AI Company Accused of Using Humans to Fake Its AI
sixthtone.com
sixthtone.com
The actual approach is a perfectly reasonable business strategy, because (A) it's really hard to work on machine learning without a lot of data, and it's often hard to get that data without actual users using the product in roughly the way it's supposed to work, and (B) the company can prove there's commercially significant demand for the user experience that they hope to be able to eventually do with some algorithms/AI technology, and thus can attract investor capital for making the thing actually work (potentially years later). In some cases, the human-powered product might actually be really good, and the economics might work in.
But it really sucks that the general hype around AI has generated a whole category of "AI companies" that don't actually use any AI technology. The dishonesty is really poor (I have a friend who's actually a machine learning expert start a job at one of these companies, and only then discover that the company did not use machine learning and no technical plans to do in the next year and actually just wanted another web developer... which obviously resulted in my friend quitting).
That said, I expect the trend to continue as long as companies get lots of extra press and investor attention for claiming to be an AI company.
Only if the AI is going to work for sure.
There was a UK company, SpinVox, they were big 10 years ago, they did voice recognition of your phone calls and sent you a text. They raised huge, like 100M pounds. Granted not AI.
Turns out they were doing most of their transcription manually with hopes of going automated.
It was a huge blow up, it boiled to fraud.
This is close to Elizabeth Holmes territory kind of storyline.
So it's possible to do as you say, but there needs to be transparency.
Also, AI is an approach, not a market ... so I think they're going to end up like transistors or some other kind of thing, the knowledge will be embedded within companies, and not so many pure AI companies.
However you have to be honest to investors, which I assume they weren't and thus commited fraud.
On the other hand, I could definitely see this being used in mock-ups, just as you might populate some UI with "fake" data to illustrate proofs-of-concept. The transition between mock-up and working application might be fuzzy, and it's a lack of honesty/transparency in this transition that leads to murky territory like you say.
Hard to rise above the background "AI noise" of well-funded marketing.
And people are already fed up with all the AI marketing bullshit; "ebbing tide sinks all boats".
One way is to offer real-time results via APIs or demos, which cannot be faked with humans due to sheer latency / volume. Another is to be active in the ML community (open source etc) to build a reputation. But there the audience is different (technical), and may not overlap with the business case much.
This will pass when funding gets tight again and/or the economy enters a recession. A very sharp recession is actually good for the valley because it helps clean out the fakers; it's no different than the first cold snap in the winter helping clear away certain pests.
(One takeaway from this is that if one wants to found a startup, and one is not a fraud, it's actually better in almost every way to do so in the first part of a recession: lower salaries, less grass-is-greener attrition, lower real estate costs, etc.)
For the customers, they need to be sure that the company is actually capable of solving the problem they are claiming to solve (before signing some sort of long term contract that doesn't actually make the claims the sales people put forward.) The customer's evaluation, however is distinct from that of an investors.
It is the job of an investor to determine if the startup is actually capable of having reasonable odds of successfully pursuing whatever they are claiming to be doing. If an investor is not, then for an actual startup doing machine learning, probably this is not they type of investor you want.
During the ICO pump and dump cycle one thing stuck out to me - nearly all of the ICOs were being run by people without any cryptography experience. Most had little to no software development experience. I'd suggest investors take their first step in evaluating "AI" startups similarly.
I get that on a purely technical level, but really... who cares about the implementation? If a company is solving problems it really doesn't matter, and if they're solving it better using some mix of ML and human intervention... again; who cares?
Trying to create a startup that can't at least compete with the existing solutions isn't a very good business strategy in my opinion.
That's what academia is for. Research that is still too far away from being profitable but that has it's merits in the long run.
I also happened to know one of the people they later hired, and right up to that point they were seriously enraged by this company's deceptive nonsense. The company didn't describe his job quite in the way the investor did, but yeah, that's pretty much what the job was. The compensation was very good.
In a sense people with strong domain knowledge and experience could view these companies as basically a way to outsource the two hardest startup filter functions (finding product-market fit, establishing timing) while at the same time preserving their option to join as highly-compensated ICs and leads.
Except for keeping this secret. Than it's called lying.
One of the biggest shocks I had going from childhood to adulthood was the sudden flip between people telling you that you have to be honest and the people now telling you to stop being honest. And the people who had previously told you that you must be honest, they tell you "oh yes, that's work", when you go back to ask them what you should do given you are being instructed by the people who pay your wages that you now have to lie.
I regret my parents never taught me we're living in softly adversarial world. People are not going to kill you, but businesses are out to fleece you.
"The secret of life is honesty and fair dealing. If you can fake that, you've got it made." - Groucho Marx
See also; the appointed Liars of the Zoon tribe of Terry Pratchett's Discworld.
And the apparent message from that debacle isn't; "be honest", rather it is; "being dishonest nets billions, try not to get caught though".
It is, as long as you have reason to assume that you will be ABLE to automate the tasks. For many of these tasks, there is NO reason to assume that.
My feeling with AI tasks is that either they work reasonably well after a short time or they will never work well. (At least not until the next AI revolution.)
If a company doesn't have a good enough AI system for their task after a couple of months of toying with it, I will be Very sceptical that they will ever have one.
Anything customer-facing and low-latency, like human language translation, image classification, or speech recognition doesn’t seem viable at even modest scale. Anything involving large amounts of data, like analyzing financial data or server logs, seems potentially doable though still unlikely.
In general, it seems like the main advantage of AI products today is either latency of output or size of input, both of which seem difficult to fake by using humans. If a customer has a small amount of input data and wants a small number of results that are indistinguishable from what a trained human would produce, surely they would have no misconceptions about the current state of AI, and would just hire the human to produce it.
Of course a company marketing an online service that sells AI-generated musical compositions could cheat by hiring experienced human composers. But if there’s any money in such a service at all, I expect the customers would want to be able to generate a huge number of compositions, something that humans would probably not be able to do at any appreciable scale. If the customer wanted a single composition that sounded like it was composed by an expert human, I’m pretty sure they would just hire the human.
One thing I certainly could believe is common is human sanity checks on AI output, but that doesn’t strike me as bad or dishonest unless the seller is specifically claiming (for some reason) that no humans are involved. I could also definitely believe that companies fake AI demos for potential investors or future customers.
What types of AI products would be best-suited for humans to be able to viably fake?
When simultaneous translation becomes a product, its makers will stop claiming they do AI, and start claiming "we used AI to create this."
This is only a problem when credulous and powerful investors (a dangerous combination) get caught up in the hype.
Delusional; if it's really hard to get it to work then it's really hard to get it to work; investors should not believe "if only we get the data"
>(B) the company can prove there's commercially significant demand for the user experience that they hope to be able to eventually do with some algorithms/AI technology, and thus can attract investor capital for making the thing actually work (potentially years later). In some cases, the human-powered product might actually be really good, and the economics might work in.
Delusional - if the economics work then everyone would be doing it, there's no money in it, the investors should only invest if they understand that rooms full of transcribers will be needed - with all the management, recruitment and security issues that brings.
This is what the last wave of voicemail transcription companies tried; they all went bust with scores of millions of loss. Be warned.
Not having sufficient data to train a machine learning algorithm is a legitimate concern. Statistical machine learning algorithms are very bad at generalising, so they need lots and lots of examples of every possible variation of a concept, in order to model the concept well.
That is why everyone's going gaga over "the new electricity" (by Andrew Ng's saying). So in some cases, yes, there are applications that are impossible to develop without "the data".
Human reading a 1GB logfile looking for key words/phrases is a lot more doable.
The closest I’ve seen to anything useful was Watson, and all it really offered was stuff we’ve been doing for years. I mean, we have 5 people working with data and building models, and Watson could maybe automate some of that, but you’d still need people to qualify what it came up with and build the things it didn’t.
We’ve run proof of concepts on stuff like traffic, and a small startup with a relationship with a university, build a route prediction thing that would make it easier to avoid jams. It worked, but it frankly didn’t work better than what our traffic department was already doing, using simple common sense. So we diverted the project to work on traffic light control for maximum flow, and it turned out that the settings our engineers had come up with were already optimal. Only our engineers had come up with the same results, using less resources.
I’m beginning to think of AI as mostly hype. It does have some uses, like for recognition. We use ML and image recognition to troll through millions of old case files for various purposes. We use speech recognition to help people with dyslexia.
But for analytics and prediction it’s been extremely useless compared to actual people.
I’d love to be convinced otherwise by the way. If it actually does something, I’d like to be on the forefront with it so I don’t have to play catch up when Deloitte sells it to the mayor. But I have to admit that after 5 years of AI bullshit bingo I’m one sales meeting away from labeling AI in the same category as blockchain.
On the other hand AI can be a powerful tool or assistant when humans are struggling with a task. AI that (for example) highlights all the suspicious places in an oncology image after an oncologist has had a look and labelled it as showing cancer could improve quality of outcome. I am deeply suspicious of people who want to chop the oncologist out though!
More prosaically prediction systems to estimate future customer churn, identify sales opportunities (or in the public service case demand for services and needs), future traffic flows, complex settings for complex devices (especially in the face of change), diagnostics of problems, reasoning over millions of records, are all valuable. Often the theme is "can we make a very narrow very specialist component that could come to the same conclusions that a reasonably smart three year old would if we could make her concentrate and do it millions of times". So a bit smart, very autistic, very fast.
What you will not do is create value from consultancy and procurement; you need to build a capability to get the value, and that is a slow, expensive and strategic game - think 4 high cost (but don't be bullshitted there are great people who will work for reasonable amounts out there so long as you don't expect them to be slaves and make bombs or crash the financial system - think doctor or lawyer cost and recruit with the support of your uni) fte for three years.
If that is too expensive then I think it's going to be a waste of your time for now, and your current stance of sceptical openness is the right thing to do - eventually someone will come with a convincing shrink wrapped offer that is right for your domains, at the right price - then all is good.
Process automations and "AI" (let's say, complex autonomous systems) are parts of a bigger picture which is the digitalization of the society. Sometime, they are sold as a drop-in replacement for current human processes, but it actually works in only a few common cases.
The issue with AI and automation is not necessarily a technical one, the issue is the shift to a different ownership model, trust model, maintenance model and faster improvement model than the ones being believed possible, or currently in place for the last decades.
It's not that software cannot replace people, rather that big corporations are resistant to changes. We don't need to introduce AI to see the internal inefficiencies of big employers...
Technologies fundamentally change the meaning and content of "work", but people cannot fundamentally change that fast. Budgets cannot change that fast, investments cannot change that fast, regulations cannot change that fast, etc. There will be a lot of preliminary work to do in changing the educational system, regulations and also changing the redistribution of wealthiness, before "AI" actually happens.
The AI came up with the optimal solution on an "already solved problem", twice, and this is a failure on the part of the AI?
Usually the problem is that companies are going to sell you an AI solution for your data, but these companies are not going to tell you that with your current data there is very little an AI system can do. I say this because I saw companies who wanted a prediction system with 2 years of data (40 datapoints per year). Any AI expert will see there is no way this will work but still the system is sold to the company. This is much worse in the public sector.
The problem for ML seem to be that we’ve been using the data for a while and that we’re already very good at it.
This is actually how wumaos are supposed to work in China as well. If they go with a blatantly pro-government message, they will be ignored by anyone critical of the Chinese government. So instead, they come in agreeing with critical positions and redirect sentiment from there. So if you ever hear someone who sounds like a wumao, they are probably just overzealous nationalists and not shills.
[1] https://foreignpolicy.com/2016/06/03/china-explained-sixth-t...
> BY BETHANY ALLEN-EBRAHIMIAN
Wow. That's an incredible title to use on a subject that is so narrow in scope.
Let's look at her other articles:
https://foreignpolicy.com/author/bethany-allen-ebrahimian/
Seems pretty one-sided.
And her twitter:
https://twitter.com/BethanyAllenEbr
Seems pretty one-sided as well.
Just to be clear, I don't intend to personally attack her. Just pointing out the bias in her reporting while we are on this topic of "wumao".
I am also not targeting Foreign Policy as a news media, as Foreign Policy itself doesn't seem to suffer from bias against China as a whole, as there are other authors that are either more neutral or biased towards China:
https://foreignpolicy.com/tag/tea-leaf-nation/
https://foreignpolicy.com/author/melissa-chan/
https://foreignpolicy.com/author/ran-jijun/
Edit: Attempt to remove personal attacks in the comment.
Edit 2: Remove the section about myself.
Wumao is basically someone who is paid by the Chinese government to subvert online public opinion. If you are just pro China and being called a womao, the term is being misapplied. Anyone who says anything critical of China is called biased against China, even if it’s just a thing or two. Few people are actually completely biased against China on everything, but in Chinese thought, there seems to be no room for nuance.
Also, there's plenty of room for nuance to judge the bias, like the other authors on FT that I pointed out in my comment. Her bias is quite clear in my own opinion, given that there are many articles by her on China in FT, and most of the content (at least first few pages) is critical of China (not just one or two).
As always in western press, bad news is much more visible than good news, leading to an appearance of bias against China when it is really just a bias for bad news that readers will take note of. The CPC is exploiting this exactly with sixthtone, realizing that their feel good only propaganda approach doesn't translate abroad.
Also, certain reporters do feel good articles, others are better at critical ones, it doesn't mean that the reporter has a bone to pick with China, just that they have a job to do. In reality, she is probably less salty than the average long term laowai.
you're either actually a wumao or are too afraid to actually take a reasoned stance. i had never heard of the term before this thread but makes sense, i'm reading a lot more content like this nowadays wondering what exactly the author is arguing about
[cit. https://paradite.com/2014/01/23/good-articletaking-a-war-of-...]
Do you know the article beforehand or you discovered it through my profile?
Ask yourself if you have to worry about your love having all their organs after an act of civil disobedience. I have and do every night. I do not know how to state that in a way that is palatable.
Not when it's done in lieu of a rational discussion of the information, which should be the majority of the argument. Hitler liking dogs does not invalidate loving dogs even one iota - so what did you gain by attacking the source (when the subject is not the source but what it says)? It's a distraction and a (dark side!) rhetorical method, attempting a shortcut through evoking emotions.
Pinscreen was also accused of faking results, including those shown off at Siggraph 2017. https://www.theregister.co.uk/2018/07/18/pinscreen_fraud_cla...
This has been happening for centuries: https://en.wikipedia.org/wiki/The_Turk
But you might also want to care about this on general level. So many bullshit companies doing fake-AI create perception that the technology is much more advanced than it really is; this influences public opinion, and political decisions. Consider e.g. the coming wave of AI-induced job loss, that was such a hot topic just a year or two ago (for some reason it seems to have died down recently). After "Humans Need Not Apply", I expected we'll see significant effects in a decade. These days, I'm much more conservative about the timeframe, because I've learned that most of the immediate fear was based on hype and false advertising of a new statistical gimmick (deep NNs), that in reality only sorta, kinda, works, for some very limited tasks.
ps. Expensify sent images with personal data to Mechanical Turkers, calls it a feature
https://arstechnica.com/information-technology/2017/11/expen...
There are others out there too and you can check out their "Privacy Policies" whether they use "data extraction teams" offshore.
The promise of AI has always been not better experiences, but similar human touched experiences for a much cheaper cost.
Beyond that, commenting on the translation a bit. They did live translation the first day of WAIC for the headline speakers. There were 2 screens, 1 was baidu and the other was iflytek. Neither were that good on the english side (it was ok..but could barely keep up with the speakers)
They claim they are still working on english. The grammar output wasn't coherent. IFlytek itself has some neat hardware they sell that is pretty good.
Beyond that, it seems like they are mainly collecting data right now. I would not be surprised they were doing this just for marketing visibility. It is easy to fake.
Happy to answer questions about the experience there if people would find it useful.
https://duckduckgo.com/?q=mechanical+turk&t=h_&iax=images&ia...
Reminds me of classic mechanical chess-playing
turk illustrations. I was about to link it here,
but all the search engines could come up using
plain text search were links to Amazon's
crowdsourcing thingYour customer cares about a business value. From my experience using buzzwords like "AI" will only get your foot in the door, but in the end, the business value will sell. IMO it's a completely valid business strategy if you are selling something at cost with the plan of reducing the cost later and you can stay solvent in the meantime.
Maybe they should've tried trading a red paperclip for a house?
This is a huge mistake that many companies are falling on, they are trying to sell us a Jarvis when actually what they have is some (good) algorithms
what
If the "AI" company lies by using manual translation without the Turk, they're lying about point B a lot. If they're using Turk, they're lying about point C and a bit less about point B.