The Age of AI, (part 3/3)
“What all of us have to do is to make sure we are using AI in a way that is for the benefit of humanity, not to the detriment of humanity” — Tim Cooke, CEO of Apple
Is AI a force for good or evil?
The number of papers published on AI is now doubling every two years. New breakthroughs are being made on a daily basis.
This raises the obvious question of what it means for us. While human-level intelligence (AGI) is still some ways off, many are concerned that once AI becomes more capable, it will start improving itself to the point where it will surpass human capabilities and overcome our ability to contain it by becoming an Artificial Super Intelligence (ASI).
This acceleration is called “FOOM” (which doesn’t stand for anything, it’s just the sound of something accelerating). The current debate sees four possible outcomes of FOOM:
- Benign AI: AI will serve the interests of humans.
- Neutral AI: AI will evaluate data and respond to questions in an objective manner (i.e. like a super-Wikipedia).
- Humans as pets: AI will keep humans around because they find us amusing.
- AI will eliminate humanity: either through malevolence or perceived lack of utility, the AI will wipe out humans.
Clearly we prefer the 1st or 2nd outcome. We therefore need to “steer” the AI to “do good.” This is known as the AI Alignment problem (how to ensure that the conduct of the AI is aligned with the interests of humanity).
At present, we use RLHF to guide AIs to “not be evil”, but this won’t scale, since it’s impossible to imagine every negative eventuality. This has created an interest in Constitutional AI, where fundamental principles are encoded at the inception of the AI to help guide its behavior. Science fiction fans will no doubt find this familiar, e.g. the 4 laws of robotics from Isaac Asimov’s “I, Robot” stories:
1. A robot may not injure a human being or, through inaction, allow a human being to come to harm.
2. A robot must obey orders given to it by human beings except where such orders would conflict with the first law.
3. A robot must protect its own existence as long as such protection does not conflict with the first or second law
4. A robot may not harm humanity, or, by inaction, allow humanity to come to harm.
Claude, the AI system from Anthropic, has a constitution derived from the UN Declaration of Human Rights and Apple’s Terms of Service (e.g. “thou shalt have no other app store before me” — just kidding!).
“With great power comes great responsibility” — Spiderman
Given the power of AI, it is not enough to hope that the companies that create them will “do the right thing” and embed good constitutional principles. In this, AI is like other powerful technologies such as gene therapy and nuclear physics: it can be used for both good and evil. For those technologies, society has chosen regulation to limit the risks. AI needs regulation.
Regulating AI
In the EU and the US there is a growing debate about the right way to regulate AI — at present there are several proposals being drafted and the time to put them in place is now, before AI systems become even more capable.
Unlike nuclear material, it’s very easy to get your hands on AI technology: The code and data to create an AI can be readily downloaded online. Regulation can’t restrict access to AI, but it can restrict the sale of applications of AI. A possible model is the regulation currently applied to medicine. To bring a new drug on the market, a company has to demonstrate its safety and effectiveness. For AI, a company that say, chooses to sell an “AI investment advisor” would have to show that it is effective at recommending investments that don’t expose customers to undue risk. A government agency (similar to the Food & Drug Administration (FDA) in the United States) would test the various claims that the creator of the AI makes about its capabilities and would grant a license (for that use case). This is the test of effectiveness — but not necessarily of safety.
A test for safety could be inspired by penetration testing that companies undertake by trying to hack into their own IT systems. A copy of the AI system would be subjected to a variety of tests (e.g. prompt injection) to see if its constitutional principles can be subverted.
The challenge with regulation is that law-making is slow, whereas technology is fast. Nonetheless, it’s probably better to put in place a regulatory system now, and evolve it as AI evolves.
So what will happen to my job?
“It’s a recession when your neighbor loses his job; it’s a depression when you lose your own.” — Harry S. Truman, 33rd US President
For now, we’re in the “pre-AGI” phase. AI is becoming more capable, but it’s not perfect. People tend to imagine binary outcomes: a) AI is useless, or b) AI will take my job. For now, the more likely outcome is c): AI will help me do more/better.
“AI will not replace you. A person using AI will” — Santiago Valdarrama
AI can already boost the productivity of writers and developers. Brian Chesky, CEO of AirBnB, expects AI to increase the productivity of his software developers by 30% this year. Therefore Chesky has a choice: he can either fire 30% of the team or do 30% more development. Most likely, he’ll (mostly) do the latter.
A recent report by the World Economic Forum on the future jobs, predicts that 40% of work will be impacted by AI. Jobs in clerical and administrative areas will decline, but other AI-related jobs will offset most of those losses.
In practice the introduction of new productivity tools takes time to trickle down to layoffs or reductions in employment. In the case of the introduction of spreadsheets, the number of bookkeepers declined steadily over 25 years (but was largely offset by increased hiring of analysts and other employees who manage data).
The impact of AI on jobs is likely to proceed as follows:
- (Some) tasks get automated — long before entire jobs are eliminated.
- AI creates new jobs: 60% of workers today are employed in jobs that didn’t exist in 1940. Just like the Internet created new jobs such as “Search Engine Optimizer”, or “Social Networking Marketer”, AI will create new jobs.
- Supply creates its own demand: the ability to sift through ever more data lets companies create more analyses/more scenarios, requiring more employees. More productive programmers can build more applications.
- AI creates new ways to waste time: (similar to when Microsoft added the game of Solitaire to Windows), AI will create (many) new ways for employees to spend time doing non-productive work, ofsetting some theoretical productivity gains.
Some jobs (e.g. taxi/rideshare-drivers) are unlikely to be entirely eliminated. These jobs require more intelligence than current AI can deliver. Waymo, Tesla, and countless other companies have spent over a decade and billions of dollars trying to create AI-enabled driverless cars. Humans are very good at understanding the difference between driving over an empty cardboard box lying in the road and one made of steel. An AI vision system that doesn’t understand that cardboard crumples and steel does not, lacks that insight. In computer science, really hard problems are known as NP-complete (non-deterministic polynomial time complete). The AI community has coined an equivalent term, AGI-complete, where to solve the problem, the answer is: “first, create an AGI, then…”. Driverless car (in the most general sense), robot surgeon, etc. are all probably use cases that are AGI-complete. In other words, we’ll need to create the equivalent of human-level intelligent AI before we can fully replace that job.
The impact of AI is somewhat unique in the history of technology because it is the first to disrupt creative and knowledge work. On a more positive note, early signs suggest that workers with moderate skill gain the most by adopting AI (experts don’t benefit as much). It also creates an exciting opportunity that reverses the centuries-old trend in education of creating ever-more narrowly specialized training: the most productive employees of the future may be generalists + AI.
For example, many Western countries are not training enough doctors for their aging populations. Rural areas are already closing clinics and hospitals due to a lack of available doctors. Emerging markets have an even greater shortage. Instead of taking 10–14 years to train a doctor, a nurse equipped with a smartphone can use AI to analyze images of skin lesions, listen to someone’s cough, analyze their gait, and create incredibly accurate diagnoses that (already now) can outperform most specialist doctors. A nurse can be trained in only 1.5 to 4 years.
Healthcare could be the “killer app” for AI. In the US, 60% of adults have one or more chronic conditions (diabetes, hypertension, etc.). When a serious health issue arises (e.g. cancer), the cost of healthcare becomes the leading cause of bankruptcy. Insurance could be made more cost effective for younger adults by starting with a very inexpensive “AI-tier”: your primary physician and mental health doctor are AIs. You can consult them anytime you want and as often as you need. For accidents or more serious ailments, clinics and the emergency room will still be available. By instantly recalling your entire medical history and by accessing data from your activity tracker or smart watch, your “AI-doc” will know you far better than a human doctor would.
Unfortunately, powerful and monied interests have a lot at stake in maintaining the very expensive current system, which means that healthcare (at least in the US) will not be disrupted by AI soon. Emerging markets could become the early adopters instead (like they did with mobile phones and mobile payments).
Why we need AI
Given the turmoil, uncertainty, and risks that AI presents, it might seem sensible to wish that the technology had never been invented. We can not erase it nor prevent its further development. We must strive to make it as safe and as useful as possible. We also need AI for humanity to thrive:
- Boosting productivity: In most western countries, populations are aging. The elderly no longer contribute to economic productivity and more of our resources will go to supporting this cohort. A loss of productivity leads to declining living standards. We therefore need productivity gains to offset this loss. AI (and robotics) holds out the greatest promise to drive productivity gains when populations shrink.
- Reinventing education: 500 years of innovation in education has consisted of replacing blackboards with whiteboards. It’s hard to think of an area of our society where technology has had less impact. We also know that (thanks to the COVID pandemic) that distance learning is not effective. AI can create interactive “tutors” that can deliver personalized learning programs adapted to each student. Students can learn new languages by speaking to an AI that will respond in their chosen language, commenting on their pronunciation and grammar.
- Developing new drugs: drug discovery today starts by testing 10,000 molecules to end up (possibly) with a single workable drug (after 10+ years). Precision medicine can speed this up dramatically by turning this process around: starting with the shape of a receptor that we want to target, AI can propose proteins that will fit. The missing link now is an “in vivo” AI that can simulate what happens to that protein as it moves around in the human body and inside the cell. This way we can discover which ones aren’t toxic and won’t be eliminated by the human immune system. One area where we need such innovation urgently is in the discovery of new antibiotics. We are almost out of such drugs as more bacteria are becoming resistant to our existing treatments.
- Staying healthy: the more we understand the complex interrelationship between genetics, nutrition, activity, and social interactions on wellness, the better we can evolve from our current “sick-care” system to a true “health-care” system. AIs can learn the specifics of our genes and our daily habits to personalize coaching advice to keep us healthy. This would let us reduce the cost of our healthcare and redeploy spending in other areas that need it (e.g. climate change, education, etc.).
- Mitigate the effects of climate change: this is an existential threat to all life on earth. Because changes in the climate occur over years, decades, and centuries (longer than a human life span) it is difficult for people to see the consequences of our activities on the environment. AI can accelerate the development of climate solutions (better batteries for electric vehicles, more efficient solar cells, improved desalination technology, etc.). The sooner we deploy these solutions, the more we can mitigate the impact of warming temperatures.
AI is one of the most promising (but also potentially dangerous technologies) ever invented. We are still in the early days of its development. It is already useful and will become much more capable very quickly. It is available to anyone who can get online, for free. We should understand it well, so that we have the greatest chance to shape its trajectory as a force for good. Our future depends on it.
