Main

January 20, 2023

New music

Nick_Cave_by_Amelia_Troubridge-370.jpgThe news this week that "AI" "wrote" a song "in the style of" Nick Cave (who was scathing about the results) seemed to me about on a par with the news in the 1970s that the self-proclaimed medium Rosemary Brown was able to take dictation of "new works" by long-dead famous composers. In that: neither approach seems likely to break new artistic ground.

In Brown's case, musicologists, psychologists, and skeptics generally converged on the belief that she was channeling only her own subconscious. AI doesn't *have* a subconscious...but it does have historical inputs, just as Brown did. You can say "AI" wrote a set of "song lyrics" if you want, but that "AI" is humans all the way down: people devised the algorithms and wrote the computer code, created the historical archive of songs on which the "AI" was trained, and crafted the prompt that guided the "AI"'s text generation. But "the machine did it by itself" is a better headline.

Meanwhile...

Forty-two years after the first one, I have been recording a new CD (more details later). In the traditional folk world, which is all I know, getting good recordings is typically more about being practiced enough to play accurately while getting the emotional performance you want. It's also generally about very small budgets. And therefore, not coincidentally, a whole lot less about sound effects and multiple overdubs.

These particular 42 years are a long time in recording technology. In 1980, if you wanted to fix a mistake in the best performance you had by editing it in from a different take where the error didn't appear, you had to do it with actual reels of tape, an edit block, a razor blade, splicing tape...and it was generally quicker to rerecord unless the musician had died in the interim. Here in digital 2023, the studio engineer notes the time codes, slices off a bit of sound file, and drops it in. Result! Also: even for traditional folk music, post-production editing has a much bigger role.

Autotune, which has turned many a wavering tone into perfect pitch, was invented in 1997. The first time I heard about it - it alters the pitch of a note without altering the playback speed! - it sounded indistinguishable from magic. How was this possible? It sounded like artificial intelligence - but wasn't.

The big, new thing now, however, *is* "AI" (or what currently passes for it), and it's got nothing to do with outputting phrases. Instead, it's stem splitting - that is, the ability to take a music file that includes multiple instruments and/or voices, and separate out each one so each can be edited separately.

Traditionally, the way you do this sort of thing is you record each instrument and vocal separately, either laying them down one at a time or enclosing each musician/singer into their own soundproof booth, from where they can play together by listening to each other over headphones. For musicians who are used to singing and playing at the same time in live performance, it can be difficult to record separate tracks. But in recording them together, vocal and instrumental tracks tend to bleed into each other - especially when the instrument is something like an autoharp, where the instrument's soundboard is very close to the singer's mouth. Bleed means you can't fix a small vocal or instrumental error without messing up the other track.

With stem splitting, now you can. You run your music file through one of the many services that have sprung up, and suddenly you have two separated tracks to work with. It's being described to me as a "game changer" for recording. Again: sounds indistinguishable from magic.

This explanation makes it sound less glamorous. Vocals and instruments whose frequencies don't overlap can be split out using masking techniques. Where there is overlap, splitting relies on a model that has been trained on human-split tracks and that improves with further training. Still a black box, but now one that sounds like so many other applications of machine learning. Nonetheless, heard in action it's startling: I tried LALAL_AI on a couple of tracks, and the separation seemed perfect.

There are some obvious early applications of this. As the explanation linked above notes, stem splitting enables much finer sampling and remixing. A singer whose voice is failing - or who is unavailable - could nonetheless issue new recordings by laying their old vocal over a new instrumental track. And vice-versa: when, in 2002, Paul Justman wanted to recreate the Funk Brothers' hit-making session work for Standing in the Shadows of Motown, he had to rerecord from scratch to add new singers. Doing that had the benefit of highlighting those musicians' ability and getting them royalties - but it also meant finding replacements for the ones who had died in the intervening decades.

I'm far more impressed by the potential of this AI development than of any chatbot that can put words in a row so they look like lyrics. This is a real thing with real results that will open up a world of new musical possibilities. By contrast, "AI"-written song lyrics rely on humans' ability to conceive meaning where none exists. It's humans all the way up.


Illustrations: Nick Cave in 2013 (by Amanda Troubridge, via Wikimedia).

Wendy M. Grossman is the 2013 winner of the Enigma Award. Her Web site has an extensive archive of her books, articles, and music, and an archive of earlier columns in this series. Stories about the border wars between cyberspace and real life are posted occasionally during the week at the net.wars Pinboard - or follow on Twitter.

September 23, 2022

Insert a human

We Robot - 2022 - boston dynamics.JPGRobots have stopped being robots. This is a good thing.

This is my biggest impression of this year's We Robot conference: we have moved from the yay! robots! of the first year, 2012, through the depressed doldrums of "AI" systems that make the already-vulnerable more vulnerable circa 2018 to this year, when the phrase that kept twanging was "sociotechnical systems". For someone with my dilettantish conference-hopping habit, this seems like the necessary culmination of a long-running trend away from robots as autonomous mobile machines to robots/AI as human-machine partnerships. We Robot has never talked much about robot rights, instead focusing on considering the policy challenges that arise as robots and AI become embedded in our lives. This is realism; as We Robot co-founder Michael Froomkin writes, we're a long, long way from a self-aware and sentient machine.

The framing of sociotechnical systems is a good thing in part because so much of what passes for modern "artificial intelligence" is humans all the way down, as Mary L. Gray and Siddhart Suri documented in their book, Ghost Work. Even the companies that make self-driving cars, which a few years ago were supposed to be filling the streets by now, are admitting that full automation is a long way off. "Admitting" as in consolidating or being investigated for reckless hyping.

If this was the emerging theme, it started with the first discussion, of a paper on humans in the loop, by Margot Kaminski, Nicholson Price, and Rebecca Crootof. Too often, the proposed policy-making proposal for handling problems with decision making systems is to insert a human, a "solution" they called the "MABA-MABA trap", for "Machines Are Better At / Men Are Better At". While obviously humans and machines have differing capabilities - people are creative and flexible, machines don't get bored - just dropping in a human without considering what role that human is going to fill doesn't necessarily take advantage of the best capabilities of either. Hybrid systems are of necessity more complex - this is why cybersecurity keeps getting harder - but policy makers may not take this into account or think clearly about what the human's purpose is going to be.

At this conference in 2016, Madeleine Claire Elish foresaw that the human would become a moral crumple zone or liability sponge, absorbing blame without necessarily being at fault. No one will admit that this is the human's real role - but it seems an apt description of the "safety driver" watching the road, trying to stay alert in case the software driving the car needs backup or the poorly-paid human given a scoring system and tasked with awarding welfare benefits. What matters, as Andrew Selbst said in discussing this paper, is the *loop*, not the human - and that may include humans with invisible control, such as someone who can massage the data they enter into a benefits system in order to help a particularly vulnerable child, or who have wide discretion, such as a judge who is ultimately responsible for parole decisions no matter what the risk assessment system says.

This is not the moment to ask what constitutes a human.

It might be, however, the moment to note the commentator who said that a lot of the problems people are suggesting robots/AI can solve have other, less technological solutions. As they said, if you are putting a pipeline through a community without its consent, is the solution to deploy police drones to protect the pipeline and the people working on it - or is it to put the pipeline somewhere else (or to move to renewables and not have a pipeline at all)? Change the relationship with the community and maybe you can partly disarm the police.

One unwelcome forthcoming issue, discussed in a paper by Kate Darling and Daniella DiPaola is the threat merging automation and social marketing poses to consumer protection. A truly disturbing note came from DiPaola, who investigated manipulation and deception with personal robots and 75 children. The children had three options: no ads, ads allowed only if they are explicitly disclosed to be ads, or advertising through casual conversation. The kids chose casual conversation because they felt it showed the robot *knew* them. They chose this even though they knew the robot was intentionally designed to be a "friend". Oy. In a world where this attitude spreads widely and persists into adulthood, no amount of "media literacy" or learning to identify deception will save us; these programmed emotional relationships will overwhelm all that. As DiPaola said, "The whole premise of robots is building a social relationship. We see over and over again that it works better if it is more deceptive."

There was much more fun to be had - steamboat regulation as a source of lessons for regulating AI (Bhargavi Ganesh and Shannon Vallor), police use of canid robots (Carolin Kemper and Michael Kolain), and - a new topic - planning for the end of life of algorithmic and robot systems (Elin Björling and Laurel Riek). The robots won't care, but the humans will be devastated.

Illustrations: Hanging out at We Robot with Boston Dynamics' "Spot".

Wendy M. Grossman is the 2013 winner of the Enigma Award. Her Web site has an extensive archive of her books, articles, and music, and an archive of earlier columns in this series. Stories about the border wars between cyberspace and real life are posted occasionally during the week at the net.wars Pinboard - or follow on Twitter.

June 17, 2022

Level two

Tesla-crash-NYTimes-370.pngThis week provided two examples of the dangers of believing too much hype about modern-day automated systems and therefore overestimating what they can do.

The first is relatively minor: Google employee Blake Lemoine published his chats with a bot called LaMDA and concluded it was sentient "basd on my religious beliefs". Google put Lemoine on leave and the press ran numerous (many silly) stories. Veterans shrugged and muttered, "ELIZA, 1966".

The second, however...

On Wednesday, the US National Highway Traffic Safety Administration released a report (PDF) studying crashes involving cars under the control of "driver-assist" technologies. Out of 367 such crashes in the nine months after NHTSA began collecting data in July 2021, 273 involved Teslas being piloted by either "full self-driving software" or its precursor, "Tesla Autopilot".

There are important caveats, which NTHSA clearly states. Many contextual details are missing, such as how many of each manufacturer's cars are on the road and the number of miles they've traveled. Some reports may be duplicates; others may be incomplete (private vehicle owners may not file a report) or unverified. Circumstances such as surface and weather conditions, or whether passengers were wearing seat belts, are missing. Manufacturers differ in the type and quantity of crash data they collect. Reports may be unclear about whether the car was equipped with SAE Level 2 Advanced Driver Assistance Systems (ADAS) or SAE Levels 3-5 Automated Driving Systems (ADS). Therefore, NTHSA says, "The Summary Incident Report Data should not be assumed to be statistically representative of all crashes." Still, the Tesla number stands out, far ahead of Honda's 90, which itself is far ahead of the other manufacturers listed.

SAE, ADAS, and ADS refer to the system of levels devised by the Society of Automotive Engineers (now SAE International) in 2016. Level 0 is no automation at all; Level 1 is today's modest semi-automated assistance such as cruise control, lane-keeping, and automatic emergency braking. Level 2, "partial automation", is now: semi-automated steering and speed systems, road edge detection, and emergency braking.

Tesla's Autopilot is SAE Level 2. Level 3 - which may someday include Tesla's Full Self Drive Capability - is where drivers may legitimately begin to focus on things other than the road. In Level 4, most primary driving functions will be automated, and the driver will be off-duty most of the time. Level 5 will be full automation, and the car will likely not even have human-manipulable controls.

Right now, in 2022, we don't even have Level 3, though Tesla CEO Elon Musk keeps promising we're on the verge of it with his company's Full Self-Drive Capability, its arrival always seems to be one to two years away. As long ago as 2015, Musk was promising Teslas would be able to drive themselves while you slept "within three years"; in 2020 he estimated "next year" - and he said it again a month ago. In reality, it's long been clear that cars autonomous enough for humans to check out while on the road are further away than they seemed five years ago, as British transport commentator Christian Wolmar accurately predicted in 2018.

Many warned that Levels 2 and 3 are would be dangerous. The main issue, pointed out by psychologists and behavorial scientists, is that humans get bored watching a computer do stuff. In an emergency, where the car needs the human to take over quickly, said human, whose attention has been elsewhere, will not be ready. In this context it's hard to know how to interpret the weird detail in the NTHSA report that in 16 cases Autopilot disengaged less than a second before the crash.

The NHTSA news comes just a few weeks after a New York Times TV documentary investigation examining a series of Tesla crashes. Some it links to the difficulty of designing software that can distinguish objects across the road - that is, the difference between a truck crossing the road and a bridge. In others, such as the 2018 crash in Mountain View, California, the NTSB found a number of contributing factors, including driver distraction and overconfidence in the technology - "automation complacence", as Robert L. Sumwalt calls it politely.

This should be no surprise. In his 2019 book, Ludicrous, auto industry analyst Edward Niedermeyer mercilessly lays out the gap between the rigorous discipline embraced by the motor industry so it can turn out millions of cars at relatively low margins with very few defects and the manufacturing conditions Niedermeyer observes at Tesla. The high-end, high-performance niche sports cars Tesla began with were, in Niedermeyer's view, perfectly suited to the company's disdain for established industry practice - but not to meeting the demands of a mass market, where affordability and reliability are crucial. In line with Nidermeyer's observations, Bloomberg Intelligence predicts that Volkswagen will take over the lead in electric vehicles by 2024. Niedermeyer argues that because it's not suited to the discipline required to serve the mass market, Tesla's survival as a company depends on these repeated promises of full autonomy. Musk himself even said recently that the company is "worth basically zero" if it can't solve self-driving.

So: financial self-interest meets the danger zone of Level 2 with perceptions of Level 4. I can't imagine anything more dangerous.

Illustrations: One of the Tesla crashes investigated in New York Times Presents.

Wendy M. Grossman is the 2013 winner of the Enigma Award. Her Web site has an extensive archive of her books, articles, and music, and an archive of earlier columns in this series. Stories about the border wars between cyberspace and real life are posted occasionally during the week at the net.wars Pinboard - or follow on Twitter.

April 29, 2022

The abundance of countries

Adam Smith-National Gallery of Scotland-PD.jpgThis week, some updates.

First up is the Court of Justice of the European Union's ruling largely upholding Article 17 of the 2019 Copyright Directive. Article 17, also known as the "upload filter", was last seen leading many to predict it would break the web. Poland challenged the provision, arguing that requiring platforms to check user-provided material for legality infringed the rights to freedom of expression and information.

CJEU dismissed Poland's complaint, and Article 17 stands. However, at a panel convened by Communia, former Pirate Party MEP Felix Reda found the disappointment is outweighed by the court's opinion regarding safeguards, which bans general monitoring, and, Joao Pedro Quantais explained, restrict content removal to material whose infringing nature is obvious.

More than half of EU countries have failed to meet the June 2021 deadline to transpose the directive into national law, and some that have simply copied and pasted the directive's two most contentious articles - Articles 17 and 15 (the "link tax") rather than attempt to resolve the directive's internal contradictions. As Glyn Moody explains at Walled Culture, the directive requires the platforms to both block copyright-infringing content from being uploaded and make sure legal content is not removed. Moody also reports that Finland's attempts at resolution have attracted complaints from the copyright industries, who want the country to make its law more restrictive. Among the other countries that have transposed the directive, Reda believes only Germany's and Austria's interpretations provide safeguards in line with the court's ruling - and Austria's only with some changes.

***

The best response I've seen to the potential sale of Twitter comes from writer Racheline Maltese: who tweeted, "On the Internet, your home will always leave you."

In a discussion sparked by the news, Twitter user Yishan argues that "free speech" isn't what it used to be. In the 1990s version, the threat model was religious conservatives in the US. This isn't entirely true; some feminist groups also sought to censor pornography, and 1980s Internet users had to bypass Usenet hierarchy administrators to create newsgroups for sex and drugs. However, the understanding that abuse and trolling drive people away and chill them into silence definitely took longer to accept as a denial of free speech rights. Today, Yishan writes, *everyone* feels their free speech is under threat from everyone else. And they're likely right.

***

It's also worth noting the early stages of the cybercrime treaty. It's now 20 years since the Convention on Cybercrime was formulated; as of December 2020 65 states have ratified it and four have signed it. The push for a new treaty is coming from countries that either opposed the original or weren't involved in drafting it - Russia in particular, ironically enough. At Human Rights Watch, Deborah Brown warns of risks to fundamental rights: "cybercrime" has no agreed definition and some states want expansion to include "incitement to terrorism" and copyright infringement. In addition, while many states back including human rights protections, detail is lacking. However, we might get some clues from this week's White House declaration for the future of the Internet, which seeks to "reclaim the promise of the Internet" and embed human rights. It's backed by 60 countries - but not China or Russia.

There is general agreement that the vast escalation of cybercrime means better cross-border cooperation is needed, as Summer Walker writes at Foreign Policy. However, she notes that as work progressed in 2021 a number of states already felt excluded from the decision-making process.

The goal is to complete an agreement by early 2024.

***

Finally....20 years ago I wrote (in a piece from the lostweb) about the new opportunities for plagiarism afforded by the Internet. That led to a new industry sector: online services that check each new paper against a database of known material. The services do manage to find previously published text; six days after publication even a free example service rates the first two paragraphs of last week's net.wars as "100% plagiarized". Even so, the concept is flawed, particularly for academics, whose papers have been flagged or rejected for citations, standardized descriptions of experimental methodology, or reused passages describing their own previous work - "self-plagiarism". In some cases, academics have reported on Twitter, the automated systems in use at some journals reject their work before an editor can see it.

Now there's a new twist in this little arms race: rephrasing services that freshen up published material so it will pass muster. The only problem is (of course) that the AI is supremely stupid and poorly educated. Last year, Nature reported on "tortured phrases" that indicated plagiarized research papers, particularly rife in computer science. This week Essex senior lecturer Matt Lodder reported on Twitter his sightings of AI-rephrased material in students' submissions. First clue: "It read oddly." Well, yes. When I ran last week's posting through several of these services, they altered direct quotes (bad journalism), rewrote active sentences into passive ones (bad writing), and changed the meaning (bad editing). In Lodder's student's text, the AI had substituted "graph" for "chart"; in a paper submitted to a friend of his, "the separation of powers" had been rendered as "the sundering of puissances" and Adam Smith's classic had become "The Abundance of Countries". People: when you plagiarize, read what you turn in!


Illustrations: Adam Smith, author of The Wealth of Nations (portrait from the National Gallery of Scotland, via Wikimedia).

Wendy M. Grossman is the 2013 winner of the Enigma Award. Her Web site has an extensive archive of her books, articles, and music, and an archive of earlier columns in this series. Stories about the border wars between cyberspace and real life are posted occasionally during the week at the net.wars Pinboard - or follow on Twitter.

October 1, 2021

Plausible diversions

amazon-astro.pngIf you want to shape a technology, the time to start is before it becomes fixed in the mindset of "'twas ever thus". This was the idea behind the creation of We Robot. At this year's event (see below for links to previous years), one clear example of this principle came from Thomas Krendl Gilbert and Roel I. J. Dobbe, whose study of autonomous vehicles pointed out the way we've privileged cars by coining "jaywalkification". On the blank page in the lawbook, we chose to make it illegal for pedestrians to get in cars'' way.

We Robot's ten years began with enthusiasm, segued through several depressed years of machine learning and AI, and this year has seemingly arrived at a twist on Arthur C. Clark's famous dictum To wit: maybe any technology sufficiently advanced to seem like magic can be well enough understood that we can assign responsibility and liability. You could say it's been ten years of progressively removing robots' glamor.

Something like this was at the heart of the paper by Andrew Selbst, Suresh Venkatasubramanian, and I. Elizabeth Kumar, which uses the computer science staple of abstraction as a model for assigning responsibility for the behavior of complex systems. Weed out debates over the innards - is the system's algorithm unfair, or was the training data biased? - and aim at the main point​: this employer chose this system that produced these results. No one needs to be inside its "black box" if you can understand its boundaries. In one analogy, it's not the manufacturer's fault if a coffee maker fails to produce drinkable coffee from poisoned water and ground acorns; it *is* their fault if the machine turns potable water and ground coffee into toxic sludge. Find the decision points, and ask: how were those decisions made?

Gilbert and Dobbe used two other novel coinages: "moral crumple zoning" (from Madeleine Claire Elish's paper at We Robot 2016) and "rubblization", for altering the world to assist machines. Exhibit A, which exemplifies all three, is the 2018 incident in which an Uber car on autopilot killed a pedestrian in Tempe, Arizona. She was jaywalking; she and the inattentive safety driver were moral crumple zoned; and the rubblized environment prioritized cars.

Part of Gilbert's and Dobbe's complaint was that much discussion of autonomous vehicles focused on the trolley problem, which has little relevance to how either humans or AIs drive cars. It's more useful instead to focus on how autonomous vehicles reshape public space as they begin to proliferate.

This reshaping issue also arose in two other papers, one on smart farming in East Africa by Laura Foster, Katie Szilagyi, Angeline Wairegi, Chidi Oguamanam, and Jeremy de Beer, and one by Annie Brett on the rapid, yet largely overlooked expansion of autonomous vehicles in ocean shipping, exploration, and data collection. In the first case, part of the concern is the extension of colonization by framing precision agriculture and smart farming as more valuable than the local knowledge held by small farmers, the majority of whom are black women, and viewing that knowledge as freely available for appropriation. As in the Western world, where manufacturers like John Deere and Monsanto claim intellectual property rights in seeds and knowledge that formerly belonged to farmers, the arrival of AI alienates local knowledge by stowing it in algorithms, software, sensors, and equipment and makes the plants on which our continued survival depends into inert raw material. Brett, in her paper, highlights the growing gaps in international regulation as the Internet of Things goes maritime and changes what's possible.

A slightly different conflict - between privacy and the need to not be "mis-seen" - lies at the heart of Alice Xiang's discussion of computer vision. Elsewhere, Agathe Balayn and Seda Gürses make a related point in a new EDRi report that warns against relying on technical debiasing tweaks to datasets and algorithms at the expense of seeing the larger social and economic costs of these systems.

In a final example, Marc Canellas studied whole cybernetic systems and finds they create gaps where it's impossible for any plaintiff to prove liability, in part because of the complexity and interdependence inherent in these systems. Canellas proposes that the way forward is to redefine intentional discrimination and apply strict liability. You do not, Cynthia Khoo observed in discussing the paper, have to understand the inner workings of complex technology in order to understand that the system is reproducing the same problems and the same long history if you focus on the outcomes, and not the process - especially if you know the process is rigged to begin with. The wide spread of move fast and break things, Canellas noted, mostly encumbers people who are already vulnerable.

I like this overall approach of stripping away the shiny distraction of new technology and focusing on its results. If, as a friend says, Facebook accurately described setting up an account as "adding a line to our database" instead of "connecting with your friends", who would sign up? Similarly, don't let Amazon get cute about its new "Astro" comprehensive in-home data collector.

Many look at Astro and see instead the science fiction robot butler of decades hence. As Frank Pasquale noted, we tend to overemphasize the far future at the expense of today's decisions. In the same vein, Deborah Raji called robot rights a way of absolving people of their responsibility. Today's greater threat is that gig employers are undermining workers' rights, not whether robots will become sentient overlords. Today's problem is not that one day autonomous vehicles may be everywhere, but that the infrastructure needed to make partly-autonomous vehicles safe will roll over us. Or, as Gilbert put it: don't ask how you want cars to drive; ask how you want cities to work.


Previous years: 2013; 2015; 2016 workshop; 2017; 2018 workshop and conference; 2019 workshop and conference; 2020.

Illustrations: Amazon photo of Astro.

Wendy M. Grossman is the 2013 winner of the Enigma Award. Her Web site has an extensive archive of her books, articles, and music, and an archive of earlier columns in this series. Stories about the border wars between cyberspace and real life are posted occasionally during the week at the net.wars Pinboard - or follow on Twitter.

April 5, 2019

The collaborative hand

Rich Walker-Shadow-2019-04-03.jpgThe futurist Anders Sandberg has often observed that we call it "artificial intelligence" only as long as it doesn't work; after that it's simply "automation". This week, Rich Walker, the managing director of Shadow Robot, said the same thing about robotics. No one calls a self-driving car or a washing machine a robot, for example. Then again, a friend does indeed call the automated tea maker that reliably wakes up every morning before he does "the robot", which suggests we only call things "robots" when we can mock their limitations.

Walker's larger point was robotics, like AI, suffers from confusion between the things people think it can do and the things it can actually do. The gap in AI is so large, that effectively the term now has two meanings, a technological one revolving around the traditional definition of AI, and a political one, which includes the many emerging new technologies - machine learning, computer vision, and so on - that we need to grapple with.

When, last year, we found that Shadow Robot was collaborating on research into care robots it seemed time for a revisit: the band of volunteers I met in 1997 and the tiny business it had grown into in 2009 had clearly reached a new level.

Social care is just one of many areas Shadow is exploring; others include agritech and manufacturing. "Lots are either depending on other pieces of technology that are not ready or available yet or dependent on economics that are not working in our favor yet," Walker says. Social care is an example of the latter; using robots outside of production lines in manufacturing is an example of the former. "It's still effectively a machine vision problem." That is, machine vision is not accurate enough with high enough reliability. A 99.9% level of accuracy means a failure per shift in a car manufacturing facility.

Thumbnail image for R-shadow-walker.jpgGetting to Shadow Robot's present state involved narrowing down the dream founder Richard Greenhill conceived after reading a 1980s computer programming manual: to build a robot that could bring him a cup of tea. The project, then struggling to be taken seriously as it had no funding and Greenhill had no relevant degrees, built the first robot outside Japan that could stand upright and take a step; the Science Museum included it in its 2017 robot exhibition.

Greenhill himself began the winnowing process, focusing on developing a physical robot that could function in human spaces rather than AI and computer vision, reasoning that there were many others who would do that. Greenhill recognized the importance of the hand, but it was Walker who recognized its commercial potential: "To engage with real-world, human-scale tasks you need hands."

The result, Walker says, is, "We build the best robot hand in the world." And, he adds, because several employees have worked on all the hands Shadow has ever built, "We understand all the compromises we've made in the designs, why they're there, and how they could be changed. If someone asks for an extra thumb, we can say why it's difficult but how we could do it."

Meanwhile, the world around Shadow has changed to include specialists in everything else. Computer vision, for example: "It's outside of the set of things we think we should be good at doing, so we want others to do it who are passionate about it," Walker says. "I have no interest in building robot arms, for example. Lots of people do that." And anyway, "It's incredibly hard to do it better than Universal Robots" - which itself became the nucleus of a world-class robotics cluster in the small Danish city of Odense.

Specialization may be the clearest sign that robotics is growing up. Shadow's current model, mounted on a UR arm, sports fingertips developed by SynTouch. With SynTouch and HaptX, Shadow collaborated to create a remote teleoperation system using HaptX gloves in San Francisco to control a robot hand in London following instructions from a businessman in Japan. The reason sounds briefly weird: All Nippon Airways is seeking new markets by moving into avatars and telepresence. It sounds less weird when Walker says ANA first thought of teleportation...and then concluded that telepresence might be more realistic.

Shadow's complement of employees is nearing 40, and they've moved from the undifferentiated north London house they'd worked in since the 1990s, dictated, Walker says, by buying a new milling machine. Getting the previous one in, circa 2007, required taking out the front window and the stairs and building a crane. Walker's increasing business focus reflects the fact that the company's customers are now as often commercial companies as the academic and research institutions that used to form their entire clientele.

For the future, "We want to improve tactile sensing," Walker says. "Touch is really hard to get robots to do well." One aspect they're particularly interested in for teleoperation is understanding intent: when grasping something, does the controlling human want to pinch, twist, hold, or twist it? At the moment, to answer that he imagines "the robot equivalent" of Clippy that asks, "It looks like you're trying to twist the wire. Do you mean to roll it or twist it?" Or even: "It looks like you're trying to defuse a bomb. Do you want to cut the red wire or the black wire?" Well, do ya, punk?


Illustrations: Rich Walker, showing off the latest model, which includes fingertips from HaptX and a robot arm from Universal Robotics; the original humanoid biped, on display at the Science Museum.

Wendy M. Grossman is the 2013 winner of the Enigma Award. Her Web site has an extensive archive of her books, articles, and music, and an archive of earlier columns in this series. Stories about the border wars between cyberspace and real life are posted occasionally during the week at the net.wars Pinboard - or follow on Twitter.

May 25, 2018

Who gets the kidney?

whogetsthekidney.jpg
At first glance, Who should get the kidney? seemed more reasonable and realistic than MIT's Moral Machine.

To recap: about a year ago, MIT ran an experiment, a variation of the old trolley problem, in which it asked visitors in charge of a vehicle about to crash to decide which nearby beings (adults, children, pets) to sacrifice and which to save. Crash!

As we said at the time, people don't think like that. In charge of a car, you react instinctively to save yourself, whoever's in the car with you, and then try to cause the least damage to everything else. Plus, much of the information the Moral Machine imagined - this stick figure is a Nobel prize-winning physicist; this one is a sex offender - just is not available to a car driver in a few seconds and even if it were, it's cognitive overload.

So, the kidney: at this year's We Robot, researchers offered us a series of 20 pairs of kidney recipients and a small selection of factors to consider: age, medical condition, number of dependents, criminal convictions, drinking habits. And you pick. Who gets the kidney?

Part of the idea as presented is that these people have a kidney available to them but it's not a medical match, and therefore some swapping needs to happen to optimize the distribution of kidneys. This part, which made the exercise sound like a problem AI could actually solve, is not really incorporated into the tradeoffs you're asked to make. Shorn of this ornamentation, Who Gets the Kidney? is a simple and straightforward question of whom to save. Or, more precisely, who in future will prove to have deserved to have been given this second chance at life? You are both weighing the value of a human being as expressed through a modest set of known characteristics and trying to predict the future. In this, it is no different from some real-world systems, such as the benefits and criminal justice systems Virginia Eubanks studies in her recent book, Automating Inequality: How High-Tech Tools Profile, Police, and Punish the Poor.

I found, as did the others in our group, that decision fatigue sets in very quickly. In this case, the goal - to use the choices to form like-minded discussion clusters of We Robot attendees - was not life-changing, and many of us took the third option, flipping a coin.

At my table, one woman felt strongly that the whole exercise was wrong; she embraced the principle that all lives are of equal value. Our society often does not treat them that way, and one reason is obvious: most people, put in charge of a kidney allocation system, want things arranged so that if they themselves they will get one.

Instinct isn't always a good guide, either. Many people, used to thinking in terms of protecting children and old people as "they've had their chance at life", automatically opt to give the kidney to the younger person. Granted, I'm 64, and see above paragraph, but even so: as distressing as it is to the parents, a baby can be replaced very quickly with modest effort. It is *very* expensive and time-consuming to replace an 85-year-old. It may even be existentially dangerous, if that 85-year-old is the one holding your society's institutional memory. A friend advises that this is a known principle in population biology.

The more interesting point, to me, was discovering that this exercise really wasn't any more lifelike than the moral machine. It seemed more reasonable because unlike the driver in the crashing car, kidney patients have years of documentation of their illness and there is time for them, their families, and their friends to fill in further background. The people deciding the kidney's destination are much better informed, and in the all-too-familiar scenario of allocating scarce resources. And yet: it's the same conundrum, and in the end how many of us want the machine, rather than a human, to decide whether we live or die?

Someone eventually asked: what if we become able to make an oversupply of kidneys? This only solves the top layer of the problem. Each operation has costs in surgeons' time, medical equipment, nursing care, and hospital infrastructure. Absent a disruptive change in medical technology, it's hard to imagine it will ever be easy to give everyone a kidney who needs one. Say it in food: we actually do grow enough food to supply everyone, but it's not evenly distributed, so in some areas we have massive waste and in others horrible famine (and in some places, both).

Moving to current practice, in a Guardian article Eubanks documents the similar conundrums confronting those struggling to allocate low-income housing, welfare, and other basic needs to poor people in the US in a time of government "austerity". The social workers, policy makers, and data scientists on these jobs have to make decisions, that, like the kidney and driving examples, have life-or-death consequences. In this case, as Eubanks puts it, they decide which get helped among "the most exploited and marginalized people in the United States". The automated systems Eubanks encounters do not lower barriers to programs as promised and, she writes, obscure the political choices that created these social problems in the first place. Automating the response doesn't change those.


Illustrations: Project screenshot.

Wendy M. Grossman is the 2013 winner of the Enigma Award. Her Web site has an extensive archive of her books, articles, and music, and an archive of earlier columns in this series. Stories about the border wars between cyberspace and real life are posted occasionally during the week at the net.wars Pinboard - or follow on Twitter.

March 23, 2018

Aspirational intelligence

2001-hal.png"All commandments are ideals," he said. He - Steven Croft, the Bishop of Oxford - had just finished reading out to the attendees of Westminster Forum's seminar (PDF) his proposed ten commandments for artificial intelligence. He's been thinking about this on our behalf: Croft malware writers not to adopt AI enhancements. Hence the reply.

The first problem is: what counts as AI? Anders Sandberg has quipped that it's only called AI until it starts working, and then it's called automation. Right now, though, to many people "AI" seems to mean "any technology I don't understand".

Croft's commandment number nine seems particularly ironic: this week saw the first pedestrian killed by a self-driving car. Early guesses are that the likely weakest links were the underemployed human backup driver and the vehicle's faulty LIDAR interpretation of a person walking a bicycle. Whatever the jaywalking laws are in Arizona, most of us instinctively believe that in a cage match between a two-ton automobile and an unprotected pedestrian the car is always the one at fault.

Thinking locally, self-driving cars ought to be the most ethics-dominated use of AI, if only because people don't like being killed by machines. Globally, however, you could argue that AI might be better turned to finding the best ways to phase out cars entirely.

We may have better luck at persuading criminal justice systems to either require transparency, fairness, and accountability in machine learning systems that predict recidivism and who can be helped or drop them entirely.

The less-tractable issues with AI are on display in the still-developing Facebook and Cambridge Analytica scandals. You may argue that Facebook is not AI, but the platform certainly uses AI in fraud detection and to determine what we see and decide which of our data parts to use on behalf of advertisers. All on its own, Facebook is a perfect exemplar of all the problems Australian privacy advocate foresaw in 2004 after examining the first social networks. In 2012, Clark wrote, "From its beginnings and onward throughout its life, Facebook and its founder have demonstrated privacy-insensitivity and downright privacy-hostility." The same could be said of other actors throughout the tech industry.

Yonatan Zunger is undoubtedly right when he argues in the Boston Globe that computer science has an ethics crisis. However, just fixing computer scientists isn't enough if we don't fix the business and regulatory environment built on "ask forgiveness, not permission". Matthew Stoll writes in the Atlantic about the decline since the 1970s of American political interest in supporting small, independent players and limiting monopoly power. The tech giants have widely exported this approach; now, the only other government big enough to counter it is the EU.

The meetings I've attended of academic researchers considering ethics issues with respect to big data have demonstrated all the careful thoughtfulness you could wish for. The November 2017 meeting of the Research Institute in Science of Cyber Security provided numerous worked examples in talks from Kat Hadjimatheou at the University of Warwick, C Marc Taylor from the the UK Research Integrity Office, and Paul Iganski the Centre for Research and Evidence on Security Threats (CREST). Their explanations of the decisions they've had to make about the practical applications and cases that have come their way are particularly valuable.

On the industry side, the problem is not just that Facebook has piles of data on all of us but that the feedback loop from us to the company is indirect. Since the Cambridge Analytica scandal broke, some commenters have indicated that being able to do without Facebook is a luxury many can't afford and that in some countries Facebook *is* the internet. That in itself is a global problem.

Croft's is one of at least a dozen efforts to come up with an ethics code for AI. The Open Data Institute has its Data Ethics Canvas framework to help people working with open data identify ethical issues. The IEEE has published some proposed standards (PDF) that focus on various aspects of inclusion - language, cultures, non-Western principles. Before all that, in 2011, Danah Boyd and Kate Crawford penned Six Provocations for Big Data, which included a discussion of the need for transparency, accountability, and consent. The World Economic Forum published its top ten ethical issues in AI in 2016. Also in 2016, a Stanford University Group published a report trying to fend off regulation by saying it was impossible.

If the industry proves to be right and regulation really is impossible, it won't be because of the technology itself but because of the ecosystem that nourishes amoral owners. "Ethics of AI", as badly as we need it, will be meaningless if the necessary large piles of data to train it are all owned by just a few very large organizations and well-financed criminals; it's equivalent to talking about "ethics of agriculture" when all the seeds and land are owned by a child's handful of global players. The pre-emptive antitrust movement of 2018 would find a way to separate ownership of data from ownership of the AI, algorithms, and machine learning systems that work on them.


Illustrations: HAL.

Wendy M. Grossman is the 2013 winner of the Enigma Award. Her Web site has an extensive archive of her books, articles, and music, and an archive of earlier columns in this series. Stories about the border wars between cyberspace and real life are posted occasionally during the week at the net.wars Pinboard - or follow on Twitter.

June 16, 2017

The ghost in the machine

rotated-patrickball-2017.jpgHumans are a problem in decision-making. We have prejudices based on limited experience, received wisdom, weird personal irrationality, and cognitive biases psychologists have documented. Unrecognized emotional mechanisms shield us from seeing our mistakes.

Cue machine learning as the solution du jour. Many have claimed that crunching enough data will deliver unbiased judgements. These days, this notion is being debunked: the data the machines train on and analyze arrives pre-infected, as we created it in the first place, a problem Cathy O'Neil does a fine job of explaining in Weapons of Math Destruction. See also Data & Society and Fairness, Accountability, and Transparency in Machine Learning.

Patrick Ball, founding director of the Human Rights Database Analysis Group, argues, however, that there are underlying worse problems. HRDAG "applies rigorous science to the analysis of human rights violations around the world". It uses machine learning - currently, to locate mass graves in Mexico - but a key element of its work is "multiple systems estimation" to identify overlaps and gaps.

"Every kind of classification system - human or machine - has several kinds of errors it might make," he says. "To frame that in a machine learning context, what kind of error do we want the machine to make?" HRDAG's work on predictive policing shows that "predictive policing" finds patterns in police records, not patterns in occurrence of crime.

Media reports love to rate machine learning's "accuracy", typically implying the percentage of decisions where the machine's "yes" represents a true positive and its "no" means a true negative. Ball argues this is meaningless. In his example, a search engine that scans billions of web pages for "Wendy Grossman" can be accurate to .99999 because the vast supply of pages that don't mention me (true negatives) will swamp the results. The same is true of any machine system trying to find something rare in a giant pile of data - and it gets worse as the pile of data gets bigger, a problem net.wars has often called searching for a needle in a haystack by building bigger haystacks in relation to data retention.

For any automated decision system, you can draw a 2x2 confusion matrix, like this:
ConfusionMatrix.png
"There are lots of ways to understand that confusion matrix, but the least meaningful of those ways is to look at true positives plus true negatives divided by the total number of cases and say that's accuracy," Ball says, "because in most classification problems there's an asymmetry of yes/no answers" - as above. A "94% accurate" model "isn't accurate at all, and you haven't found any true positives because these classifications are so asymmetric." This fact does make life easy for marketers, though: you can improve your "accuracy" just by throwing more irrelevant data at the model. "To lay people, accuracy sounds good, but it actually isn't the measure we need to know."

Unfortunately, there isn't a single measure: "We need to know at least two, and probably four. What we have to ask is, what kind of mistakes are we willing to tolerate?"

In web searches, we can tolerate a few seconds to scan 100 results and ignore the false positives. False negatives - pages missing that we wanted to see - are less acceptable. Machine learning uses "recall" for the fraction of true positives in the set of results, and "precision" for that of true positives in the entire set being searched. The various ways the classifier can be set can be drawn as a curve. Human beings understand a single number better than tradeoffs; reporting accuracy then means picking a spot on the curve as the point to set the classifier. "But it's always going to be ridiculously optimistic because it will include an ocean of true negatives." This is true whether you're looking for 2,000 fraudulent financial transactions in a sea of billions daily, or finding a handful of terrorists in the general population. Recent attackers, from 9/11 to London Bridge 2017, have already been objects of suspicion, but forces rarely have the capacity to examine every such person, and before an attack there may be nothing to find. Retaining all that irrelevant data may, however, help forensic investigation.

Where there are genuine distinguishing variables, the model will find the matches even given extreme asymmetry in the data. "If we're going to report in any serious way, we will come up with lay language around, 'we were trying to identify 100 people in a population of 20,00 and we found 90 of them." Even then, care is needed to be sure you're finding what you think. The classic example here is the the US Army's trial using neural networks to find camouflaged tanks. The classifier fell victim to the coincidence that all the pictures with tanks in them had been taken on sunny days and all the pictures of empty forest on cloudy days. "That's the way bias works," Ball says.

Cathy_O'Neil_at_Google_Cambridge.jpgThe crucial problem is that we can't see the bias. In her book, O'Neil favors creating feedback loops to expose these problems. But these can be expensive and often can't be created - that's why the model was needed.

"A feedback loop may help, but biased predictions are not always wrong - but they're wrong any time you wander into the space of the bias," Ball says. In his example: say you're predicting people's weight given their height. You use one half of a data set to train a model, then plot heights and weights, draw a line, and use its slope and intercept to predict the other half. It works. "And Wired would write the story." Investigating when the model makes errors on new data shows the training data all came from Hong Kong schoolchildren who opted in, a bias we don't spot because getting better data is expensive, and the right answer is unknown.

"So it's dangerous when the system is trained on biased data. It's really, really hard to know when you're wrong." The upshot, Ball says, is that "You can create fair algorithms that nonetheless reproduce unfair social systems because the algorithm is fair only with respect to the training data. It's not fair with respect to the world."


Illustrations: Patrick Ball; confusion matrix (Jackverr); Cathy O'Neil (GRuban).

Wendy M. Grossman is the 2013 winner of the Enigma Award. Her Web site has an extensive archive of her books, articles, and music, and an archive of earlier columns in this series. Stories about the border wars between cyberspace and real life are posted occasionally during the week at the net.wars Pinboard - or follow on Twitter.

October 5, 2012

The doors of probability

Mike Lynch has long been the most interesting UK technology entrepreneur. In 2000, he became Britain's first software billionaire. In 2011 he sold his company, Autonomy, to Hewlett-Packard for $10 billion. A few months ago, Hewlett-Packard let him escape back into the wild of Cambridge. We've been waiting ever since for hints of what he'll do next; on Monday, he showed up at NESTA to talk about his adventures with Wired UK editor David Rowan.

Lynch made his name and his company by understanding that the rule formulated in 1750 by the English vicar and mathematician Thomas Bayes could be applied to getting machines to understand unstructured data. These days, Bayes is an accepted part of the field of statistics, but for a couple of centuries anyone who embraced his ideas would have been unwise to admit it. That began to change in the 1980s, when people began to realize the value of his ideas.

"The work [Bayes] did offered a bridge between two worlds," Lynch said on Monday: the post-Renaissance world of science, and the subjective reality of our daily lives. "It leads to some very strange ideas about the world and what meaning is."

As Sharon Bertsch McGrayne explains in The Theory That Would Not Die, Bayes was offering a solution to the inverse probability problem. You have a pile of encrypted code, or a crashed airplane, or a search query: all of these are effects; your problem is to find the most likely cause. (Yes, I know: to us the search query is the cause and the page of search results if the effect; but consider it from the computer's point of view.) Bayes' idea was to start with a 50/50 random guess and refine it as more data changes the probabilities in one direction or another. When you type "turkey" into a search engine it can't distinguish between the country and the bird; when you add "recipe" you increase the probability that the right answer is instructions on how to cook one.

Note, however, that search engines work on structured data: tags, text content, keywords, and metadata all going into building an index they can run over to find the hits. What Lynch is talking about is the stuff that humans can understand - raw emails, instant messages, video, audio - that until now has stymied the smartest computers.

Most of us don't really like to think in probabilities. We assume every night that the sun will rise in the morning; we call a mug a mug and not "a round display of light and shadow with a hole in it" in case it's really a doughnut. We also don't go into much detail in making most decisions, no matter how much we justify them afterwards with reasoned explanations. Even decisions that are in fact probabilistic - such as those of the electronic line-calling device Hawk-Eye used in tennis and cricket - we prefer to display as though they were infallible. We could, as Cardiff professor Harry Collins argued, take the opportunity to educate people about probability: the on-screen virtual reality animation could include an estimate of the margin for error, or the probability that the system is right (much the way IBM did in displaying Watson's winning Jeopardy answers). But apparently it's more entertaining - and sparks fewer arguments from the players - to pretend there is no fuzz in the answer.

Lynch believes we are just at the beginning of the next phase of computing, in which extracting meaning from all this unstructured data will bring about profound change.

"We're into understanding analog," he said. "Fitting computers to use instead of us to them." In addition, like a lot of the papers and books on algorithms I've been reading recently, he believes we're moving away from the scientific tradition of understanding a process to get an outcome and into taking huge amounts of data about outcomes and from it extracting valid answers. In medicine, for example, that would mean changing from the doctor who examines a patient, asks questions, and tries to understand the cause of what's wrong with them in the interests of suggesting a cure. Instead, why not a black box that says, "Do these things" if the outcome means a cured patient? "Many people think it's heresy, but if the treatment makes the patient better..."

At the beginning, Lynch said, the Autonomy founders thought the company could be worth £2 to £3 million. "That was our idea of massive back then."

Now, with his old Autonomy team, he is looking to invest in new technology companies. The goal, he said, is to find new companies built on fundamental technology whose founders are hungry and strongly believe that they are right - but are still able to listen and learn. The business must scale, requiring little or no human effort to service increased sales. With that recipe he hopes to find the germs of truly large companies - not the put in £10 million sell out at £80 million strategy he sees as most common, but multi-billion pound companies. The key is finding that fundamental technology, something where it's possible to pick a winner.


Wendy M. Grossman's Web site has an extensive archive of her books, articles, and music, and an archive of all the earlier columns in this series.


February 18, 2011

What is hyperbole?

This seems to have been a week for over-excitement. IBM gets an onslaught of wonderful publicity because it built a very large computer that won at the archetypal American TV game, Jeopardy. And Eben Moglen proposes the Freedom box, a more-or-less pocket ("wall wart") computer you can plug in and that will come up, configure itself, and be your Web server/blog host/social network/whatever and will put you and your data beyond the reach of, well, everyone. "You get no spying for free!" he said in his talk outlining the idea for the New York Internet Society.

Now I don't mean to suggest that these are not both exciting ideas and that making them work is/would be an impressive and fine achievement. But seriously? Is "Jeopardy champion" what you thought artificial intelligence would look like? Is a small "wall wart" box what you thought freedom would look like?

To begin with Watson and its artificial buzzer thumb. The reactions display everything that makes us human. The New York Times seems to think AI is solved, although its editors focus, on our ability to anthropomorphize an electronic screen with a smooth, synthesized voice and a swirling logo. (Like HAL, R2D2, and Eliza Doolittle, its status is defined by the reactions of the surrounding humans.)

The Atlantic and Forbes come across as defensive. The LA Times asks: how scared should we be? The San Francisco Chronicle congratulates IBM for suddenly becoming a cool place for the kids to work.

If, that is, they're not busy hacking up Freedom boxes. You could, if you wanted, see the past twenty years of net.wars as a recurring struggle between centralization and distribution. The Long Tail finds value in selling obscure products to meet the eccentric needs of previously ignored niche markets; eBay's value is in aggregating all those buyers and sellers so they can find each other. The Web's usefulness depends on the diversity of its sources and content; search engines aggregate it and us so we can be matched to the stuff we actually want. Web boards distributed us according to niche topics; social networks aggregated us. And so on. As Moglen correctly says, we pay for those aggregators - and for the convenience of closed, mobile gadgets - by allowing them to spy on us.

An early, largely forgotten net.skirmish came around 1991 over the asymmetric broadband design that today is everywhere: a paved highway going to people's homes and a dirt track coming back out. The objection that this design assumed that consumers would not also be creators and producers was largely overcome by the advent of Web hosting farms. But imagine instead that symmetric connections were the norm and everyone hosted their sites and email on their own machines with complete control over who saw what.

This is Moglen's proposal: to recreate the Internet as a decentralized peer-to-peer system. And I thought immediately how much it sounded like...Usenet.

For those who missed the 1990s: invented and implemented in 1979 by three students, Tom Truscott, Jim Ellis, and Steve Bellovin, the whole point of Usenet was that it was a low-cost, decentralized way of distributing news. Once the Internet was established, it became the medium of transmission, but in the beginning computers phoned each other and transferred news files. In the early 1990s, it was the biggest game in town: it was where the Linus Torvalds and Tim Berners-Lee announced their inventions of Linux and the World Wide Web.

It always seemed to me that if "they" - whoever they were going to be - seized control of the Internet we could always start over by rebuilding Usenet as a town square. And this is to some extent what Moglen is proposing: to rebuild the Net as a decentralized network of equal peers. Not really Usenet; instead a decentralized Web like the one we gave up when we all (or almost all) put our Web sites on hosting farms whose owners could be DMCA'd into taking our sites down or subpoena'd into turning over their logs. Freedom boxes are Moglen's response to "free spying with everything".

I don't think there's much doubt that the box he has in mind can be built. The Pogoplug, which offers a personal cloud and a sort of hardware social network, is most of the way there already. And Moglen's argument has merit: that if you control your Web server and the nexus of your social network law enforcement can't just make a secret phone call, they'll need a search warrant to search your home if they want to inspect your data. (On the other hand, seizing your data is as simple as impounding or smashing your wall wart.)

I can see Freedom boxes being a good solution for some situations, but like many things before it they won't scale well to the mass market because they will (like Usenet) attract abuse. In cleaning out old papers this week, I found a 1994 copy of Esther Dyson's Release 1.0 in which she demands a return to the "paradise" of the "accountable Net"; 'twill be ever thus. The problem Watson is up against is similar: it will function well, even engagingly, within the domain it was designed for. Getting it to scale will be a whole 'nother, much more complex problem.

Wendy M. Grossman's Web site has an extensive archive of her books, articles, and music, and an archive of all the earlier columns in this series.


September 1, 2006

The elephant in the dark

Yesterday, August 31, was the actual 50th anniversary of the first artificial intelligence conference, held at Dartmouth in 1956 and recently celebrated with a kind of rerun. John McCarthy, who convened the original conference, spent yesterday giving a talk to a crowd of students at Imperial College, London, on challenges for machine learning, specifically recounting a bit of recent progress working with Stephen Muggleton and Ramon Otero on a puzzle he proposed in 1999.
Here is the puzzle, which expresses the problem of determining an underlying reality from an outward appearance. Most machine learning research, he noted, has concerned the classification of appearance. But this isn't enough for a robot – or a human – to function in the real world. "Robots will have to infer relations between reality and appearance."

One of his examples was John Dalton's work discovering atoms. "Computers need to be able to propose theories," he said – and later modify them according to new information. (Though I note that there are plenty of humans who are unable to do this and who will, despite all evidence and common sense to the opposite, cling desperately to their theory.)

Human common sense reasons in terms of the realities. Some research suggests, for example, that babies are born with some understanding of the permanence of objects – that is, that when an object is hidden by a screen and reappears it is the same object.

Take, as McCarthy did, the simple (for a human) problem of identifying objects without being able to see them; his example was reaching into your pocket and correctly identifying and pulling out your Swiss Army knife (assuming you live in a country where it's legal to carry one). Or identifying the coin you want from a collection of similar coins. You have some idea of what the knife looks and feels like, and you choose the item by its texture and what you can feel of the shape. McCarthy also cited an informal experiment in which people were asked to draw a statuette hidden in a paper bag – they could reach into the paper bag to feel the statue. People can actually do this with little difference than if they can see the object.

But, he said, "You never form an image of the contents of the pocket as a whole. You might form a list." He has, he said, been trying to get Stanford to make a robotic pickpocket.

You can, of course, have a long argument about whether there is such a thing as any kind of objective reality. I've been reading a lot of Philip K. Dick lately, and he had robots that were indistinguishable from humans, even to themselves; yet in Dick's work reality is a fluid, subjective concept that can be disrupted and turned back on itself at any time. You can't trust reality.

But even if you – or philosophers in general – reject the notion of "reality" as a fundamental concept, "You may still accept the notion of relative reality for the design and debugging of robots." Seems a practical approach.
But the more important aspect may be the amount of pre-existing knowledge. "The common view," he said, "is that a computer should solve everything from scratch." His own view is that it's best to provide computers with "suitably formalized" common sense concepts – and that formalizing context is a necessary step.

For example: when you reach into your pocket you have some idea of the contents are likely to be. Partly, of course, because you put them there. But you could make a reasonable guess even about other people's pockets because you have some idea of the usual size of pockets and the kinds of things people are likely to put in them. We often call that "common sense", but a lot of common sense is experience. Other concepts have been built into human and most animal infants through evolution.

Although McCarthy never mentioned it, that puzzle and these other examples all remind me of the story of the elephant and the blind men, which I first came across in the writings of Idries Shah, who attributed it to the Persian poet Rumi. Depending which piece of the elephant a blind man got hold of, he diagnosed the object as a fan (ear), pillar (leg), hose (trunk), or throne (back). It seems to me a useful analogy to explain why, 50 years on, human-level artificial intelligence still seems so far off. Computers don't have our physical advantages in interacting with the world.

An amusing sidelight that seemed to reinforce that point. After the talk, there was some discussion of building the three-dimensional reality behind McCarthy's puzzle. The longer it went on, the more confused I got about what the others thought they were building; they insisted there was no difficulty in getting around the construction problem I had, which was how to make the underlying arcs turn one and only one stop in each direction. How do you make it stop? I asked. Turns out: they were building it mentally with Meccano. I was using cardboard circles with a hole and a fastener in the middle, and marking pens. When I was a kid, girls didn't have Meccano. Though, I tell you, I'm going to get some *now*.

Wendy M. Grossman’s Web site has an extensive archive of her books, articles, and music, and an archive of all the earlier columns in this series. Readers are welcome to post here, at net.wars home, at her , or by email to netwars@skeptic.demon.co.uk (but please turn off HTML).