Category: articles

  • Why Linus’ Speech Matters (to Jews)

    But honestly, I appreciate the Peanuts special more for how it looks directly at the Christianity of it all. It’s not trying to hide the ball. It is …

    Why Linus’ Speech Matters (to Jews)

    Shared by my good friend Shanan; a wonderful read.

  • Distribution in digital products

    Distribution in digital products

    This was a great story about how Google’s search distribution deal dramatically increased cost of entry for Neeva, a putative competitor.

    This also reminded me of the Monster/AB InBev deal mentioned in “The Little Book that Builds Wealth:”

    To be fair, it is occasionally possible to take the success of a blockbuster product or service and leverage it into an economic moat. Look at Hansen Natural, which markets the Monster brand of energy drinks that surged onto the market in the early part of this decade. Rather than resting on its laurels, Hansen used Monster’s success to secure a long-term distribution agreement with beverage giant Anheuser-Busch, giving it an advantage over competitors in the energy-drink market.

    Anyone who wants to compete with Monster now has to overcome Hansen’s distribution advantage. Is this impossible to do? Of course not, because Pepsi and Coke have their own distribution networks. But it does help protect Hansen’s profit stream by making it harder for the next upstart energy drink to get in front of consumers, and that’s the essence of an economic moat.

    Once you find product-market fit you need to quickly scale distribution to own as much of the market as possible and preempt new entrants.

  • Mastodon, ActivityPub, the Fediverse and how standards actually work on the internet.

    This is a great article from Ars Technica that explains how a bunch of published and defacto standards come together – more or less – to make Mastodon go and create a decentralized social web.

  • Mastodon for Fun and Profit: The Fediverse for Brands

    Mastodon for Fun and Profit: The Fediverse for Brands

    Twitter and Facebook appear to be in trouble. Twitter has been hemorrhaging users, with the Elon deal as a tipping point. Facebook just laid off thousands in an effort to re-focus on their core business and improve profitability. In light of that, and the sudden traction of Mastodon and the Fediverse, I started thinking about what the Fediverse could look like for brands.

    The Fediverse is decentralized. That makes it different.

    • The Fediverse is a collection of “social media feed” services that use the ActivityPub protocol.
    • Mastodon is an open-source microblogging server that looks and feels a lot like Twitter.
    • The main difference is that there is no one Mastodon “service” – there are thousands of servers managed independently, running servers that speak ActivityPub.
    • There is no one organization or company that manages the site, creates recommendation algorithms, validates users or does moderation.

    It’s different – and these differences mean that brands have to think differently.

    But it’s also like Twitter, Facebook, Instagram, Snapchat and TikTok

    It’s like a legacy social network:

    • Individuals, brands and publishers identify themselves with “handles” and “profile pages”
    • There’s “verification,” although it means something different, and works differently.
    • People “follow” accounts, tag their posts with hashtags and engage with followers.

    By default, everything that gets posted on one server gets pushed to the entire network. Brands, publishers and individual users can even set up their own servers – just like email – for extra assurance.

    With a dash of podcasting

    Podcasts are distributed across a federated network using standards like RSS and MP3. Podcasts are consumed in proprietary apps like Spotify. But there’s no built-in back channel as there is on the Fediverse.

    It’s all earned and owned media

    Just like with email, you can own your own subscriber list. You can set-up an instance of Mastodon (or any other ActivityPub server) and then you effectively become the admin – not Twitter or Meta.

    So instead of being @nike on twitter or Instagram, or facebook.com/nike, you can be @nike@nike.com (or @lebron@nike.com) – and your posts will be visible on every server on the Fediverse. If ActivityPub works, I can see brands signing up for hosted mastodon services just like they do email or web hosting.

    You can share text posts, photos, live streams and recorded videos in native experiences from the same account

    Today’s social networks are media specific: twitter for short text, Instagram for photos, twitch for live video and so on. The Fediverse offers a single network that any media format can (and does) ride on. This also means that your identity works across every format.

    There is no one algorithm

    Algorithmic discovery (e.g. trending) is not built in. Different Fediverse services can implement it in different ways. Different client applications handle discovery differently. Some servers let you follow hashtags, some don’t. It’s a mess. It’s also an opportunity for brands and publishers to build their own client experiences. Like a Facebook page turned inside out.

    Paid advertising isn’t quite a thing

    Although some servers do accept advertising, ads do not flow across the network. Every server in the Fediverse can set its own terms of use. Most instances are run by volunteers. Allowable content rules vary.

    Admins (server administrators) control content by blocking specific accounts and servers. This needs to get standardized.

    I could imagine some sort of “brand-safe” ad network that servers and client applications could participate in, provided that they also use an approved moderation service.

    It’s less about “decentralization,” and more about “unbundling”

    • Brands will need to break their social strategy down another level. It won’t just be Instagram, twitter and Facebook. It will be which sub-brands get get visibility, how they are affiliated, and how reputations are managed. It’s more like the “old” web, with website URL, search position and email deliverability all being factors.
    • Facebook, TikTok, Snapchat and Twitter bundle together moderation, brand safety, and ad sales. This is unbundled on the fediverse, more like email or podcasts.
    • This could actually give brands and publishers even more power; they can pick and choose the bundles of services, audiences and media types that make sense for them.

    Get to work!

    The most important thing is to start experimenting with the Fediverse, now. Create a personal or stealth account, find the people who are joining Mastodon who are influencers in your domain. Follow them, let them follow you. Cross-post your personal twitter account. Try to figure out the neighbourhood and it’s norms. If you do that, you will be well positioned to figure out an experimentation strategy for your brand (or your publisher), and then start to roll that out if you see value in it.

    Good luck!
    Joseph (https://mastodon.social/@josephby)

  • What I Learned from Interviewing 85 Product Manager Candidates in 18 Months

    What I Learned from Interviewing 85 Product Manager Candidates in 18 Months

    This post was originally published on Inside Q4, stories and lessons learned from the Q4 Inc. R&D team.

    The requirement was extreme: hire 15 Product Managers (and one Director of Product) in the next year and a half.

    When I arrived at Q4 in October of 2021, we were entering a period of hyper-growth in Product Design, Engineering and Product Management. We had huge ambitions and needed to scale our team — fast. I had hired talent before and had conducted dozens of interviews in my career, but the task ahead of me was daunting. As a new hire myself, Q4 was a domain that was still fairly new to me. Plus, the company serves a specialized clientele and isn’t exactly a household name — and the job market is insanely competitive!

    With the help of our CTO, my peers, our amazing talent acquisition team and under the guidance of our inspirational and very applicable company values (Grind, Hustle, Iterate, Compete, Care), we reached our goal. Here are some of the things that I learned as the hiring manager on this assignment.

    Cross-optimize your Hiring Process for Speed and Fairness

    Hiring is a flow, not a project. The point is to run candidates through a process and hire quality, diverse candidates in a fair and equal way as fast as possible — but no faster. Go in expecting that it will take time and unsexy work to do it well, but it will all pay off in the end.

    Create a process: align on it, refine it, repeat

    • Write a good job description for each role; ensure that all the interviewers know what’s in it. A good job description can help turn a mediocre interviewer into an amazing one.
    • Write a good rubric (‘marking key’) for the interviews and take care to set it up properly in your applicant tracking system. Refine it between searches so that you’re constantly applying what you learn in a hiring round to make your process better.
    • Learn from your mistakes! When you have a mis-hire — and we had a few — take another look at your questions, process, and rubric and improve on it for next time.

    Optimize for speed:

    • Have a well-defined interview process for each role, with backup interviewers on call to be added to an interview loop within a day’s notice to keep the process flowing.
    • Use your application tracking system and hold up your end by entering feedback immediately after the interview. This allows your talent acquisition partner to move right into the next step with a candidate (setting up next interviews, generating offers, closing, etc.)
    • Have a fast, scheduled window of time for interviewers to align on a candidate (if needed) and resolve any conflicting hiring decisions. For example, in the case where three interviewers rate a candidate “hire” and one rates the same candidate as “no hire,” use this window to quickly come to consensus.
    • Make it easy for talent acquisition and candidates to book your calendar — Calendly, or Google Calendar Appointment Windows, are great tools to keep all necessary parties organized.
    • Follow up with candidates quickly — especially when an offer goes out!

    Optimize for fairness:

    • Build a question bank! Over the years, I have tried to document every interview I’ve been in as a candidate. That includes writing down any really good questions that I was asked. Good artists copy, great artists steal.
    • Have a consistent process, scoring rubric, skills criteria and question bank that all interviewers can draw on, so that each candidate has the same experience and opportunity.
    • Write. Things. Down. I try to fully transcribe every interview I conduct so that if I need to go to bat for a candidate — or reject a candidate — I can go back to the specifics and have a fruitful conversation with the other interviewers based on facts, not memory or impression.
    • Manage the flow of feedback between interviewers such that you don’t bias each other, but can still use subsequent rounds of interviews to fill gaps in understanding.
    • Create a great applicant experience. Get back to people quickly!

    Never stop recruiting

    Always be thinking about where a good person could fit with your team in the future. Circumstances change often, and quickly, so don’t ever assume the team you have today is the one you’ll have tomorrow.

    • Go back to the well: Past interviews, past applicants, and people who got an offer in the past but declined, are a good source of candidates for future hiring; mine your applicant tracking system for candidate “gold” you may have missed last time around.
    • Do your own outreach: Like us, you may have a strong talent acquisition team, but you’re the one with the in-role expertise. Buy and expense a LinkedIn premium account and send InMail to people you think could be a fit in the near future.
    • Do informational interviews: 15 minutes of your time is worth it if you can meet and nurture a potential future candidate. Team building is a marathon, not a sprint. The older I get, the more I enjoy, appreciate and learn from folks who are just starting out as PM. If a person is too junior today they may be a fit for a new role in six months, or a more senior role in 18. Build your rolodex: the value of this investment compounds over time.
    • Know (and love) your external recruiters: I was brought into Q4 by a top-shelf recruiter who had carefully built a relationship with our CTO over years and has turned into a trusted resource and advisor to me, too. He touches base every four or five months, we catch up briefly, he gives me a read on the market and periodically brings me amazing candidates. Relationships like this propel careers.
    • Focus on team composition, not “fit.” When hiring Product Managers, remember that you’re not just hiring individual contributors, you’re hiring a key peer leader for your product team. When I think of team composition, it’s not in terms of some arbitrary cultural fit but rather of the particular talents that needed to be added to bring the team to the next level.
      For example, I think about the emotional energy the team needs. Is the team a “hair on fire” sort that panics easily and could benefit from a calming presence? Or is the team more in need of a high-energy product manager to get them excited? Is it a creative product role? An analytical one? A highly technical one? How much executive exposure will this PM have? What skills will they need to spike on to do well for themselves and the team?

    You’re going to have interviews where the candidate is amazing but not for the specific role. When that happens, be up front with the candidate. Let them know they’re not going to get the offer for the position, and tell them why. Then, shop the person around internally and keep in touch for future opportunities.

    Of the 16 people we hired, three have already been promoted from individual contributors to managers. I’ve also heard some really nice things from colleagues about our evolved team. Hearing this makes me proud. Hiring processes are a lot of work for everyone involved. We put a lot of thought and energy into our part of it, and it’s rewarding to see it come together well.

  • Watching Your iPhone Work to Protect You from Covid-19

    Much has been written about the Apple + Google Covid-19 Exposure Notification framework. This is the software that is now part of Android and iOS (13.5+) and powers Covid-19 detection apps for Android and iPhone like COVID Alert (much of Canada), COVIDWISE (Virginia) and dozens of other jurisdictions around the world .

    I’m in Ontario and use COVID Alert on my iPhone 8 Plus. The apps are fantastic pieces of work from the Canadian Digital Service and its private sector partners Shopify and BlackBerry. That said, I have always wished for more feedback from the app itself. Something that gives me a sense of it actually working. I’m the first to admit that this isn’t a rational need. When you open COVID Alert here is what you see:


    Great! You’re active! But what does that mean?

    I’m grateful that no exposure has been detected! But the app doesn’t look like it’s doing anything. I know that that’s not the case. I know that it is doing stuff but that’s because I’m a nerd and because the Canadian Digital Service maintains the source for both the Android and iPhone COVID Alert apps on GitHub .

    But how can I see it actually doing stuff?

    Well here’s one way. Both iPhone and Android allow you to see a log showing each time COVID Alert has downloaded a list of exposures from the COVID Alert server.

    On iPhone you can see the log in Settings -> Exposure Notifications -> Exposure Logging Status -> Exposure Checks .

    What I believe this means is that in that one Exposure Check done at 10:09am ET COVID Alert downloaded 246 Tracing Keys (“device IDs”) of devices that had had a positive Covid-19 test reported over the past 14 days. It also determined that my iPhone did not get close enough to any of those phones, for a long enough period of time, to warrant me getting a Covid-19 test. It’s pretty cool to see the app at work.

    What else could it do?

    I would also love the app to help me understand:

    1. How risky is my current behaviour?
      How many devices did my phone see in the past 24 hours? How many rolling proximity identifiers (RPIDs) did my phone log? I know that you are not supposed to be able to derive a Tracing Key from an RPID, but could the system run a function over a set of RPIDs and estimate the number of unique Tracing Keys they represent?
    2. How effective is the app at warning people about potential exposure?
      We had 625 new cases of Covid-19 reported yesterday in Ontario. How does that compare to the 246 Tracing Keys my phone received? Do the time frames line up? Can I compare them? What’s the effective penetration of the app?

    Closing thoughts

    You can’t tech your way out of a policy or political problem. That said, I strongly agree with the what Apple, Google, and the Government of Canada have done here. If the policy decision is to continue to deploy these decentralized, anonymous exposure notification applications on a voluntary basis then we need to keep looking for ways to make them more effective and more compelling to download and use. Sharing more useful information with people could be a way to get more people to use the app and better inform public health authorities on what to do next.

  • Winning Communication in Remote Teams

    Man signalling with flare and coloured smoke
    Hello? Is there anybody out there. Photo by Hugo Jehanne on Unsplash

    Distributed teams are hard. Distributed teams, where some people are in office and others at home, are harder. Widely distributed teams with people working across countries and time zones are exceptionally hard. Widely distributed teams in a pandemic are damn near impossible to get right.

    Culture clash alone is an enormous challenge. I once had a fantastic American product manager piss off an equally strong director of engineering (in India) by complaining that the tool the PM needed “had fallen and can’t get up.” The director—and the engineers on the thread—took it as a tremendous insult, even though the product manager was trying to inject a little levity. The PM should have linked to the commercial:

    Apparently Lifecall isn’t big in India.

    Distance, culture, and time zone problems make communicating hard in remote teams.

    We have spent trillions of electrons this year on how to lead remote teams, so I’ll stick to something more prosaic: communication tools.

    Slack, Jira, Microsoft Teams, Zoom – the communication tools you use matter but not as much as how your organization uses them.

    Over the past twelve years I have worked in four different distributed organizations that ranged from bootstrapped start-ups to huge, traditional corporations. I often found a lack of alignment caused by the inconsistent use of communication tools. Solving this required us to be way more intentional in how we communicate.

    To communicate as a distributed team, you need to establish norms. Brainstorm norms as a group, winnow them down as leaders, align on them as a team and then communicate them out. When it works you get a written, jointly owned artifact that you can use to orient new team members and bolster the confidence of quieter team members. Above all leaders must model the right behavior; they need to adhere to the norms. When the rules are set up front, collaboratively, and followed by leaders, then the team is more likely to follow them. This eliminates the need for awkward corrections down the road.

    My List

    So here’s my list of communication tools and how to use them.

    Face to Face Meetings

    Good for: Well, really, everything. As my friend Dave Feldman put it, “all other things equal, face-to-face meetings are better for everything because they have really high emotional bandwidth. But they’re the most interruptive, hardest to coordinate, and can be wasteful of people’s time.” Face-to-face is the best way to have hard conversations. Generative exercises. Team building. Anything that started not face-to-face and got heated. Things that require dialogue. Design meetings, planning meetings, betting tables.

    Bad for: Status meetings. Networking conversations. Routine discussions.

    Video or Audio Conferences

    Good for: Things that require synchronous attention but less dialogue. Structured discussions. Ideally, decisions that are less contentious. Prioritization exercises, design reviews, presentations of an analysis. Collaboration where a small number of people are presenting to a larger number of people.

    Also good for: Not getting Covid 19. Right now it’s usually as close as we get to face-to-face. That said, video conferences have unavoidable limitations. We need to acknowledge them and mitigate.

    Bad for: Energy levels. Video is draining. Use it sparingly. Find other ways to ensure that all participants are “present” and attentive. You know your team is healthy if it is present without being threatened by the green camera LED. If the meeting is small enough be sure to “go around the room” once and check in with each attendee, one by one, asking them what they have to add. Make sure that everyone feels heard.

    Email

    Good for: Asynchronous, text-heavy communication. Wide distribution groups. Not time sensitive. Easy to read, so long as you write them carefully and put important content high in the email body. Things that people may want to consider and respond to. Monthly status updates to investors or senior leadership. Summaries of organizational changes (after those organizational changes have been announced synchronously to those directly affected).

    Major 🔑: Put the most important information into the subject line and lede (more here). Make sure your emails are easy to read on mobile devices.

    Phone and Text

    Good for: Getting a hold of someone after hours or while they are away from keyboard (e.g. in transit) and you need a dialogue. Always text first.

    Chat

    Patterns: Establish clear norms as to when people are / are not expected to be available on Slack; use threads to allow people to follow topics. Be aware of timezones; encourage use of Away messages. Train people on how to manage their notifications, alert settings, and Do Not Disturb. Don’t assume they’ll figure it out, or will feel comfortable turning it off. Use public channels, created for topics, teams and projects.

    Anti-patterns: This is going to need a bulleted list.

    • Remember that text loses tone, and that low-fidelity communication tools like chat can amplify management mistakes. Give people the benefit of the doubt when they seem clueless but do not tolerate behavior that makes people feel unsafe, particularly by those with power.
    • Chat transcripts ≠ documentation because is almost impossible to find things in history and often unclear what decisions were made. If you agree on a behaviour or solution to a problem then document it somewhere and link to it from the conversation, then link the document or ticket back to the conversation.
    • Chat messages ≠ to-dos or workflow tasks. If someone agrees to do something in slack, create a ticket in the relevant tracking tool and link to it from the conversation.
    • Avoid using #general. Real conversations should happen in dedicated channels. Slack channels are analogous to email threads.
    • Avoid using group messages: they’re hard to find, not categorized by topic, and carry an unclear expectation of privacy.
    • Most channels should be public: You should only make a channel private if you have a solid, clear reason to make it private, like a channel for peer managers to discuss performance issues across their collective team. If people only feel safe in private team channels something bigger is wrong with your organization.

    Finally, chat can make bad managers worse. The combination of immediacy, reach, lack of nuance, and lack of non-verbal feedback can encourage some incredibly stupid behaviour that can be hard to undo.

    Wiki

    Good for: Documentation, references, working documents; Version tracking important; High-fidelity (fancy formatting) not important

    Bad for: Things that imply sequence. Workflows, etc. Require a lot of maintenance. Prone to getting out of date.

    Bug Tracking or Workflow Management Tools

    Good for: Things that have a deadline, dependencies, multi-step tasks, etc.

    Bad for: Requirements, specs, designs – anything more than a few lines shouldn’t live in the bug, but should be linked out to a doc from the bug. Bug tracking tools are hell to search.

    Documents

    This includes Google/Word Docs, Slides/Powerpoint/Keynote and Sheets/Excel

    Good for: Narratives or analyses; work that needs a lot of formatting and high fidelity. Work that changes slowly. Read-aheads for meetings. Group editing. Tools like Google Docs and Dropbox Paper have great commenting tools although Confluence has come a long way.

    Bad for: Requirements. Reference documents. Documents that are expected to be maintained over a long period of time. After a project is over things like Word Documents are typically lost in a shared folder, hidden away, and never used again.

    Major 🔑: Tuck documents away neatly after a project. If they are no longer being updated, add a prominent note to that effect in the top of the document, with a link to other relevant, and more recent information (i.e. a Wiki). Manage the zombies.

    Things to remember

    Avoid holy wars. Focus on picking the smallest number of tools that cover the greatest number of use cases, weighted by importance. Don’t let the perfect be the enemy of the good, but be aware of tools with high switching costs (bug tracking, chat) and be careful.

    Remember that the process is as important as the outcome. You want to have broad support for the list so that influential team members feel ownership and will actively help to improve communication. That support should be backed up by top-down direction. And finally, don’t be cheap. Buy the right version of the tools and spend the money needed to get them configured and maintained property. If well paid software engineers are cutting and pasting tickets between bug tracking systems, I will find you.

    Above all, expect change. How an organization communicates will change. Plan for it.

    TL;DR

    In a distributed organization, leading well requires intentional communication. Intentional communication requires the effective, consistent use of digital tools. To achieve this:

    • take an inventory of the communication tools in use,
    • crowd source and align on the specific tools you are going to use and how you’re going to use them,
    • write patterns, anti-patterns, and starting points for each–make norms explicit by writing them down,
    • share them widely, and finally,
    • revise them often to keep them useful.

    One last thing. Communication is good, but too much communication is not a healthy signal. People need time to think, plan and work. They should not be thrashing endlessly between Slack, Jira and Docs. So make sure that people know how to protect their time and are able to set boundaries, get work done, and feel as though they are accomplishing things.


    Special thanks to my friends Michael Masouras and Dave Feldman for their ideas and their thoughtful help polishing this. Updated with ideas and edits from Dave on September 7, 2020.

  • Ramblings on No Code and the Permanent Revolution

    Since leaving Borealis I’ve spent time some time getting to know “no code.” Low Code Application Platforms (LCAPs), or “low” or “no code,” “No Code,” seem to have broken through. The promise is that non-programmers can point-and-click their way through building mobile/web Apps and deploy them straight to Google Play, the App Store, or a corporate app store. Last November Gartner assumed that “by 2024, three-quarters of large enterprises will be using at least four low-code development tools for both IT application development and citizen development initiatives” and “low-code application development will be responsible for more than 65% of application development activity.” Stuff is happening (waves hands).

    There is Nothing New Under The Sun

    We’ve been promised programming for non-programmers since electronic computers were invented. Just Google “4GL” or read the history of PowerBuilder. That said, as with most technologies a big enough difference in degree generates a difference in kind. So what’s different now is:

    1. Cloud and continuous deployment

    It’s now “normal” to have a developer deploy live code to production with a click. Ten years ago you’d still expect to have a sysadmin setup physical machines, flow code through a staging environment, whatever. Now, some 22 year old makes a commit, it runs through automated testing, and minutes later the code is live in multiple data centres for some or all end users. Magic.

    2. Every company is digital first

    Companies now, primarily interact with customers through digital channels. Banks, of course, still have branches, and retailers still have physical stores, but these are really just meatspace user interfaces that sit on top of software. Software ate the world. So anyone who works in a large organization who has some customer responsibility is now responsible for software, regardless of their stated job description.

    3. The data is sitting there

    Thanks to cloud, and some mind-numbing-but-vital data fabric building many “citizen developers” can now read-and-write important data in more or less real-time. If you layer in human-in-the-loop stuff like Figure 8, would-be developers are one click away from any bit of data that they could possibly need to build an app.

    4. Thinner IT organizations with an ever-growing dependence on outsourcing

    She who can get the most done, in the least amount of time, and claim the most credit will win. Low-Code lets you move fast as it takes the availability of specialized, highly-skilled specialized technical labour off the critical path.

    Oh you’ll still need specialized labour — people with a formal computing background who can untangle spaghetti code and keep user data out of unit 61398 but low-code lets you, as a middling business leader, move fast and break things (cringe). The time to proof value can be way shorter than with traditional programming — and in many cases you can throw dozens of 22 year olds from $consultingFirm to show traction and secure your bonus. Then you can get a proper budget in the next annual cycle, hire specialists, and build your empire.

    If you’re a BigCo manager you don’t need approval or budget to start a low-code project; you just need access to whichever LCAP(s) your IT and security people have approved. You can cobble things together with existing staff, get a bit of traction (or at least make a sexy demo) and then marshal resources for a real launch. It’s Lean Startup for Enterprise (TM).

    The point is, there are technical (cloud, digital first) and business (agility in the face of rigidity) factors that make Low-Code possible, and logical. It holds at least some promise of moving companies into a state of glorious Permanent (Digital) Revolution.

  • Connecting Humans with What They Need

    A year ago I wrote that 2016 would be the year that consumer AI went mainstream. In one sense, I was wrong. The average consumer still doesn’t interact with an AI application on a typical day. Siri, Cortana, Google Home, and Allo are making inroads but still have small reach when compared to, say, Android or iOS as a whole.

    Look Deeper

    Looking back, I realize that my angle of attack was wrong. “Consumer artificial intelligence” will not mostly be about putting AI in end-user devices. Siri, chatbots, and other natural language interfaces are a piece of the picture. However, the really interesting stuff is happening just below the surface.

    Last week, The New York Times published a long, breezy piece about AI. It centered on the migration of Google Translate from “conventional software” to neural networks. The article is a wonderful ramble through the AI countryside and I highly recommend it.

    What was done to Google Translate is a great illustration of where the action is with AI right now. Rather than transforming how we interact with technology, AI is, instead, first transforming what happens behind the scenes in applications like web search and machine translation. This is analogous to what happened with other, historic new technologies.

    Horseless Carriages, then Cars

    In the early steam age, manufacturing plants were still laid out in the same way that they were when they were powered by water wheels. When workplaces were electrified, machines were still placed as though they had to be connected to a steam engine and desks were left near large windows, and so on. It takes time to learn how to apply a new technology. But we are learning.

    The Machine Intelligence Landscape

    For the past three years, Shivon Zilis and the team at Bloomberg Beta have been mapping the Machine Intelligence Landscape. 2016 saw an explosion of startups applying artificial intelligence to specific business problems. This is because it is easy to estimate, capture, and charge for the value an AI solution creates for a business. It is also relatively easy to identify and process training data.

    Because of this, AI is becoming ubiquitous in business software. As a result, two years from now you won’t talk about “AI SaaS companies,” or “AI technology companies selling to business”— you’ll just assume that every piece of software marketing to business incorporates AI appropriately, just as that software uses a relational database or runs on the internet.

    We’ll have optimized the heck out of what’s inside the little boxes of a business (sales, marketing, etc.) and will move on to the interesting stuff — creating new things that couldn’t happen without AI, and that blow up the boxes altogether.

    But Back to the Consumer

    This is where things get interesting; what happens when these new companies — with their blown up boxes, and their AI powered businesses — interact with consumers, directly through their devices? We were given a glimpse of this with Amazon Go a few weeks ago — a grocery store that uses machine vision and AI, along with other technologies, to end the checkout line. The thin end of the wedge were the Google Now and iPhone prompts that tell you when you need to leave to get to your best appointment, or warn you about a transit delay.

    A tough question: What do we even want?

    If we take a step further back — what do we, as consumers, ultimately want AI to do? Clearly, automating routine work and improving our safety is important. But then what? The answer is definitely not “generate more notifications on my phone.” Or “give me another awful pseudo-human to talk to,” one that’s just as frustrating as the typical Comcast customer-service representative.

    AI will break down barriers and connect people to what they need

    So now we truly begin to apply AI to the long project of connecting humans with the things that they value. The things they need to complete their day, raise their families, and run their lives. We need to create new connections from human needs, across time and space, going deep into the organizations that can satisfy those needs. Re-think our businesses, re-think our lives, embrace what is possible, and blow up the little boxes. To create revolutionary change with accelerating, incremental change.

  • Why Consumer AI Goes Mainstream in 2016

    In October of 1994, Netscape released Netscape Navigator 1.0, the first commercial web browser. Over the next decade, the web went mainstream as it became increasingly usable.

    In October of 2011, Apple announced that Siri — the mobile personal assistant it acquired in 2010 — would ship on the iPhone 4S. Siri has continued to improve, as have Google Now, Amazon Echo, and a host of other solutions. 2016 will be the year that consumer AI goes mainstream.

    The Future Arrives

    For twenty years voice recognition had been “the future.” While Watson and Wolfram Alpha captured the attention of the press, they both had negligible impact on consumers. While the world obsessed over screen size and Angry Birds, Siri kept getting better.

    With WatchOS2 and an iPhone 6, Siri finally feels usable. I find myself tapping the phone less, and talking to my wrist more. How did this happen, and what can it tell us about the near future?

    • Experience effects. Like most machine intelligence, the more heavily Siri is used, the better it gets. Apple’s acquisition of VocalIQ, the world’s first self-learning dialogue system, should further increase the rate of improvement.
    • Moore’s law. AI problems are often bound by processing speed. CPUs double in performance every two years, creating a “natural” rate of improvement for AI systems. So, for the same CapEx, Apple now has four times as much processing power to throw at the problem. AI performance is further improved by algorithmic enhancements, like migrating processing work away from general-purpose CPUs and onto GPUs (graphical processing units), which are faster but more difficult to work with.
    • Apple built a search engine. Apple built a search engine to make it easy and worthwhile for app developers to extend Siri — without even knowing it. The Search API lets developers to describe their app to Apple (and Siri) with simple markup. Search is Apple’s way to evolve Siri into a tightly controlled platform.

    Siri is Now A Toy that Works

    Siri is now the main way I handle use cases where I can articulate a clear question (“what is the weather tomorrow?”) or command (“remind me to buy milk tonight”). While I emphasize ‘clear question’, the future will likely behold Siri’s ability to handle increasingly complex natural language questions that deliver an optimal solution to a problem. For example, “Given the tastes of my dinner guests, what meals should I prepare?”

    It Keeps Getting Better

    The ecosystem around Apple’s AI implementation is strengthening every day. Developers are exposing more and more functionality to Siri through the Search API. There are, however, bottlenecks imposed by Apple’s privacy policies that prevent it from having access to the rich user-generated datasets it helps create.

    The Kids Think It’s Normal

    And by “kids” I don’t mean “millennials,” I mean the toddlers running about my house. The older one now orders Siri around. She expects to be able to talk to a computer. Remember when kindergarteners suddenly expected all screens to be touch sensitive? Generational shifts like these are great leading indicators of what’s next.

    Conclusion: Next Year…

    By the end of next year, consumer AI will be everywhere. Operating systems will expose key features of installed applications or replace them altogether.

    Facebook M, Operator, WhatsApp, WeChat, Slack, Kik, and every service with a natural language interface is, at its heart, an AI platform. Algorithms can either establish a direct rapport with users or monitor what is being said in a privacy-sensitive way, collecting intelligence and offering assistance.

    For example, Apple will integrate Siri within Messages and Mail, Emu-style.

    …and Beyond

    Consumer AI will continue to improve by a factor of two every two years. This sustained, exponential improvement will bring startling results. Incremental projects will overtake attempts at big-bang disruption. Think less “Google Self-Driving Car” and more “Tesla Autopilot.”

    Incumbent players will accelerate the acquisition of consumer AI applications to bolster their teams, and to defend their positions. Nobody wants to have done to them what Google did to Yahoo.


    Updated on December 16, 2016 based on feedback from Nathan Benaich and Ahmad Nassri, as well as what I learned at the Machine Learning and the Market for Intelligence hosted by the @creativedlab and the University of Toronto. Inspired in part by Shivon Zilis’s excellent work in the space.