Happy Sunday and welcome to the latest edition of Inside AI! For those of you who are new, I'm Rob May, CEO of Talla and active investor in the A.I. space. If you have an early stage A.I. startup looking for investment, I'd love to hear from you. (Omelas and Phiar are two recent investments.)
This week's issue is awesome. If you agree, I hope you will forward this to a friend so they can subscribe too.
Or if you really want to support us, go Premium and get a special Wednesday edition each week, plus extra analysis and reports. Thank you to Ascent and Kylie for being corporate sponsors. (email firstname.lastname@example.org if you want to become a sponsor)
-- Big Idea --
The most interesting thing I read this week was a piece in Quanta Magazine about "How Nature Solves Problems Through Computation." It looks at the work of Jessica Flack, who studies computation in nature. The comment that really jumped out at me was discussing neurons in a monkey's brain - "We found that as the monkey initially processes the data, a few single neurons have strong opinions about what the decision should be. But this is not enough: If we want to anticipate what the monkey will decide, we have to poll many neurons to get a good prediction of the monkey’s decision. Then, as the decision point approaches, this pattern shifts. The neurons start to agree, and eventually each one on its own is maximally predictive." As best we know, this does not appear to be how neural networks are working today. It shows that we are still very far away from having systems that accurately mimic biological models of computation. Although, there is some debate about whether or not those are need for progress. But, either way, nature provides interesting ideas for us to try.
-- Must Read Links --
Opening The Deep Learning Black Box. Science.
This article looks at research using counterfactual probes to understand what individual neurons are doing in deep learning. Incredibly promising approach that, if it spreads, will have big implications for our understanding of these systems.
AI Progress Measurement. Electronic Frontier Foundation.
This project from the EFF is a comprehensive notebook tracking the progress made in AI and machine learning. The notebook has collected problems and datasets from AI research and created a resource for people who want to report results or build new data projects. Very interesting stuff to follow.
Technical Debt In Machine Learning. Medium.
How feedback loops, correction cascades, and "hobo" features can slow down a fast-paced project by creating technical debt. Machine learning has tremendous upside, but with that comes the downsides mentioned here.
Our Sexual Future With Robots. Responsible Robotics.
This is an overview of the current state-of-the-art sex robots and sextech, as well as the implications of the technology: their use as sex workers, the effect on sex crimes, the idea of sexual intimacy and relationships, and other societal impacts. Fascinating stuff.
Neuromorphic Bureaucracies: Space, Time, and Cost. Medium.
An interesting examination on the way that neural circuits can manipulate the trade-offs between space, time, and cost, and how these ideas can be extrapolated to "neuromorphic management structures" in organizations.
-- Industry Links --
The Darker Side of Machine Learning. Semiconductor Engineering.
How Bots Win Friends and Influence People. IEEE Spectrum.
Where Machine Learning Meets Rule Based Verification. Foretellix Blog.
The Machines Are Getting Ready To Play Doctor. MIT Tech Review.
Machine Ethics and Artificial Moral Agents. CyberTales.
Is The Government Ready For A.I.? FCW.
How Fake News Could Get Even Worse (from A.I.). Economist.
7 Myths about AI that are holding your business back. Venture Beat.
My Top 5 favorite AI Experiments. Medium.
Neural Networks: Innumerable Architectures, One Fundamental Idea. Medium.
Make Work Great Again; First, Unplug All The Robots. Salisbury Post.
By 2025, blockchain, IoT, machine learning will converge in healthcare. Health IT Analytics.
Probabilistic Programming From Scratch. Oreilly.
Bots — Why Should Enterprises Bet on Them? Medium.
Using History to Chart the Future of AI: An Interview with Katja Grace. Future of Life.
Rise of the machines: Is a universal basic income the answer for mass unemployment? ABC.
Google Stakes Its Future on Tensorflow. MIT Tech Review.
Introduction to Genetic Algorithms. Medium.
Mark Zuckerberg touts universal basic income as a bipartisan idea. Fast Company.
The Future of Productivity: AI and Machine Learning. Entrepreneur.
Learning from DeepMind's Data Privacy Woes. Architecht.
AI in marketing: Selling through robots. Medium.
The Evolution of Bots. Entrepreneur.
Robotic Knee Surgery Competition Heats Up. Scientific American.
Many thanks to Inside AI's corporate supporters. Please go check them out!
(email email@example.com to become a sponsor)
-- Research Links --
DeepStory: Video Story QA by Deep Embedded Memory Networks. Link.
Restricted Causal Inference Algorithm. Link.
-- Commentary --
My friend Michael Yamnitsky at Work-Bench recently published the 2017 Enterprise Almanac, which is a fantastic report focused on the forces impacting the enterprise software world at the moment. You should definitely go check it out. A large section of the report is dedicated to machine intelligence, and so I asked Michael to explain some of his thoughts on that topic below.
1. What does Work-Bench do, and what is your role there?
Work-Bench is a New York based venture capital firm investing in early-stage enterprise tech startups. Our expertise lies in strategic and tactical go-to-market assistance for deeply technical companies aiming for F500 customers. I won’t go into too much of a sales pitch, but we hustle to be the best bang for your cap table buck.
I’m a Venture Partner based in SF at the moment where I’m focused on investing in machine intelligence, cybersecurity, and enterprise infrastructure startups at the seed and Series A level. Some investments I’ve worked on include Cockroach Labs, CoreOS, x.ai, and Alluvium.
2. Your report focuses on Machine Intelligence as one of the key pillars impacting enterprise technology, but A.I. has a lot of skeptics. Do you think A.I. has staying power this time, or is this another temporary wave?
Yes, because only now is it easily accessible. Pundits can argue all day about the economic and societal impacts of AI software, but there’s no doubt AI technologies will be core to software development moving forward.
Frank Chen from a16z made a pithy comment recently that machine intelligence will be like databases — ubiquitous components of all software. I couldn’t agree more. And if databases became ubiquitous via SQL, AI will become ubiquitous via tools like Tensorflow and Keras.
3. You highlight Greylock's post about A.I. and "systems of intelligence" in your report. How are systems of intelligence different from systems of record?
System of Record is an old school term popularized by IBM and Forrester Research for the authoritative data source for a given domain or department within an organization. For example, Salesforce is the System of Record for CRM. System of Intelligence is the AI intelligence layer that makes sense of the information in order to glean insights and perform automation.
Note that a single company can perform both of these functions, or focus on one (i.e. you can build the intelligence and own the data, or just play nice with all the data owners). We’ll likely see success stories in both categories. The key difference between the two is the types of offensive tactics you can use against your competitors.
4. Do you think systems of intelligence will be largely verticalized like SaaS, or more horizontal like email?
I’d speculate mostly vertical. AI training and data moats tend to work within the constraints of specialized domains. This will also put pressure on today’s horizontal apps to become more purpose built. For example, email may break up into email for sales, email for executives, etc. due to these dynamics.
On the other hand, ML models are somewhat scalable across different parts of the same company. That insight drove our recent investment in enterprise data science collaboration platform Algorithmia.
5. Workbench has seen a ton of early stage A.I. deals. Are most A.I. entrepreneurs thinking about their products as systems of intelligence, or are they missing the boat on A.I. product strategy?
More and more are, but as a cohort AI startups often lack deep application domain experts in the founding team.
6. You point out that tech giants are playing defense with A.I. acquihires - do you think this trend will continue as A.I. talent becomes more widespread and A.I. tools abstract away much of the details so more engineers can tap into these concepts? In other words, do you see A.I. talent remaining a rare commodity?
I believe they will remain a rare commodity for two reasons:
#1. We’re still in the early innings. As fast as AI tooling abstract the details, bleeding edge new advances with mass scale practical application be driven by the experts.
#2. Using Tensorflow isn’t hard, but using Tensorflow to solve difficult, domain-specific challenges is.
7. Your report talks about "invisible apps" in A.I., what does that mean and why do you think they will be so valuable?
Invisible apps are UI-less invisible software that perform several consecutive steps in a larger business process without human intervention. They will be valuable for two reasons:
- They are widely applicable. Invisible apps are often built on cloud based systems of record like gmail, salesforce, and automate common business tasks like calendar scheduling, sales lead generation, and employee on boarding. Basically, they apply to all industries and company sizes.
- They will be adopted bottoms up. The same ambitious and highly empowered early adopters of Dropbox and Evernote will buy up AI-powered apps that help them do their job more effectively. Because invisible apps ride on top of cloud-based SORs, there’s no implementation hurdle to pass and the software can be sold bit sized to fit into employee expense reports.
8. The prediction I was surprised to see, but wholeheartedly agree with, is the protracted time to product-market fit for A.I. applications. No one else is talking about that. Explain why this is happening.
With SaaS you can do agile development and fast A/B tests. The AI development process and customer feedback loop is much longer. Also, AI apps have the chicken-and-egg problem where you need customers to use software in order to make it effective. All this leads to longer time-to-value for customers and thus longer time to product-market fit.
Startups can avoid this by thinking very carefully about pricing and packaging early on. AppDynamics did an amazing job scaling a complicated infrastructure tool with a two-step freemium sales motion. AI startups should do the same by breaking off whatever capability of their solution minimizes time-to-value and selling it as a “thin edge of the wedge.”
9. Are the deals you see at Workbench taking this protracted phase into account in their fundraising?
Not so directly. It’s more that price/competition is bidding round size up to a point that promising AI startups have cushion should things inevitably move slower than expected.
10. You point out that most of the A.I. talent is locked up at big companies focused on getting us to click on more ads. But that isn't where the opportunities are. What does that mean for the ecosystem? Do you see IBM becoming more competitive in industry verticals because it is the only large tech company with significant A.I. talent but without core advertising businesses?
We need more talented people excited about B2B/enterprise. I’m hopeful AI will be the catalyst for this.
I’m less bullish on IBM and more bullish on ITSM vendor ServiceNow, now a public company with an $18B market cap. It uses ML to power workflow automation much like the startups.
11. You point out the value of labeled data as a moat - is data annotation capability the competitive advantage of the early A.I. winners?
Yes it is, but moving forward data annotation is getting faster and the Systems of Intelligence moat is really about the balance and coordination of data, AI, domain expertise, and product design.
12. If you think A.I. is going to crash a bit, then create real value on the other side of the crash, how do recommend entrepreneurs play that?
Raise money now while it’s good because you never know how long the good times will last. Your job as the CEO is to finance the company opportunistically taking full consideration of market demand.
13. We've talked several times in this newsletter about whether A.I. is the sport of kings. After researching this report, do you feel more optimistic or more skeptical about the ability of A.I. startups to be successful?
I feel more optimistic than ever. When you size up AI from a technical perspective, it’s natural to see it as a shiny new weapon for the Kings to deploy. But through the lens of Systems of Intelligence, it’s clear the processes and barriers to entry for AI-powered businesses are vastly different. With the Innovator’s Dilemma in mind, I’m optimistic that new startups will snag the cheese before the Kings reorient themselves towards Systems Of Intelligence.
That's all for this week. Thanks again for reading. Please send me any articles you find that you think should be included in future newsletters. I can't respond to everything, but I do read it all and I appreciate the feedback.
-- ABOUT ME --
For new readers, I'm the co-founder and CEO of Talla, I'm also an active angel investor in A.I. I live in Boston, but spend significant time in the Bay Area, and New York.