AI and Economic Productivity: Expect Evolution, Not Revolution
spectrum.ieee.org
spectrum.ieee.org
Just look at Hong Kong’s recent law outlawing masks.
NLP (GPT-2/BERT) is also extremely powerful and is evolving at a frightful pace. Just google news over the last month. I read a lot of generated text and non generated text and more and more I really can’t tell the difference. Yes, it doesn’t work in isolation, but human guided NLP is stunning in what it can do. Self-attention applied to many other areas is huge as well, such as Software engineering.
Drones is another revolutionary technology which has already completely upended how war is executed.
There are reasons why leading technology companies such as google, Microsoft, Amazon are dealing with employee blow back around the military and government. They recognize the possibilities here and are uncomfortable with with potential contributions to the disruptive downsides of what can be done by “smarter” algorithms (or whatever you want to call them).
Software truly is eating the world.
> Take facial recognition for example (there are many many others, but this is very illustrative). The possibilities of revolutionizing society with facial recognition is legion. Imagine having a camera attached to your jacket which is hooked up to your phone and immediately gives a compete profile of everyone you meet. This tech is possible today. It probably won’t work with total strangers, but will with people you are likely to meet given your social circles.
I mean, I can namesearch the internet for people I meet via my social circles without [admittedly much improved] facial recognition AI: the innovation's all in the hardware, culture of broadcasting one's life via social media and search tech. Hell, I could and did check out people's more basic social profiles online two decades ago when Kurzweil was predicting that by 2020 AI would have pushed life expectancy over 100, eliminated financial crises, economic deprivation and road accidents and simulated human personalities so convincingly that human-robot relationships were a normal thing. The facial recognition is neat, to the degree that it's accurate, but probably less revolutionary for social profiling than the cultural shift towards putting more information online.
Similarly, AI startups are far from the only types of business where employees have ethical qualms about selling to the military and certain areas of law enforcement, and one of the biggest concerns employees have is what happens when bodies take action based on vastly overestimating the AI's conclusions...
Generally consisting of people who know what they're talking about rather than marketers or science fiction fans.
Hong Kong's mask law has nothing to do with tracking people. It has to do with protesters committing violent acts while masked. All you need to do is track people is get them to carry a cell phone which can be tracked via its SSID, IMEI and ICCID.
Laughing at the idea of being afraid of BERT. Reminds me of Hofstadter thinking ATNs might soon achieve sentience.
The pace of advancements is unnerving though. May you live in interesting times indeed.
There's a reason why that stuff has remained a niche for people interested in trying to implement Vannevar Bush's "As We May Think".
Perceived pace of progress when seen from the inside of a field is significantly slower than the perceived pace of progress from the outside. That said, I will agree that the progress in AI and face recognition since 2011 has been phenomenal.
You can set up a useful automated virtual business card system without AI. There's no particular benefit to the facial recognition, and even less benefit to having a full profile to chew through.
What you really need is an AI with an awareness of social context, and for that be useful you need an entire biographical information collection, filtering, categorising, and sharing, which doesn't exist today.
But then there are privacy concerns. Do you really want complete strangers to know about your personal and professional resume (successes + failures) parking and speeding tickets and criminal record, sexual preferences, credit worthiness, major illnesses and minor health issues, and so on?
Of course Google, FB, and China's Social Credit system are working towards this already. And we'd better hope they fail, because creating this kind of personal biographical infrastructure without oversight gives corporations and governments the potential to exercise an unholy level of personal and professional control over individuals.
The point being - your supposedly "obvious" facial recognition application is in fact a demonstration of how thinking about these projects as if they're trivial hackathon demo pieces can lead to a culture of horrors.
The real issues with all AI are political, not technical. And the problems can only be solved by strengthening limits on the exploitation of personal information by corporations and governments. Because otherwise we're going to be living in a culture where the primary application of AI is personality control - no matter if it looks like advertising, or media, or art, or political debate.
Black mirror episodes, some of them are very dead on and immediately relevant.
Then there is the whole 'attention economy' we have shifted the majority of our economic endeavors to. Just like some examples in the article, 'efficiency gains' in that space are just a red-queen race in a zero-sum game.
This is not to say AI will not have major impacts on those whose livelihoods are swept up in its evolutionary or revolutionary (isn't that mostly just a difference in timescale) jaws. And even AI might open up new pathways of improvement and genuine innovations for mankind, and at the very least it is exiting and intellectually stimulation to advance. But to look at 'economic gains' might not be the right paradigm. That might just be like trying to improve the health of a morbidly obese cat by feeding it more.
Either the concept, or how the terminology you're using arises.
There is a huge lack of doctors, nurses, architects, etc in many of those areas. In some countries, a doctor may need to see 100+ patients a day as a matter of course. We can imagine how the quality of care suffers as a result.
As examples, AI, including NLP and computer vision technologies, can complement doctors and nurses by:
- performing initial querying of symptoms,
- regular pre-screening for diseases at minimal costs (thus reducing costs that incur with delayed diagnosis and treatment),
- automatic checking for medicine side effects or, in future, recommend suitable medicine for particular patients given diagnosis and test results,
- etc
If deployed properly, the technologies will improve care, reduce healthcare worker’s overload, and enable greater access from rural/poor areas.
The same applies to several other expertise that is not lacking in well-off regions.
Few people seem to notice that "computer revolution" has largely been taking jobs from white collars way more than blue collars:
Lawyers: Google.com devastated the paralegal profession, and seem to be now taking on more "retail" law services.
Marketing: Marketings is the biggest IT services spender in a lot of companies now, but the amount of people involved is less and less with each year. Marketing execs seem to be first C-level people to feel pressured by the computer industry. Operations and strategy that previously were a job for 100K a year man now being taken on by marketing interns for 30k
Operations and admin staff: very clearly, email, Google Calendar and MS Word has allowed people to do more fancy documents, and organisational work without aid of professional typists, and secretaries
Project management professionals: no explanations needed, MS Project, no matter how badly it is normally used, have made PM work available to less trained staff without formal PM training
Finance: the entire class of "retail fund management" shops are now being driven out of business simply by banks with better GUI apps. The entire need for them was previously dictated by having somebody on the phone who was ready to do all financial formalities for the millionaire class.
Investment: same as above, cheaper brokerages with online GUIs have displaced a lion share of brokers. Even well moneyed M&A people now often see 30 something junior partners taking their work with use of Google and ctrl-F
What I discussed above is about the domains and geographical areas where the expertise is lacking resulting in human suffering or underdevelopment and the situation either improves too slowly or worsens (due to population aging, for example).
In domains where complex expertise is required, people will not be replaced unless some kind of human-level AI is invented (when it happens, at least some years/decades away, the ramifications will be vast and outside the scope of this post). Doctors, nurses, architects, etc perform diverse, often creative tasks—many of which no current AI is even at an amateur level.
Expertise is an "inferior good" economics wise. Sadly, but a pretty much a fact for me. You don't need "superhuman AI" for this. When a business gets a lot of money to throw on people, the preference will go to more easily available cadres, even if net cost employing them will be higher.
See, how IT outsourcing went. By far, computer programming is the hardest of intellectual professions, yet people with money are more fine ability to hire legions of come and by people in South Asia, than one man with a worth of 100 back home.
You will also be surprised that this is also the case in such "intellectual" industries as management consulting, finance, and fund management. Thirty something MBAs doing googling, ctrl-Fing, and some times using purpose made analytical software is what makes most of workforce in "big and serious" investment groups.
A local streaming company tries to compete with Netflix and its app is not up to par on latency and reliability. Soon people catch on, cancel subscriptions in drove, and tell others in public forums.
Now do your judgement on their skill proficiency.
Not to say that Google and other akin companies still do employ droves of people in ultra cheap countries.
Healthcare workers can be almost an order of magnitude cheaper than doctors and take much less time to train.
In a middle-income country I know very well, there is no lack of such workers, but doctor shortage, esp specialists, is acute in many rural areas.
This is unrealistic in the developing countries (low income) I am familiar with. The density of clinics is so low that clinics need to cover vast geographic areas. Clinical workers are easier to train, but still in extremely short supply in most rural areas of the developing nations I know well.
Would telemedicine plus AI-enabled UI, such as voice interface, help at all? For example, it could remind patients of chronic diseases to take medicine regularly or take better care of their diet. Obviously, technologies alone would not solve such a major problem, but they are enablers and accelerators.
Technologies are some of the more easily scalable and lower-cost solutions available. I have written more about that in other comments as well as provided links to some very recent advances in NLP. I would also enjoy hearing some of your proposed solutions that are practical, plausible, and scalable. I think we are all ears for constructive comments. :)
None of the points you name strikes me as a good use case for modern machine learning techniques:
- Siri and co can hardly understand everyday things. Using NLP for symptoms seems like a massive downgrade from speaking with a human.
- how does machine learning play a role in pre-screening? Pre-screening can also increase costs because you find a lot of harmless things as well
- I hope you will Put out a product telling if something is safe or not using machine learning. There are already systems out there and they are much more akin to experts systems which makes much more sense.
The only area where I see a big impact is medical imaging where you can already see great research going on.
I work in NLP/NLU and foresee some of future advances that will enable them.
Moreover, BERT et al models that were just invented over the last 1-2 years may already allow the implementation of basic versions of these human language technologies. In many domains, they are on-par with humans in answering multiple-choice/span selection questions. See, for example:
https://gluebenchmark.com/leaderboard
http://www.qizhexie.com/data/RACE_leaderboard.html
I also agree that technologies are not silver bullets. They are enablers/accelerators of improvement in process and efficiency that can increase human well-being.
Nobody thought to pause and question whether startups had an incentive to overstate their impact? Watch one episode of Shark Tank and you can see the fault in this methodology
First, 2035 is very far away. The first modern smartphone came out 12 years ago. I bet in 2007 you could have analyzed the mobile startup space and come to the same conclusion. Projecting what benefits a revolutionary technology will bring us in the next 16 years by analyzing what startups exist right now does not seem like an effective way to make predictions. Uber (arguably the largest mobile-first company), came into existence only 10 years ago. The AI startups that will shape the industry of 2035 probably don't even exist yet.
Then there is the lack of deep analysis of the three divisions in the current wave of AI, each of which have very different trajectories. There is:
1. Computer-world interaction. The work that focuses on giving computers new ways to interact with the world - speech recognition/synthesis, natural language understanding, computer vision, interpreting more specialized sensor data such as lidar or 3d imaging.
2. Data as an economy of scale. The work that focuses on using data to make better spending and planning decisions by leveraging vast amounts of historical data.
3. Exploratory research. Speculative research towards autonomous agents and AGI. Stuff like most reinforcement learning and other pure research that is not ready for production.
Computer-world interaction is where most of the current set of startups seem to live. You can build on top of publicly available data and then finetune it for your needs with a relatively low amount of data, which means the barrier to using AI is not insurmountable. Example successful products using this type of AI - smart speakers, L2 autonomous driving (e.g. Autopilot), Pixel night camera, medical diagnosis. This is still in the early stages of commercialization, but the technology is improving at a ridiculous rate. This is mostly still in the late R&D/early commercialization phase so I would not expect many of the players in this space to be focused on profitability. What matters for the long-term projections here is whether the technology is advancing towards being viable for many real-world usecases within the next 5 years (IMO it is), whether the technology will be cost effective (less obvious but it looks promising with advancements in hardware and more efficient network designs) and whether it will be accessible to many companies (lots of good work happening here between AutoML research and MLaaS companies).
Data as an economy of scale is where most of the near-term financial benefits of AI are being demonstrated. The successes I've seen are primarily confined to big companies since they have the data and resources needed to train and benefit from these models. There are some startups in the space (maybe the security startups mentioned in the article?), but I'm not sure exactly how they would build useful AI models without the large amounts of historical data that established companies have. It's possible startups in this space are sometimes guilty of overselling their 'AI'. However, big companies are finding a lot of success here - better logistics decisions, better fraud detection, better content/ad recommendations. Some of the techniques have existed for quite a while and we are only seeing benefits now because it has taken a long time to build up the data and ML infrastructure to apply those techniques. Other techniques such as graph neural nets are still in their early phases but are showing great promise. This is completely missed in the analysis done by the article because it only looked at startups.
Finally there is exploratory research. Some of this is quite promising - RL as a technology is incredible, although RL as a useful tool for business still appears to be a ways away. While any advancements here would be great for the impact of AI, it doesn't really matter if none of them pan out - the current set of proven techniques will still have a huge impact. 'Companies' in this area are in many ways just research labs. I would not include them in my analysis - DeepMind and OpenAI.