The Age of AI by Kissinger, Schmidt and Huttenlocher

It is already beating masters (and other AI) at chess, helping invent new vaccines, helping us parse long meetings, nudging our creativity, acting as a sounding board, flying military aircraft, helping me write this review and so much more.

AI can be considered a form of intelligence that humans cannot achieve, something that will be outside our reach. The authors make broad predictions about what is going to happen next:

We’ll be using AI a lot to aid our creativity, decision making, knowledge and so on. So, apart from humans making decisions (common) and machines making decisions (getting common) there will be humans and machines working together to make decisions. In fact in some areas it will be considered immoral (maybe even illegal) to not consult AI before taking a decision. A humbling time for humanity as we concede our position as the smartest beings on the planet.

We’ll probably not have AI overlords in the near future because the current technology needs humans to define the problem first and also the constraints within which a solution can be called “right” solution.

That is, AGI is not coming anytime soon.

But the pace of innovation will increase and things we thought difficult/near impossible earlier will become manageable. AI will impact almost every part of our lives. “It is reasonable to expect that over time, AI will progress at least as fast as computing power has, yielding a millionfold increase in fifteen to twenty years.“.

Just like with any other technology, there will be people who do not adopt AI. Although since it is a general purpose technology, the lack of adoption will be an act of commission rather than omission and those who do so will likely do it at significant cost. The “AI Amish” if you will.

“As societies develop their own human-machine partnerships — with varying goals, different training models, and potentially incompatible operational and moral limits with respect to AI — they may devolve into rivalry, technical incompatibility, and ever greater mutual incomprehension.“.

Lots of ethical and moral questions coming our way like this one: “… if AI recommended that a commander in chief sacrifice a significant number of citizens or their interests in order to save, according to the AI’s calculation and valuation, an even greater number? On what basis could that sacrifice be overridden?“.

And although AI of today cannot be called “conscious”, we already see cases of intelligent people treating it as a fellow being.

As the behaviour of AI gets more indistinguishable from that of humans, these discussions are only going to get muddier.

And it seems that the number of people who are working on predicting and preempting such questions are far lesser compared to the numbers of people working to keep AI at the cutting edge of technology.

Whenever we observe something that our current models of thinking cannot explain, we develop new models and that marks the start of a new epoch.

The authors do a quick run through of how we’ve evolved over millennia in our thinking about the universe, how it works and our place in it.

Notably, this is all from a collectively “western”/European point of view (keep that in mind whenever I say “we”, “us” and so on in the following points) and does not represent how thinking evolved in, say, Ancient India.

The history lesson starts with Ancient Grecian and Roman thinkers who placed central emphasis on the quest for knowledge (the “what”).

There were things that we could not explain, so we ascribed them to the divine. Notably there was no single entity, instead “… deities of a thousand groves … possessed, … their local … influence“.

So, the nature of divinity and the divinity of nature were both explored open-mindedly but people rarely settled on a single, definitive figure or force that could be worshipped.

Then came the church that thought that these open ended explorations about the nature of the divine weren’t going anywhere useful, so they posited that all of divinity could be explained by a single word: “God”.

So, anything for which you could not offer an explanation (like the changing of seasons) was now ascribed to “the workings of the Lord”.

This God, apart from being all-knowing, all-powerful and so on, was also the most VIP of all VIPs and so you could not really call him up on his phone.

Instead, the church was to act as the mediator between hoi-polloi and the bearded man in the sky.

And because there were a lot of things at the time we needed to talk to God about (like the frequent disease, injury, death plus not enough food to go around) the church accumulated a lot of money, power and a “… near monopoly on scholarly inquiry for centuries, guiding individuals through sacraments toward an understanding of scripture that was both written and preached in a language few laymen understood.“.

If you did everything the church said when you were alive, you’d go to heaven after you died.

This went on for a long time (about a thousand years), a time we now call the Middle Ages.

Eventually, Gutenberg improved the printing press, this loosened the church’s control on dissemination of knowledge, and exposed corruption within the Church) and the Protestant Reformation told people they could call up God directly.

So, we were now getting back to learning about reality (like in Ancient Greece).

This blossomed during the Renaissance when Europe started exploring the world. It soon realised there are other people living on it with entirely different cultures, beliefs and ways of looking at reality than theirs.

Many good things happened during this time like the rise of humanism, scientific method (and its attendant benefits like “… keeping time, navigating the ocean, synthesizing useful compounds …”), literary explosion, and artistic flourishing.

Such was the pace of progress that it left us a little confused – Because for the longest time the Church had told us that all that was to be known was known, and the rest was God’s doing, and here we found ourselves getting to know all sorts of new things many of which were earlier ascribed to God (like germ theory).

This period, which we now call the Enlightenment, placed emphasis on explaining reality (not just learning about it). “In this atmosphere of intellectual challenges, once axiomatic concepts — the existence of physical reality, the eternal nature of moral truths — were suddenly open to question.“. Faith sort of took a backseat during this time (since we had a better way of explaining things), but it was not completely forgotten (since there were many things we still could not explain).

Kant came along and said that reason could not explain everything and that it should literally use its own devices to understand its own limitations (totally meta). His position was “… the human mind relies on conceptual thinking and lived experience, it could never achieve the degree of pure thought required to know this inner essence of things. At best, we might consider how our mind reflects such a realm … but this does not constitute true knowledge of it.“.

Tradition had dominated why we certain did things for a long time, but now reason started to become, well, the reason we did things.

“Armed reason” it is called when reason goes up against tradition with a revolutionary mindset, i.e. when it does not accept anything until it has been vetted rationally.

BTW, Romanticism arose in response to Enlightenment, Romantics felt these values downplayed the importance of emotions, imagination, and individual experience.

Nevertheless, the ideals of Enlightenment have shaped modern society and thinking as we know it for the last three hundred years. We’ve known for a long time that the way humans perceive reality is not objective but till now we’ve been the only ones around unearthing reality and so we’ve not really had a counterpoint.

Until now, because AI is becoming the counterpoint, it may be said that AI is “experiencing” reality that humans do not have access to, maybe they never will.

Essentially, modern AI came to be when we gave up trying to teach it things and let it learn for itself.

Initially we tried teaching AI by showing it ideal versions of things and expecting it to identity these things in reality in all their various forms. This did not work. “For example, to teach AI to identify a picture of a cat, developers created abstract representations of the various attributes — whiskers, pointy ears, four legs, a body — of an idealized cat. But cats are far from static: they can curl up, run, and stretch, and manifest various sizes and colors.“.

So then we were like (paraphrasing) “AI, here are a bunch of images, all of them are cats, do what you will to learn how to identify cats”.

And this was a big deal, because with this approach AI became less precise (i.e. it could make mistakes) but more general (it could be applied to whole lot more of things). Until this point, the closest approximation to AI we had, computer algorithms, were very precise but also very constrained to the what they could be applied to. “Using machine learning to create and adjust models based on real-world feedback, modern AI can approximate outcomes and analyze ambiguities that would have stymied classical algorithms.“.

AI can learn in broadly three ways: supervised learning, unsupervised learning, and reinforcement learning.

AI today has limitations that we should all know about and it is not up to a single entity whether government or private to guide its development, we’re all in this together and there is still time:

AI is not sentient, it cannot reflect on what it discovers. So it is up to us to tell it when it encounters something “good” versus when it sees something “bad”.

It can suffer from:

• Dataset bias (incomplete or inaccurate data)
• Human bias (that skew the training data or algorithm)
• Rigidity/brittleness (AI unable to adapt to novel encounters)
• Hallucinations (dreaming up things that never happened)

Also, we have little idea of what the AI is learning because machine learning drives AI. This may be addressed by having AI take “exams” like we humans have to do in order to prove that we’ve really learnt something.

This will be more critical for AI that is going to be employed in areas like self-driving vehicles, drug discovery, electricity grid management and so on.

Basically things that are important to humanity should see respective AIs go through comprehensive exams before they’re allowed to take up the job.

But it is already deep into global network platforms: You know, those platforms that benefit from the network effect like Facebook, WhatsApp and Instagram, but also Google, YouTube, Maps, Instacart and Amazon. These platforms are already so large that their decisions have real impact on real people often in real time.

These platforms generally champion the cause of western democracy like freedom of speech and ideas although their self-defined guidelines profoundly impact what gets discovered and amplified, and what gets censored and forgotten. This can have “not great” repercussions even thought that wasn’t the intent.

And when you add AI to the mix (which you have to because scale is better handled by computers) for things like content recommendations and comment moderation it becomes important that these algorithms are checked for biases and limitations I mentioned above.

A kind of tacit partnership has formed for many of us who use network platforms and the AI that power them.

We rely on the AI to help us sort though of the information deluge and choose what is right for us. We come to largely trust its recommendations. In return the feedback we provide helps the AI make the network better for others.

None of us knew that we were signing up for this, but now that we have, there’s not a lot to complain about.

We don’t stop to think about the processes that the AI is following to get to the result as long as the result is good.

These AIs are commonplace: YouTube AI decides what I watch, Amazon/Swiggy AI decides what I buy, Apple Music/Podcasts AI decides what I listen to, Google Maps AI decides how I get to places, Gemini AI decides how to explain things to me, Facebook AI decides what kind of news I consume and so on.

And this is expected to only increase in the times ahead. The question is: Who decides what I should be told? And how do I know they’ve got it right?

These networks platforms are influential to the point of swaying public policy and elections.

So should the government have a say in how the AI of these networks operates? Or even consider nationalising large network platforms?

If a small mistake in an AI program leads to millions getting harmed, is that the government’s fault? Since traditionally it was only the government that operated at that scale.

If an AI driven network platform from one country becomes indispensable in another country. Will the country now have to treat execs from the network platform at the same level as ambassadors and diplomats?

Will the platform’s country of origin wield the platform on the international stage in the same way as they wield their arms and money to sway global geopolitics?

Will a government’s decision to ban a platform from another country on reasons of national security make them lose elections next time due to the disgruntlement it creates among its people?

Should the senior management of network platforms have civic obligations and be “re-elected” if they’ve delivered social “good” apart from just making money for their employer?

Policymakers would like to believe that “victory” means complete control over a platform, esp. one that has originated in another country.

Cyberweapons, they are unique because they can be made opaque (difficult to trace back to the attacker, making it hard to pinpoint who launched a cyberattack), deniable (countries can deny responsibility for cyberattacks, creating ambiguity about who is behind them).

Traditionally, militaries have well-defined doctrines outlining how they use force. However, the passage argues that cyberwarfare creates a problem because there are no clear doctrines for its use. This lack of clear rules makes it difficult to predict how cyberattacks will unfold or how countries will respond.

Due to the opacity and ambiguity surrounding cyberwarfare, countries might develop cyberwarfare strategies without publicly acknowledging them. This creates an additional layer of uncertainty in international relations. And this is before you add AI to the mix:

“Even if nations refrain from the widespread deployment of so-called lethal autonomous weapons — automatic or semiautomatic Al weapons that are trained and authorized to select their own targets and attack without further human authorization – AI holds the prospect of augmenting conventional, nuclear, and cyber capabilities in ways that make security relationships among rivals more challenging to predict and maintain and conflicts more difficult to limit.“

AI does have a lot to offer when it comes to defence and national security.

The authors mention how it can fly planes better, be used to analyse enemy strategy and counter it, be used to condense and transmit key information about personnel or civilians located in crisis zones much faster than a human associate could and much more.

No country can afford to ignore AI when it comes to national security though the US, China and Russia have taken a head start. The authors opine that the better approach will not be to call for disarmament, instead “United States not cede these fields — indeed, the United States should endeavor to shape them.“. Another important force in this discussion are nuclear weapons, these helped bring anxious peace to the world when nations essentially mutually deterred each other.

But this peace is not everlasting, and needs each new generation of world leader’s to appreciate the importance of the status quo.

Trigger finger reactions by nations are common during times when there is a lot of uncertainty, and well, we happen to be living in those times.

Unlike conventional weapons where disclosure does not materially diminish the weapon’s power (in fact may deter conflict), disclosing cyber weapons and AI makes them less effective as the other side can prepare defences against them (like patching loopholes in code).

Which means powers will not disclose them (unlike what happened between US and Russia during the nuclear arms race), leading to uncertainty.

Also, unlike conventional weapons where there are codified rules of conduct, cyber weapons do not have any such code, which makes an attack on civilian systems a real risk.

Even the definition of what is a true act of war in the cyber sense is unclear. “Some activities, such as intrusions into networks to collect information, may be analogous to traditional intelligence gathering – though at new scales. Other attacks – such as the election-interference campaigns on social media undertaken by Russia and other powers – are a kind of digitized propaganda, disinformation, and political meddling with a larger scope and impact than in previous eras.“.

In absence of clear vocabulary it can be said that “… major countries are engaged in a kind of cyber conflict now, though one without a readily definable nature or scope.“. Finally, unlike traditional weapons, cyber weaponry does not need anything more than a computer and a determined actor to write the lines of code to mount an attack. It becomes hard to identify the bad guy when he could be literally anyone.

Oh, and AI is emergent so even its creators beyond a point do not know what it is up to or can do. “How does one develop a strategy – offensive or defensive – for something that perceives aspects of the environment that humans may not, or may not as quickly, and that can learn and change through processes that, in some cases, exceed the pace or range of human thought?“.

National relations are like walking a tightrope and war is hugely uncertain; AI and cyber add a layer of complexity to this. Since humans like certainty and AI helps get to semblance of certainty, we may find ourselves delegating serious military decisions to it. This would be a risky idea, and a downright bad one if we do it without thoroughly understanding the potential downsides first.

It is high time to start the dialogue between major powers, esp. US and China, about how we’re going to deal with AI’s impact on the military. A degree of openness in revealing what capabilities one has will be important (though the authors cede that there will still some key secrets kept). US and China should literally enter into agreement that they will not get into a cyber war.

Major leaders should speak openly and often discuss about kinds of conflict they will not like to get into. Leading powers should define their operating manuals for cyber and AI warfare, and make them clear for other nations to see. Important that we never design an autonomous AI without a kill switch.

We humans are the most intelligent folks around here, “not anymore!” says AI.

Like the loan officer whose carefully crafted expertise over years of experience on credit extension/denial is dispensed with in a fraction of a second by AI. Like the researcher who is summarily told by AI that his research will not succeed even before he has started it. Like the MBA who is presented with superior insights within moments of feeding in the data to AI. The future is going to be a time of collective humbling of humanity.

Humanity as a whole can either take a proactive or reactive stance to the incoming wave of AI. Love it or hate it, we cannot ignore it. The authors recommend being proactive and deciding the parts we are okay giving to AI and which ones we’d like to keep for ourselves while there is still time.

Of course, any powerful technology is inherently divisive, and AI, with its potential to reshape society, may be the most divisive of them all. We can expect to see a spectrum of viewpoints emerge, ranging from complete rejection to full-fledged embrace.

Generally, those who are not tech. savvy, those who are managing others, those who are short of time should benefit from AI. Those who are in the creative field, those who make a living reviewing/interpreting data, those who make a living drafting documents into standard forms should be ready for loss of employment. Individually and at a societal level we need to think about all of the extra hours we’d have because of AI and what to do with all these people that AI just made redundant.

I heard once on a Tim Ferriss podcast that the “future belongs to the man who can work the machine“, the “machine” being AI. Indeed it is expected the specialists of tomorrow will not only be experts in their domain but also adept at transmuting their knowledge into a form that can be most effectively used by AI to solve problems.

It is also interesting to note that as humans, we now not only have the job of explaining phenomena in the real world but also in the virtual world. “Across the biological, chemical, and physical sciences, a hybrid partnership is emerging in which Al is enabling new discoveries that humans are, in response, working to understand and explain.“.

Perhaps no one will be as uniquely impacted as the children that will grow up with AI as commonplace as cars and street shops. Adults today have a worldview that pre-dates AI through which they can try to conceive of a place for AI in their lives.

But for babies and toddlers today, who have not yet formed theories about the world, AI will not just be a cool gimmick or even a useful assistant but rather an indispensable friend, advisor, guide, mentor or even, parent.

How do we deal with this? Will AI be programmed with specific values and biases that could influence children? How will it handle complex emotions or moral quandaries? Do we need regulations to ensure our kid’s safety and education on how to interact with AI responsibly? What happens when you add humanoid robots to the mix?

“The irony is that even as digitization is making an increasing amount of information available, it is diminishing the space required for deep, concentrated thought.“. Absolutely, there is so much information/content out there (not to mention how rapidly AI itself can generate it) that we’re up to our noses in it. Our trusted AI companions are likely to help us sift and sort through the deluge and present only what we “need to know”. But who decides? A trusted AI will know us better than our own mothers, which opens gates for manipulation and misguiding, while an AI designed to keep its user happy will quickly become an echo-chamber.

It’s “damned if you do, damned if you don’t” because “… when the majority of people in a society accept AI intermediation, either as a default or as the price of powering network platforms, those pursuing traditional forms of personal inquiry through research and reason may find themselves unable to keep pace with events. They will certainly find their ability to shape them progressively limited.“.

Look, we may initially try to keep all the important stuff in our control and relegate AI to purely a support function, but the pull is strong. And since we are designed to be energy conservers and selfishness has proven to be evolutionarily beneficial, yes, soon there will be people and organisations using AI for a lot more than support. So it’s foolhardy to think otherwise.

Therefore, better to be clear at the outset what roles need to be “human only” or “AI only”, and carve out regulation to keep it that way. An example could be that a nation mandates that administration of justice be the sole province of humans while also outlawing driving cars yourself because self-driven cars would be a lot safer.

And then there will be a large middle ground where humans and AI “collaborate”. The authors name a few really important things that should be kept exclusively the business of humans: democracy, elections, justice, free speech.

AI is going to open up new “vistas of knowledge and understanding.“, and at the same time make us less patient, less willing to tax our own minds and apply the critical thinking/reasoning skills that got us here. Quite the dilemma, don’t you think?

It will be invisible, like any general purpose technology, embedded in commonplace objects.

Many will be fearful of AI and think of it as “big brother”, while the developers of AI will (mostly) try and assuage concerns by being transparent about the process and the intent.

Some will treat AI’s advice as the word of a superior being and make the mistake of following through without at least trying to apply reason.

Some will fall in love with their AI. Expect there to be a lot of well meaning advice in the future about why not to trust your AI blindly.

Deepfakes and disinformation will be a real challenge. What actually happened and what was generated by AI will become harder to figure out, this will be a time of distrust.

Imagine getting a call from your Mom that she is in danger unless you transfer a large sum of money to a stranger’s back account. Are you going to bite? (keep family codewords BTW)

Almost all will be surprised with the things AI comes up with even when we try our best to define its guardrails. BTW humans are far from perfect, rational beings – this must be kept in mind and programmed in the AI so that it knows it does not need to act when an emotional human starts asking it to act irresponsibly.

A lot of AI specific roles will start emerging – we already know of the AI developer, the prompt engineer – we soon start seeing the AI ethicist, AI philosopher, AI trainer, AI explainer and more.

Thus, a lot of regulation around AI will come as well.

We’ll start seeing situations where our current rules do not apply: like if an AI driven car runs someone over, who do you sue? The owner? The primary passenger? The car manufacturer? The manufacturer of the car’s radar and other components? The telco whose 6G network the car was using?