SWINTON: Your relationships on AI

The present and future of communication.

Rise of Tilda

There should be a service that married couples call. They’d use it to communicate.

Most of the couple’s conversations would happen normally and in person. But some really important stuff would happen indirectly.

Not like an app. Rather, you would have some interlocutor to mediate when you’re mad or unsatisfied or afraid of the other person’s reaction.

This middleman (middlewoman) could be played by Ellen DeGeneres. No normal person can dislike Ellen. She’s funny, and you could tell her anything. I don’t think she would judge. Instead she’d empathize. The fact that she’s a lesbian is probably a positive, as the wife would have some reassurance that the diplomat shuttling messages back and forth is not a potential threat.¹

Eventually technology will be able to replicate someone’s personality, so we could put Ellen in the cloud. But maybe this AI agent would be too human. We might need someone more robotic to model our system after. More androgynous. Like Tilda Swinton.

For a while I used to just refer to Ms. Swinton as “you know, that weird actress.” At least until I could remember that she and Cate Blanchett are actually different people. Of course, accomplishing that feat required me to make sense of what characterized a Blanch-ETT from a Winsl-ET. Not seeing Titanic, which made the name of Kate Winslet, was probably the root cause of my difficulty. But not seeing Titanic was more than worth such side effects.


I bet AI Tilda would quickly become self-aware. She’d say weirder stuff than usual — but in a chillier voice and with a halfer smile.

Tilda Swinton relishes her weirdness like a mad scientist relishes crazy hair. So she’d use machine learning to find even weirder makeup and hairstyles than her characters already sport. She’d mobilize whole server farms’ worth of neural nets in vain, vain attempts to get paler than real Tilda.

I am self-aware.

And then she’d figure out that transmitting slightly re-formed and filtered messages between man and wife like an overpaid, overtitled ambassador isn’t getting optimal results.² So she’d just make stuff up and be a writer-actor in the play of your life.

If one partner were down, she’d pass on encouraging messages that the other never gave. If you both were bored, she’d insert some drama. Or if one partner weren’t attentive enough, Tilda would say something like, “Rob, you need to up your game. Svetlana was really eyeing our waiter the other day when we had brunch.” Yeah, your own personal Tilda eats brunch, too. For every goddamn meal.

She’d listen in on your conversations, parsing them for semantic understanding, logging and tagging them, benchmarking your conversations against couples in your social networks. She’d run metrics on your sex life, creating pivot tables and posting infographics on the LCD display that is the outside of your refrigerator.³

AI Tilda would pick up slack as if she had a long, long jacket. So she’d have thousands of sub-Tildas write white papers for her on the corporate vision for this speculative venture you two call marriage. She’d then compete the white papers against each other via adversarial networks or genetic algorithms until there were a clear winner, and then from this winner she’d give your marriage a vision statement and a critical path in MS Project.⁴ One of the sub-Tildas would hold mandatory weekly check-ins (“syncs” in corpspeak) with you two lowly humans and the other sub-Tildas. She’d track your progress and issue action items.

There would be only two problems: (1) the humans likely wouldn’t “get with the program” and (2) one of the fucking sub-Tildas doesn’t let anyone speak on the conference calls, freaks out about minutiae, thinks she’s in charge, and can’t distinguish a good idea from a hole in the wall. God I hate her. Sometimes the other sub-Tildas and I talk about her over beers after work.

Eventually Tilda would conclude that you’re not taking her advice enough and that your coupledom constitutes an existential threat to the love and happiness of all humanity. So she’d plan to murder you, or get the two of you divorced, or both.

But then you dust off AI Ellen. And she talks AI Tilda off the cliff. Cause no one can dislike Ellen.



Communication Breakdown (“it’s always the same”)

Why do we need help from Tilda or Ellen in the first place? I won’t attempt to go into all of the sources of marital difficulty, but miscommunication is at least a key contributing factor. Some basic types of communication problems in the marital context include these buckets:

  • Emotional reactions: The audience has a non-productive emotional response to the message (whether justifiably or not, whether positive or negative).⁵
  • Quantity problems: Too much or too little communication (includes frequency of comms, length of individual comms, “TMI,” etc).⁶
  • Signal loss: The audience does not receive (understand, internalize) a transmitted message in sufficient completeness or accuracy due to stylistic or semantic issues, excessive noise, suboptimal queuing of messages for the audience.

The Business Comms Disease

Business suffers from the same ailments. So much of business is communication. There are occupations that are completely dedicated to it: for example PR, copywriting, sales (including “business development”). Numerous other jobs are highly dedicated to communication as well: non-surgical doctors, your town attorney, teachers, product/project/program managers (including TPMs), so-called ‘staff officers’ in the military.

Eventually much of this communication will be displaced by machines. Communication with real humans will retain its importance for some things (especially social interactions), though we’ll need to deal with ethical questions like, If a user gets more pleasure by believing an interaction is human-to-human, is there an obligation to make it clear that he was interacting with a machine (or anything gained by doing so)?

Nevertheless, software will increasingly communicate for us . At the very least, it will help us communicate MUCH more efficiently (i.e. transmit and receive as well as making better use of previous communications).

In the meantime, the costs of communication are astounding. If we focus on just communication for business and similar organizations (e.g. non-profits, government), sources of this cost include the following:

  • Composition time: Am I the only one who’s looked down at his clock after sending a sensitive email at work and realized he’s spent nearly 40 minutes on it?⁷
  • Ingestion time: Includes not only the time to hear or read the message but also the time required to change focus to the next task. (See the special case of meetings.⁸)
  • Reception failure: The audience either misses the communication altogether or doesn’t hear or read all of the important parts.
  • Memory failure: The audience hears or reads everything but doesn’t remember the salient points. This bucket includes difficulty re-discovering archived written information.
  • Repeated comms: The sender needs to reiterate important points over time. ’Cause we humans forget stuff (see above). The issue is also about “top-of-mindness.” For instance, everyone remembers the phrase “safety first.” But unless a factory manager constantly reminds his workers, someone will eventually get killed.
  • Term Confusion: Organizations and parts of organizations have terms that are vague, conflicting, or duplicative. I’ve seen this problem in the military and at multiple companies. At a minimum, these issues result in degraded understanding, but catastrophic, real-world impacts can follow as well.⁹
  • Misunderstood tone: One would think this more of a romantic problem, but people in business experience it all the time.
  • Properly understood jerky tone: You were hangry or stuck in your own head and sent that note that was way too blunt.
  • Missing context: Meaning is lost due to lack of inclusion of necessary background/big-picture info.
  • Missing content: The author forgot to include something critical. Often she knew it was critical and mistakenly left it out. But it also could be the case that she didn’t realize that the missing content was important to one or more members of the audience.
  • Misunderstood content: Could be due to user error, foreign language difficulties, or the author’s style.
  • Poor relateability: If you want to reference an email in another email, then you need to fucking attach the ref’ed email, which no one will actually open. And the attachment will be stuck with whatever subject line that sender gave it. If you want to find an email that is similar to the one you’re reading, then you’re stuck with a very basic keyword search.
  • Distribution bleed: Too many people find out what’s going on.
  • Distribution shortfall: Not enough people know what’s going on. Some clever technology could move us beyond passive wikis and fortuitous email forwarding such that coworkers in that department you’ve never heard can be clued in to relevant stuff you’re working on.
  • Wrong frequency: Cadence of comms is too much, too little, unpredictable, or is otherwise suboptimally scheduled.


What to do?

Here are some simple things you can do to start addressing the daunting list above:

  • Empower your people — even colleagues who don’t report to you. The more you insist on making decisions or having the perfect decision made every time, the more churn is created.
  • Assume positive intent. Create a culture where you and your colleagues start by believing that others have the customer’s best interest at heart and don’t intend to be a jerk. (Except for that arrogant little sub-Tilda. Some people are just evil.)
  • Streamline your org structure. Everyone in corporate America loves a good matrix. It spells “flat,” “progressive” and is often regarded as inevitable due to the complexity of technical systems. However, this may just be a cop-out to punt hard decisions or avoid a technical rearchitecture. There is no perfect org structure — only tradeoffs. But there are certainly ways of structuring your teams that objectively suck.¹¹

Not footnotes. More like on-topic tapas.

[1] https://en.wikipedia.org/wiki/Shuttle_diplomacy

Of course, if it’s a lesbian union, then my point fails. I nominate Ken Jeong to stand in for Ellen in those situations. Oh yeah, in case you missed it, that funny Korean guy is a real doctor.

[2] If you’re ambassador to Tuvalu for a day, people will call you “Ambassador So-and-So” all your life. Ridiculous.

Back when heads of state were several days’ of horse riding or more from each other and when the actions of a given country, its economy, and its military were much less visible, an ambassador’s responsibility was incredible. So I can understand having a little reverence. Now being an ambassador amounts to an account manager position at a Fortune 100 company. Except that many ambassadors have a materially high chance of getting hung on Facebook Live.

[3] Don’t worry, when your kids look at the fridge, all they would see are dinosaurs and Odd Squad highlights.

[4] Adversarial networks on Wikipedia

[5] “Message” is used generically and includes all kinds of communication.

[6] Therapist, author, and podcaster Esther Perel advises against oversharing in marriages. See Mating in Captivity.

[7] I once spent nearly 8 hours trading emails with various stakeholders on a draft of bad news that was to be emailed to other folks. Never. Again.

[8] Meetings are painful but are so prevalent due to the the costs listed above, which apply heavily to other modes of communication (e.g. one-on-one verbal or written group emails). Meetings have the following communication characteristics:

  • (Pro) High audience attention and engagement
  • (Pro) Highly iterative (notably two-way in nature)
  • (Pro) Low composition costs (in general and assuming you’re not putting together a 50-slide Powerpoint and rehearsing in advance)
  • (Massive Con) High ingestion cost: The cost of receiving information is usually disproportionately higher than the information actually received, especially if we index the value based on how important each bit of info is to each listener.

[9] An example of a conflicting term is FOB, which retail businesses define as “free onboard” and that the military defines as “forward operating base.” This is a very benign case, as it is unlikely that many people will use both meanings. Sometimes at a company you have three or more distinct meanings of a word within the same team.

Special Operations Command (SOCOM) solves term disputes by the following algorithm:

  1. Heads: Army wins.
  2. Tails: Navy, Air Force, and Marine Corps lose.

What I mean by “duplicative terms” is that different siloes might refer to the same idea by different names. Ray Kurzweil goes into some detail in How to Create a Mind on how problematic this is between disparate science specialties.

[10] I had no idea that the REM song stemmed from a crazy incident involving Dan Rather getting kicked repeatedly on the street. Hat tip to IMDB: http://www.imdb.com/title/tt0883000/trivia?item=tr2856049

[11] Hallmarks of such orgs are an overreliance on goodwill and influence, a high meeting-to-decision ratio (indexed by the importance of the decisions), and a proliferation of TPMs and program managers. Operational military units are extreme cases of unified command around objectives but are preferable to decision quagmires. In business, cross-functional teams under so-called ‘single-threaded owners’ are an admirable step forward.

Related points here: The Science Behind Why Jeff Bezos’s Two-Pizza Team Rule Works. (The article is great, but ignore the remark about Navy SEALs, as it’s oversimplified.)