that’s what we need to talk about now.
PART THREE: THE RISING
Most of what we hear in the media about machine learning and artificial intelligence is how they’re going to put all of us out of work.
That’s less true than we fear.
What tends to happen — what we’ve seen so far — is that machine learning and AI make people better at what they do.
An oncologist using IBM’s Watson for Oncology is a better oncologist.
A lawyer using ROSS Intelligence is a better defender.
A cargo driver using Premonition is a better driver.
This is what we see across the board. These new tools make us better at whatever we do.
Does this mean we’ll need fewer professionals — or does it mean those professionals will be focused on more interesting work, spend more time working directly with people and less time focused on the mechanics of their professions?
These are the questions we will be asking over the next billion seconds, as we make the transition to a culture where machine learning is being applied at every point.
Make no mistake — as telecoms were to the 20th century — connecting everything, everywhere — machine learning is to the 21st, bringing capacity and capability to every point. And to us.
That’s the overall trend.
You’re going to want to learn how to lean into that trend. Because more and more of the investments you’ll be making over the next decades will have some component of machine learning.
Eventually, all of them will, because it will simply be The Way Things Work.
But in the transition, well, that’s the opportunity for huge disruptions.
So people will be coming to you with all sorts of pitches for all sorts of crazy ideas for all sorts of interesting ways machine learning will solve a problem.
And you need to ask yourself some questions:
- How does this proposal incorporate machine learning?
- How does it use its own errors to improve its performance?
- How does it use simulation to increase its speed and accuracy?
If they can’t answer these questions — in clear and simple terms — I’d advise that you pass on the investment.
Machine learning isn’t snake oil.
It should be easy to understand why something is learning.
Where it’s not clear, either the folks involved don’t understand — or they don’t want you to understand.
And it’s important that you understand, because the real winners over the next decade, the best investments, are the businesses that use machine learning to solve really hard problems.
Maxwell MRI is in that category — though they’ve told me they suffer from an interesting problem.
They’re fundraising now, and when they go to the tech funds, those funds say, “You’re a biomedical company, go talk to those funds.” But the biomedical funds say, “You’re a tech company, go talk to the tech funds.” So they fall through the cracks.
That confusion will become commonplace, as every business of every kind incorporates machine learning to improve its performance, and solves the hard problems.
Broaden your thinking: yes, machine learning is a technology — but so was mains electricity a hundred and twenty years ago -and we don’t think of a business that uses electricity as a technology business.
Intelligence is in the same category.
It’s going to be everywhere. It’s going to be embedded into everything, from smarter smartphones to smarter services to smarter professionals to smarter businesses. Everywhere.
That’s the scale of this opportunity. Because right now very little of the world has the benefit of this rising intelligence. But all of it soon will.
So look for the companies that are learning how to learn from their own mistakes to make people better at what they do.
Those businesses are going to be the winners over the next decades.
Look to yourselves: It’s easy to imagine someone developing machine learning tools that make you better investors, better mentors, and better board members.
Increase your own capacity, using the kinds of systems that allow you to learn from your own mistakes.
The middle years of the 21st century are dominated by rising intelligence and increasing capacity.
We’re all going to need these tools to do our jobs. We’re all going to need these tools to get smart and stay well.
You have the unique capacity to identify, mentor and capitalise those making these tools. You won’t always make the right decisions — but you can learn from your mistakes.
This is the learning curve, and we’re all on it together.
