Beyond the AI Winter

Logline and treatment. As the story gets told, the tech industry is on the verge of sweeping transformations based on AI.

My job is about learning. However, my career in the tech industry began with AI studies; followed by a string of related projects at IBM, NASA, Bell Labs; subsequently, several years of neural network R&D; then roles leading data teams; eventually moving on to a deeper study of the nature of learning. People learning, machine learning, many aspects of the word learning tugging at each other. I’m particularly interested in the directions of AI, thrilled to see some of its arguments come full circle, and quite curious about its current public discourse.

Opening Scene

Introduce the protagonist, establishing stakes, and foreshadow the first plot point. Machine learning played a crucial role in successful disruptive forces of the past decade-ish, ushering us into the dawn an AI era. Consider a litany of use cases throughout Google, Amazon, LinkedIn, Facebook, Twitter, Netflix, Airbnb, Udacity, SpaceX, Tesla, Uber, Pinterest, Etsy, etc. We’ll even toss in OkCupid for flavor. Old economy dinosaurs dwindle amidst the tumult of globalized financialization, while an ineffable Next:Economy emerges as an extensive partnership between people and machines. Uber’s valuation now exceeds that of GM, Ford, Honda, or any auto manufacturer other than Toyota.


Hook moment, create a question that compels the audience to crave an answer, possibly tangential to the protagonist’s needs. Name ten high-flying tech start-ups that lack any application of machine learning on their roadmaps. Can you do that, off the top of your head? Probably not. It’s become pervasive. Part of the DNA. Microsoft penned an edict, along with compelling examples of the beauty of machines and humans working in tandem.

Inciting incident, game-changer event that leads to a decision at the first plot point. Thanks to tenacious researchers working for many years on artificial neural networks, we now have a technology called Deep Learning. Its magic somehow ups the ante of machine learning, enabling the likes of Google, Apple, IBM, Facebook, Amazon, Microsoft, Baidu, etc., to commercialize AI. Economic outlooks now abuzz with personal assistants and chat bots promise to hurtle us toward the future.

Debate, point of no return, protagonist must make a choice. Pitch decks marched up and down Sand Hill Road pivoted abruptly from the rhetorical trope of “Like Uber for XYX” to a more contemporary figure of speech “XYZ with AI” instead. Do we place our bets on old or next? Meanwhile, news stories throughout the webosphere report fantastic leaps in the science of mingling automation and cognition. People are pleased, frightened, ecstatic, disgusted; dancing to the intro narrative of a NIN remake of a popular Queen song.


First pinch point, a reminder of the antagonist unfiltered by the protagonist’s POV, directly visible to the audience. Something nagged at the back of my mind, the obscured remains of some lingering memory. Arguments about AI transforming economies near-term have merit. I buy much of that, despite decades to contrary. The thing that continued to nag seemed more about the public dialog becoming warped. Alarmingly so, and not in a particularly Elon Musk kind of way.

Circa 1970s and 1980s, arguments about “Hard AI” loomed large in academe, then fell precipitously during the AI winter. One of my favorite examples before the fall was Eurisko. Rather curiously at the time, much the same AI community had all but rejected Eurisko as they rushed headlong into their winter. I attended a portentous lecture at Stanford in 1984 by Eurisko’s author Doug Lenat where other researchers dismissed notions of machine learning as “overly pragmatic” hacks. Nonetheless, some became quite wealthy through a Stanford spin-off that leveraged machine learning a decade later.

Part of what had been gnawing at the fringes of my memory was exactly that narrative arc: from machine learning as a derided hack to machine learning as an engine for wealth creation. Something about the opportunism, what had been left unsaid.

Going back to the math, contemporary notions of ML describe a collection of methods used to generalize patterns from examples in data. It’s a subset of techniques under the general category of optimization, mostly used for pattern recognition albeit with a strong dose of dimensional reduction. Clustering algorithms, decision trees, support vector machines, random forests, neural networks, and many more. We automate the training of models based on these algorithms, where learners optimize for a given data set, generally solving for some loss function plus a regularization term to reproduce patterns effectively. Relatively recent innovations combine layers of neural networks into more complex organizations, aka Deep Learning. An entire generation of engineers and scientists has grown up regurgitating this definition. Useful stuff, smarter uses of computing overall — though ML is not in itself embodying much that could be called AI. Yet do these learners form the proverbial ghosts within the machine?


Midpoint, new info changes the context for the protagonist and audience, some catalyst activates a different course. Prior to the mid–1980s, pattern recognition was a footnote. Circa projects like Eurisko the phrase “machine learning” included a much broader range of technologies: planning systems, schedulers, heuristics for generalized search, blackboards. In general, we could characterize those as control systems that leverage learning.

Even before the invention of digital computers, control systems — which represent a considerably larger and more complex area of optimization — provided the early context for AI research. Norbert Wiener and others explored the first wave of cybernetics in the early 20th century. For example, control systems were used to link new radar systems with anti-aircraft weapons during WWII. After the war, Wiener made what was considered at the time a strange move by inviting biologists McCulloch and Pitts to MIT. There they conducted some of the earliest research on artificial neural networks, the ancestral grounds for Deep Learning.

Close In

Second pinch point, another reminder of the antagonist’s forces, which up their game against the protagonist. Cybernetics was about the automation and optimization of control systems. We can reach back a century and a half earlier to gauge the impact of mingling automation and control: consider the enormous implications of the cotton gin. On the surface that one invention gave a huge boost to agricultural productivity. Yet it became a major factor in the demand for slave labor in the United States: a terrible lesson from history that we can never afford to forget.

A crucial point here is that the first-order cybernetics of the early 20th century was about automation: the artifice, manufacturable. Through the addition of digital computers and software, we gained optimization along with automation. As a result the industrial transformations of the 1960s and 1970s in particular were startling. Even so, AI is about a combination of artifice and intelligence. Weiner set the stage for that by inviting McCulloch’s team into the fold. They expanded the scope of the discourse beyond control systems and toward a cognitive dimension, through neural networks.

Flash forward several decades. Judging by tech talks and case studies from Silicon Valley start-up circa mid–2010s, one might believe that machine learning as pattern recognition drives business. It does not. In particular, ML does little to act upon insights gained. Let’s consider Uber as an example. ML use cases may help detect patterns: traffic, drivers, commuters, etc. Those can help indicate where value could be “harvested” within the system: opportunities for action. Even so, the dispatcher at the center of Uber works to schedule and optimize rides. It operates as a control system at the heart of the business. Those kinds of control systems may leverage ML to detect patterns, etc., but there’s much more involved. Notably, determining which offers to sell, scheduling resources to deliver on those customer promises, handling contingencies etc. Manipulating the supply chain is where a business earns profit. Patterns only play minor parts, while control is center stage.


Second plot point, final injection of new info, doesn’t need to be fully understood by the protagonist yet, quest gets accelerated. Microsoft recently expressed a big notion — and evidence — about the potential for machines and humans working in tandem. Google, IBM, Apple, Amazon, etc., are on the same vibe. A magic decoder ring for that, which I highly recommend, is the PyCon 2016 keynote by Lorena Barba (see video and slides). She examines “language action perspective” from Understanding Computers and Cognition by Terry Winograd and Fernando Flores. Notably:

We don’t communicate to share information, we communicate to change the world … we coordinate in conversation, offering promises, building trust, developing relationships … All of these actions are linguistic, they happen in conversations … the actual doing of the work is outside of the graph.

Dr. Barba analyzes this in the context of the semantics of GitHub pull requests. Looking in almost an evolutionary perspective, she proposes that open source culture serves as a training ground for teams effectively coordinating actions through computational platforms, ostensibly including the roles of machines. In turn, speech acts represent powerful language rituals that build trust and coordinate actions.

I haven’t felt quite that many intellectual tingles since I first read René Thom’s Structural Stability and Morphogensis. Especially in the “capture” morphology of predators manipulating their POVs vis-a-vis that of their prey, as an articulation of evolutionary pressure for cognition. Thom and Barba both resonated on a frequency remarkably close to what I’d heard Lenat invoke so many years before. A particular beauty of this connection is found in the history of these writings.


Even Rocky had a montage. Peeling back the layers, Winograd and Flores had partnered on language action perspective after Stanford and others helped get Flores released from prison. Following Kissinger’s war against Chile, Flores had become a political prisoner for his part in Project CyberSyn, helping to automate and optimize aspects of the Chilean economy under Allende. Flores had employed an expert named Humberto Maturana, who in turn connected the project with leading cybernetics experts. Maturana had been a graduate student at Harvard, collaborating with McCulloch and Pitts.

Punctuate the syntactic suspense, full stop, then roll forward. Maturana went on to partner with one of his grad students, Francisco Varela, on the theory of autopoiesis in second-order cybernetics. Notably, the observer becomes part of the system observed, engages with it cybernetically. I’ll point to one of my most treasured books, Autopoiesis and Cognition by Maturana and Varela. They present criteria for analysis of living systems, through a linguistic domain as Barba described. Watch the Barba video, read Autopoiesis and Cognition, with Thom as extra credit — then reconsider that Uber example above. Outlines of practical AI emerging within the context of Next:Economy begin to take shape, whereas they do not with ML/Deep Learning merely as the examples. We see conversations outside of the graph of the actual work being performed, where speech acts commit what ML could only suggest — e.g., optimizing supply chains, making offers, etc., in the Uber example.

I’d be remiss in my job if this didn’t tie back into how people learn. Lorena Barba’s keynote also hints at a taxonomy for a kind of pedagogy, a structure for acquiring intelligence. Consider the stages of evolution of participation in an open source project:

  • initiates first must learn to program, manipulating language (linguistic domain) to cause actions through computational platforms
  • advanced actors learn to distinguish key patterns amidst ambiguity, as a rough approximation of the role of machine learning
  • experts distinguish themselves as second-order cybernetic “control systems”: optimizing for energy, recognizing when to commit, guiding the projects, etc.

That’s quite close to a cognitive model that I use to understand the communities involved in learning at O’Reilly Media.


Act III, combine the A story and B story to reveal a solution. The big reveal, so to speak, for me was quite appropriately made in cinema. Circa late 1960s, Kubrick produced an amazing vision through 2001: A Space Odyssey, exploring notions of intelligence from several camera angles. I recommend reading all three books in the series, plus the additional 2–3 books about “the making of”, plus the original short story that Clarke adapted. Several times. While taking notes. Then watch the movie. Repeatedly, if you can. Not long after I’d learned to ride a bicycle, I rode downtown to watch that movie, alone as a kiddo, in what I’d describe as a near-religious experience of connecting the dots, foreshadowing my life-long interests in AI, cognition, etc. Kubrick holds the cinematic world’s record for a 4 million year jump cut: from an early hominid first inventing an externalized weapon (an animal’s leg bone) thrown spiraling up into the sky, cut to an orbiting nuclear weapon. Sanitized versions of the narrative describe those as “tool” and “satellite” respectively, though Kubrick followed an arc not too different from Thom’s predator morphology.

I used the phrase “near-religious”: there are indications that Kubrick was attempting to create the world’s first “religious text” on the silver screen. Not in the sense of any monotheistic representation, but a narrative about self-reflection, transcendent, numinous. Throughout that, Kubrick thoroughly explored notions of automation, control systems, speech acts, etc., in the context of intelligence. Key plot points concern control of nations over public information, control among nations over lunar territory, people controlling a ship, an AI controlling the same ship, people controlling information that the AI lacks until a pivotal moment, the AI attempting to control the people as a kind of predation to accomplish its encoded mission, people gaining control over the AI, etc. Ultimately, the pull-push of these scenarios represents the “fun and games” of the film’s story beats. Meanwhile, an unfathomably-evolved alien intelligence has crafted the entire set up as a kind of “baby alert” rube goldberg machine for tracking the evolution of emerging intelligent species — the IoT version of SETI. Were the aliens manipulating outcomes the entire time? Had they become cybernetically engaged with the observed system? What a beautiful system of interconnected metaphors for discussing AI.

Some humor arrives in how Kubrick was asked so many times by interviewers to give explicit answers about the meaning of the film. Eschatological pronouncement? A Nietzschean allegory? Tenaciously, Kubrick appears to have explained patiently to each interviewer that the interpretation was intended to be highly subjective for each viewer. Or, as I’d tend to paraphrase, enabling an autopoietic enactivism as participation by the audience. Then again, I’m an animist. But I digress.

Another literary reference helps explore where we are now with respect to AI: the science fiction novel Neuromancer by William Gibson. My all-time favorite literary character, Marie-France Tessier, constructs a dual system of opposing AIs on behalf of her family: Wintermute and Neuromancer. These AIs become principal characters in the novel. One of the better commentaries on Marie-France that I’ve been able to find examines Wintermute vis-a-vis Heidegger’s essay “The Question Concerning Technology”. That provokes a question: per Heidegger, in the Next:Economy do today’s workers become mostly a standing reserve for the AIs to manipulate?

Tying Up Lose Ends

Finale, wrap-up while dispatching all the bad guys in ascending order. Back to the definitions of ML circa mid–1980s, there were planners, schedulers, etc. Not just pattern recognition. One area which was all but forgotten by the popular dialog about AI is genetic programming. GPs evolve many generations of programs to solve a given problem. Back to what Barba described as speech acts, GPs determine actions in the linguistic domain. For example, serving as planners. I wrote a tutorial about running a GP as an example of implementing a Mesos framework.

Taking a cue from precursors such as Eurisko, the field of probabilistic programming is showing promise. Keep an eye open about Gamalon, among others. What if entire workflows could evolve and restructure themselves, based on available data?

Another interesting area is in fuzzy logic, which made a blip in the early 1990s, then dove for cover during the AI winter. I’m seeming more than blips reemerge. For background, check out Fuzzy Thinking by Bart Kosko.

Pulling these threads together, recently a gaming AI defeated a US Air Force expert combat pilot in simulated combat. As the pilot explained,

It seemed to be aware of my intentions and reacting instantly to my changes in flight and my missile deployment. It knew how to defeat the shot I was taking. It moved instantly between defensive and offensive actions as needed.

The system ran on a $500 PC, employing a “genetic-fuzzy” approach.

Lest these examples cast too much of a shadow on AI in literature and practice, here’s a recommended antidote, Adventures in Narrated Reality by Ross Goodwin. Money quote:

The fictional definitions it created for real words were also frequently entertaining. My favorite of these was its definition for “love” — although a prior version of the model had defined love as “past tense of leave,” which I found equally amusing.

Also check out the inexplicably entertaining short film Sunspring, authored by an AI named Benjamin. That leverages a recurrent neural network approach called long short-term memory (LSTM ) which takes abductive reasoning to new levels. While more of a hack though hugely entertaining, I’ve recently enjoyed playing with this gem from Machinamenta.

Roll Credits

Final image, opposite of opening image, show what changed. Conjecture: we’re seeing rudimentary AIs emergence as control systems for the supply chains of a new kind of company. These automate what in previous times could have been handled by layers of management, directing a work force of employees. Instead, now we’ll get an AI directing some pool of “resources”, i.e., contractors who’ve probably signed complex NDAs and strict arbitrage agreements. That’s a key point. We have highly codified legal structures for how people employ and manage people, i.e., labor laws. For example, California has labor laws that are arguably tilted to benefit the employees — at least more so than in “right to work” states such as Texas or Missouri. One might argue how this has been a distinguishing factor for Silicon Valley to collect and retain top talent in technology. Or not? Labor laws and labor forces represent some of the main expenses in business. They are entities squarely within the linguistic domain, something that AIs can understand and manipulate.

While there’s much public discourse about the risk of AIs displacing workers, the lesson of the cotton gin indicates otherwise. I have a hunch that the bulk of those receiving pink slips due to automation will not be workers, but mid-level managers. To some extent, that effect is already in progress.

Another definition of corporate structures, perhaps one that’s not widely shared, holds that the laws, practices, cultures, etc., that arise within publicy held corporations in effect serve to dissipate risk while perpetuating wealth on behalf of their shareholders. That might also be called “self-reproduction” — a property of autopoiesis, in second-order cybernetics. Putting on an SF writer’s hat in place of an engineer’s pocket protector, a view that emerges is where a wealthy 1% of executives leverage AIs to coordinate the employees. Lessons from cloud, DevOps, Google SREs, Amazon AWS, underscore the many-to-one nature of the displacement via automation. Consider how systems engineers could begin, circa 2005-ish, to manage several thousand servers instead of previously managing a hundred or so. An autopoietic system is self-producing in terms of both information and complexity: in other words, these systems are inherently more complex than their environments. One can imagine byzantine tangles of legal structures and supply networks, overseen by a few executives while executed by a large work force as a standing reserve.

Anywho, next time anybody begins to equate machine learning with artificial intelligence, please call BS on them. Control and evolution are much more apt.

If you’re interested in more of this kind of material, I cover some of the history of machine learning, the math-without-tears, and hands-on examples of related software technology in the video Just Enough Math, advanced math for business leaders to recognize the contexts for where and when to leverage big data. Also check out a fantastic new compilation, The Future of Machine Intelligence by David Beyer. See you at our new O’Reilly Artificial Intelligence conference in NYC, September 26–27.