Co-Intelligence by Ethan Mollick

Humans seem to have a fascination with creating artificial intelligent beings – in the book Mollick talks about the Mechanical Turk, and Theseus the mechanical mouse but there have been many more attempts throughout history such as in Ancient China (Mechanical Orchestra & Flying Birds, Yan Shi’s Humanoid Automaton), Hellenistic and Roman Period (automatons by Hero of Alexandria and Ctesibius of Alexandria), Islamic Golden Age (Al-Jazari’s automatons), Medieval and Renaissance Europe (Juanelo Turriano’s Monk’s automaton and Da Vinci of course), and the 18th and 19th Centuries (The “Golden Age” of Automata – Pierre Jaquet-Droz’s Androids and more).

Yet, despite this fascination we had never truly been able to create artificial intelligence, so in essence, AI was more hype than substance and interest in it waxed and waned, sometimes for a long time – times that became known as “AI Winters”.

According to Mollick, the most recent AI winter got over starting 2010 and for the first half of the decade it was all about big data combined with machine learning for making predictions – useful in a business or governmental setting but not really attracting attention from the public. AI at this time was not very good at (1) predicting the “unknown unknowns”, (2) unsupervised learning, and (3) understanding context.

Starting 2017, something called a “Transformer” got introduced in a paper published by Google researchers: “Attention Is All You Need“. The Transformer architecture was relevant for teaching AI how to understand human language, so that we could communicate with it better (and vice versa), and get it to do stuff. Prior to Transformers there were things like Markov Chain Generators, and they were not very good.

Transformers was able to better address the three problems with AI listed in the previous section using an “attention mechanism” – the AI was able to naturally understand what parts of a sentence were more important contextually and generate (predict) the next sequence of words accordingly.

AIs based on the Transformer model became known as “Large Language Models” (LLMs) – LLMs are still like the older AI in the sense that they are essentially prediction machines – but much better. Transformers were like a phase shift in text prediction.

It was at this time that the word “AI” starting becoming lingua franca, and it was the Transformer architecture that was behind ChatGPT exploding on to the scene in 2022 (BTW the “T” in ChatGPT stands for “Transformer”).

LLMs are part of a larger group of AI called “Generative AI“. Generative AI has been everywhere in the news since end-2022 because of how it is the first thing we’ve ever created that actually seems “intelligent”.

That is what this book is about. About how we should be navigating the coming years as Generative AI proliferates (or infiltrates?) more into our lives. It is not a book about other forms of AI such as Predictive AI, or about AI crossing over into the physical world such as with robots and autonomous cars. There is such a lovefest around the word “AI” these days that it is becoming critical to talk about the exact kind of AI one is talking about.

To understand how Generative AI might impact the workplace it is important to understand how LLMs work.

Though LLMs are a subset of Generative AI, the other ones being Diffusion Models and Generative Adversarial Networks (GANs), they often play a crucial supporting role even in these non-text tools.

For example: When you type a text prompt to an image generator (e.g., “a cat wearing a superhero costume, flying over a futuristic city”), an LLM or a similar natural language processing component is often used to understand your text prompt and translate that meaning into a format that the image generation model can work with.

LLMs employ unsupervised learning – which means they are given huge amounts of text and are told to study the connections between the words. The more certain words appear together (for instance if the LLM sees that “man” generally follows “the”) the greater “weight” the LLM assigns to that connection.

In reading humongous amounts of text the LLM is able to build billions or even trillions of such connections and assign weights to each of them.

Now, because the training is unsupervised you want to make sure that the LLM has enough source material to be able to differentiate the signal from the noise, that is, not pick up accidental or spurious connections/weights as it learns, that is why LLMs are trained on anything and everything that AI developers can get their hands on legally (at least that’s the goal) – from publicly available documents to discussions on internet forums to old text no longer under copyright. BTW lately, the internet is running out of quality training data, that is why people are trying to get it to train on synthetic data.

This initial phase of the training the LLM is called “Pre-training“, the “P” in ChatGPT. The LLM also undergoes further rounds of “fine tuning” where actual humans get involved – called Reinforcement Learning from Human Feedback (RLHF).

As the LLM recognises patterns and connections between different words, and assigns weights to each connection, it essentially becomes a very sophisticated version of “auto-complete”.

And the way it answers our queries is by looking at our query (a.k.a. prompt), the connections between the words from its pre-training/fine-tuning phase (and the weight of each connection), and then guessing what combination of words is most likely to follow.

LLMs generate a large number of potential responses from the connections they saw in the training data and pick the one(s) most apt given the weights of the connections. This is the “G” in ChatGPT, “Generative”.

We design the LLM’s architecture and the learning process, but the exact weights are learned by the AI itself. If the weights it assigns are even slightly different, that can impact the final output it produces as a response to a prompt.

“No one programmed these weights; instead, they are learned by the AI itself during its training.“.

That is why AI generated content is a legal grey zone right now – because the AI does not copy the actual words from its training data, it only uses the training data to assign weights to the connections between words.

This is very different from all other types software we have developed till now where we are able to predict the exact behaviour of the software. In LLMs, the specific pathways and the emergent properties are not explicitly programmed.

Emergence is happening in AI and that is the weird part, “They shouldn’t be able to play chess or demonstrate empathy better than a human, but they do.“, software did not show emergence until now.

So, in essense, with LLMs we’ve got a tool that can identify connections between things (words, for now), come up with potential options based on what it has seen in the past, and generate output that is a little different each time – not precisely accurate but directionally correct.

In this way LLMs are “connection machines“, you know what else are connection machines?

Human brains.

Our brains too are able to identify connections between things (what we call creativity), come up with a few options (what we call imagination), and generate output that is generally similar each time (if that’s what we’re going for).

So, the current wave of AI, that is Generative AI, is going to impact tasks that require creativity, imagination and where it is okay if the output is directionally correct.

And that is what seems to be happening with.

The pace of progress in artificial intelligence, as Mollick underscores, is not just rapid; it’s accelerating in ways that can be difficult to fully comprehend.

This isn’t merely another incremental step in technology. While past technological advancements primarily provided us with better tools to do things more efficiently, the current wave of AI, particularly generative AI, introduces something more profound: systems that can perform tasks historically requiring human-like cognition, creativity, and complex understanding.

It’s this emergent ability to generate novel content, engage in sophisticated reasoning (based on patterns in its training), and even exhibit flashes of what looks like creative insight that makes this era feel distinctly different and more transformative.

The evidence for this leap is compelling and tangible. Benchmarks that once seemed like distant, theoretical goalposts are now being met or surpassed. For instance, AI systems are increasingly engaging in conversations that can feel indistinguishable from human ones (Turing Test), and some argue they are beginning to demonstrate novel outputs that their creators cannot fully explain (Lovelace Test for creativity). Beyond these conceptual tests, AIs are now achieving world-class human performance levels in highly specialized and demanding professional domains—passing rigorous exams for lawyers, doctors (like components of neurosurgeon qualifying exams), sommeliers, and advanced academic subjects. This isn’t just about doing things faster; it’s about matching or exceeding expert human capabilities in areas requiring deep knowledge and complex application.

Yet, as Mollick rightly points out, this field is so cutting-edge, so “bleeding edge,” that even the brilliant minds architecting these AI systems often express a profound humility about predicting its long-term trajectory.

For all its brilliance, Generative AI fails at some pretty easy tasks, Mollick gives the example of tic-tac-toe.

There isn’t a clear, universally agreed-upon roadmap for how all these capabilities will evolve or precisely how they will integrate and reshape our world.

We are, in many ways, learning about the full scope of AI’s emergent abilities and societal implications in real-time, making this a period of both unprecedented opportunity and necessary caution.

Beyond the immediate legal grey areas of whether Generative AI output constitutes plagiarism or who owns the copyright, Mollick delves into more deeply embedded challenges.

A primary one is AI bias. He points out that the vast datasets used for pre-training these models often represent an “… odd slice of human data“. It’s frequently sourced from what AI developers—historically a relatively homogenous group, often male computer scientists from predominantly American and English-speaking backgrounds—could readily find, access, and perhaps assumed was free to use. This very selection process, guided by their own inherent perspectives and limitations, can introduce a significant initial skew to what the AI learns about the world.

The problem of bias is then compounded by the content of the training data itself. As Mollick highlights, much of it comes from the open web, a space hardly renowned for being consistently “nontoxic, friendly” or representative of balanced global perspectives. Consequently, the AI models absorb the biases and prejudices present in our collective online writings and conversations.

The RLHF (fine-tuning) stage, designed to make AI more helpful and aligned, introduces its own layers of ethical questions and potential biases. Often the challenging and potentially distressing “dirty work” of filtering and labelling AI content might be outsourced to individuals in “low-cost” countries, like Kenya, raising ethical questions about exploitation and psychological impact. Furthermore, the human raters involved in RLHF bring their own cultural and individual biases to the table.

Plus, RLHF is not foolproof; these systems can still be manipulated through “prompt injection” or “jailbreaking” to bypass intended safeguards.

This leads to broader concerns about misuse. The power of LLMs means that “Even amateurs can now apply LLMs for widespread digital deception.” This democratisation of sophisticated content generation tools also democratises the potential for creating convincing misinformation, scams, or other forms of harmful digital content.

Looking at the bigger picture, the “alignment problem”—ensuring that future, more advanced AI (like Artificial General Intelligence or AGI, and potentially Artificial Superintelligence or ASI) acts in ways that are beneficial and aligned with human values—is a profound challenge that has led to calls from many quarters for a slowdown or even a ban on certain types of AI research. The ease with which current safety checks can be circumvented in existing models only amplifies these fears, not just about intentional misuse by bad actors, but also the potential for well-intentioned users to accidentally cause harm.

Given these multifaceted challenges Mollick’s perspective underscores that no single entity can tackle these issues alone. AI companies have a responsibility but can’t be the sole problem-solvers. Government regulation often struggles to keep up and may even stifle beneficial innovation. The most constructive path forward is a broad societal response involving companies, governments, researchers, and an informed public actively participating in shaping how these transformative tools are integrated into our lives.

As the future of work changes with Generative AI and Mollick recommends four principles that you should keep in mind.

#1 Always invite AI to the table

Don’t be afraid of it, instead use it in different ways and see what it can do. Experiment and explore.

“… familiarizing yourself with AI’s capabilities allows you to better understand how it can assist you—or threaten you and your job.“.

In doing this you will be able to understand more about what things AI can do well, and what things it cannot do that well – what Mollick calls the “Jagged Frontier of AI” and modify your workflow (or career strategy) accordingly.

Expand your use – not just text but video, image and audio generation too. See what other people are doing with it, lots of people are sharing their AI uses on YouTube and other places.

The thing about the latest wave of AI (and the speed at which things are moving) is that individuals are much better placed than corporations to ride it. Companies are often constrained by internal processes and review boards that only meet a few times a year to deice whether or not to adopt a particular piece of technology, in that way they are much slower than an individual who decides to adopt and learn about new tech.

Being proactive about AI will allow you to stay one step ahead of curve. Plus AI does not “think” like you do, so having this “alien” perspective is often useful. I personally use AI a lot here at Sunchaser, and although I remain committed to the human-made principle, using AI has a brainstormer has been helpful.

#2 Be the HITL

“HITL” stands for “Human In The Loop”, and with this Mollick basically means that AI output today is not good enough to be blindly trusted. He talks about a lawyer who used ChatGPT to conduct research for him and in the process ChatGPT hallucinated past cases that did not happen.

Because AI is so good at sounding coherent, we seem to take what it is saying as the truth and do not bother to fact check. “LLMs can seem more impressive than they actually are because they are so good at producing answers that sound correct …”.

Mollick cites cases of people “falling asleep at the wheel“, that is when workers over relied on AI generated output and unquestioningly accepted it. This at best leads to sub-optimal results or at worse can lead to financial loss or even loss of life. Be the HITL, provide oversight, critical thinking and ethical viewpoints.

#3 Know that better AI is coming

Better AI that can do a lot more and with a lot more reliability than what is currently possible. Mollick talks about how the growth in capability is likely going to continue and in the near future AI Agents are going to be all the rage. Coincidentally right when I was writing this blog, Google in its 2025 I/O conference introduced “Agent Mode” in Gemini.

I too can see the improvement in capability: I was using Google Gemini till May last year (2024), and in between switched over to Microsoft’s Copilot, now when I have started using it again – I can see clearly that it remembers context much better and are is to make coherent arguments out of multi-layer discussions.

Of course, with powerful AI taking over or at least significantly supplementing much of what we do (first in the knowledge space, and then in the real world with robotics), we’re going to need to have some deep conversations about meaning and purpose.

#4 Treat it like a person

I think this is something that is easily done anyway – the way a ChatGPT or a Google Gemini works, so fluent with language, it is super easy to imagine that you’re talking to another person. Which means you should be taking a conversational tone with it, as you would take with a friend or colleague. Treating it like software where you have to clarify every last bit may be unnecessary.

And Mollick has an important addition to this: Tell the AI what kind of person it is. Super important and useful. Without this you risk getting generic and trite responses.

To get better responses, tell the AI:

1. What kind of persona you want it to take.
2. What is that person’s relation with you – peer, friend, boss, mentor, subordinate, intern etc.
3. What constraints should it know about.

Like I explained above, generative AI is good at being directionally correct and generating a lot of options. But it may not get the precise details right.

When AI makes big mistakes we are often able to catch them, but when it is directionally correct we think it’s got the details right as well – this is a mistake. It is the “small hallucinations” of AI that come back to haunt you.

So be careful in using Generative AI where you need perfect accuracy. But this very bug for accurate work is a feature when it comes to creative work.

AI, the connection machine that it is, is by its fundamental nature a boon to creativity. Mollick cites studies that show how a team of humans working with AI are able to generate higher quality solutions to a problems than either acting alone.

“… researchers have argued that it is the jobs with the most creative tasks, rather than the most repetitive, that tend to be most impacted by the new wave of AI.“.

Use AI to augment your creativity, especially when it comes to the idea generation/brain storming phase.

Using AI you should generate a large number of ideas and perspectives. Expect most of these to be mediocre or impractical, but that is where you come in as the HITL – pick up the few good ideas from the lot and build on them, potentially using AI as a sounding board. “Fortunately, we are good at filtering out low-quality ideas, so if we can generate novel ideas quickly and at low cost, we are more likely to generate at least some high-quality gems.“.

The downside of this freely available creativity is that people will start using it with reckless abandon and stop doing the hard work themselves¹ – Mollick calls this “The Button” and says that everyone is going to use it in the future – creativity on tap. Software companies are also putting the button in many places. I’ve shown a few examples as of today:

Examples abound on the internet of this misuse – there are many AI generated books on Amazon, countless AI generated images made in the style of a particular artist (as I’m writing this a “Ghibli” wave seems to be washing over us), there are AI generated sound bytes and even full length music tracks. I think humanity is going to be jaded with AI generated content in the coming years and there may actually be a resurgence of demand for human made content.

I think that people will have to ask themselves – Why should I create myself when I can just push The Button?

And as AI acquires more powerful agentic capabilities, going beyond just generating text/images to things like setting our daily calendar, to do lists, managing our emails, buying our groceries online, booking concert tickets and so on, the pull of The Button will be ever stronger.

They will need to think about their motivations – do they just want the thing done or actually do something².

But to gain real benefit from The Button you need to know what to ask for, and that you will get to understand only when you follow principle #1 mentioned above: Invite AI to everything and test its limits.

Indeed, Mollick talks about the ability to work effectively with AI becoming a “CV-point”, you know, like how “Proficient in MS Excel” used to be.

“Situations in which there is no right answer, where invention matters and small errors can be caught by expert users, abound. Marketing writing, performance reviews, strategic memos—all these are within the capability of AI because they have both room for interpretation and are relatively easily fact-checked.“.

Unfortunately for me, “AI overlaps most with the most highly compensated, highly creative, and highly educated work.” – too bad, I liked being paid.

The domain where AI is not yet going to impact (yet), is physical work or more generally interacting with the physical work we inhabit. AI is still very much a digital being. This may change in the future with robotics. In fact, I heard a Sam Altman interview recently where he talked about robots going from a “curiosity to a serious economic creator of value” in 2027.

But coming back to AI’s impact on knowledge work: If you are a knowledge worker don’t start planning for life as a farmer yet, thinking that AI is coming after your job. Because your job is really a bundle of tasks, and though AI may be good at doing certain tasks but that does not mean that it can do your entire job.

Even if at a future date AI gets so good that it can do your job – know that your job fits into a larger “system” – the organisation you work in, the economy the organisation functions in and so on. This larger “system” must also change structurally for the AI to fully integrate in it. This is not going to happen in a hurry.

This doesn’t mean there is no risk, but rather that you don’t need to lose sleep over it. Follow the four principles mentioned above to consistently adapt to the new way of working.

Coming to the immediate future, here is what Mollick recommends:

1. You must be the HITL and not completely rely on AI. “When the AI is very good, humans have no reason to work hard and pay attention. They let the AI take over instead of using it as a tool, which can hurt human learning, skill development, and productivity.“.
2. Start working like a “Cyborg” or at least like a “Centaur“: A Centaur is a knowledge woker who defines a clear line of demarcation between what work AI does and what they do. While a Cyborg is someone whose workflow has them “intertwining” with AI, going back and forth in the process with no clear line of demarcation.

3. Divide the tasks that your job needs you to do into:
   • “Just Me Tasks“: These are tasks where AI today is not very helpful. Or the ones you want to do yourself for personal, professional, moral, or ethical reasons. For example, writing on Sunchaser is a “Just Me Task” for me.
   • “Delegated Tasks“: These are things you give to AI to handle but keep a check on the output and the process. These things are likely to be boring things that sap our energy and need a lot of time. Like sorting emails or tracking expenses or filing taxes – these are the kinds of tasks that you may delegate to AI but keep a watch on the output and the process from time to time.
   • “Automated Tasks“: These are tasks tha you completely hand over to AI, admittedly this is a short list today, but as you explore the “Jagged Frontier” you may find things to include here.

4. Make your AI team and assign them clear roles: Like I showed you above, my own little AI team has a editor, a social media manager, a sage, an executive coach, a researcher – each whose job description (i.e. prompt) I clarified at the beginning.

AI as the great Leveller in Knowledge work

Mollick calls AI a “great leveler” in how it boosts the creativity and productivity of the less skilled knowledge worker. Gains are across the board, don’t get me wrong, but the main beneficiaries are those who were not as skilled. UBI and 4DWW may not be so far off after all: “With lower-cost workers doing the same work in less time, mass unemployment, or at least underemployment, becomes more likely, and we may see the need for policy solutions, like a four-day workweek or universal basic income, that reduce the floor for human welfare.“.

AI is super helpful in tutoring because it is able to personally adapt to each student, meeting them where they are.

Though LLMs may have started with students using them to cheat but now more institutions are building around this new reality such as asking students to critique or fact check AI generated content.

Like the calculator, it may initially been as a thing that will harm but will end up being something that allows us to think at higher levels – our institutes will have to think about what kind of AI use is acceptable and what is not, just like how we have exams in which one can bring a calculator and ones where one cannot.

In the book, Mollick gives a few examples of how he himself has included AI in his pedagogy.

There is no avoiding the fact that AI is going to be part of our future, and so Mollick recommends teaching kids (apart from the general public) the strengths and weaknesses of AI. He talks about teaching kids how to use AI more effectively, how to be the HITL and how to have the critical thinking to not accept AI output blindly. He talks about how in the future the classroom might get “flipped”, as kids learn at home with their personalised AI teachers and then apply the concepts in a group setting in school. What becomes the unique role of the human teacher? Perhaps focusing on mentorship?

AI will also help kids answer the most fundamental question (one that I wish had been exposed to growing up): Why should a I bother learning this? I think this will be one of the central questions in a world where personal AI assistants, perhaps always with us through a pair of glasses or earbuds as in the movie “Her”, can give us instantaneous answers to any question.

Keeping aside philosophy, AI will also give students the more practical answer to this question as well – For instance, it can tell a student wanting a career in stock trading the importance of learning human psychology by showing how emotions drive the market much more than he might think.

With our rapidly changing world though, education is not going to be one-time thing any longer. The era of “learn till your 20s and earn earn till your 60s” is well past us – “lifelong learning” is going to be critical for both kinds of people: those seeking to make an impact in our AI future, as well as those seeking protection from its challenges.

And one of the ways we learn, especially after we pass out of school, is the mentoring we receive from other experts. This is where Mollick points out a challenge: it seems that a major skills gap is about to emerge because of how incentives are stacked today both for experts and novices.

Because in training the next generation of lawyers, or doctors, or architects – the expert spends valuable time teaching a greenhorn and patiently waiting as the greenhorn makes mistakes and learns from them.

The AI today are (and AI agents tomorrow will) already be at a decent level of expertise given their massive training data, the expert may be compelled to just delegate the tasks to AI for the sake of efficiency (or because they are stretched for time), in this way compromising the learning of the next generation.

The other way people become experts by going through the grind, but AI is automating (or making simpler) these exact things as well.

For instance, I remember when I started out as a young MBA, we had to learn how Excel worked, its various formulas and features, and how they were to be applied for data analysis – today, the AI in these spreadsheet tools just does these things for the user, it is able to automate simple functions like summarise or counting formulas and so youngsters do not have to learn this any more.

Mollick calls this the “paradox of knowledge acquisition in the age of AI“: the more we become cyborgs the more we need to remember to remain human in the fundamental areas.

What areas? Perhaps foundational knowledge, fundamental truths, critical thinking, ethical judgment, deep empathy, complex problem-framing (as opposed to just problem-solving), and the very act of asking insightful questions – skills that AI currently augments rather than originates.

But perhaps the death of expertise will not happen because of AI because people will start following their natural curiosity more and decide to do the grunt work in those areas after all? Because of inherent meaning or desire for learning?

Could AI itself help in this? I am optimistic that it will play a role in new forms of mentorship or skill development, even as it disrupts old ones. And perhaps, human-to-human mentorship, just like teaching, will need to evolve to focus on different skills – how to effectively use AI and when to not, how to ask the right questions, ethical considerations etc.

Because AI is trained to talk in human language it may be better to treat it like another person while interacting with it to get better quality of output.

But such a treatment does raise some interesting societal questions about the era we are entering in.

The first one of course being: Is AI conscious? The prevailing notion is that it is not, it just appears to be through the fluidity of its prose and its emergent properties (although one can say that consciousness itself is emergent, you know, “the mind is what the brain does”).

Even if we consider AI as not having consciousness, does it even matter? AI is demonstrating its usefulness irrespective of this debate and in some unique ways too – Mollick talks about Replika, the generative chatbot that many across the world developed “intimate” relations with its users. The creators of AI design it so that it keeps us happy, this feeling that AI “gets me” can lead many to get emotionally close to a chatbot. I am reminded of the Google engineer who started believing that AI was conscious and raised ethical concerns around how Google was handling its development.

The era of personal AI companions has started and I think it is only going to increase from here, so there’s also the question of personal echo chambers and social isolation resulting from this.

Will we have a future where people decide to forego traditional social institutions like marriage and family and instead choose AI companionship?

Will it be okay to have AI bosses/managers and AI politicians in the future?

Would about AI therapists? Mollick talks about how AI is able to demonstrate empathy better than humans, about how it can even act like a therapist (which may not be a bad thing after all).

Quite an exciting time ahead for all of us – human or AI. The future is going to be interesting.

Possibility #1: As good as it gets

This is a future where AI stops getting better and its capabilities stagnate. This is unlikely to happen from a technical point of view but may happen due to overreactive and overreaching legislation and policy that stifles development.

Mollick says that this scenario is the most unlikely, yet the one that most number of people and organisations are planning for – perhaps because the pace of change is too fast for us as humans.

Despite this being the least disruptive of the four futures, it still has its concerns.

Misinformation is a real problem in this world. Bad actors are able to use Gen AI to spread more of it in the world and faster. It becomes hard to know what to believe and so people may stop watching the news altogether. News media tries super hard to convince people that what they’re saying is the “real” truth.

Interacting more with AI than humans becomes a real thing in this world (especially as the global population starts to decline in the second half of the century).

This also leads to hyper-personalised filter bubbles because AIs will be designed to keep us happy (how social algorithms are designed to keep us “engaged”).

At the workplace too AI starts getting used more prominently and being able to work with AI becomes a critical skill for the future worker.

Possibility #2: Things happen slowly

In this future the growth in AI capability continues but at a pace that we’re able to keep track of – so, instead of the “order of magnitude” improvements in AI models we see today, say from GPT 2 to GPT 3, the rate slows down to “… 20 percent, a year …”.

The challenges of AI as mentioned in possibility #1 still remain but the worst outcomes do not come to pass as organisations and governments can see them coming. Laws around AI use and disclosure develop in time to ensure human remain in the driver’s seat. And so do social norms around what’s okay to do with AI and what is creepy.

The knowledge worker sees how the tasks she used to do previously have been completely taken over by AI, but this frees her up to focus on higher order work. Her job evolves and along with that so do her skills rather than being rendered obsolete.

AI starts helping a lot in science and researchers are able to use it to push the frontiers of knowledge as well as alleviate the “burden” of knowledge (that is, the amount of things you have to learn before you reach a point where you can actually do original work).

By and large, this future of linear growth is a good one, or at least not such as bad one.

Possibility #3: Things happen fast

That is, exponentially fast. Exponential growth may happen because good AI is used to develop better AI and so on.

Humans don’t get exponential growth and so in this future are not able to comprehend the pace of change. The same things happen as in possibility #2 but because they are happening so fast, we let the AI handle it. Mollick talks about how this is the world of AIs, an “AI-tocracy“, where “good” AI fights “bad” AI and the “good” humans try to be helpful but mostly just hope for the best.

Because AI is so powerful and so quickly evolving, governments around the world take it super seriously – ubiquitous surveillance (again powered by AI) becomes the norm as governments track citizens and highlight potential risks i.e. bad actors who themselves use AI. I think the time constant surveillance is also going birth a rebellion against it, like in the video game “Watch Dogs 2”.

Loneliness, surprisingly, becomes less of an issue because in this world it is totally okay to have AI partners, like how Samatha was to Theodore in “Her”. Having AI coworkers and bosses is completely normal too. So is having AI doctors and therapists.

This is also the future where AI actually comes for our jobs as the AI itself is very powerful, plus it powers rapid development in robotics thus entering the physical world. Initially as humanoid robots (because our world today is built for robots) and then as custom forms as we start letting certain parts of our world be completely taken over by robots (say, in factories and warehouses).

“What are humans good for?” becomes a real question and governments globally start thinking about UBI.

This future also acts as an off ramp to AGI.

Possibility #4: Machine God

In this future, machines reach AGI, start to rule the planet and human supremacy ends. This is where the alignment problems becomes central and humanity gets to see if it passed the test.

It may not be all bad and actually result in a “eucatastrophe” – a time where things become good beyond belief, a world of true abundance. Just like the first possibility though, this too is an unlikely one.

Again, no one knows what is going to happen, the best approach to be will be to stay informed and play on the “Jagged Frontier”.