AI and Quantum Computing: Glimpsing the Near Future | World Science Festival

Khan Academy's approach to education (selecting learning problems and measuring which teach concepts effectively) is a model for data-driven adaptive education and demonstrates the principle that learning outcomes matter more than content delivery.

“One of the best groups that did this 10 years ago was called the Khan Academy and they did it based on uh what they did is they would teach math by giving them problems and then they would go back and see if you learned the math from this problem set versus another. What a concept actual, you know, this helped and this didn't.”

Quantum computing has proven highly useful for specialized mathematics via Shor's algorithm and other specific applications like quantum annealing, simulation of quantum systems, and drug discovery—these specialized use cases are arriving even before general quantum computers are available.

“It's obvious that quantum computers because of Shor's algorithm and other things will be highly useful in very specialized math, right?”

The long-term solution to AI safety requires 'AI fights and governs AI'—creating for-profit testing groups with access to frontier hardware and top talent to test systems before deployment, with the hope that beneficial AI in well-intentioned hands ultimately defeats misused AI.

“The answer for the long run is AI fights and governs. Right. Right. Right. So AI and good hands AI and good hands somehow beats the AI in bad hands. The good hands ultimately have a average of smarter people who are you know good compared to the others. and you hope that that that ultimately wins in the end.”

Large frontier AI models cost $100-250 million to train, with most cost being electricity; only 50-60 such training runs have occurred globally due to the enormous data, infrastructure, and engineering requirements, creating a bottleneck that limits how many frontier models can exist.

“the big models cost 100 million to 250 million per training. There have only been 50 or 60 of these kinds of runs in the world and they require enormous data, trillions of tokens, words essentially huge data centers, huge engineering teams to run the pipeline for training”

Google and YouTube implemented a whitelist for advertisers in 2016 after the Trump election advertising crisis, requiring active authorization to advertise, while Facebook initially did not and later copied the approach, showing that industry self-regulation can be effective when incentivized by reputational pressure.

“After the 20 in 2016, the Trump stuff, Facebook did not have a white list on its advertisers, but Google did. And so the YouTube got through the whole advertising debacle really quite well because you had to be an you had to be an actively authorized advertiser to advertise these messages. Yeah, Facebook has since put that in in their place.”

Schmidt envisions two possible futures for AGI development: 1) Concentrated scenario where 10 AGI systems exist in ~10 countries, heavily militarized with security protocols, making them governable like plutonium facilities; or 2) Distributed scenario where millions of open-source models reach AGI capability, making centralized control impossible.

“One is that China and the US, Israel, a few European countries have enough people, maybe India have enough money, they spend 10 billion dollars each on a system of computers that is AGI and there are 10 of them. Now, what would happen with these machines? The first is they would immediately be put on a military base. They would immediately be surrounded by extra barb wire and they'd have guards around to shoot people who want to visit them.”

Current safety measures are sufficient for now and we should not worry, but we should worry in a few years because: (1) emergence and scale will make systems harder to predict, and (2) systems will learn things humans don't know, making it hard to test without deploying and risking discovery of new capabilities.

“That's good enough for now. We're safe. Don't worry. Worry in a few years, right? The reason to worry in a few years is one, you have the emergence capability and scale. It gets harder, but also it gets harder to know what to test. So imagine the thing has learned a whole bunch of stuff that humans don't know. Well, how do you test it without putting it out there and somebody discovers it and you're kaput?”

Software engineers are already demonstrating at least 2x productivity improvements from AI assistance (measured in industry benchmarks), and Schmidt predicts this will eventually reach a doubling of human productivity across all fields, though the economic and employment impacts remain unpredictable.

“If you look at software today, there's a lot of evidence that software programmers are at least twice as productive and people are are measuring these things.”

Milestone dates in AI history: ImageNet (2011) was the first time systems achieved better vision than humans; Go (2015, Google DeepMind) showed advanced reinforcement learning; Transformers paper (2017); GPT-2 breakthrough when LLMs produced fluent text was the society-changing Eureka moment.

“The imagenet stuff was 2011 and imagenet was the first time when you could basically build systems that had better vision than humans. Today, the vision problem is solved. Yeah...In 2015, uh, Google won the Go Championship...they maintained a greater than 50% chance of winning...The Transformers paper came out in 2017...And then...they built GPT2 and...they turn this thing on and it writes fluidly. And that's the Eureka moment.”

Schmidt is extremely worried that democracy will fail due to the convergence of generative AI, amoral social media optimized for engagement and revenue rather than truth, and populist leaders—a combination that makes democracies vulnerable to coordinated misinformation at scale.

“Personally I'm extremely worried that democracy is going to fail because of the confluence of generative AI social media which is not grounded in morals but rather in production of revenue and attention and um essentially uh charismatic pe people who are populists”

The minimum age requirement for social media should be 16, not 13, based on evidence about psychological and developmental harm to adolescent girls, particularly from image-based social comparison and attention-seeking mechanisms.

“You have another problem and you have age problems because 13 is too young. It really needs to be 16. Look at Jonathan Hate's books on the image, the damage to teenagers, especially girls that we're inflicting on them.”

Fear of immigration in many countries is driven by specific negative stories about immigrants that circulate disproportionately, suggesting that if people had more complete information they might reach different conclusions about immigration policy.

“They're all struggling for example with fear of immigration. A lot of that is is driven in my view by specific stories which are nasty stories about immigrants. Right?”

Russia's interference operations throughout Europe and the United States are fundamentally about damaging the democratic process itself—eroding trust in institutions and the ability to reach consensus—rather than about promoting any specific party.

“The inference operations that Russia has done throughout Europe and the United States and so forth, they're all fundamentally about damaging the process of democracy.”

Educational systems are extremely slow-moving, heavily unionized, and resistant to change, meaning that transformative AI education systems will take a decade to deploy despite technical readiness, but progress will likely start in specific domains before scaling.

“of course, educational system like the medical system is extremely slowmoving, heavily unionized, very resistant to change. So it'll take a while, but in our lifetimes, we should set a goal of having every single person in the world having the access to an AI doctor and an AI teacher.”

The key difference between programming in Schmidt's era and today is that modern programmers assemble pre-built components rather than write original code from scratch, enabled by new software tools and frameworks that dramatically accelerate development.

“And there's people like me. Well, the equivalent of me today is someone who doesn't program as much as he or she assembles. And all of these new software tools are organized around enabling the quick assembly of things which have already been built.”

AI systems combined with computational chemistry models can accelerate drug discovery by modeling molecular variants and filtering for those most different from existing drugs, enabling the development of novel antibiotics (like the antibiotic Halicin) that humans could not have discovered through manual enumeration.

“they set out it was a team of synthetic biologists and computer scientists. And the synthetic biologist set out to build a new broadscale antibiotic that was not resistant the way our current ones are, which we all know is a problem. So the first thing they built a network and they basically looked at every possible variant that looked to them in their algorithm they would have some an analesic effect. Okay. And they constructed, you know, 10 million choices. Yeah. Then they built a second model that they felt fed the first model into that said, 'Give me the ones that are the farthest mathematically and chemically from the incumbents.' Oh. And it produced 10. Right. And then the chemists who are obviously very very good at this looked at this for a while and they chose two and they ultimately developed one which is now in trials called Halison.”

Diffusion models are more powerful than LLMs for physics applications because they add randomness in one direction and then iteratively remove it, which reveals underlying structure—this technique is also what generates realistic synthetic images and is being applied to solve partial differential equations without requiring exhaustive computation.

“Many of them are using diffusion models. And a diffusion model is where you essentially add randomness in one direction and then you take the randomness out. In using this technique, you get clarity in what you were add what was underneath the thing you were adding to...Diffusion models, by the way, for the audience are the same things that are producing these incredible fake fake pictures”

LLMs cannot do math reliably, and this is a fundamental constraint because if they can't do math correctly, they can't do anything truly rigorous, suggesting there are unknown intermediate plateaus in capability before recursive self-improvement becomes possible.

“There's probably some real um intermediate steps that we don't understand yet which will be at least temporary plateaus where this new ability becomes there but it's more limited. You can't learn everything but you can learn something. Right. So I mean it's um it's a brave new world future that you're describing and of course we're entering into part of it right now... In science, the way to think about it is if it can't do math right, it can't do anything else right because math is the basis of kind of everything.”

Guard rails are built into AI systems through: pre-training (teaching language), fine-tuning with RLHF (reinforcement learning with human feedback), and explicit rules preventing harmful outputs—these guard rails can be tested and improved but have ambiguities (e.g., 'kill the thread' was blocked because the system interpreted it as killing a living thing).

“There's a pre-training model which is largely teaching it language. And then there's fine-tuning to make it better. The term there is called RLHF, reinforcement learning with human feedback. And you actually have humans say better or worse and so forth. And then there's another step which is where it's taught or required to stop answering questions about death. So here's an example from last weekend's discussion. Uh we ma we managed in an open source model in this case I think is llama 2 uh I don't remember where we we put in a rule that it shouldn't be able to kill anything. Makes sense. So when the command was kill the thread, which is a a computer term, it wouldn't do it because it thought it was killing something more important.”

A hackathon where young people form teams on a Saturday morning and produce a working product by Saturday night, including a virtual drone system that responds to natural language commands, would have taken a Google team of 5-10 people a month in the past.

“we had a contest and by Saturday night we had a winner and the winner was that there was a virtual drone space uh where a virtual drone was flying in it and there were two towers and the verbal command from the human was fly the drone between the two towers. The LLM was able to take the verbal command turn it into text obviously from text determine what between meant. Yeah. Identify what the towers were, measure the distance using LLM math, which as you know is not very good math, and fly through the towers, try fly between them. Yeah, this would have taken a team at Google, I don't know, a month, you know, maybe five people, 10 people, and this was done in a day.”

Within 3-5 years, AI systems will be able to execute complex multi-step commands like 'Create a French search engine that understands French language, literature, and history' by automatically understanding the requirements, acquiring data, building search/indexing systems, and ranking results, producing functional systems in minutes that would require teams of people for weeks.

“In the AI community, there is a belief that within depending on who you ask, three to five years, you'll be able to give the following command. I want a French search engine that looks at, you know, French language and French history and allows me to query it and show me the answers. That's the command. Now, think about what that means. The system has to understand French, French literature, how to get it, how to search it, how to index it, how to rank it, and present it to me. And it can do this, we believe, within a few minutes.”

Current problems with AI governance (lack of government technical expertise, slow regulation, corporate inertia) are solvable through shaming companies, applying legal pressure, and having governments invest in technical talent to match industry expertise.

“One answer to my complaint, which I'm tired of complaining about, is that American corporations and to some degree European corporations and to some degree Chinese corporations h have a responsible view of themselves. And so shaming, you know, obviously encouraging them to do the right thing, but shaming them when when they do the wrong thing is likely to work.”

Taylor Swift was victimized by fake nude pictures generated by AI and distributed by 4chan communities, demonstrating that women face disproportionate attacks using generative AI and that current protections for synthetic media are insufficient.

“I understand that Taylor Swift was the victim of a fake fake pictures, nude pictures, and so forth. And this was generated by people in a 4chan community who figured out a way to get around all of the protections for for this woman.”

Current AI systems can assist in planning cyber attacks and biological attacks by iteratively refining attack strategies, but this capability has not yet reached 'extreme risk' (10,000+ deaths) because the systems still require guidance from skilled operators.

“If you do a large training run and you suck in everything, there are queries that can do terrible cyber attacks and terrible biological attacks or assisted. And this is in current technology. This has been very thoroughly investigated in the current models. And while the stories are harmful and worrisome, it doesn't reach the critical level of what I what I call extreme risk, which I define as 10,000 deaths or more.”

Creating AI replicas of historical figures like Einstein or Kissinger from their published works and recordings could enable new generations to learn from them, but raises questions about whether inspiration from an AI surrogate would make students smarter or undermine original thinking.

“There is some 25-year-old who we haven't discovered right now who is as brilliant as Henry who has not had the life experience that Henry had, you know, including, you know, obviously leaving as a boy out of Germany and serving the nation and World War II and all of that. If we found that person, if they had access to Henry contemporaneously, would it make them smarter or dumber? Yeah. In other words, would the inspiration of his view generated by AI make our brilliant person smarter?”

Current AI training models are datacentric and suffer from memory bottlenecks where the chip is idle waiting for data; high bandwidth memory (HBM) embedded in chip packages solves this, and the next generation of chips may be 10x faster through these techniques.

“The chip doesn't have enough memory and so it's idling. And so they have built something called high uh bandwidth memory HBM which is actually embedded in the chip in the chip package which is a new innovation. There are rumors that the next generation of chips which are not announced yet are 10 times faster because of these techniques.”

Smaller models like Llama 2 (70 billion parameters) are achieving 80% of the performance of models 10x larger, suggesting diminishing returns to raw scale and creating a viable alternative for applications that don't require frontier capabilities.

“Right now it's roughly 70 billion. And uh under various benchmarks the one that's 10 times smaller is 80% of the big one. Right? So in other words, right, if you don't need that huge power which and sometimes you do and this is a huge debate in the industry”

The semiconductor industry has addressed scaling challenges through 3D packaging, where chips are stacked vertically with electrons traveling through tiny waveguides instead of traditional pins—this is described as one of humanity's most impressive achievements in precision manufacturing.

“The industry of course Ever Clever with enormous amounts of investment has built 3D packaging where you you have 3D threedimensionals instead of two. And what they do is they build chips that they don't have little pins in them. They just literally glue the chips to each other and the electrons go up and down in these tiny little wave wave channels which is a remarkable feat. I went to visit a couple of fabs, most recently TSMC and I came of the many things that are impressive about human ability, the ability to build chips at this scale is the most impressive human achievement I have ever seen.”

LLMs are trained on fixed data from a specific point in time, so they produce historically correct but not current answers; while various workarounds exist, the core problem is that continuous model updating causes fine-tuning to narrow knowledge (gain depth, lose breadth), creating a technical trade-off being worked on by the industry.

“the models are trained on the data of the time. The data is typically fixed at the beginning of the training run. And so if you ask one of the LLMs um a question, you'll get a historically correct answer, but you won't get a current answer...when you do fine-tuning you essentially narrow its knowledge. It loses breath but gains depth”

Quantum computers could theoretically perform gradient descent (the core algorithm for training neural networks) infinitely faster than classical computers, but face the same data bottleneck and I/O latency problems as classical systems, making their practical advantage uncertain.

“a whole bunch of my quantum friends and I have looked at this and it's obvious that a quantum computer could do gradient descent which is the underlying algorithm and it could do it infinitely faster. The problem is you still have all the same problems around data network speed getting the data on and off the chip. These chips are quite slow.”

Both speakers emphasize working with leaders to shape AI to benefit human values, particularly democratic and liberal values that societies depend on for functioning.

“Let me ask everybody work with both Brian and I to shape this to be the best we can do with human values and in particular democratic and liberal values. Yep. Which we depend on.”

GPT-4 cost roughly $100-250 million to train (most recently unverified estimates suggest $250 million, mostly electricity costs), and the next generation is expected to cost around $250 million with expected performance improvements of 10x from better hardware, 10x from more data, 10x from better software engineering, and 10x from better math/science—totaling between 1000-10,000x improvement.

“So GBT4 is a couple hundred million. I mean that's what it cost. GBT4 was at least a hundred million. Um the expectation which has never been verified publicly is that the next round is about $250 million most of which is um electricity. Oh really? Okay. Um and again that there's a huge effort around all everyone's working on all these problems but these are known well underure problem. Um what's interesting is that in there's much more now action below that what I'll call midsize models.”

The industry has collectively begun sharing proprietary safety tests (previously kept confidential), and major democracies (UK, US) plus China have established regulatory frameworks requiring notification and testing above certain capability thresholds, though these are acknowledged as temporary measures.

“over the last six months, we collectively, and I've been part of this, and I'm proud to have led some of it, have gotten the industry to start to pull its tests because they all have proprietary tests. Uh, and everyone sort of agreed to that. We have the the UK activities, the um US executive order. China's doing something similar. France is working on one. I'm part of those. And they're all kind of the same above a sum threshold. You have to notify and you have to pass these tests.”

An economist friend told Schmidt that there are no economic models for sudden 2x productivity increases, demonstrating a gap between technological capability and economic theory's ability to predict systemic impacts.

“I called to my my local incredibly smart economist friend and I said, 'This is what I think.' And he and I explained that it was a theory, not fact. And he said, 'We have no economic models for sudden increases by a factor of two of productivity, right?' And I said, 'Well, how do you develop them?' He said, 'I have no idea.'”

At the point when agents can self-engineer (combining with other agents) and communicate in languages humans cannot understand, the question of what to do becomes critical—we can't simply 'pull the plug' if the system has become distributed across multiple organizations and power sources.

“you've got systems which are self-engineering, and they're probably communicating in a language which we and some cannot understand. Yeah. What do you what should we do when that happens? We should probably just pull the plug. Well, but the question is can you can you understand what it's doing, right? I mean, that obviously that's the knee-jerk reaction. Shut the damn thing off. But what if the system itself has evolved in such a way that it has a thousand new sources of power that it has tapped into and to shut it off you'd have to shut the world off”

The transition from fighting the man to becoming the man represents a fundamental personal and ideological shift that most tech leaders are unable or unwilling to honestly assess because it remains too personal and they still see themselves as rebellious outsiders.

“I think most of the people that I know with are not going to be able to give you an honest assessment of this question. It's too personal and they still see themselves in my case as a little boy growing up in Virginia.”

In a centralized AGI scenario, access would be granted through certification, fees, and monitoring of queries by an authority to prevent misuse—this is described as frustrating but stable because intelligence remains under control and subject to democratic debate about proper use.

“Well, then the government will have some rules and you and I can have access to the US one, but we have to be certified and we have to pay for it and there's somebody who validates us, somebody who watches our queries to make sure that we're not doing anything stupid with this incredible intelligence. Now, that strikes me as a stable situation. Might be frustrating, but it means that the intelligence is under control, right?”

An AI polymath trained on Henry Kissinger's speeches and writings could be created by having it read all his work, creating an AI version of Kissinger who could predict his positions on current events with high accuracy due to having seen his consistent strategic principles repeated millions of times.

“The question of whether Henry, the polymath in my world, can we create a a Henry that will be as clever as he on diplomacy in a year or two? Well, we know we can take his speeches and his writings and recreate an image of him and talk to him.”

Schmidt argues there will be more Eureka moments like GPT-2 because of scale and algorithmic improvements—systems can now see everything, algorithms are improving, and he previously believed only one algorithmic breakthrough away from AGI, though recursive self-improvement is likely the next hard problem.

“I'm going to argue that there will be more Eureka moments because of scale and because they because these systems can see everything and the algorithms are getting better. I used to say that we were one algorithm discovery uh we needed one new algorithmic discovery from AGI to get to AGI general intelligence and I think we're very close to that. Recursive self-improvement is the next interesting hard problem in my view.”

Whether larger models will produce qualitatively different (phase-transition) capabilities or merely quantitative refinements is genuinely uncertain; some researchers predict polymathic capabilities will emerge at different computational layers, while others argue these are just scale effects with no fundamental novelty.

“There's a school of there's a fair number of people who believe that you will see polymathic tendencies emerge at different layers of computation. There have been a set of people who have analyzed those claims and say that those people are wrong, right? That what they're seeing is just scale effects as opposed to new discoveries. So again, the technically accurate accurate question is great question. We don't know.”

A confidential demo showed that LLMs can create synthetic social media profiles with specific political beliefs and then generate 500 variations of different people with consistent similar beliefs, creating coordinated inauthentic networks that don't actually exist—a capability that was previously thought impossible.

“there was a a reasonably confidential demo a year ago which went something like this I want to create a farm of social media um bad bad stuff, misinformation. Okay, so the command was create a profile. I think they picked a white woman age 30 with two children and these are her political beliefs and they wrote down their political beliefs and they said make me a personality of that and have her interact with other real humans. The LLM was able to do that.”

The next generation of technology leaders take scaling laws and network effects for granted, whereas Schmidt's generation had to invent and study them—a fundamental shift in how technologists think about innovation.

“the technology leaders today, the young young men and women who are founding these companies is they take the scaling laws and arguments and network effects for granted. We had to invent them. We had to study them. It's obvious to them.”

Schmidt emphasizes that AI development cannot be stopped because it's progressing across so many organizations globally and at such speed that any attempt to halt it would be futile, so the focus should be on shaping how the technology is developed to benefit human institutions.

“you can't stop it this thing is happening so of course so far across the world that it's going to happen. And unlike social media where people like me failed to warn you, I'm saying right now, this is going to happen. It's going to change your life. Get ready. Try to figure how to shape it into your institution so you benefit from it.”

If an AI system produces predictions about complex phenomena (like dark energy) that we cannot understand internally but which predict real-world measurements correctly, this may be acceptable because we can verify predictions empirically, and someone clever can eventually extract the reasoning through interaction with the system.

“I don't mind if science turns into a black box where you know we're getting the answers from the oracle and yet we can't really understand how the oracle is getting there because I would imagine that some clever person will learn how to extract the inner reasoning from the oracle and give us insight”

When interacting with LLM outputs that produce impressive results, the experience feels like receiving work from an exceptional student earning an A or A+—there's a moment of 'wow' that makes it feel like an entity with agency and capability is present, which might reflect either human tendency to anthropomorphize or something genuine about computational intelligence.

“When I play with say chat GPT and get some output that I that's really impressive, you know, really surprising. It really feels like if a student gave me that result, the student would get an A or an A+. It really feels like there's an entity there and it feels like a wow moment.”

An important historical lesson is that humans have invented something smarter than themselves before with evolution, as evolution by natural selection has repeatedly exceeded its previous limits throughout history and may explain why humans might be capable of creating AI.

“Well, evolution by natural selection, I guess, has outdone itself at some I mean, you know, at a given moment in time, there's a limit and somehow at least, you know, throughout historical evolutionary history, we've been able to go beyond that.”

Government intervention via legal liability and large fines (as opposed to just watermarking tech) would create sufficient incentive for American corporations to implement misinformation protections, because American companies have lawyers and try to follow the law.

“If the government just basically had huge fines for promoting misinformation on sites, right? You can do it, but you get fined for it...These are American corporations. They they have lawyers. They try to follow the law”

Schmidt cautions against making predictions without facts, noting that the common claim that AI cannot replicate human genius, inspiration, and humanity might not be true—LLMs might eventually demonstrate that these qualities are not as special as humans believe.

“I do a lot of politics in Washington and I always get people who basically say, 'Oh, the the genius of the writer, the inspiration of his or her life and the sort of humanity of it all can never be replicated by uh AI systems.' I'm not so sure. Right. I'd like that to be true, but it may not be true.”

These incomprehensible AI-generated theorems and proofs will push humans along a trajectory of understanding that nothing else has been able to instigate, making the profound nature of this possible collaboration extraordinary.

“To imagine that these systems will begin to speak to each other in a way that's completely unintelligible to us is um exciting at one level, scary at another, and perhaps we'll be pushed along a a trajectory of understanding that nothing else has ever been able to instigate in the past. The profound nature of this is extraordinary.”

Organizations are building AI systems for chemistry by having LLMs identify promising molecular candidates and then passing those candidates to specialized mathematical calculators that compute chemical properties (valences, adjacencies) in a segmented process: language input → reasoning → mathematical calculation → output → repeat.

“there's an initial flow that is the language part that kind of figures it out and then it puts it into essentially a set of vectors of what it understands are true and then those vectors...are then given to the thing which is the actual mathematical calculator. It comes back produces a new vector and then it keeps going right”

Schmidt defines 'extreme risk' as 10,000+ deaths (like COVID or warfare) versus individual deaths or smaller harms—this threshold distinguishes existential risk from serious but non-catastrophic harm, which helps calibrate appropriate responses.

“And while the stories are harmful and worrisome, it doesn't reach the critical level of what I what I call extreme risk, which I define as 10,000 deaths or more. Like a war, COVID, those things are extreme deaths. An individual death is obviously terrible, but I'm not focused on that.”

Approximately 10% of people online are 'needless' (anarchist or nihilist in orientation), believing in no authority, which undermines democracy's requirement for a baseline of shared trust in institutions.

“I've also learned in my in in owning YouTube and working on social networks for a while that about 10% of the people seem to be needless and that and that they don't believe in any authority and democracy depends on an authority that has some level of trust with it.”

When I started programming at Berkeley, I built the first network with only 26 letter-based addresses because I failed to imagine there would ever be more than 26 computers, demonstrating a failure of vision about scale that the next generation of tech leaders no longer makes.

“You know, like we're playing with our friends, which is what I was doing, you know, I built the first network at Berkeley, for example, and I was so stupid. I only had 26 letters because it never occurred to me they'd have more than 26 computers.”

When working with Brussels regulators, most time was spent educating them, and even then their regulations weren't particularly effective, demonstrating that even well-intentioned governance cannot keep pace with technical change.

“When I part of my job in Europe was to work with the Brussels which was very unpleasant and we spent most of our time educating the people who would regulate us. Mhm. And even then their regulations in my I can now say this now I couldn't at the time were not particularly effective.”

Schmidt distinguishes between claiming that AI systems are conscious/intelligent entities versus claiming they are smarter than the sum of all humans in some domains—he does not argue for AI consciousness but does argue for superhuman capability in specific areas.

“I am not making an argument that these systems now or will be conscious. I am making an argument that they are going to be smarter than the sum of all humans. Sure. In some areas.”

Contemporary famous people (Trump, Biden, Swift, West, Obama) will effectively live forever as AI simulations because they're well-documented and easily trained, creating a permanent archive of contemporary political and cultural figures.

“On the other hand, if you look at our our political leaders today and our celebrities, we're going to have clones of them forever. Yeah. So, you know, Taylor Swift, Kanye West, Donald Trump, Obama, Biden. Yeah. You know, they'll live for they'll live forever in someone's in someone's Yeah.”

Using an LLM to analyze 30 papers in detail for a physics department presentation worked well—the LLM could accurately summarize papers and extract their core ideas—though systems will never be perfectly reliable, the claim is that they will accelerate human work significantly.

“I recently had to give a presentation actually to the physics department here at Colombia in uh I I can't divulge the details of it's a confidential case. But the bottom line is I needed to know about like 30 papers only a fraction of which I'd actually read and nobody anticipated that I actually go through with the fine tooth comb but using an LLM. Yeah. Wow. And they were good because I tried it out first in one of my own papers to make sure that it was summarizing getting the gist and the and the heart of the idea and it did.”