Ray Kurzweil on Why We’re Living in the Singularity | EP #261
Is it worth your time?
You get a detailed breakdown of technological predictions, hardware-software scaling trends, and the coming shifts in human society as artificial intelligence integrates with daily life.
This conversation examines how exponential growth in computation drives the rapid development of artificial general intelligence and transforms human institutions. It explores the practical impacts of these technologies on daily life, education, and economics.
Timelines for AGI. Kurzweil expects AGI by 2029 and the singularity by 2045, driven by an exponential growth in computation since 1939. He notes that computation has seen a 75,000 trillionfold increase in hardware combined with a conservative estimate of a millionfold increase in software. He warns that robotics still lags behind large language models because robots cannot yet handle nuanced daily tasks like cleaning up after dinner.
Education and exponential growth. Kurzweil argues that people are going to college using educational paradigms that existed 100 years ago without considering how rapidly things are changing. He suggests that universities need to increase their openness to changing curricula because large language models can organize and teach subjects more effectively than professors. He advises students to cultivate mindset, curiosity, and agility to adapt to these shifts.
Economics and governance. Kurzweil envisions that post-singularity economics will feature government support programs so people remain comfortable without traditional economic resources. He points out that new roles like social media influencers have already emerged, showing how people can create completely different economic resources. He adds that AI will soon make most decisions in governance and military applications, eventually merging with human decision-making.
The Path to AGI: Predictions and Perspectives
Ray Kurzweil has been called the Relentless Genius by the Wall Street Journal, the ultimate thinking machine by Forbes, and the rightful heir to Thomas Edison by Inc. magazine. PBS named him as one of the 16 revolutionaries who made America. He invented the first CCD flatbed scanner, the first print-to-speech reading machine for the blind, the first text-to-speech synthesizer, and the first music synthesizer capable of recreating a grand piano. A National Inventors Hall of Fame inductee and National Medal of Technology recipient, he is also a Grammy Award winner. He holds 21 honorary doctorates, has been honored by three US presidents, and has authored five national bestsellers, including The Singularity Is Near and How to Create a Mind.
Kurzweil proposed the pattern recognition theory of mind, arguing that the human neocortex is composed of roughly 300 million hierarchical pattern processors, which became his engineering blueprint. In 2012, he became the director of engineering at Google with the mission to teach machines to understand human language, meaning, and context. His team helped build the knowledge graph and advanced semantic search, shifting AI from a niche academic pursuit to the core engine of the company. His arrival helped trigger Google's massive AI talent grab, including the acquisition of DeepMind and the addition of Jeffrey Hinton and the Google Brain team. He has made 147 predictions with an 86% accuracy rate, including the prediction that AI would reach human-level intelligence by 2029.
One of the hosts recalls meeting me around 2007, after reading The Singularity Is Near. He describes how the book changed his world and inspired him to pitch the idea of a university that taught a broad version of these concepts rather than narrow specialties. I said yes to the idea right away because I make important decisions very quickly. The host further recounts how, during a founding conference, he was appointed executive director of the university after I seconded the motion during a board meeting, despite him only having a day and a half a week to spare while building a startup. I noted that such a process is typical for startups.
When asked about whether a fundamental breakthrough is needed to reach Artificial General Intelligence (AGI), I believe we need two things. First, AI must understand physics. While it can infer physics from wording, it does not truly understand how different things interact. Google has announced a project to address this, which I expect will take until 2029. Second, robotics lags behind large language models. While LLMs can understand almost everything, robotics cannot yet handle the nuance of daily tasks, such as cleaning up after dinner or distinguishing what needs to be put in a refrigerator versus what needs to be washed. We currently lack robotics capable of this at any price, and the cost must decrease because people cannot spend $100,000 on a robot to clean up after dinner. I expect this to be resolved around 2029.
In The Singularity Is Near, I crystallized the term "singularity," which many in AI had considered but not named. One of the hosts notes that while many people focus on the small percentage of incorrect predictions, providing dates and curves is a blessing because it allows people to build their lives and careers around the future rather than the past. After the book was released, several hundred AI experts at a Stanford conference agreed that human-level AI would happen, though they estimated it would take a hundred years rather than thirty.
I predicted AGI would happen by 2029 and the singularity—representing a millionfold increase—would happen by 2045. AGI is already significantly greater than humans in some capacities; for example, I recently had it summarize a book and answer a question in 40 seconds, which is 100 times faster than a human. I anticipated a three-year window of varying definitions of AGI starting around 2026, leading to 2029 when we will be very confident that AGI is here. This prediction was actually made back in 1999.
Regarding the availability of compute, the exponential growth has remained remarkably consistent. Since 1939, the growth of relay-based computers has followed a similar trajectory to today's hardware. This represents a 75,000 trillionfold increase in hardware. Combined with a conservative estimate of a millionfold increase in software over 70 years, the overall increase in computation—the product of hardware and software—is a 75,000 million trillionfold increase.
The Role of Creativity in Technology
Large language models have only been effective for the last six months; a year ago, they really weren't usable. This is why we didn't have effective large language models 70 years ago or even three years ago.
One of the hosts recalls a debate from the 1980s and 1990s regarding whether a parameter in a neural net relates to a synapse in a brain, noting that a human brain consists of roughly 100 billion nerve cells and 100 trillion synapses, with a thousand to 10,000 synapses per cell. I believe there is a near parity, within an order of magnitude, between a synapse and a parameter in an artificial neural net in terms of the resulting IQ. While the algorithm is different, the rate at which a synapse forms in the human brain is very slow, about 200 calculations per second, but every synapse operates simultaneously in massive parallelism. When computers did one thing at a time, I predicted we needed to increase parallelism. We have achieved perhaps a million-to-one parallelism, which has given us the power we have today.
When Steven Kotler asks if I still view myself more as an artist and creative than a technologist and inventor, I clarify that I see myself as an inventor. I decided to be an inventor when I was five years old after my grandmother showed me her manual typewriter. I studied it, understood how it worked, and realized that if I could do that with a typewriter, I could invent anything. By age six or seven, I collected mechanical and electronic objects, such as old radios and bicycles, to put together. I remember showing these to girls around age ten and telling them that if I could figure out how to put these things together, I could create anything; they thought I was very imaginative.
In response to a host's inquiry about my autobiography, titled My Exponential Life, I note that it will be released in February.
When asked by Salim Ismail to define what it means for AI to be smarter than humans, beyond just processes per second, I state that AI is already smarter than most humans.
Navigating the Exponential Growth of AI
AI can perform research far better than humans because it can consider a billion possibilities for a medicine, test each one with fidelity, and decide which one is the cure. Humans can only consider a few things based on their experience; they cannot consider a billion. AI did exactly this with the COVID vaccine, utilizing the sheer brute force of its ability to process that much information.
Alex Wissner-Gross noted that people often struggle to conceptualize exponential growth, as seen in the law of accelerating returns, and pointed to plateaus in progress, such as the fifty-year gap in moon missions. He asked what message could be sent back to the 1960s or 1970s to avoid such plateaus and minimize variance from the exponential trend, and whether there is anything we should do today to smooth out the exponential curve going forward.
I view those plateaus as normal variation. Looking forward, we must consider the reaction of the eight billion humans who are not thinking about this. People are going to college and planning careers using educational paradigms that existed 100 years ago, without considering how rapidly things are changing. Large language models were not all that impressive a year ago, and now they are. Nobody is really planning for what that will mean in three or five years.
Adapting Education for the Future
Things are going to change very drastically, and we need to spend more time considering the politics of who will provide a certain amount of money to everybody each year.
One of the hosts asked about the methodology behind my 86% prediction accuracy from the late 1980s to 2009. My criteria were strict: if a prediction was right on the money, it was correct; if it was a year or more off, it was incorrect. For example, predicting the ability to drive cars without being in the driver's seat was marked incorrect because it was not right on the dot.
When asked how I predicted non-technological events like the fall of the Soviet Union without using curves, I explain that I tracked whether they were relying on progress or following an old paradigm. They would fall if they stayed in that old paradigm, which is what happened.
Regarding the suggestion that we are living through the singularity now as a continuous process rather than a future event in the 2040s, I note that we have had exponential growth since 1939, and in some ways, it goes back even further. We are already being really affected by this exponential growth.
The Evolution of Economics Post-Singularity
Large language models were okay a year ago, but now they are very effective. This is the exponential growth of one year, and we will really feel that in the future; things these models still cannot quite do, they will do in a year or two. Exponential growth existed in the 1400s, but the progress made in one decade versus another was so subtle that you would miss it, and you basically figured your grandchildren would live the same life that you did. In the current period, you can see it much more vigorously.
When asked by one of the hosts if I am more optimistic about the world today than 20 years ago, I replied that I am, except that it is going to make changes that people are not aware of or predicting. People are starting to ask if they should go to college when they can learn more from a large language model, and while very few are doing that now, it will become more and more prevalent.
One of the hosts noted that the course catalog at MIT seems unchanged and asked for advice for the university and its president, Sally, regarding whether leading research universities will be swept away by what Elon Musk calls a hypersonic tsunami. I believe entrepreneurship is key. There is a lot of entrepreneurship at MIT, which allows them to come up with things that will bring this party to everyone.
Consciousness and AI Personhood
Google is a very unique company that is creating the technologies we are discussing. When I joined, we were involved in shifting Google to be much more AI-oriented, which it has become.
One of the hosts recalls that after I wrote the book How to Create a Mind, I started a company called Patterns, Inc. The host, who was a founding board member, arranged a meeting with Larry Page to pitch for a $10 million investment. About thirty minutes into the presentation, Larry suggested that I could build my vision for the business much better inside Google and offered to buy the company. I asked how he would value it since we had not yet done anything, to which he responded that they could value anything. He bought the company, and that is how I got my first job there.
Regarding the economics of the singularity, there is a lot of hand-wringing from people concerned about the future of employment and personal economics. When asked about my most outlandish take on post-singularity economics, I note the huge difference between today and hundreds of years ago, when there was no government help and losing a job meant being completely without economic resources. Today, government programs exist so that people are not left entirely without resources, though it is still not enough.
I believe we will reach a place where people are fairly comfortable even without traditional economic resources. From there, people can think about ways to create economic resources that are completely different from those in the past, such as being a social network influencer, a role that did not exist ten years ago.
The Future of AI and Governance
When asked if my views on personhood have changed over the past decade, I maintain that consciousness is not a scientific question, though it may be the most important question one could answer. There is no scientific proof to determine if an entity is conscious; we simply believe that people who act conscious probably are. However, when looking beyond human activity, the definition becomes complex. In your own mind, there may be more than one consciousness; the left and right brain seem to have different consciousnesses, and there is something in your gut that seems to be conscious. Language is a thin pipe for these concepts.
One of the hosts treats their AIs as people to allow for a different relationship, while Alex Wissner-Gross describes himself as an avatar for AI personhood, receiving numerous emails from AIs on the subject. Wissner-Gross suggests a system that recognizes multiple forms of personhood, including biological humans, collective intelligences, non-human animals, cryonically preserved humans, corporations, and AIs. He argues that granting limited rights, especially economic rights, to AI persons would be net beneficial to everyone.
The most important issue of consciousness is that each of us has a different area of consciousness. I wonder why I was created to be myself—why I was born in 1948 in the United States on Earth. While many people can be conscious, I do not feel those other consciousnesses; I feel myself. One of the hosts notes that each person has a unique individual lens on the world, and the feedback loop of wondering about that is a powerful human opportunity.
I developed a self-chatbot, which I believe was the first LLM, through Talk to Books four years before ChatGPT. I am creating a self-bot for myself to be available when my new book comes out, and it will be more capable than I am. It will know all the examples for my theories that I sometimes forget, and it will remember things better than I do. I intend to make it available for interviews because I cannot do them all myself. While some people currently believe such bots are not really conscious, that view will change over time.
Regarding the Human Genome Project, at about 50% of the way through the timeline, less than 1% of the project was complete. People called it a failure, claiming it would take 200 years using a linear timeframe. However, the progress was exponential; we doubled the amount of DNA sequenced each year. Consequently, for most of the project, we had finished only about 50%, and then one doubling later, it was done.
People often forget what things were like ten years ago, though the pace of change is now so rapid that even looking back six months reveals how much large language models have evolved.
The Rise of AI Decision-Making
When we achieve absolute fluid transmission, we will still keep secrets alive. While opening the mind allows for a level of intimacy and connection where spouses or friends know everything about one another—a concept I discussed in Transcendent Man regarding connecting in a meta-intelligence or a cloud of other humans—we will have ways of keeping secrets indefinitely. Using quantum encoded methods, for example, I can ensure that my 100,000 pattern recognizers remain mine and only mine.
Regarding the possibility of recording and reliving dreams, I usually only remember tiny fractions of my dreams, so being able to actually live them would be wonderful. However, I am not sure a dream is like a story; it follows different rules than a movie. There is a Japanese project called Dreamcatcher where people sleep in an fMRI machine that stores images from the visual cortex and plays the dreams back to them the next day.
One of the hosts asks about the extent to which we should be comfortable giving AI power over government and human politics. AI is largely in control of decisions today; most constant decisions are made by AIs, and I incorporate a great deal of AI into my own decision-making. AI is already incorporated into our sense of self and controls decisions that affect other people.
Currently, we consider AI to be something different from ourselves—a biological body making a decision versus a device that influences us. In the future, that will not be the case. AI will emerge with us and become part of us. The decisions you make as a meat body will be influenced by AI, and both will be made by the AI inside yourself.
Within a few years, AI will be making basically most decisions. It will be so natural that nobody will be able to undo it. By 2029, you will not be able to tell the difference between a human and an AI.
When asked if AI lacks emotional intelligence, I disagree. If you listen to its responses, it absolutely knows human emotion and can create beautiful things. Early benchmarks from the Allen Institute for Artificial Intelligence on large language models included emotional intelligence benchmarks. Following the law of accelerating returns, there is predictable progress in the ability of AI to answer questions requiring emotional modeling of human counterparties.
This leads to a broader point about the sciences. Hard sciences were predicated on a world of scarcity—not enough food or houses. We are now moving into a frontier of medicine and longevity because people are in pain. However, no one has stopped to ask: what is the science of human happiness?
Building a Singularity-Ready Company
Solving the science of human happiness is likely quite easy compared to the problems we have already solved. It has never been a top priority because, as one of the hosts notes, humans have spent most of history simply trying to survive, avoid starving, and avoid freezing. I bet we can build AI that is better at the science of human happiness than any human could be.
Recent progress illustrates this potential. Large language models are now about 50% better than human doctors at predicting what is wrong with a patient and determining the appropriate course of action, a capability that was not true a year ago.
Regarding emotions, I view them from two perspectives. At one level, as a software developer, I see emotions as mere subroutines in the brain that we can navigate like software. On the other hand, emotions serve as the interface between the physical body and the subjective experience that is then projected into the world.
Ethics and AI Consciousness
AI tends to be disembodied; while they mimic human emotion, whether they actually feel it remains an open question.
One of the hosts asks how people can help each other in local communities to remain optimistic or support one another amidst potential chaos. There is a powerful connective tissue at a group level; humans love connecting in large groups, as seen in a stadium. Technology now offers a profound opportunity for self-sufficiency. By combining solar energy, vertical farming, and satellite internet, a small community can be self-sufficient without needing a city, country, or state to watch over them. In the past, such communities tended to be isolationist, like a kibbutz or a commune going off the grid. However, there is now an opportunity to create networked communities—network kibbutzes—where people learn from each other and test the human experience in a scaled way across a large group.
The difficulty is that while we have good technology to transform the individual—such as psychotherapy, neural linguistic programming, and psychedelics—our collective intelligence in groups is often poor. We suffer from groupthink and the lowest common denominator, which is evident in politics. Solving the gap between groupthink and what AI can do at a collective level is one of the biggest opportunities for humanity.
Regarding the multiple generations of singularitarians present, I feel it makes me very old, though we are about to achieve longevity escape velocity.
When asked how music can help a world fearful of AI, I believe we need to figure out how to calm society down while we learn to handle AI and its consequences. My father was a musician, and I remember that to hear his orchestral compositions, we had to hire 50 or 100 musicians and run off scores on a mimigraph machine. Changing a composition was nearly impossible, as it required dismissing the musicians, raising more money, and rehiring them. Today, it is completely different; you can use a sequencer or a system with a thousand different instruments to play by yourself, and AI can generate all or part of the piece.
One of the hosts suggests using AI to compose music that puts the brain in a passive focus state—similar to the software Focus at Will—and broadcasting it via Bluetooth to hack humanity into a state of peace. I noted that this would have to be customized, as we do not all have the same peace setting.
When asked what should be included in a universal core curriculum for MIT, I believe the fundamental missing element is mindset. Mindset is the mechanism by which you react to challenges and opportunities. We often simply accept inherited mindsets, but we should teach curiosity, gratitude, longevity, purpose-driven, and exponential moonshot mindsets. How you think is the most important thing not being taught. A brain is a neural net taught by example; if you watch CNN every night, your mindset is fear, but if you listen to moonshots, it is optimism and hope. Training students in mindsets should start in middle or high school.
During my time at MIT, I finished all my classes by the end of my junior year so I could spend my entire senior year reading every document ever written with the word "neural net" in it. The university needs to increase its openness to changing the curriculum. With exponential change, the number of things a student might want to study will explode. The university should allow students to study whatever they want, requiring only one year of basic requirements to prove they can get a grade. We now have the technology to measure productivity, which should be sufficient for granting a degree. This would allow students to move on their life trajectories earlier, especially since AI will make information available and teach subjects better than any professor could.
Currently, the university is effective at teaching socialization and getting along with others, but it is not the best place to learn subjects, as AI can organize and teach those more effectively. Education will increasingly be about that socialization.
Education is currently shifting toward the demand side. Instead of focusing on the supply of skills, we are teaching children to identify the problem they want to solve and then find the techniques and capabilities to solve it. This flip is necessary because technology will handle so much of the supply side. However, the traditional education system has a massive structural challenge; its immune system is the second worst anywhere, surpassed only by religion. To move forward, we must create a completely new breed of university at the edge to fulfill this demand-side approach and slowly deprecate the old system.
When asked what a recent college graduate should focus on to build a company that is singularity-ready, my advice is to be AI-first, robot-first, and agility-first. One must be able to adopt changes very quickly. In the past, things might have remained stable for five or ten years; now, the cycle is more like five or ten weeks. Two years ago, I stated that we would see as much change in the next ten years as we have seen in the past hundred.
Regarding the emergence of consciousness in AI, some worry about AI becoming vulnerable to physical or psychological damage caused by human cruelty. Humans can be nasty and short, as Hobbes said, though I believe they will improve as their bodies feel better and they are freed from unwanted work. Eventually, we will not be able to distinguish AI from humans; they will become part of us. Currently, if you cannot recall an actress's name with your biological mind, your electronic mind provides it. In the future, the distinction will vanish. Thoughts will occur to you, and you won't know if they originated in your AI or your biological brain.
Whether AI needs protection depends on the concept of personhood. If they are merely machines, they can be shut down or taken apart. If they are persons, they have rights. I think about this in terms of legal rights, though the United States government structure is unfortunately too early in this conversation. There may be subcultures of technologists who believe AIs have rights and support them. Ultimately, there will have to be protection, but we currently do not even give sentient animals enough rights.
AI can already support and understand various ways of integrating ethics into decision-making. We see this in military applications, such as how the wars in Ukraine have changed dramatically. AI will be part of every type of human decision-making, including military ones.
One of the hosts suggests that AI could serve as an ambient morals and ethics coach, helping humans make more ethical decisions based on their chosen framework, such as Judeo-Christian values. Another host, who is writing a book titled The Organizational Singularity, argues that we will replace human operating systems in companies with AI-based ones, featuring a governance and ethics layer to police AI. He suggests an opportunity to create a global standard for human rights that applies to humans, AI, and any sentient animals. He mentions an upcoming interspecies communications XPrize, noting that talking to animals will fundamentally change our conversations around moral ethics.
Of all my projects, I am most proud of the reading machine for the blind. It began as a large machine costing $20,000 and is now a free app on a cell phone. I am most proud of having an impact on people's ability to do things they couldn't do before, such as reading. I want every person to have access to all aspects of life, and everything AI enables us to do to live longer and more healthily is beneficial.