#29 On the referendum & #4c on Expertise: On the ARPA/PARC ‘Dream Machine’, science funding, high performance, and UK national strategy

Post-Brexit Britain should be considering the intersection of 1) ARPA/PARC-style science research and ‘systems management’ for managing complex projects with 2) the reform of government institutions so that high performance teams — with different education/training (‘Tetlock processes’) and tools (including data science and visualisations of interactive models of complex systems) — can make ‘better decisions in a complex world’.  

This paper examines the ARPA/PARC vision for computing and the nature of the two organisations. In the 1960s visionaries such as Joseph Licklider, Robert Taylor and Doug Engelbart developed a vision of networked interactive computing that provided the foundation not just for new technologies but for whole new industries. Licklider, Sutherland, Taylor et al provided a model (ARPA) for how science funding can work. Taylor provided a model (PARC) of how to manage a team of extremely talented people who turned a profound vision into reality. The original motivation for the vision of networked interactive computing was to help humans make good decisions in a complex world.

This story suggests ideas about how to make big improvements in the world with very few resources if they are structured right. From a British perspective it also suggests ideas about what post-Brexit Britain should do to help itself and the world and how it might be possible to force some sort of ‘phase transition’ on the rotten Westminster/Whitehall system.

For the PDF of the paper click HERE. Please correct errors with page numbers below. I will update it after feedback.

Further Reading

The Dream Machine.

Dealers of Lightning.

‘Sketchpad: A man-machine graphical communication system’, Ivan Sutherland 1963.

Oral history interview with Sutherland, head of ARPA’s IPTO division 1963-5.

This link has these seminal papers:

  • Man-Computer Symbiosis, Licklider (1960)
  • The computer as a communications device, Licklider & Taylor (1968)

Watch Alan Kay explain how to invent the future to YCombinator classes HERE and HERE.  

HERE for Kay quotes from emails with Bret Victor.

HERE for Kay’s paper on PARC, The Power of the Context.

Kay’s Early History of Smalltalk.

HERE for a conversation between Kay and Engelbart.

Alan Kay’s tribute to Ted Nelson at “Intertwingled” Fest (an Alto using Smalltalk).

Personal Distributed Computing: The Alto and Ethernet Software1, Butler Lampson. 

You and Your Research, Richard Hamming.

AI nationalism, essay by Ian Hogarth. This concerns implications of AI for geopolitics.

Drones go to work, Chris Anderson (one of the pioneers of commercial drones). This explains the economics of the drone industry.

Meditations on Moloch, Scott Alexander. This is an extremely good essay in general about deep problems with our institutions.

Intelligence Explosion Microeconomics, Yudkowsky.

Autonomous technology and the greater human good. Omohundro.

Can intelligence explode? Hutter.

For the issue of IQ, genetics and the distribution of talent (and much much more), cf. Steve Hsu’s brilliant blog.

Bret Victor.

Michael Nielsen.

For some pre-history on computers, cf. The birth of computational thinking (some of the history of computing devices before the Turing/von Neumann revolution) and The crisis of mathematical paradoxes, Gödel, Turing and the basis of computing (some of the history of ideas about mathematical foundations and logic such as the famous papers by Gödel and Turing in the 1930s)

Part I of this series of blogs is HERE.

Part II on the emergence of ‘systems management’, how George Mueller used it to put man on the moon, and a checklist of how successful management of complex projects is systematically different to how Whitehall works is HERE.

On the referendum #28: Some interesting stuff on AI/ML with, hopefully, implications for post-May/Hammond decisions

Here are a few interesting recent papers I’ve read over the past few months.

Bear in mind that Shane Legg, co-founder and chief scientist of Deep Mind, said publicly a few years ago that there’s a 50% probability that we will achieve human level AI by 2028 and a 90% probability by 2050. Given all that has happened since, including at Deep Mind, it’s surely unlikely he now thinks this forecast is too optimistic. Also bear in mind that the US-China AI arms race is already underway, the UK lost its main asset before almost any MPs even knew its name, and the EU in general (outside London) is decreasingly relevant as progress at the edge of the field is driven by coastal America and coastal China, spurred by commercial and national security dynamics. This will get worse as the EU Commission and the ECJ use the Charter of Fundamental Rights to grab the power to regulate all high technology fields from AI to genomics — a legal/power dynamic still greatly under-appreciated in London’s technology world. If you think GDPR is a mess, wait for the ECJ to spend three years deciding crucial cases on autonomous drones and genetic engineering before upending research in the field…

Vote Leave argued during the referendum that a Leave victory should deliver the huge changes that the public wanted and the UK should make science and technology the focus of a profound process of national renewal. On this as on everything else, from Article 50 to how to conduct the negotiations to budget priorities to immigration policy, SW1 in general and the Conservative Party in particular did the opposite of what Vote Leave said. They have driven the country into the ditch and the only upside is they have exposed the rottenness of Westminster and Whitehall and forced many who wanted to keep the duvet over their eyes to face reality — the first step in improvement.

After the abysmal May/Hammond interlude is over, hopefully some time between October 2018 — July 2019, its replacement will need to change course on almost every front from the NHS to how SW1 pours billions into the greedy paws of corporate looters via its appallingly managed >£200 BILLION annual contracting/procurement budget — ‘there’s no money’ bleats most of SW1 as it unthinkingly shovels it at the demimonde of Carillion/BaE-like companies that prop up its MPs with donations.

May’s replacement could decide to take seriously the economic and technological forces changing the world. The UK could, with a very different vision of the future to anything now proposed in Whitehall, improve its own security and prosperity and help the world but this will require 1) substantially changing the wiring of power in Whitehall so decisions are better (new people, training, ideas, tools, and institutions), and 2) making scientific research and technology projects important at the apex of power. We could build real assets with much greater real influence than the chimerical ‘influence’ in Brussels meeting rooms that SW1 has used as an excuse to give away power to Brussels where thinking is much closer to the 1970s than to today’s coastal China or Silicon Valley. Brushing aside Corbyn would be child’s play for a government that could focus on important questions and took project management — an undiscussable subject in SW1 — seriously.

The whole country — the whole world — can see our rotten parties have failed us. The parties ally with the civil service to keep new ideas and people excluded. SW1 has tried to resist the revolutionary implications of the referendum but this resistance has to crack: one way or the other the old ways are doomed. The country voted for profound change in 2016. The Tories didn’t understand this hence, partly, the worst campaign in modern history. This dire Cabinet, doomed to merciless judgement in the history books, is visibly falling: let’s ‘push what is falling’…

For specific proposals on improving the appalling science funding system, see below.

*

The Sam Altman co-founded non-profit, OpenAI, made major progress with its Dota-playing AI last week: follow @gdb for updates. Deep Mind is similarly working on Starcraft. It is a major advance to shift from perfect information games like GO to imperfect strategic games like Dota and Starcraft. If AIs shortly beat the best humans at full versions of such games, then it means they can outperform at least parts of human reasoning in ways that have been assumed to be many years away. As OpenAI says, it is a major step ‘towards advanced AI systems which can handle the complexity and uncertainty of the real world.’

https://blog.openai.com/openai-five-benchmark-results/

RAND paper on how AI affects the chances of nuclear catastrophe:

https://www.rand.org/content/dam/rand/pubs/perspectives/PE200/PE296/RAND_PE296.pdf

The Malicious Use of Artificial Intelligence:

https://img1.wsimg.com/blobby/go/3d82daa4-97fe-4096-9c6b-376b92c619de/downloads/1c6q2kc4v_50335.pdf

Defense Science Board: ‘Summer Study on Autonomy’ (2016):

http://www.acq.osd.mil/dsb/reports/2010s/DSBSS15.pdf

JASON: ‘Perspectives on Research in Artificial Intelligence and Artificial General Intelligence Relevant to DoD’ (2017)

https://fas.org/irp/agency/dod/jason/ai-dod.pdf

Artificial Intelligence and National Security, Greg Allen Taniel Chan (for IARPA):

Artificial Intelligence and National Security – The Belfer Center for …

Some predictions on driverless cars and other automation milestones: http://rodneybrooks.com/my-dated-predictions/

Project Maven (very relevant to politicians/procurement): https://thebulletin.org/project-maven-brings-ai-fight-against-isis11374

Chris Anderson on drones changing business sectors:

https://hbr.org/cover-story/2017/05/drones-go-to-work

On the trend in AI compute and economic sustainability (NB. I think the author is wrong on the Manhattan Project being a good upper bound for what a country will spend in an arms race, US GDP spent on DoD at the height of the Cold War would be a better metric): https://aiimpacts.org/interpreting-ai-compute-trends/

Read this excellent essay on ‘AI Nationalism’ by Ian Hogarth, directly relevant to arms race arguments and UK policy.

Read ‘Intelligence Explosion Microeconomics’ by Yudkowsky.

Read ‘Autonomous technology and the greater human good’ by Omohundro — one of the best things about the dangers of AGI and ideas about safety I’ve seen by one of the most respected academics working in this field.

Existential Risk: Diplomacy and Governance (Future of Humanity Institute, 2017).

If you haven’t you should also read this 1955 essay by von Neumann ‘Can we survive technology?’. It is relevant beyond any specific technology. VN was regarded by the likes of Einstein and Dirac as the smartest person they’d ever met. He was involved in the Manhattan Project, inventing computer science, game theory and much more. This essay explored the essential problem that the scale and speed of technological change suddenly blew up assumptions about political institutions’ ability to cope. Much reads as if it were written yesterday.  ‘For progress there is no cure…’

I blogged on a paper by Judea Pearl a few months ago HERE. He is the leading scholar of causation. He argues that current ML approaches are inherently limited and advance requires giving machines causal reasoning:

‘If we want machines to reason about interventions (“What if we ban cigarettes?”) and introspection (“What if I had finished high school?”), we must invoke causal models. Associations are not enough — and this is a mathematical fact, not opinion.’

I also wrote this recently on science funding which links to a great piece by two young neuroscientists about how post-Brexit Britain should improve science and is also relevant to how the UK could set up an ARPA-like entity to fund AI/ML and other fields:

https://dominiccummings.com/2018/06/08/on-the-referendum-25-how-to-change-science-funding-post-brexit/

 

Effective action #4a: ‘Expertise’ from fighting and physics to economics, politics and government

‘We learn most when we have the most to lose.’ Michael Nielsen, author of the brilliant book Reinventing Discovery.

‘There isn’t one novel thought in all of how Berkshire [Hathaway] is run. It’s all about … exploiting unrecognized simplicities…Warren [Buffett] and I aren’t prodigies.We can’t play chess blindfolded or be concert pianists. But the results are prodigious, because we have a temperamental advantage that more than compensates for a lack of IQ points.’ Charlie Munger,Warren Buffett’s partner.

I’m going to do a series of blogs on the differences between fields dominated by real expertise (like fighting and physics) and fields dominated by bogus expertise (like macroeconomic forecasting, politics/punditry, active fund management).

Fundamental to real expertise is 1) whether the informational structure of the environment is sufficiently regular that it’s possible to make good predictions and 2) does it allow high quality feedback and therefore error-correction. Physics and fighting: Yes. Predicting recessions, forex trading and politics: not so much. I’ll look at studies comparing expert performance in different fields and the superior performance of relatively very simple models over human experts in many fields.

This is useful background to consider a question I spend a lot of time thinking about: how to integrate a) ancient insights and modern case studies about high performance with b) new technology and tools in order to improve the quality of individual, team, and institutional decision-making in politics and government.

I think that fixing the deepest problems of politics and government requires a more general and abstract approach to principles of effective action than is usually considered in political discussion and such an approach could see solutions to specific problems almost magically appear, just as you see happen in a very small number of organisations — e.g Mueller’s Apollo program (man on the moon), PARC (interactive computing), Berkshire Hathaway (most successful investors in history), all of which have delivered what seems almost magical performance because they embody a few simple, powerful, but largely unrecognised principles. There is no ‘solution’ to the fundamental human problem of decision-making amid extreme complexity and uncertainty but we know a) there are ways to do things much better and b) governments mostly ignore them, so there is extremely valuable low-hanging fruit if, but it’s a big if, we can partially overcome the huge meta-problem that governments tend to resist the institutional changes needed to become a learning system.

This blog presents some basic background ideas and examples…

*

Extreme sports: fast feedback = real expertise 

In the 1980s and early 1990s, there was an interesting case study in how useful new knowledge jumped from a tiny isolated group to the general population with big effects on performance in a community. Expertise in Brazilian jiu-jitsu was taken from Brazil to southern California by the Gracie family. There were many sceptics but they vanished rapidly because the Gracies were empiricists. They issued ‘the Gracie challenge’.

All sorts of tough guys, trained in all sorts of ways, were invited to come to their garage/academy in Los Angeles to fight one of the Gracies or their trainees. Very quickly it became obvious that the Gracie training system was revolutionary and they were real experts because they always won. There was very fast and clear feedback on predictions. Gracie jiujitsu quickly jumped from an LA garage to TV. At the televised UFC 1 event in 1993 Royce Gracie defeated everyone and a multi-billion dollar business was born.

People could see how training in this new skill could transform performance. Unarmed combat changed across the world. Disciplines other than jiu jitsu have had to make a choice: either isolate themselves and not compete with jiu jitsu or learn from it. If interested watch the first twenty minutes of this documentary (via professor Steve Hsu, physicist, amateur jiu jitsu practitioner, and predictive genomics expert).

Video: Jiu Jitsu comes to Southern California

Royce Gracie, UFC 1 1993 

Screenshot 2018-05-22 10.41.20

 

Flow, deep in the zone

Another field where there is clear expertise is extreme skiing and snowboarding. One of the leading pioneers, Jeremy Jones, describes how he rides ‘spines’ hurtling down the side of mountains:

‘The snow is so deep you need to use your arms and chest to swim, and your legs to ride. They also collapse underfoot, so you’re riding mini-avalanches and dodging slough slides. Spines have blind rollovers, so you can’t see below. Or to the side. Every time the midline is crossed, it’s a leap into the abyss. Plus, there’s no way to stop and every move is amplified by complicated forces. A tiny hop can easily become a twenty-foot ollie. It’s the absolute edge of chaos. But the easiest way to live in the moment is to put yourself in a situation where there’s no other choice. Spines demand that, they hurl you deep into the zone.’ Emphasis added.

Video: Snowboarder Jeremy Jones

What Jones calls ‘the zone’ is also known as ‘flow‘ — a particular mental state, triggered by environmental cues, that brings greatly enhanced performance. It is the object of study in extreme sports and by the military and intelligence services: for example DARPA is researching whether stimulating the brain can trigger ‘flow’ in snipers.

Flow — or control on ‘the edge of chaos’ where ‘every move is amplified by complicated forces’ — comes from training in which people learn from very rapid feedback between predictions and reality. In ‘flow’, brains very rapidly and accurately process environmental signals and generate hypothetical scenarios/predictions and possible solutions based on experience and training. Jones’s performance is inseparable from developing this fingertip feeling. Similarly, an expert fireman feels the glow of heat on his face in a slightly odd way and runs out of the building just before it collapses without consciously knowing why he did it: his intuition has been trained to learn from feedback and make predictions. Experts operating in ‘flow’ do not follow what is sometimes called the ‘rational model’ of decision-making in which they sequentially interrogate different options — they pattern-match solutions extremely quickly based on experience and intuition.

The video below shows extreme expertise in a state of ‘flow’ with feedback on predictions within milliseconds. This legendary ride is so famous not because of the size of the wave but its odd, and dangerous, nature. If you watch carefully you will see what a true expert in ‘flow’ can do: after committing to the wave Hamilton suddenly realises that unless he reaches back with the opposite hand to normal and drags it against the wall of water behind him, he will get sucked up the wave and might die. (This wave had killed someone a few weeks earlier.) Years of practice and feedback honed the intuition that, when faced with a very dangerous and fast moving problem, almost instantly (few seconds maximum) pattern-matched an innovative solution.

Video: surfer Laird Hamilton in one of the greatest ever rides

 

The faster the feedback cycle, the more likely you are to develop a qualitative improvement in speed that destroys an opponent’s decision-making cycle. If you can reorient yourself faster to the ever-changing environment than your opponent, then you operate inside their ‘OODA loop’ (Observe-Orient-Decide-Act) and the opponent’s performance can quickly degrade and collapse.

This lesson is vital in politics. You can read it in Sun Tzu and see it with Alexander the Great. Everybody can read such lessons and most people will nod along. But it is very hard to apply because most political/government organisations are programmed by their incentives to prioritise seniority, process and prestige over high performance and this slows and degrades decisions. Most organisations don’t do it. Further, political organisations tend to make too slowly those decisions that should be fast and too quickly those decisions that should be slow — they are simultaneously both too sluggish and too impetuous, which closes off favourable branching histories of the future.

Video: Boxer Floyd Mayweather, best fighter of his generation and one of the quickest and best defensive fighters ever

The most extreme example in extreme sports is probably ‘free soloing’ — climbing mountains without ropes where one mistake means instant death. If you want to see an example of genuine expertise and the value of fast feedback then watch Alex Honnold.

Video: Alex Honnold ‘free solos’ El Sendero Luminoso (terrifying)

Music is similar to sport. There is very fast feedback, learning, and a clear hierarchy of expertise.

Video: Glenn Gould playing the Goldberg Variations (slow version)

Our culture treats expertise/high performance in fields like sport and music very differently to maths/science education and politics/government. As Alan Kay observes, music and sport expertise is embedded in the broader culture. Millions of children spend large amounts of time practising hard skills. Attacks on them as ‘elitist’ don’t get the same damaging purchase as in other fields and the public don’t mind about elite selection for sports teams or orchestras.

‘Two ideas about this are that a) these [sport/music] are activities in which the basic act can be seen clearly from the first, and b) are already part of the larger culture. There are levels that can be seen to be inclusive starting with modest skills. I think a very large problem for the learning of both science and math is just how invisible are their processes, especially in schools.’ Kay 

When it comes to maths and science education, the powers-that-be (in America and Britain) try very hard and mostly successfully to ignore the question: where are critical thresholds for valuable skills that develop true expertise. This is even more a problem with the concept of ‘thinking rationally’, for which some basic logic, probability, and understanding of scientific reasoning is a foundation. Discussion of politics and government almost totally ignores the concept of training people to update their opinions in response to new evidence — i.e adapt to feedback. The ‘rationalist community’ — people like Scott Alexander who wrote this fantastic essay (Moloch) about why so much goes wrong, or the recent essays by Eliezer Yudkowsky — are ignored at the apex of power. I will return to the subject of how to create new education and training programmes for elite decision-makers. It is a good time for UK universities to innovate in this field, as places like Stanford are already doing. Instead of training people like Cameron and Adonis to bluff with PPE, we need courses that combine rational thinking with practical training in managing complex projects. We need people who practice really hard making predictions in ways we know work well (cf. Tetlock) then update in response to errors.

*

A more general/abstract approach to reforming government

If we want to get much higher performance in government, then we need to think rigorously about: the selection of people and teams, their education and training, their tools, and the institutions (incentives and so on) that surround and shape them.

Almost all analysis of politics and government considers relatively surface phenomena. For example, the media briefly blasts headlines about Carillion’s collapse or our comical aircraft carriers but there is almost no consideration of the deep reasons for such failures and therefore nothing tends to happen — the media caravan moves on and the officials and ministers keep failing in the same ways. This is why, for example, the predicted abject failure of the traditional Westminster machinery to cope with Brexit negotiations has not led to self-examination and learning but, instead, mostly to a visible determination across both sides of the Brexit divide in SW1 to double down on long-held delusions.

Progress requires attacking the ‘system of systems’ problem at the right ‘level’. Attacking the problems directly — let’s improve policy X and Y, let’s swap ‘incompetent’ A for ‘competent’ B — cannot touch the core problems, particularly the hardest meta-problem that government systems bitterly fight improvement. Solving the explicit surface problems of politics and government is best approached by a more general focus on applying abstract principles of effective action. We need to surround relatively specific problems with a more general approach. Attack at the right level will see specific solutions automatically ‘pop out’ of the system. One of the most powerful simplicities in all conflict (almost always unrecognised) is: ‘winning without fighting is the highest form of war’. If we approach the problem of government performance at the right level of generality then we have a chance to solve specific problems ‘without fighting’ — or, rather, without fighting nearly so much and the fighting will be more fruitful.

This is not a theoretical argument. If you look carefully at ancient texts and modern case studies, you see that applying a small number of very simple, powerful, but largely unrecognised principles (that are very hard for organisations to operationalise) can produce extremely surprising results.

We have no alternative to trying. Without fundamental changes to government, we will lose our hourly game of Russian roulette with technological progress.

‘The combination of physics and politics could render the surface of the earth uninhabitable… [T]he ever accelerating progress of technology and changes in the mode of human life … gives the appearance of approaching some essential singularity in the history of the race beyond which human affairs, as we know them, could not continue.’ John von Neumann

As Steve Hsu says: Pessimism of the Intellect, Optimism of the Will.


Ps. There is an interesting connection between the nature of counterfactual reasoning in the fast-moving world of extreme sports and the theoretical paper I posted yesterday on state-of-the-art AI. The human ability to interrogate stored representations of their environment with counter-factual questions is fundamental to the nature of intelligence and developing expertise in physical and mental skills. It is, for now, absent in machines.

State-of-the-art in AI #1: causality, hypotheticals, and robots with free will & capacity for evil (UPDATED)

Judea Pearl is one of the most important scholars in the field of causal reasoning. His book Causality is the leading textbook in the field.

This blog has two short parts — a paper he wrote a few months ago and an interview he gave a few days ago.

*

He recently wrote a very interesting (to the very limited extent I understand it) short paper about the limits of state-of-the-art AI systems using ‘deep learning’ neural networks — such as the AlphaGo system which recently conquered the game of GO and AlphaZero which blew past centuries of human knowledge of chess in 24 hours — and how these systems could be improved.

The human ability to interrogate stored representations of their environment with counter-factual questions is fundamental and, for now, absent in machines. (All bold added my me.)

‘If we examine the information that drives machine learning today, we find that it is almost entirely statistical. In other words, learning machines improve their performance by optimizing parameters over a stream of sensory inputs received from the environment. It is a slow process, analogous in many respects to the evolutionary survival-of-the-fittest process that explains how species like eagles and snakes have developed superb vision systems over millions of years. It cannot explain however the super-evolutionary process that enabled humans to build eyeglasses and telescopes over barely one thousand years. What humans possessed that other species lacked was a mental representation, a blue-print of their environment which they could manipulate at will to imagine alternative hypothetical environments for planning and learning…

‘[T]he decisive ingredient that gave our homo sapiens ancestors the ability to achieve global dominion, about 40,000 years ago, was their ability to sketch and store a representation of their environment, interrogate that representation, distort it by mental acts of imagination and finally answer “What if?” kind of questions. Examples are interventional questions: “What if I act?” and retrospective or explanatory questions: “What if I had acted differently?” No learning machine in operation today can answer such questions about actions not taken before. Moreover, most learning machines today do not utilize a representation from which such questions can be answered.

‘We postulate that the major impediment to achieving accelerated learning speeds as well as human level performance can be overcome by removing these barriers and equipping learning machines with causal reasoning tools. This postulate would have been speculative twenty years ago, prior to the mathematization of counterfactuals. Not so today. Advances in graphical and structural models have made counterfactuals computationally manageable and thus rendered meta-statistical learning worthy of serious exploration

Figure: the ladder of causation

Screenshot 2018-03-12 11.22.54

‘An extremely useful insight unveiled by the logic of causal reasoning is the existence of a sharp classification of causal information, in terms of the kind of questions that each class is capable of answering. The classification forms a 3-level hierarchy in the sense that questions at level i (i = 1, 2, 3) can only be answered if information from level j (j ≥ i) is available. [See figure]… Counterfactuals are placed at the top of the hierarchy because they subsume interventional and associational questions. If we have a model that can answer counterfactual queries, we can also answer questions about interventions and observations… The translation does not work in the opposite direction… No counterfactual question involving retrospection can be answered from purely interventional information, such as that acquired from controlled experiments; we cannot re-run an experiment on subjects who were treated with a drug and see how they behave had then not given the drug. The hierarchy is therefore directional, with the top level being the most powerful one. Counterfactuals are the building blocks of scientific thinking as well as legal and moral reasoning…

‘This hierarchy, and the formal restrictions it entails, explains why statistics-based machine learning systems are prevented from reasoning about actions, experiments and explanations. It also suggests what external information need to be provided to, or assumed by, a learning system, and in what format, in order to circumvent those restrictions

[He describes his approach to giving machines the ability to reason in more advanced ways (‘intent-specific optimization’) than standard approaches and the success of some experiments on real problems.]

[T]he value of intent-base optimization … contains … the key by which counterfactual information can be extracted out of experiments. The key is to have agents who pause, deliberate, and then act, possibly contrary to their original intent. The ability to record the discrepancy between outcomes resulting from enacting one’s intent and those resulting from acting after a deliberative pause, provides the information that renders counterfactuals estimable. It is this information that enables us to cross the barrier between layer 2 and layer 3 of the causal hierarchy… Every child undergoes experiences where he/she pauses and thinks: Can I do better? If mental records are kept of those experiences, we have experimental semantic to counterfactual thinking in the form of regret sentences “I could have done better.” The practical implications of this new semantics is worth exploring.’

The paper is here: http://web.cs.ucla.edu/~kaoru/theoretical-impediments.pdf.

*

By chance this evening I came across this interview with Pearl in which he discuses some of the ideas above less formally, HERE.

‘The problems that emerged in the early 1980s were of a predictive or diagnostic nature. A doctor looks at a bunch of symptoms from a patient and wants to come up with the probability that the patient has malaria or some other disease. We wanted automatic systems, expert systems, to be able to replace the professional — whether a doctor, or an explorer for minerals, or some other kind of paid expert. So at that point I came up with the idea of doing it probabilistically.

‘Unfortunately, standard probability calculations required exponential space and exponential time. I came up with a scheme called Bayesian networks that required polynomial time and was also quite transparent.

‘[A]s soon as we developed tools that enabled machines to reason with uncertainty, I left the arena to pursue a more challenging task: reasoning with cause and effect.

‘All the machine-learning work that we see today is conducted in diagnostic mode — say, labeling objects as “cat” or “tiger.” They don’t care about intervention; they just want to recognize an object and to predict how it’s going to evolve in time.

‘I felt an apostate when I developed powerful tools for prediction and diagnosis knowing already that this is merely the tip of human intelligence. If we want machines to reason about interventions (“What if we ban cigarettes?”) and introspection (“What if I had finished high school?”), we must invoke causal models. Associations are not enough — and this is a mathematical fact, not opinion.

‘As much as I look into what’s being done with deep learning, I see they’re all stuck there on the level of associations. Curve fitting. That sounds like sacrilege, to say that all the impressive achievements of deep learning amount to just fitting a curve to data. From the point of view of the mathematical hierarchy, no matter how skillfully you manipulate the data and what you read into the data when you manipulate it, it’s still a curve-fitting exercise, albeit complex and nontrivial.

‘I’m very impressed, because we did not expect that so many problems could be solved by pure curve fitting. It turns out they can. But I’m asking about the future — what next? Can you have a robot scientist that would plan an experiment and find new answers to pending scientific questions? That’s the next step. We also want to conduct some communication with a machine that is meaningful, and meaningful means matching our intuition.

‘If a machine does not have a model of reality, you cannot expect the machine to behave intelligently in that reality. The first step, one that will take place in maybe 10 years, is that conceptual models of reality will be programmed by humans. The next step will be that machines will postulate such models on their own and will verify and refine them based on empirical evidence. That is what happened to science; we started with a geocentric model, with circles and epicycles, and ended up with a heliocentric model with its ellipses.

We’re going to have robots with free will, absolutely. We have to understand how to program them and what we gain out of it. For some reason, evolution has found this sensation of free will to be computationally desirable… Evidently, it serves some computational function.

‘I think the first evidence will be if robots start communicating with each other counterfactually, like “You should have done better.” If a team of robots playing soccer starts to communicate in this language, then we’ll know that they have a sensation of free will. “You should have passed me the ball — I was waiting for you and you didn’t!” “You should have” means you could have controlled whatever urges made you do what you did, and you didn’t.

[When will robots be evil?] When it appears that the robot follows the advice of some software components and not others, when the robot ignores the advice of other components that are maintaining norms of behavior that have been programmed into them or are expected to be there on the basis of past learning. And the robot stops following them.’

Please leave links to significant critiques of this paper or work that has developed the ideas in it.

If interested in the pre-history of the computer age and internet, this paper explores it.