‘The discourse is unhinged’: how the media gets AI wrong
theguardian.com
theguardian.com
We see an article, click a link, by a paper, but we are never committing to more than a second or two of attention at a time. The headline makes you read the byline. The byline gets you reading the second sentence. At every step, a journalist loses most of her readers. The winning strategy is to increase "story," play to existing opinions, be sensational, use bait... There are flavours and degrees, but by and large writing like this is n inevitability of the medium, of journalism.
If you are reporting on programs that "developed a type of machine-English patois to communicate between themselves.." you are reporting on whether or not the skynet singularity is coming, almost certainly.
If we want something different, I think we'll need a medium change.
Personally, I was hoping e-readers would catalyze a new medium subtype: 30-100 page mini books. A lot of "news cycle.." the drip, drip, sensation of the day journalism just isn't a good way of understanding anything.
What happened between North Korea and the US this year? What's been happening with the Syrian war? Brexit as of July 2018...? I would be very happy to exchange the daily/weekly news bulletins for monthly/quarterly mini books.
Nautilus is another mag that I've been happy with for more science-oriented stuff. They put out monthly issues with a cohesive theme.
Instant mass media has shown us that people want to consume media about other real people -- that they admire in some fashion -- going through life. Not reporting, explanation, facts, or science.
So reporters have responded with tweets and instant feedback on anything that's happening. Many times even when nothing is happening, they know they have to put out some kind of personal emotional response to keep the audience.
Newspaper and in-depth reporting worked for the most part because nobody really cared about the author. They cared about the material. That's flipped upside down now. The material doesn't matter as much as the author. Even the reall good occasional long-form material you might find is all geared to get you into a personal, minute-by-minute "relationship" with the author. To become part of the tribe.
I wouldn't put it down to some moral failing of humanity. The problem is more of a systemic one than a fundamental one.. I think.
The way we use our phones/computers, our time and such... it adds up to a system where decisions are instant and reflexive. It adds up to a medium, with all its trappings.
Yet a large deal of my millennial colleagues are the most progressively scientific, rejecting religion, embracing global warming legislation, very strongly against anti-vax movements. How many people really read the newspaper back in the day? Every period piece shows the adult male reading the paper. No one else. Do you think 26 year olds back then just read the news all the time and now it's different and no one does?
> So reporters have responded with tweets and instant feedback on anything that's happening.
It has more to do with managers and companies pushing reporters and journalists to feed page views. That's it. It's not reporters responding to the end user. It's reporters responding to the fact they don't have good jobs anymore and have to generate ad revenue.
Anyway, that was my only writing that really went viral, and it was not about a person. It was about a corporation, a device, and a set of security policies.
Anyway, I've written incendiary things, informative things, and I guess some bad things.
I now understand the basic math behind "stochastic gradient descent." Maybe after some practice, I can help.
Of course the New Yorker has restaurant reviews, events, and other sections that are New York-specific, but these are remarkably easy to skip regardless of medium.
https://www.newyorker.com/magazine/2017/10/23/welcoming-our-... (title is in jest) https://www.newyorker.com/magazine/2018/06/25/the-reputation... (astroturfing and politics) https://www.newyorker.com/magazine/2018/07/23/how-e-commerce...
Originally described in Crichton's "Why Speculate?" speech, the Gell-Mann amnesia effect labels a commonly observed problem in modern media, where one will believe everything they read from a journalist even after they come across an article about something they know well that is completely incorrect.
The conclusions found and perspectives portrayed by the author are entirely erroneous, often times flipping the cause and the effect. Crichton notes these as "wet streets cause rain" stories.
In short, most eloquently put by Thomas L. McDonald, the Gell-Mann amnesia effect defines the idea that "I believe everything the media tells me except for anything for which I have direct personal knowledge, which they always get wrong."’
It's really hard - you are told or assume, you internalize it, build on top of it, and then when you realize that you've built on falsehood it's easy to remove that piece without understanding what load it was bearing in your tower of thought. I've found myself doing this all the time when doing designs - I've had to train myself to go back frequently and retrace the logic I followed, being careful not to fall into the same pitfalls and re-introduce the same flaws -that the tower still holds up to inspection when those components were never introduced.
Most people were't exactly reading [insert name of reputable publication] for their daily news, it was low quality tabloids with a helping of TV and radio pundits.
Outside of NYC, the US has basically no daily tabloids.
> Most people were't exactly reading [insert name of reputable publication] for their daily news
Most people in the US that were reading any daily news were probably reading the dominant local newspaper or one of the latter reputable metropolitan or national papers, because that covers pretty much all the daily written news outlets they would have had access to.
Unfortunately, it seems coverage of journalism and its decline and what not is very US centric, which likely confuses the hell out of people from other countries where their media outlets got worse much earlier on/didn't have a golden age in general. Kind of like how US media outlets and YouTubers assume the Great Video Game Crash was some worldwide thing, whereas it really only affected the US market and left the European and Japanese ones almost untouched.
It may also explain some of the comments on local journalism, since that stuff over here in the UK isn't exactly fantastic itself. Maybe there's more ambition in local US newspapers than in UK ones, in that they actually did some journalism at some point rather than merely covering what local businesses were up to or what not.
Huh, guess there's a cultural divide in how the news is presented too.
Our disagreement could be a dialect thing though, when I say tabloid I mean a printed paper that consists entirely of stories about UFOs and lurid celebrity gossip. I believe in other parts of the world tabloid has something to do with the format of the paper and the content is considered somewhat but not totally disreputable.
I'm not sure which meaning the other posters are using.
Both common local tabloids (which are often have basic journalistic standards, though they frequently have...interesting...editorial viewpoints) and the supermarket tabloids like the Enquirer are weeklies.
Dieticians lost trust because they kept changing their advice about what to eat and clearly aren't reliable either. That's unrelated to newspapers.
However, even if I think someone is lying 99.99% of the time when they say a Tiger in a Top hat is behind the door that still raises the odds immensely.
It would be more interesting if we could understand what happens in those black boxes, and synthesize it. All we see is showcase projects and services, but never something you can run on a client computer.
It really looks like machine learning is just brute forcing problems until you have a partial solution key to a problem, without understanding how it works internally. Granted, it's progress, but why isn't it possible to use the data of a trained ML network for further analysis? I see many models of learning, but not a lot of analysis or simplification of the resulting data.
I would have thought AI would help science understand what intelligence is, but obviously it's always money first, science later. You often see a lot of tools and models, but not a lot of good insights.
But saying its just bruteforce is missing the larger point, much of evolution is bruteforce.
Of course its money or power first thats how you pay for these things to begin with, with utillity we wil get more and more conscious ai but its obviously going to take time but a blip compared the time evolution have used.
And we don't have to, unless you're insisting on something that might be referred to more aptly as "Artificial Human Intelligence". Generally speaking though, in the AI field, there's no particular belief that AI must work the same way human intelligence works. If AI research leads to discoveries that help us better understand human intelligence, that's a nice perk, but nobody treats that as the ultimate goal.
That said, of course it makes sense to try and model after human intelligence to the extent we can, since we are currently the best example of intelligence we have available to use as a template.
It is ignorant conceit to think we will just magically stumble upon the answer with zeitgeist machine learning techniques.
Who is out there suggesting that we will "just magically stumble upon the answer with zeitgeist machine learning techniques"? From where I sit, it seems that most contemporary researchers who are focused on Deep Learning / Deep Reinforcement Learning / etc. are not talking much at all about "Artificial Intelligence" in the general sense. And the people out there talking specifically about "Artificial General Intelligence" (Ben Goertzel, Marcus Hutter, Pei Wang, etc.) certainly aren't claiming that all we need is the currently faddish ML techniques. See, for reference:
http://agi-conf.org/2017/?page_id=20
http://agi-conf.org/2016/schedule/
http://agi-conf.org/2015/schedule/
etc..
I remember watching Andrew Ng's course on ML and he was often talking about "AI dream".
I think that the goal of AI is to build machines that are progressively more intelligent. To build those machine you have to build an artificial form of intelligence, to further analyze and research what intelligence really means.
I don't think humans are really able to visualize a form of intelligence other than human or mammal/earthly life forms. Our intelligence comes from an evolutionary need to figure out things in order to survive, but it's an earthly version of how we evolved. With that said, I think we can already say that our definition of intelligence will always biased because we put our human intelligence on a pedestal, and worst than that, we won't be able to detect other forms of intelligence because of those reasons.
Hypes are not necessarily wrong. I also think in the end you won't get AI winters anymore. If it is advanced enough AI will be used and you won't get the genie back in the bottle.
The complaints in the article seem to apply to crapy sensationalist journalism in general and there is much more danger from that applying to immigrants, politics war and the like. If the Sun prints some nonsense Facebook's bots does it really matter? Nonsense about the enemies WMDs on the other hand can costs thousands of lives and billions of dollars.
The problem is two fold a) Lawyers completely underestimated the decision making abilities of software even without AI b) People who are not lawyers that completely overestimate the complexity of legal resolution.
(a) Happens because the software we lawyers use is basically... well... crap and most lawyers are clueless when it comes to technology, even ones that specialize on it. (b) Happens because of TV and Movies have created this fantasy legal world where lawyer, especially expensive ones can prove that they are elephants because of their virtue to win arguments.
Surprisingly and this was also a surprise to me when I started to study law, law and coding is much closer than people think because there is a lot more logic than there is tv drama in a real court room.
Not only AI should not have problem resolving legal issues , judging from its current achievements, but it can help with the most valuable skill for a lawyer which is pattern recognition. Our profession bombards us daily with tons of data that is extremely hard to organize and keep track of. This applies more on civil than criminal law , but most of law is civil law anyway (economical issues), and this data is documents used as evidence, case law (court cases that have created a precedence) and of course legislation.
Also there is no much of an option really, AI is pretty much unavoidable because the ever evolving immense complexity of modern society has made legal resolution so complex that court cases take up to decades to be fully resolved which of course is not a viable solution.
An example is the IT law which has been a huge suffering for courts and legislators to keep up with its rapid evolution in a profession where court cases and legislations take decades to move forward. In IT decades are in legal terms , centuries.
AI will replace lawyers , for that I have no doubt, cause law is a dying profession anyway for the reason I explained above. Obviously lawyers will still be around for a long long time but yes AI will fundamentally change the profession. The profession is in desperate need of modernization as it has barely evolved that last few thousands of years.
The problem was never what AI can do, it can do amazing thing, the problem is supply and demand. AI is a field of huge demand and minimal supply but then this is a problem that has rampaged the coding profession which is why freelancer coders make more money than lawyers.
Instead of turning law into a computer game where the company with the most TPUs wins, why not simplify or reform the system so that human beings can understand it? Isn't the point of the legal system to resolve conflicts between people, not computers?
I'm worried that we're driving off a cliff of incomprehensibility, where things happen but nobody can understand why. Or even if they do, they don't have the authority to override the system which is making the decision. Reforming the system outright is always too risky -- it's been working OK so far, right? But what happens when it stops working? How can you fix a system you don't understand?
You are absolutely correct though that we are indeed driving off a cliff of incomprehensibility , I cannot count the times when I have caught , including myself, lawyers and coder not understanding even basic concepts like OOP or legal responsibility under the influence of drugs and alcohol. It's not that the concepts are hard to understand but they are so numerous it becomes so easy to get lose track of where you are , where you were and most importantly where you are going.
When I started coding back in the end of 80s coding in Assembly was not that hard. After 30 years of coding I decided to go back to Assembly I just got blown away how much more complex it became, though obviously not surprised, and of course I discovered that even Assembly coder mostly use C libraries cause well, otherwise it gets insane really fast.
My solution to this may sound insane but in life I have learned that when I have a crazy idea ,usually, I end up being correct. I do believe that AI wont replace us but rather augment us. I am not talking about cyborgs , singularity and these nonsense I am talking about software that helps you navigate through the chaos of information. And when I say AI I mean it in the most vague way possible obviously the technology will change in the future in so many ways.
The only viable solution for the human being is to either find news ways to take advantage of the potential of his own intelligence or augment himself in some way.
Afterall its not a secret that AI is already used to construct AI and this opens the door to a ton of potential.Afterall is it not coding all about automated decisions making ? It's not as if we have not being trusting automated machines for thousands of years. But nonetheless humans are terrified of technology. The marvel of the human condition.
Look at GDPR. It's impossible to know what it really means. Huge efforts are put into action with no idea of whether it will be considered good enough or not. That's not a problem you can fix with ai. You need better law (in this case, no law would be better)
I've a friend who abandoned her engineering career at 30 to become an IP lawyer (she is incredibly senior in her firm now) who says the law is just a program that you run on a judge.
Generally I agree with many of your points, especially that tech w i l l transform law. However, I'm not so sure how soon that will be.
I also studied law and I am a coder and I see similarities between both. However, I think it will be a long time until a computer will truly replace a lawyer. There is one thing that I think is especially hard: Not all rules are clear cut (like "If the company has a turnover >5 Million, x applies"). Many rules have an element of judgement (in Germany we call it "unbestimmte Rechtsbegriffe", which roughly translates as vague legal concepts) that require context knowledge, which is something that has not yet been achieved by machines (or at least I wouldn't know about it). While logic is undoubtly an important part of law, judgement (especially one that requires broad context knowledge) and interpretation it another one, and from what I experienced, the latter can unfortunately often not to be solved with the rules of logic and thus it will be hard to bring a computer to do it.
The law (at least, litigation) involves marshalling precedent and facts to achieve persuasion. Circa 1990s tools like WestLaw are still the gold standard for legal research. And technological tools for collecting, organizing, and synthesizing facts are basically non-existent.
There is, in fact, market validation of the idea that legal technology is not particularly useful. You might argue that defense lawyers have disincentives to adopt technology that would reduce billable hours, but what about the other side of the v.? Plaintiffs' lawyers working on a contingency basis have enormous incentives to minimize effort invested in each case. Yet they don't.
The US has a vast oversupply of lawyers (sure, star litigators make lots of money because they are in short supply, but they aren't the people whose labor legal technology aims to replace), which makes the value of reducing legal gruntwork low, plus to replace legal grunts using WestLaw or similar tools, you've got to either build on top of one the handful of such tools or duplicate it's corpus of annotated data and the massive infrastructure dedicated to keeping it current before you even get to the novel part of your tool, or you won't have anything usable.
So, it's a super high barrier to entry for a market where you are competing with cheap, abundant human labor.
AI winter will be caused, once again, by the failure of the technology to do what the researchers and practitioners claim it can do. This time, tragically, with fatalities.
I don't think that is necessarily a bad thing. I did my masters' thesis in 2004 using evolutionary algorithms. No ordinary person would talk to me about that, even worse, I was easily labeled an outsider and a nerd. Today, people having AI skills are the cool kids on the block. That's the good side of the hype: it's bringing all these nerd topics into the mainstream.
Meanwhile, in reality, they went from 'This is an idea' -> 'Hey, we can solve a few interesting problems with this' and then split. The people who know how to develop technology built some interesting limited applications, and then scaled into more powerful stuff. The people who don't really know anything about technology will have bought into overhyped projects, they'll never go anywhere and get canned.
The second one is a pretty reasonable way for technology to evolve, the problem is that people aren't looking at what's happening, they're looking at 'influencers' and 'commentators' who are more interested in selling their opinions than talking about the technology.
I think it's a mistake to conflate the two, they are unrelated and orthogonal to each other. AI has a whole raft of problems waiting to be solved as soon as it develops enough but blockchain is very much a solution in search of a problem that only it can solve. Every proposed application of it eventually requires a "trusted third party" and then the whole house of cards comes tumbling down. In a former role I sat in on many fintech pitches where painfully earnest people from outside the industry proposed solutions to problems that they only imagined existed... I mean at the level of claiming that only blockchain can do something that actually the Medicis were doing in the 15th century...
AI is as dangerous as we decide it can be, for example we can create gun with camera and shoot to people based on their looks. Law and common sense should not allow that, that's criminal activity.
I think AI should grow, and take over boring and repetitive tasks. This will free up many people form dull jobs and new jobs on top of that will be created, i.e. jobs to tune and organize AI units of computations and talk to other people about results.
It’s more a different kind of intelligence which happens to be able to do some of the same tasks. It already far exceeds human ability, at the tasks that particular flavor of intelligence is good at. We don’t need to be looking at things people do and asking how deep neural nets can do those as well, we need to be looking at deep neural nets and asking which tasks they’re uniquely capable of doing.
I remember seeing a post here a while ago, talking about essentially this process. The poster was disappointed that AI researchers tend to get to the point where a new approach starts bearing fruit, and then get sidetracked finding applications for the new approach and forget about the search for 'real' AI. IIRC this was also suggested as an explanation for the "once we can do it it doesn't count as AI" phenomenon.