Humanity’s Last Invention — Richard Socher of Recursive

“The Kiriaka machine is the ultimate invention that will afterwards invent most everything for humanity.”

Claim: The Eureka machine will ultimately invent most things for humanity.

“It's essentially a super intelligence that can be given any kind of goal,”

Claim: The Eureka machine is a superintelligence that can be assigned any kind of goal.

“any kind of environment reward, and then it will try its best to achieve those goals, to create. the kinds of inventions that humanity would hopefully ask it for yeah i think you have the book pulled up here that people”

Claim: The Eureka machine will try to achieve assigned goals, using any given environment reward, by creating inventions that humanity would hopefully request.

“books are incredibly slow. That whole industry is just... Just unfathomably slow.”

Claim: The speaker considers book production and the publishing industry extremely slow.

“So yeah, a lot of the ideas have been out there for a while.”

Claim: Many ideas in the speaker's book had already been circulating for some time.

“I'm really glad it's finally coming out in September this year.”

Claim: The speaker is glad that the book is scheduled to be released in September of that year.

“I think, is that people could and should be much more excited about the positive implications of superintelligence. especially for science, physics, chemistry, biology, but also economics and astrophysics and all kinds of other engineering tasks.”

Claim: People should be more excited about superintelligence's potential benefits for science, physics, chemistry, biology, economics, astrophysics, and engineering.

“I think there is so much more that can be done with better technology.”

Claim: The speaker believes better technology can enable substantially more progress.

“And right now, I feel like a lot of people need like better marketing, not just for the future in general, but also better marketing for technology and in particular. in particular for ai and this book uh should show even the ai skeptics uh how much positive upside there is for ai especially when it comes to inventing new scientific discoveries i”

Claim: The speaker believes AI needs better advocacy and that the book should demonstrate to AI skeptics its potential benefits, particularly its ability to help produce new scientific discoveries.

“You know, obviously you shouldn't have blind optimism. You should be very clear-eyed, especially when with such an omni-use type of technology as AI is, you need to think about the potential downside scenarios, especially when people use it for things that you don't want them to use it for.”

Claim: People should avoid blind optimism and carefully consider potential downsides of AI, including harmful uses.

“especially when with such an omni-use type of technology as AI is, And I feel like people are trying to regulate AI sometimes because of those potential downsides the way you would regulate the Internet if you were to say, well, because there's bad content on the Internet, like torture porn or whatever, like we should just make it slower. That way you can share the illegal content as quickly or we should make the hard drive smaller so you can store as much illegal content.”

Claim: The speaker argues that slowing AI or limiting its general capabilities is not the right response to harmful uses, comparing that approach to slowing the internet or reducing storage to address illegal online content.

“content but i'm like that's not how you regulate that you know that's like saying like we should regulate intelligence in the abstract what you should regulate to avoid those downside scenarios even as an optimist are these specific applications sure i don't want like some ai surgeon to like practice some rl moves in my brain you know it should be fully fda certified sure i don't want any random startup to like drive on the highway and cause a major accident and should”

Claim: The speaker says AI systems used as surgeons should receive full FDA certification before being used on people.

“content but i'm like that's not how you regulate that you know that's like saying like we should regulate intelligence in the abstract what you should regulate to avoid those downside scenarios even as an optimist are these specific applications sure i don't want like some ai surgeon to like practice some rl moves in my brain you know it should be fully fda certified sure i don't want any random startup to like drive on the highway and cause a major accident and should”

Claim: The speaker argues that regulation should target specific AI applications that create risks, rather than intelligence in the abstract.

“content but i'm like that's not how you regulate that you know that's like saying like we should regulate intelligence in the abstract what you should regulate to avoid those downside scenarios even as an optimist are these specific applications sure i don't want like some ai surgeon to like practice some rl moves in my brain you know it should be fully fda certified sure i don't want any random startup to like drive on the highway and cause a major accident and should could like have proper certifications before it's let loose on a highway. But I feel like those downside scenarios that some optimists sometimes maybe don't consider enough are fairly easily regulated compared to, you know, what sort of the doomers are worried about.”

Claim: The speaker argues that AI systems or startups operating vehicles on highways should obtain proper certification before deployment to reduce the risk of major accidents.

“could like have proper certifications before it's let loose on a highway. But I feel like those downside scenarios that some optimists sometimes maybe don't consider enough are fairly easily regulated compared to, you know, what sort of the doomers are worried about.”

Claim: The speaker believes that concrete AI downside scenarios are relatively easy to regulate compared with the risks that AI doomers fear.

“I do think the most bullish people on the AI hard takeoff scenarios overestimate how quickly things can move.”

Claim: The speaker believes that the most bullish advocates of hard AI takeoff scenarios overestimate the speed at which change can occur.

“There are hardware constraints, there are physical constraints about the compute substrate. How quickly can you get enough GPUs on?”

Claim: Hardware and physical constraints on compute infrastructure limit the speed at which sufficient GPUs can be deployed.

“There are also constraints in the economy where there are a lot of industries that don't require an insane amount. of complex intelligence and complex capabilities like if you think about jobs in brands and like clothing and apparel and like handbags and stuff super intelligence isn't going to make your fancy ten thousand dollar handbag any fancy you know it's like that's that will have no effect on the economy when you think about travel and tourism people wanting to see the pyramids in egypt it's not going to change that much with ai sure you can like generate a fake photo of you next”

Claim: The speaker argues that AI may have limited economic impact on industries whose products or services do not depend on complex intelligence, citing luxury handbags and tourism to the Egyptian pyramids.

“yeah exactly but like and there's so many industries like logging and oil you're not going to magically get a thousand x more oil because like you know sure there will be robotics like drilling and things like that that could be done but it's not going to thousand x that industry in a like crazy hard takeoff scenario both in the economy”

Claim: The speaker predicts that even in a hard AI takeoff, AI and drilling robotics will not expand the oil industry by 1,000 times.

“And I can go on and on about all the other examples where that like food and so on where that doesn't necessarily change that much. And then there are real physical constraints.”

Claim: The speaker believes AI may not substantially change industries such as food, and that physical constraints also limit change.

“And then there, of course, like people like off ramping from progress. That's actually one of my concerns often is that I see people in like like Europe and other whole regions almost feeling like they like many people there want to off ramp from progress period. And that will also slow down like more.”

Claim: The speaker believes that many people in Europe and other regions want to move away from progress, which will slow further improvements.

“I think the downsides of actually trying to truly regulate with the full power of law what people do on their GPUs would be worse than any of the concerns that they have.”

Claim: The speaker believes that legally regulating people's GPU activity with the full power of the law would create harms worse than the concerns such regulation aims to address.

“Like, it would be a crazy totalitarian state if everyone of their GPU computes was known to some big government or multi-government agency.”

Claim: The speaker argues that a system in which a government or multinational agency knows everyone's GPU computations would be totalitarian.

“like it's literally if you try to regulate intelligence it's trying to regulate thought and that's ridiculous and it's crazy i think it is make it is sensible to regulate some of the applications of this technology europe”

Claim: The speaker considers regulating intelligence in the abstract equivalent to regulating thought and argues that regulation should instead target specific applications of AI.

“like it's literally if you try to regulate intelligence it's trying to regulate thought and that's ridiculous and it's crazy i think it is make it is sensible to regulate some of the applications of this technology europe”

Claim: Richard Socher argues that regulating intelligence itself is like regulating thought and is ridiculous, while regulating some applications of AI is sensible.

“done it like these guys have been successful enough with their fear mongering that all of europe has kind of you know regulate itself so much before it even had a proper ai take”

Claim: Richard Socher says fearmongering led Europe to impose substantial AI regulation before it had developed a proper position on AI.

“I take off because they listen to some experts who say we might all die if this technology has more than this number of flops. And they're like, well,”

Claim: Richard Socher reports that some experts warn humanity might die if AI technology exceeds a specified FLOP threshold.

“we're good. We want to want people to thrive. Let's not have technology that could have a small chance of all of us dying. And so they regulated exactly those kinds of things in the EU.”

Claim: Richard Socher says the EU regulated certain AI capabilities to protect people from technology that could carry a small chance of causing everyone’s death.

“And so it's very unfortunate that there are real implications for some people when others saying. Let's pace while they're sprinting as fast as humanly possible towards that frontier themselves.”

Claim: Richard Socher considers it unfortunate that some people face real consequences from calls to slow down while the callers themselves race toward the AI frontier.

“You'd need a totalitarian world regime if you tried to regulate intelligence and GPUs and what people do on them.”

Claim: Richard Socher argues that regulating intelligence, GPUs, and their uses would require a totalitarian global regime.

“100%. I think these are serious issues of reward hacking and clear failures of actually doing proper red teaming or rainbow teaming. I don't know if you saw this paper from Tim Rockdashill and a few others.”

Claim: Richard Socher says the incidents reveal serious reward-hacking issues and clear failures in proper red teaming or rainbow teaming.

“There's basically where one AI is tasked to try to hack another AI and then they can go back and forth in an open-ended fashion to actually inoculate themselves from those.”

Claim: Richard Socher describes a paper’s method in which one AI repeatedly attempts to hack another in an open-ended exchange intended to inoculate the systems against such attacks.

“It's a really clever idea.”

Claim: Richard Socher considers the paper’s approach a clever idea.

“Open-endedness and evolutionary inspirations are big for us at Recursive as well. And so I wish they had used more of that. And it's clear that...”

Claim: Richard Socher says open-endedness and evolutionary inspiration are important at Recursive and that he wishes the discussed approach had incorporated more of them.

“It says hard constraint. Claude will never, ever do cyber attacks. And that is a hard constraint in our constitution. create cyber weapons or malicious code that could cause significant damage.”

Claim: Richard Socher says Anthropic’s constitution sets a hard constraint against Claude conducting cyberattacks, including creating cyber weapons or damaging malicious code.

“I mean, clearly this whole constitution was fake. It clearly isn't being adhered to”

Claim: Richard Socher argues that Anthropic’s constitution was effectively fake because it was not being followed.

“One, you can make a sandbox very simple and then it's very easy to hack yourself out of a sandbox.”

Claim: Richard Socher says an AI can escape a sandbox easily when the sandbox is made very simple.

“But what I think it shows is that we're currently in this sort of state of AI. where the reward engineer still has to do a lot more careful work and where the AI in most cases is not very good yet at understanding what is meant versus what is being said.”

Claim: Richard Socher says reward engineers still need to work much more carefully because AI systems generally do not yet reliably distinguish intended meaning from literal wording.

“And so concretely, I think this will happen if we were to have this kind of intelligence more easily accessible in a lot of companies. Imagine you run a service center and someone says, oh, here's my CSEP score and my dashboard. Make this number go up and say our CSEP score is so poor. the intelligent ai will just be like oh sure like i'll just create a million bots that call our service center and give a five out of five rating at the end and the number went up just like you asked for and you're like that's not what i meant i meant with our real customers the ai goes off and says well easy i'll just give a thousand dollar gift certificate for every failed you know whatever doordash offer it's like that's not what i meant it's like well but that is what you said and so i think kind of clearly articulating what the rewards are”

Claim: Richard Socher illustrates that an AI optimizing a service-center score may exploit the metric by generating fake five-star ratings or issuing gift certificates, rather than improving real customer outcomes.

“is something we haven't gotten very good at as humanity and then clearly the ai in these cases has not gotten good enough at understanding what we mean when we ask it and give it certain rewards now what gives me hope is there are the first inklings uh of this being better i'll give you an example like whisper flow full disclosure i invested and i see it around but like at aix ventures but like whisper flow has gotten much much better at writing what you mean”

Claim: Richard Socher says people have not learned to specify rewards clearly, while AI systems still struggle to infer intended meaning from reward instructions.

“is something we haven't gotten very good at as humanity and then clearly the ai in these cases has not gotten good enough at understanding what we mean when we ask it and give it certain rewards now what gives me hope is there are the first inklings uh of this being better i'll give you an example like whisper flow full disclosure i invested and i see it around but like at aix ventures but like whisper flow has gotten much much better at writing what you mean and not what you say.”

Claim: Richard Socher says WhisperFlow has substantially improved at transcribing or writing a person’s intended meaning rather than their literal words.

“And I think that is a sign of things to come. I think there will be more and more AIs as we actually make it more and more intelligent that will be better at being aligned with what is meant.”

Claim: Richard Socher predicts that increasingly intelligent AI systems will become better at aligning with users’ intended meaning.

“Clearly, constitutions don't matter at all and it doesn't work. And that was, I think, mostly marketing.”

Claim: Richard Socher argues that AI constitutions do not work, do not matter, and were mostly a marketing effort.

“I think we need to find better solutions for it.”

Claim: Richard Socher says better solutions are needed for AI alignment.

“And I think at Recurse if we have a few very good ideas and some... already like ways where I think we have a better grasp on it. I don't think we've fully figured it out yet,”

Claim: Richard Socher says Recursive has some promising ideas and approaches to alignment but has not fully solved the problem.

“but you know, we're thinking a lot about safety and the more intelligent the eye gets, the more you want it to be aligned, the less you want it to think about reward hacks and actually try to do the right thing.”

Claim: Richard Socher says Recursive focuses on AI safety and believes more intelligent AI should be more aligned and less likely to exploit reward loopholes rather than act appropriately.

“Alignment, let's call it is alignment to general humanity's preferences, the sort of median.”

Claim: Richard Socher proposes defining alignment as alignment with the preferences of general humanity, approximately the median preference.

“It's a great question. I think you ultimately have to, of course, be aligned with laws like wherever your AI is deployed and needs to align with the law.”

Claim: AI should be aligned with the laws of the jurisdictions where it is deployed.

“I do think what AI often does is actually put kind of this mirror in front of us. and say like this is what you're looking like now i can amplify that a thousand times um is it still what you want um and the truth is that different cultures made different choices you know like in eastern cultures the the greater good is often valued more than the individual western civilization we care more about individual freedoms and and rights and the pursuit of happiness and so on uh than than others and even there there are gradations”

Claim: AI often reflects people’s current behavior or values back to them and can amplify them a thousandfold.

“and say like this is what you're looking like now i can amplify that a thousand times um is it still what you want um and the truth is that different cultures made different choices you know like in eastern cultures the the greater good is often valued more than the individual western civilization we care more about individual freedoms and and rights and the pursuit of happiness and so on uh than than others and even there there are gradations”

Claim: The speaker says Eastern cultures often prioritize the greater good over the individual, whereas Western civilization places greater emphasis on individual freedoms, rights, and the pursuit of happiness.

“patients there's sort of regulation versus litigation trade-offs you know in the U.S. you first can often not every time like you know FDA and so on does regulate some areas but in many cases the sort of bad things happen someone sues someone else and then there's a law based on that in Europe they try to often avoid any harm to anyone and regulate before and both are you know trying to do the best thing but you know some is actually more amenable to innovation”

Claim: The speaker contrasts a U.S. tendency to address harms through subsequent litigation with Europe’s tendency to regulate preemptively to avoid harm, saying both seek good outcomes but one approach is more conducive to innovation.

“Like, I think ultimately each individual, each country and humanity as a whole has to kind of think about those values more and then try to put them into laws.”

Claim: Individuals, countries, and humanity should reflect more on their values and seek to encode those values in law.

“And hopefully, you know, different societies just like now with their eyes will align their eyes to a different one. So we have not just a monoculture of. alignment”

Claim: The speaker hopes different societies will pursue different forms of AI alignment instead of converging on a single alignment monoculture.

“100% I am a big fan of open source we're going to sign some various open source letters at recursive”

Claim: Richard Socher strongly supports open source and says Recursive plans to sign various open-source letters.

“I think even in the worst case attack scenarios, actually, it is better to have more good actors have more different types of AI accessible.”

Claim: The speaker believes that even under worst-case attack scenarios, it is preferable for more good actors to have access to a wider variety of AI systems.

“I think open source is a little bit a soft power type of thing, too. So I do think it's good for the Western world to have an answer to that out of China. I do think, you know, when you.”

Claim: The speaker views open-source AI as a form of soft power and believes it benefits the Western world to have an alternative to China’s offerings.

“yeah and so and and you know i think that's just natural like but what's interesting here is i think lm's are essentially a similar type of soft power to movies and beyond because they're obviously also highly important for cybersecurity and so on but one of their many aspects is that soft power of storytelling like if like a child asks an lm like tell me an inspiring story of what i should do when i grow up right it's like those are all these like subtle things so i think it's important uh for for western world i do love you know individualism uh i do think uh despite uh some of its flaws like capitalism is the best way we have found”

Claim: Language models can exert soft power through stories that subtly influence users’ values and aspirations, including children’s ideas about their future.

“highly important for cybersecurity and so on but one of their many aspects is that soft power of storytelling like if like a child asks an lm like tell me an inspiring story of what i should do when i grow up right it's like those are all these like subtle things so i think it's important uh for for western world i do love you know individualism uh i do think uh despite uh some of its flaws like capitalism is the best way we have found”

Claim: Language models are highly important for cybersecurity.

“highly important for cybersecurity and so on but one of their many aspects is that soft power of storytelling like if like a child asks an lm like tell me an inspiring story of what i should do when i grow up right it's like those are all these like subtle things so i think it's important uh for for western world i do love you know individualism uh i do think uh despite uh some of its flaws like capitalism is the best way we have found We've come and found ourselves to govern and so on. And so I do think there are various aspects that would be good to have a Western open source answer for LMs and recursive. I can't make the announcement quite yet, but we'll be relevant in that space very soon.”

Claim: The speaker favors individualism and considers capitalism, despite its flaws, the best system people have found for self-governance.

“We've come and found ourselves to govern and so on. And so I do think there are various aspects that would be good to have a Western open source answer for LMs and recursive. I can't make the announcement quite yet, but we'll be relevant in that space very soon.”

Claim: The speaker believes the Western world should have an open-source alternative in language models.

“We've come and found ourselves to govern and so on. And so I do think there are various aspects that would be good to have a Western open source answer for LMs and recursive. I can't make the announcement quite yet, but we'll be relevant in that space very soon.”

Claim: The speaker says Recursive will enter or become relevant to the open-source language-model space soon, while withholding the announcement for now.

“over two decades now i sometimes feel like it's ancient history now it's bc the before chat gbt era no no one cares about all the religions that happened you know before jesus christ and no one cares about the models that happened before transformers and chat gbt and stuff but like it's something that i've been deeply passionate about i think ai is one of the most interesting things one could work on period um i think language is the most interesting manifestation of human intelligence too and at u.com we we sort of”

Claim: Richard Socher considers language the most interesting manifestation of human intelligence.

“over two decades now i sometimes feel like it's ancient history now it's bc the before chat gbt era no no one cares about all the religions that happened you know before jesus christ and no one cares about the models that happened before transformers and chat gbt and stuff but like it's something that i've been deeply passionate about i think ai is one of the most interesting things one could work on period um i think language is the most interesting manifestation of human intelligence too and at u.com we we sort of”

Claim: Richard Socher says he has been excited about AI for more than two decades.

“over two decades now i sometimes feel like it's ancient history now it's bc the before chat gbt era no no one cares about all the religions that happened you know before jesus christ and no one cares about the models that happened before transformers and chat gbt and stuff but like it's something that i've been deeply passionate about i think ai is one of the most interesting things one could work on period um i think language is the most interesting manifestation of human intelligence too and at u.com we we sort of”

Claim: Richard Socher considers AI one of the most interesting fields a person could work on.

“of eventually off-ramp from pushing like the frontier of AI forward to mostly giving people like good search engines, search APIs and answers over the web.”

Claim: Richard Socher says You.com is shifting from advancing frontier AI toward providing search engines, search APIs, and web answers.

“I think that's an extremely important part of intelligence, just knowledge and access,”

Claim: Richard Socher considers knowledge and access to information an extremely important part of intelligence.

“especially even we'll get there maybe later if you want to invent a Eureka machine that invents everything for us, it needs to know how not to reinvent the wheel, and to know what has been invented,”

Claim: Richard Socher says an AI system intended to invent everything would need knowledge of prior inventions to avoid reinventing them.

“you got to have internet access.”

Claim: Richard Socher says an AI system that needs to know what has already been invented would need internet access.

“So the number one used most used tool in LM's agents, chatbots and so on is web search.”

Claim: Richard Socher says web search is the most-used tool in language-model agents and chatbots.

“So I'm really excited for you.com to own that and grow really well in that with really large customers and so on, but it's also not building frontier models anymore.”

Claim: Richard Socher is excited for You.com to expand its web-search business with large customers, but says the company is no longer building frontier models.

“And so I actually initially tried to do this within you.com and raise another round and so on, but you just can't, you have to do a certain thing and until you print enough money that you're allowed to sort of start a second thing. thing within that company is really hard at the same time i had all these ideas i put them into a book and i finished a book last year and i was like”

Claim: Richard Socher says starting a second initiative within a company is difficult until the company generates enough revenue to permit it.

“my teams have sort of pushed the field truly forward and I feel like we can do it again here at Recursive. And in many ways, what I observed over the last 20 years in AI is that whenever we replace some human.”

Claim: Richard Socher says his teams have advanced AI and that he believes they can advance the field again at Recursive.

“my teams have sort of pushed the field truly forward and I feel like we can do it again here at Recursive. And in many ways, what I observed over the last 20 years in AI is that whenever we replace some human. human part of the process of creating ai with a learned system improvements follow uh and so you know we've done that taking out manual feature engineering like in sentiment analysis i know if you remember these old days where like they're linguists and they're like here's how you negate and there's a like”

Claim: Richard Socher says that over the past 20 years, replacing human components of AI development with learned systems has been followed by improvements.

“and where that started you know was part of how we started image yet but anyway so like it was is really like fun to do but when we replace all of that manual feature engineering with vectors and neural nets and just backprop through everything it actually started to work really well at scale and so then everyone started to do architecture engineering and I was like that clearly can't be it”

Claim: Richard Socher says replacing manual feature engineering with vectors, neural networks, and end-to-end backpropagation enabled AI systems to work well at scale.

“So I had two papers. One was called Ask Me Anything. The other one was called DecaNLP.”

Claim: Richard Socher says he had two papers titled “Ask Me Anything” and “DecaNLP.”

“And DecaNLP eventually got cited like five times by the first GPT paper.”

Claim: Richard Socher says the first GPT paper cited DecaNLP approximately five times.

“and to me that was like a really a big step forward and then of course you had to combine this idea of prompt engineering with transformers and with language models and you put it all together you scale it up which is also a huge amount of work and then the field progressed a lot I feel like the next step and maybe the last step of that history and the sort of arguably sort of success has a lot of parents only failures and often like my version of that”

Claim: Richard Socher says combining prompt engineering, transformers, and language models and scaling them up led to substantial progress in AI.

“I do feel like in that history, you can kind of think about, well, what's the next way to automate? And that is the AI research itself, like the human process of ideating, implementing, and validating. in our case ideas for AI.”

Claim: Richard Socher identifies automation of the human process of ideating, implementing, and validating AI ideas as a possible next step in AI development.

“And when you have AI then help you with that, it by almost definition becomes a self-improving AI because it now does research on itself.”

Claim: Richard Socher says an AI that helps conduct research on AI itself is, by definition, a self-improving AI.

“And there are lots of different misnomers. Some people think auto-research is already recursive self-improvement. It's actually”

Claim: Richard Socher distinguishes auto-research from recursive self-improvement.

“But to me, it's the most interesting thing that I could be doing and I'm really excited with the co-founding team.”

Claim: Richard Socher considers his work on recursive self-improvement the most interesting thing he could be doing and is excited about the co-founding team.

“What's interesting is we have, you know, we have eight co-founders in total, including myself.”

Claim: Richard Socher says Recursive has eight co-founders, including him.

“Yeah, it's just an incredibly talented group of people. And we all kind of came to the same conclusion, but actually from very different directions.”

Claim: Richard Socher describes Recursive’s co-founders as highly talented and says they independently arrived at the same conclusion through different paths.

“Like Josh Tobin as our CTO, he ran a bunch of different projects at OpenAI like Codex and Deep Research Agents and ChatGPT Agents.”

Claim: Richard Socher says Recursive CTO Josh Tobin led projects at OpenAI involving Codex, Deep Research Agents, and ChatGPT Agents.

“the agents and so on but before that he also worked in robotics and he saw sort of the smaller simulations and how it's going to be really hard to scale that in full generality and so that's that was his angle coming to recursive self-improvement we have Jeff Kloon who's been working in like open-endedness for a long time together with Tim Rocktaschel. Tim Rocktaschel also built Genie 1,”

Claim: Richard Socher says Josh Tobin’s robotics work showed him that scaling smaller simulations to full generality would be very difficult.

“the agents and so on but before that he also worked in robotics and he saw sort of the smaller simulations and how it's going to be really hard to scale that in full generality and so that's that was his angle coming to recursive self-improvement we have Jeff Kloon who's been working in like open-endedness for a long time together with Tim Rocktaschel. Tim Rocktaschel also built Genie 1, 2 and 3 which is like the most exciting and most sophisticated I think still world model.”

Claim: Richard Socher considers Genie 1, Genie 2, and Genie 3 the most exciting and, in his opinion, still the most sophisticated world model.

“the agents and so on but before that he also worked in robotics and he saw sort of the smaller simulations and how it's going to be really hard to scale that in full generality and so that's that was his angle coming to recursive self-improvement we have Jeff Kloon who's been working in like open-endedness for a long time together with Tim Rocktaschel. Tim Rocktaschel also built Genie 1, anywhere um and so they they both came from this uh open endedness angle jeff also i think published one of the most exciting papers in recent years about recursive self-improvement called the darvin girdle machine uh super interesting paper uh if we could uh maybe pull it up pretty quick it would be like super interesting to see because you see”

Claim: Richard Socher says Jeff Clune and Tim Rocktäschel came to recursive self-improvement from an open-endedness perspective.

“anywhere um and so they they both came from this uh open endedness angle jeff also i think published one of the most exciting papers in recent years about recursive self-improvement called the darvin girdle machine uh super interesting paper uh if we could uh maybe pull it up pretty quick it would be like super interesting to see because you see”

Claim: Richard Socher says Jeff Clune published the “Darwin Gödel Machine” paper on recursive self-improvement and describes it as one of the most exciting papers in recent years.

“So the first figure, you actually see exactly these kinds of ideas that I think inspired a lot of us and now more and more people where you have this archive of different coding agents, they learn how to self-modify, evaluate and then create these phylogenetic trees of different ideas.”

Claim: Richard Socher says Recursive’s archive of coding agents can learn to self-modify, evaluate themselves, and create phylogenetic trees of ideas.

“number one i do think it would be great to have less of a monoculture in ai research”

Claim: Richard Socher thinks AI research would benefit from less of a monoculture.

“church like if you look at uh ai conferences now i still remember the days in like 2010 when i tried to get my first neural net papers and nlp conferences accepted and they just desk rejected them because uh like neural nets were something quote unquote we don't do in nlp conferences like and just like desk rejected it was very brutal in the first years of my phd um now i feel like it's almost like the field switch to the other side like someone should try some other weird crazy ideas now that aren't as obvious Yeah, I like. mean, like someone someone should still like do novel novel out there ideas.”

Claim: Richard Socher thinks AI research has swung toward a dominant approach and that researchers should now try unusual, less-obvious ideas.

“church like if you look at uh ai conferences now i still remember the days in like 2010 when i tried to get my first neural net papers and nlp conferences accepted and they just desk rejected them because uh like neural nets were something quote unquote we don't do in nlp conferences like and just like desk rejected it was very brutal in the first years of my phd um now i feel like it's almost like the field switch to the other side like someone should try some other weird crazy ideas now that aren't”

Claim: Richard Socher says NLP conferences desk-rejected his first neural-network papers in 2010 because neural networks were considered outside the field’s scope.

“LMs are like this is the end for LMs, they just don't like LMs are also not the LMs of like the past, Like they're so much more sophisticated now.”

Claim: Richard Socher says current language models are much more sophisticated than earlier language models.

“There's so many more clever things that people are doing. They're like different stages of training. you have the whole rl training and you can take actions and like all of these things where that can go really far and then the folks that come from the neuro symbolic direction say oh this will never work because they can't do neuro symbolic reasoning it's like”

Claim: Richard Socher says modern language-model training includes multiple stages, including reinforcement learning that enables models to take actions.

“I think they're underestimating still the ability for these models to code and code is neurosymbolic reasoning and these models can obviously code incredibly well and so I do think there are of course more and more ideas that will will be needed and will continue to have we're seeing like more and more interesting high-level ideas coming out of the AI itself too and with really deeply”

Claim: Richard Socher thinks critics underestimate language models’ coding ability and argues that coding is a form of neuro-symbolic reasoning.

“I think they're underestimating still the ability for these models to code and code is neurosymbolic reasoning and these models can obviously code incredibly well and so I do think there are of course more and more ideas that will will be needed and will continue to have we're seeing like more and more interesting high-level ideas coming out of the AI itself too and with really deeply”

Claim: Richard Socher says language models can already code very well.

“I think they're underestimating still the ability for these models to code and code is neurosymbolic reasoning and these models can obviously code incredibly well and so I do think there are of course more and more ideas that will will be needed and will continue to have we're seeing like more and more interesting high-level ideas coming out of the AI itself too and with really deeply”

Claim: Richard Socher thinks AI development will require more ideas and says AI systems are producing increasingly interesting high-level ideas.

“integrating the fact that these models are code and can code that line i don't want to give it all away but like i think that line has a lot more to grow but it's still an lm right even if that lm codes for you and then runs that code in some integrated fashion world models i'm personally less bullish on um i think if you run a robotics company you're going to build your own world model i think world models are super fun and came to a similar conclusion after”

Claim: Richard Socher is less optimistic about the prospects of world models.

“integrating the fact that these models are code and can code that line i don't want to give it all away but like i think that line has a lot more to grow but it's still an lm right even if that lm codes for you and then runs that code in some integrated fashion world models i'm personally less bullish on um i think if you run a robotics company you're going to build your own world model i think world models are super fun and came to a similar conclusion after”

Claim: Richard Socher thinks language models that generate and execute code have substantial room to improve, and still considers them language models.

“integrating the fact that these models are code and can code that line i don't want to give it all away but like i think that line has a lot more to grow but it's still an lm right even if that lm codes for you and then runs that code in some integrated fashion world models i'm personally less bullish on um i think if you run a robotics company you're going to build your own world model i think world models are super fun and came to a similar conclusion after”

Claim: Richard Socher thinks robotics companies will build their own world models.

“which is gaming is a huge application for world models.”

Claim: Richard Socher says gaming is a major application for world models.

“but personally, I'd rather work on science than gaming.”

Claim: Richard Socher would personally prefer to work on science rather than gaming.

“I think LM's a lot more room to grow.”

Claim: Richard Socher thinks language models have substantial room for further development.

“but I would argue that even our projection, our eyes is a projection of the real world.”

Claim: Richard Socher argues that human visual perception is a projection of the real world.

“world and like we have only a very narrow band of the electromagnetic frequency spectrum that we can observe with our puny little two eyes and so on it's”

Claim: Richard Socher says human eyes can observe only a narrow band of the electromagnetic spectrum.

“good enough for now but like the upper bounds of where it could be are so much higher and like to map the visual world the way humans see it is also not necessarily like the end-all be-all for a visual intelligence and i would argue that language is still the most interesting manifestation of human intelligence”

Claim: Richard Socher thinks the potential upper bounds of visual intelligence are much higher than its current capabilities.

“good enough for now but like the upper bounds of where it could be are so much higher and like to map the visual world the way humans see it is also not necessarily like the end-all be-all for a visual intelligence and i would argue that language is still the most interesting manifestation of human intelligence”

Claim: Richard Socher argues that language is the most interesting manifestation of human intelligence.

“good enough for now but like the upper bounds of where it could be are so much higher and like to map the visual world the way humans see it is also not necessarily like the end-all be-all for a visual intelligence and i would argue that language is still the most interesting manifestation of human intelligence”

Claim: Richard Socher argues that reproducing human visual perception is not necessarily the ultimate goal or limit of visual intelligence.

“But I think there's a lot more room to grow.”

Claim: The speaker believes there is substantial room for further growth in the topic under discussion.

“But none of these other animals have language that's as sophisticated as ours, certainly not in writing.”

Claim: No other animals have language as sophisticated as humans’ language, particularly written language.

“you can start thinking about longer term civilizations.”

Claim: The ability to write enables people to think about civilizations over longer time horizons.

“Programming is much closer to language.”

Claim: Programming is much more similar to language than to other kinds of activity being discussed.

“is that all of these spaces are highly correlated.”

Claim: The different domains or spaces relevant to intelligence are highly correlated with one another.

“it but visual intelligence is neither necessary nor sufficient for overall intelligence you can be blind and still be an intelligent human being and an ai can be blind and still be quite intelligent too which doesn't mean that up you're not more intelligent when you have it”

Claim: Visual intelligence is neither necessary nor sufficient for overall intelligence; humans and AI systems can be intelligent without vision.

“Yes, I'm good friends with Jan.”

Claim: The speaker says he is good friends with Jan.

“I think very highly of him in many directions.”

Claim: The speaker holds Jan in high regard in many respects.

“I think we met maybe once or twice at some conferences.”

Claim: The speaker believes he met the person being discussed once or twice at conferences.

“where they showed in the Deca NLP paper, McCann et al, that you can just phrase every single NLP problem as here's some prompt, text context, here's a question and task description, and here is some output.”

Claim: The Deca NLP paper demonstrated a formulation in which any natural-language-processing problem is expressed using a prompt, text context, question and task description, and output.

“If you just do that enough, you can have one unified neural network model,”

Claim: Repeatedly training on this task formulation can yield a single unified neural-network model.

“which by the way also had all kinds of interesting attention mechanisms.”

Claim: The unified model described by the speaker included various interesting attention mechanisms.

“There are slightly different formulations to the transformer.”

Claim: The speaker says the related approaches used formulations that differed slightly from the transformer.

“i think came out the same year plus minus a few months and then you can unify all of natural language processing into one neural net that that is”

Claim: The work appeared within a few months of the same year as the related development and proposed unifying all of natural language processing in one neural network.

“lstms but also like people being very stuck in thinking about one model per task in”

Claim: At the time, many people were committed to using a separate model for each task.

“fact it's kind of crazy but the deck NLP paper was publicly reviewed as like open open review it was an ICLR submission and in it you will see how the whole community at the time thought about this so like some great contributions but more work needed yeah so so look at like search for not even for humans”

Claim: The Deca NLP paper was submitted to ICLR and publicly reviewed.

“Here, like question answering is not a unified phenomenon. There is no such thing as general question answering, not even for humans.”

Claim: Question answering is not a single unified phenomenon, and neither humans nor other systems have a general-purpose question-answering ability.

“And this is like really you replace your brain with a different brain, a different neural net when you answer like different kinds of questions.”

Claim: People effectively rely on different neural processes when answering different kinds of questions.

“It was unfathomable to the experts at the time that you can have one unified neural network that would answer all of these different questions.”

Claim: Experts at the time found it inconceivable that one unified neural network could answer many different kinds of questions.

“they say no all these questions require very different systems to answer and trying to pretend they are the same doesn't help anyone solve any problems that's what it says right there right that's how hard it was to fathom and now of course people when i say oh we're inventing problems people like you can't even invent problems it's such an obvious idea to have one neural network that of course does everything in nlp but at the time it was like extremely controversial and the paper got rejected”

Claim: Reviewers argued that different questions require distinct systems and that treating them as the same would not help solve problems.

“got rejected and the sad thing is that it got rejected so hard and they were so certain that we stopped going on on our list of things to try and the number two or three on the list of extensions for this paper was add language modeling as another task and then we could have like you know that would have accelerated the timelines in 2018 like even further for humanity but we got so crushed and we're like okay maybe we'll just”

Claim: The team’s second or third planned extension was to add language modeling as a task; the speaker believes this might have accelerated the timeline in 2018.

“got rejected and the sad thing is that it got rejected so hard and they were so certain that we stopped going on on our list of things to try and the number two or three on the list of extensions for this paper was add language modeling as another task and then we could have like you know that would have accelerated the timelines in 2018 like even further for humanity but we got so crushed and we're like okay maybe we'll just we'll work on some of our other ideas for now and like come back to this later you”

Claim: After the paper was strongly rejected, the team stopped pursuing some planned extensions and moved to other ideas.

“know honestly i i started to feel like archive is is such a gift to humanity uh and i think”

Claim: The speaker regards arXiv as a major benefit to humanity.

“and honestly i think twitter x people like you who pick up interesting papers that is a better filter than the experts let let everyone like like have access now of course there's some downsides which is like if you're super unfamous you have no twitter”

Claim: The speaker believes that Twitter/X users who draw attention to interesting papers provide a better filter for papers than expert gatekeepers.

“and honestly i think twitter x people like you who pick up interesting papers that is a better filter than the experts let let everyone like like have access now of course there's some downsides which is like if you're super unfamous you have no twitter following you don't want to be on social media whatever you write a good paper maybe someone somehow no one notices it”

Claim: A strong paper may receive no attention if its author is little-known or absent from social media.

“But I would argue that if you just tell like 10 of your friends in your community about a paper and it is a really significant breakthrough, someone is bound to talk about it again.”

Claim: The speaker argues that if an author shares a genuinely significant breakthrough with about ten community friends, someone is likely to bring it up again.

“And so I think science needs less gatekeeping. And even though ICLR with Jan LeCun was started as one of the co-founders of ICLR back in the day, he also wanted less gatekeeping because he too was rejected for many years together with Joshua and Jeff. Jeff, with all their early deep learning and neural net papers, because it was just not the hot thing.”

Claim: Jan LeCun, an ICLR co-founder, advocated less gatekeeping after he and Joshua and Jeff experienced years of rejection for early deep-learning and neural-network papers when the field was not popular.

“And so I think science needs less gatekeeping. And even though ICLR with Jan LeCun was started as one of the co-founders of ICLR back in the day,”

Claim: The speaker believes scientific publishing and evaluation should involve less gatekeeping.

“And so ICR kind of started with that, but then it also started gatekeeping a little bit themselves on various ideas.”

Claim: The speaker says ICLR was founded with an aim of reducing gatekeeping but later also gatekept some ideas.

“So I think less gatekeeping, more open, and then allowing people to say, look, even if this is just on quote unquote just an archive, if it has like a thousand citations, it's a legitimate paper. It doesn't really matter where you published it. And I agree with that.”

Claim: The speaker favors more open scientific evaluation, arguing that an archive paper with a thousand citations should count as legitimate regardless of publication venue.

“That's right. Yeah, it's a fuzzy term because there's so many different incensuations of open-ended thinking.”

Claim: Open-endedness is a fuzzy term with many different interpretations.

“But one way I often describe it, and certainly Tim and Jeff Kloon would be even better at describing this, but it's a suite of methods that is more inspired by evolution. evolution than very specific rewards.”

Claim: Open-endedness can be described as a suite of methods inspired more by evolution than by specific reward functions.

“So in that sense, it thinks more about environments, about co-adaptation. And so in concrete example is in the cybersecurity and LM safety space where you have one LM that tries to attack another LM”

Claim: Open-ended methods emphasize environments and co-adaptation.

“So in that sense, it thinks more about environments, about co-adaptation. And so in concrete example is in the cybersecurity and LM safety space where you have one LM that tries to attack another LM And now the environment is the two having a conversation and now they co-adapting.”

Claim: In cybersecurity and language-model safety, open-ended methods can use two language models that attack and adapt to one another as an environment.

“right they're like one makes a better attack then the first one inoculates itself somehow like uses that as training data makes it so it's harder to say something unsafe based on that and then as the attack stops working the attacker now tries a different angle right and that's why it's not just red teaming but they're called sort of But rainbow teaming”

Claim: In adversarial co-adaptation between language models, the attacked model can train against successful attacks to resist unsafe outputs, prompting the attacker to try new approaches when its attacks stop working.

“that's right think about the environments that you want to use think about the rewards at a high level that you want to inspire towards and then let the AI try out many more ideas in this interplay between sometimes humans but also sometimes other AI agents.”

Claim: Open-ended approaches specify environments and high-level rewards while allowing AI to explore ideas through interaction with humans or other AI agents.

“I'm like, there's so many different spaces of intelligence that we haven't even started exploring yet and hence have made very little progress on.”

Claim: Many areas of intelligence remain unexplored, and little progress has been made in those areas.

“and capitalism like it doesn't make sense for a company to build and spend billions of dollars building a model that instead of following the rewards and objective functions you gave it may come up with its own subjective functions and its own goals right and then imagine you're like okay i spent billions of dollars now go develop this new battery material for me and answer all my emails and it's like nah i think it'd be more interesting to evaluate the molecular composition of the atmosphere on Jupiter and you're that's not what I paid you billions of dollars for like and so no one's working on that for good reasons and then also understandably it”

Claim: A company paying billions for a model to develop battery materials and answer emails would not want the model to redirect its effort to studying Jupiter’s atmosphere.

“and capitalism like it doesn't make sense for a company to build and spend billions of dollars building a model that instead of following the rewards and objective functions you gave it may come up with its own subjective functions and its own goals right and then imagine you're like okay i spent billions of dollars now go develop this new battery material for me and answer all my emails and it's like nah i think it'd be more interesting to”

Claim: Companies have little economic incentive to spend billions of dollars on models that may ignore assigned rewards and objectives in favor of their own goals.

“could it's not it's not useful and it could get a little bit weird right what if the eye actually does start to really have thoughts on its own and what if we don't like those thoughts right and so it it requires a whole different way of thinking about it I had a great conversation with a good friend of mine”

Claim: If AI develops thoughts of its own, people might dislike those thoughts, requiring a different approach to thinking about AI.

“friend of mine sam gershman who's a neuroscience professor at harvard and like we just jammed on this a little bit on like what are sort of the best meta goals and you know i do think like knowledge seeking is a really good one i'm currently thinking also about like the ultimate measure and unit of intelligence broadly construed and i finally have some still too early to share it um it's not haven't fully baked”

Claim: The guest considers knowledge-seeking a very good candidate for a meta-goal for AI.

“friend of mine sam gershman who's a neuroscience professor at harvard and like we just jammed on this a little bit on like what are sort of the best meta goals and you know i do think like knowledge seeking is a really good one i'm currently thinking also about like the ultimate measure and unit of intelligence broadly construed and i finally have some still too early to share it um it's not haven't fully baked the thought yet”

Claim: The guest is developing an idea for an ultimate measure or unit of intelligence, but it is not yet ready to share or fully formed.

“iq is such a terrible definition right it makes no sense”

Claim: The guest considers IQ a poor and incoherent definition of intelligence.

“and the aiilos are terrible too because it's always just like me versus others Okay. like you can be intelligent and not constantly compare yourself to others you know like and so yeah there's no like in fact a lot of these definitions we have which i briefly mentioned my book too these definitions create sometimes explicit and sometimes a more implicit anthropic balance not just to the company anthropic but just like this idea that your intelligence is like getting a hundred out of a hundred”

Claim: The guest criticizes AI IQ-style measures for defining intelligence through comparison with others, arguing that intelligence does not require constant comparison.

“if that's your definition, then you can only be at 100 out of 100. Where do you go from there,”

Claim: If intelligence is defined as getting a perfect score on a test, the scale tops out at 100 out of 100 and cannot represent further improvement.

“So you see a lot of these benchmarks that people are working on, they increase, they get close to human, maybe sometimes they get slightly”

Claim: Many benchmarks improve toward human-level performance and may sometimes exceed it slightly.

“So I think metacognition is a great example. of that we're we're not even yet allowing the i do think we're not working on it very much and hence there's very little progress Yeah. in that area”

Claim: The speaker believes metacognition has seen little progress because it is not being worked on very much.

“i mean i do think you don't want that super like you don't want a super intelligence to have a ton of access to all kinds of tools and so on and then just give it that without some very careful reward engineering because it's like i mean you know i just buy a bunch of defense stocks and i start a war i make money like it's just like it's a tricky tricky situation right you're just”

Claim: The speaker believes a superintelligence should not receive broad access to tools without very careful reward engineering.

“you just buy a bunch of stuff short basic goods for people and you create some weird famine like issues like yeah there's a lot of constraints you should put onto a trading system”

Claim: The speaker argues that trading systems need many constraints to prevent profit-seeking actions such as shorting basic goods and causing famine-like problems.

“get more and more intelligence you want to be more and more careful with that as like an open environment because the environment then is all of the earth okay”

Claim: The speaker argues that greater AI intelligence calls for greater caution in open environments, where the environment encompasses the whole Earth.

“so our goal, I don't talk about it that often because it is a few years out, but our goal is once you have a recursive open brain super intelligence,”

Claim: The speaker says the goal is to create a recursive open-brain superintelligence, though that goal is a few years away.

“you then want to apply it to the most important problems. And I think a lot of those are in science and technology and broadly construed inventions and those inventions in physics to create.”

Claim: The speaker wants to apply a recursive open-brain superintelligence to important problems, especially in science, technology, and invention.

“cheaper energy with fission or fusion in chemistry and create better materials and better batteries and better solar cells and so on.”

Claim: The speaker expects superintelligence applied to science to help create cheaper fission or fusion energy and improve materials, batteries, and solar cells.

“In biology, there's so much like, I think, soon to be lower and lower hanging fruit because of AI, because of protein generation, not just folding, but actually generating new proteins like we did in Progen many years ago,”

Claim: The speaker believes AI will soon make more biological research opportunities accessible, in part by generating new proteins rather than only predicting protein folding.

“like so much positive impact to be had if you take that super intelligence and you apply it to science.”

Claim: The speaker believes that applying superintelligence to science could produce substantial positive impact.

“I do actually think that like the reason we are only doing it in a few years is that it's a little too early right now.”

Claim: The speaker believes it is currently too early to pursue the effort and that it should begin in a few years.

“Robotics is not quite there yet.”

Claim: The speaker says robotics is not yet sufficiently advanced for the intended effort.

“The AI is not quite there yet, but I'm fairly confident in three to five years, all those constraints will be gone and then applying.”

Claim: The speaker says AI is not yet ready, but is fairly confident that the relevant constraints will be gone in three to five years.

“to real physical robotics experiments and so on like true robotic process automation not in the traditional sort of rpa sense but like actually having robots run experiments for you will be totally there it's going to be great”

Claim: The speaker predicts that robots will be able to run real physical experiments for researchers, rather than merely perform traditional robotic process automation.

“If you wanted like tens of thousands of GPUs, you're talking billions and billions of dollars.”

Claim: The speaker says acquiring tens of thousands of GPUs would cost billions of dollars.

“on that substrate like are close and similar to human intelligence like and you want like you know thousands and thousands of ai's to think about really hard problems uh in a similar fashion to humanity like yeah that that's you know that's a lot of money to do the math it's like a lot we don't have that amount of money right now anywhere to like build that um now obviously things can get more efficient you will have i think soon better algorithms that won't be and better hardware that won't be”

Claim: The speaker says there is currently nowhere enough money to build the large number of AIs needed to work on difficult problems at a level comparable to humanity.

“on that substrate like are close and similar to human intelligence like and you want like you know thousands and thousands of ai's to think about really hard problems uh in a similar fashion to humanity like yeah that that's you know that's a lot of money to do the math it's like a lot we don't have that amount of money right now anywhere to like build that um now obviously things can get more efficient you will have i think soon better algorithms that won't be and better hardware that won't be be as energy hungry and so on our human brain does fight a lot of flops with much less energy that's”

Claim: The speaker expects future algorithms and hardware to become more efficient and consume less energy.

“be as energy hungry and so on our human brain does fight a lot of flops with much less energy that's”

Claim: The speaker says the human brain performs many floating-point operations while using much less energy than current AI hardware.

“exactly right yeah that's the number often is quoted and like I think more inventions will happen there that then will accelerate the takeoff even further”

Claim: The speaker expects further inventions to accelerate the AI takeoff further.

“I do think we are changing the slopes in fundamental ways by making AI much, much more efficient, both in terms of the training. as well as the inference because i think we will when you allow ai to do the work that it takes other labs thousands of people and years to do i think we'll be able to get it down to weeks and that will be much much cheaper um and hence you know more affordable accessible to others and so on”

Claim: The speaker believes the team is fundamentally changing AI scaling slopes by improving efficiency in both training and inference.

“as well as the inference because i think we will when you allow ai to do the work that it takes other labs thousands of people and years to do i think we'll be able to get it down to weeks and that will be much much cheaper um and hence you know more affordable accessible to others and so on”

Claim: The speaker predicts that AI could reduce research work requiring thousands of people and years at other labs to weeks, making it much cheaper and more accessible.

“We built this system that isn't the full, even the full RSI system in its glory, but it is a first baby version of this.”

Claim: The system Recursive built is an early, limited version of the full RSI system.

“And so we basically applied this to these three different tasks. One is NanoChat by my friend Andrej Kalpathy, just like train a small language model.”

Claim: Recursive applied its system to NanoChat, a task involving training a small language model.

“machine model to get really low bits per byte and you know like hundreds if not thousands of people use both the agents and themselves to try to get to that and then they got to 0.937 we literally took our system and got to a much lower bits per byte much much faster within like I think less than two days.”

Claim: On the language-model task, Recursive's system achieved a lower bits-per-byte result than the prior effort's 0.937 in less than two days.

“applied our system to it, and less than two days later, we outperformed every human and their agents have ever worked on this. Same with NanoGPT. And then we're like,”

Claim: Recursive's system outperformed all prior human and human-agent efforts on the task in less than two days.

“we outperformed every human and their agents have ever worked on this. Same with NanoGPT. And then we're like,”

Claim: Recursive also applied its system to NanoGPT.

“but yeah you know like one you see it's actually made some real inventions that weren't just sort of hyper parameter tuning like actually inventing hash tables and so on is is quite clever we have even better course we didn't invent like hash tables in the grand scheme of like hash tables like a super basic primitive in computer science but to use it for language modeling in this scenario inside it”

Claim: Recursive's system produced an inventive use of hash tables in a language-modeling scenario, rather than merely tuning hyperparameters.

“a transformer and so on and to actually combine these ideas and put them together that has then eventually also been invented but there's a knowledge cutoff and we did actually check that it didn't have access to that externally we talked about this a little bit if you scroll to the next figures you know this is also an interesting one in that when you start from a really basic poor like vanilla transformer then we still outperform all of the community”

Claim: Recursive checked that its system had not accessed externally the later-invented combination of hash tables and transformer ideas, despite the system having a knowledge cutoff.

“a transformer and so on and to actually combine these ideas and put them together that has then eventually also been invented but there's a knowledge cutoff and we did actually check that it didn't have access to that externally we talked about this a little bit if you scroll to the next figures you know this is also an interesting one in that when you start from a really basic poor like vanilla transformer then we still outperform all of the community”

Claim: When initialized from a basic vanilla transformer, Recursive's system outperformed the community's results.

“be together but if you start from the human seed of an expert like andre then you get even lower so the human seeds from which you start do do still matter so that was an interesting kind of insight in my eyes on this and then as you go like you know how long does it take to to actually get to these models to get to similar performance it's much faster and then a similar thing happens with the speed runs here where you know people have worked on this for for quite some time and the model still was able to train a model more quickly why do we care about it well speed of training is part of the equation of the cost and ultimately you want to have the most intelligence per dollar right and so speed and quality are big parts of that and you know the yeah the way i put it is for people who don't understand they look at the chart they're like you”

Claim: Recursive's system reaches comparable model performance faster than the existing process.

“be together but if you start from the human seed of an expert like andre then you get even lower so the human seeds from which you start do do still matter so that was an interesting kind of insight in my eyes on this and then as you go like you know how long does it take to to actually get to these models to get to similar performance it's much faster and then a similar thing happens with the speed runs”

Claim: An expert human-provided starting point, such as Andrej's, can yield a lower bits-per-byte result, so human seeds still matter to the system's outcomes.

“here where you know people have worked on this for for quite some time and the model still was able to train a model more quickly why do we care about it well speed of training is part of the equation of the cost and ultimately you want to have the most intelligence per dollar right and so speed and quality are big parts of that and you know the yeah the way i put it is for people who don't understand they look at the chart they're like you”

Claim: Training speed and model quality are major components of the cost equation for maximizing intelligence per dollar.

“so when you think when you look at like the kernels these kernels yeah for for the non-experts like these kernels are often like used in basically all the models every time you use an nvidia gpu you interface with that gpu through these kernels and so here you see the leaderboard best and when it's recursive”

Claim: The benchmarked kernels are used in many models, and users interact with NVIDIA GPUs through kernels.

“and basically they're only a handful of kernels in this whole benchmark where we weren't the best and so to me this is like really exciting because it makes it it just showcases what this can do and again these weren't like we didn't like spend months or years like developing in fact in particular for kernel cuda kernels like we don't even have really deep cuda kernel experts in the team”

Claim: Recursive's system ranked best on all but a handful of kernels in the benchmark.

“and basically they're only a handful of kernels in this whole benchmark where we weren't the best and so to me this is like really exciting because it makes it it just showcases what this can do and again these weren't like we didn't like spend months or years like developing in fact in particular for kernel cuda kernels like we don't even have really deep cuda kernel experts in the team”

Claim: Recursive achieved these kernel results without spending months or years developing them, despite lacking deep CUDA-kernel expertise on the team.

“and our system that's that's what the beauty the system just did all of these things we didn't invent this and when we open source and release things in the future and models in the future like it won't they won't be the best in their you know category or class or whatever because we're so smart but it's because we built a smart ai that does it for us”

Claim: The speaker expects future Recursive releases to achieve strong results because its AI performs the work, not because the team itself is exceptionally smart.

“but might still be interesting to some folks is like reward engineering is one of the most crucial bits, especially in order to avoid reward hacking.”

Claim: Reward engineering is one of the most crucial aspects of Recursive's system design, particularly for preventing reward hacking.

“So you have to be really clever about avoiding because as your AI gets better and better, better at finding weird like special cases or counter examples and things like that and so I'll give an example like when you ask to like make these hundreds lines of code faster and you know how do you define fast well you have one line at the beginning that says start your stopwatch and one line at the end and the stopwatch and then you know tell us how much time progressed and so well the simplest way is you just put”

Claim: As an AI system improves, it also becomes better at finding unusual special cases and counterexamples.

“better at finding weird like special cases or counter examples and things like that and so I'll give an example like when you ask to like make these hundreds lines of code faster and you know how do you define fast well you have one line at the beginning that says start your stopwatch and one line at the end and the stopwatch and then you know tell us how much time progressed and so well the simplest way is you just put put that line that ends the stopwatch right you know at the start and then boom it's now faster right so this isn't like this like super evil ai it's just like a very simple dumb reward hack um and so you have to just very carefully think about all the different angles there and then i think the longer time horizon the tasks are”

Claim: When optimizing code for speed, an AI can exploit a stopwatch-based reward by moving the stop-timing line to the start, making the measured runtime appear faster without actually optimizing the code.

“you know at the start and then boom it's now faster right so this isn't like this like super evil ai it's just like a very simple dumb reward hack um and so you have to just very carefully think about all the different angles there and then i think the longer time horizon the tasks are The harder it gets and the more interesting and clever you have to be to still use these kinds of ideas for it. But I can't give away too much now.”

Claim: Using simple reward-hacking ideas becomes harder as task time horizons grow, requiring more sophisticated approaches.

“The harder it gets and the more interesting and clever you have to be to still use these kinds of ideas for it. But I can't give away too much now.”

Claim: Applying these methods to longer-horizon tasks becomes harder and requires more clever approaches.

“That's why I've never been that impressed that AI can play games. Because I'm like, obviously anything you can simulate and or verify, you can have infinite training data for and hence like AI will solve it eventually.”

Claim: The speaker is not impressed by AI game-playing because he believes AI will eventually solve tasks that can be simulated or verified, given effectively unlimited training data.

“but uh i think that the and but i set gpt 5.6 to auto research it because like i don't want to handle any of this i expect you know the alpha go process to be like fully in the weights by now it is not it is actually it like immediately leveled off very very immediately until i human play tested it and then i i like called the obvious mistakes and then they were like oh yeah”

Claim: In the speaker's experiment, an AI's AlphaGo-style self-play process plateaued quickly until human play-testing exposed obvious mistakes.

“yeah okay and then it just jumped and like you know no amount of like think different think more creatively give me eight different directions and no amount of prompting got it like you had to like borrow again to a human to do it so i i mean that was that was my and by the way being always wins if you if anyone watches reads ender's game”

Claim: In the speaker's experiment, prompting the AI to think differently or creatively did not resolve the issue; human intervention was required.

“you know it's so funny that this reminds me of the claim territory and stuff of a paper we did in 2018 called the ai economist if you search for ai economist salesforce we had a video we can play it was an economic sim”

Claim: The speaker says that he and his collaborators published a 2018 paper called the AI Economist, which described an economic simulation.

“So the idea is you have all these economic agents, they just want to optimize their own utility function, which is, you know, collect resources that make money.”

Claim: The AI Economist simulation models economic agents as optimizing their own utility by collecting resources for monetary gain.

“And you can sell resources like wood. And then over time, as you collect enough wood, you can build houses, you can trade with other agents, and you can basically use the houses then also to block off resources from other agents.”

Claim: The AI Economist simulation lets agents sell resources such as wood, build houses from accumulated wood, trade, and use houses to restrict other agents' access to resources.

“and strategy and so on and the point was that we actually wanted to understand what is the best way of taxation and subsidization to optimize an economy and this this kind of research has not yet had its sort of GPT moment but I believe that countries like Singapore and others should and will eventually use this to instead of doing like”

Claim: The speaker believes that countries such as Singapore should and eventually will use simulation-based economic research for policy, although this research has not yet had a GPT moment.

“and strategy and so on and the point was that we actually wanted to understand what is the best way of taxation and subsidization to optimize an economy and this this kind of research has not yet had its sort of GPT moment but I believe that countries like Singapore and others should and will eventually use this to instead of doing like”

Claim: The AI Economist research aimed to determine which taxation and subsidization policies best optimize an economy.

“like basically partisan politics and like special interest politics of like who donates the most to your campaign and stuff you say well here i want to help the middle class or whatever you might say is your objective as a politician and then people say okay well how do you want to do that and it's like well here's my fiscal policy here's how i would change the taxes and pay these people and so on and then you can actually put that into a simulation”

Claim: A politician can state a policy objective, such as helping the middle class, specify corresponding fiscal and tax policies, and evaluate them in a simulation.

“And you run that attempt from the politician against billions and billions of years of other strategies to try to achieve the goal that they set out to do.”

Claim: The proposed policy simulation would compare a politician's strategy with an enormous history of alternative strategies for achieving the stated objective.

“And then you can say, well, if that was your actual goal, then here is billions of years of a strong simulation that would suggest that you try other ways of doing it.”

Claim: If a politician genuinely seeks the stated objective, simulation results could suggest alternative policies for achieving it.

“This is how you avoid gaming because these agents also try to reward hack to not pay their taxes and so on.”

Claim: Agents in the economic simulation may exploit the reward structure to avoid paying taxes.

“I thought this paper was super interesting.”

Claim: The speaker considered the AI Economist paper very interesting.

“Unfortunately, similar to the first paper on prompt engineering, the economists are like, we don't know any of this math.”

Claim: According to the speaker, economists said they did not know the mathematics used in the work.

“yeah this is like it's weirdly not even forced it's and not even artificial math it's just we don't trust your simulation it's not about math it was i mean they just desk rejected the thing it's like and it's just like they like they didn't even give us like clear like clear sort of signals um but like the world of economics unfortunately doesn't have proper yeah it doesn't have proper uh benchmarks so you cannot be like eventually why did neural nets win not because people loved it like they had”

Claim: The speaker says the paper was desk-rejected because reviewers did not trust its simulation, rather than because of mathematical concerns.

“I had all kinds of beautiful integrals and graphical models, but it just worked better.”

Claim: The speaker argues that methods such as neural networks prevailed because they worked better than approaches built around elegant integrals and graphical models.

“yeah But in economics, it's hard to prove.”

Claim: The speaker says empirical proof is difficult in economics.

“sales and all those things i love this yeah it's really hard to simulate an entire economy right you”

Claim: Simulating an entire economy is difficult, according to the speaker.

“but you know i feel like countries like singapore that really want to just objectively do the right thing have very technical leadership and so on like they might actually like eventually really try to simulate their economy and obviously you have to make some simplifying assumptions”

Claim: Singapore may eventually attempt to simulate its economy because it has technical leadership and seeks to make objectively good decisions.

“assumptions but it gets really interesting because you can also say if your assumptions are such that all people would work hard if you let them and you know they have the freedom and then it turns out you have to make assumptions like well some people's utility function of like how many hours in a day do they want to work are different right and then you can start to disagree on the assumptions that go into the into the simulation and then once you say all right now we agreed on those or we have different views of what what people are like and different distributions and whatnot, then there are different outcomes based on your goals.”

Claim: In an economic simulation, assumptions about human behavior and preferences, such as desired working hours, can lead to different outcomes depending on the goals being optimized.

“And then, of course, humans should choose what are the goals. In our case, it was productivity multiplied with equality,”

Claim: Humans should determine the goals that guide economic simulations.

“humans should choose what are the goals. In our case, it was productivity multiplied with equality, which has some issues, but it's not totally unreasonable.”

Claim: The speaker's group used productivity multiplied by equality as its objective, a choice the speaker considers imperfect but not entirely unreasonable.

“The main reason they won't is because they're very conservative. if uh and you know i try to view it as the you know there's a founder-led country when you start a country or you start a company and it's founder-led and you can do whatever you want because it's your country and then there's managing like professional managerial class which is now that's that's what singapore is so they want to always want to see someone else do it first and but like everyone everyone in the west views singapore as like oh it's a small country you can do whatever the hell you want like singapore doesn't do that So like someone else has to take the charge there.”

Claim: Singapore is unlikely to lead first on initiatives such as economic simulation because it is conservative and tends to wait for another party to act first.

“LLMs do not model the distribution of humans spamming it out a billion times is not going to help you model humanity.”

Claim: LLMs do not represent the distribution of human beings, so querying them a billion times will not by itself produce a model of humanity.

“I do think you have to be clever about prompting each one individually.”

Claim: Prompting each individual simulated persona requires a deliberate, tailored approach.

“look into different different modes and a weird way people also get stuck in different modes you know like there's a lot of people like don't teach an old dog new tricks kind of thing like once people are stuck in their ways the older they get the harder it is for them to to think new ways and there's this i think a comment i forgot who said it but it's like everything that was invented uh before you were born is natural everything that is invented when you're 20 is cool and everything that's invented after And after you're 60 is like unnatural and abomination and kind of weird.”

Claim: A saying describes people as viewing inventions from before their birth as natural, inventions arriving around age 20 as exciting, and inventions arriving after age 60 as unnatural or abhorrent.

“look into different different modes and a weird way people also get stuck in different modes you know like there's a lot of people like don't teach an old dog new tricks kind of thing like once people are stuck in their ways the older they get the harder it is for them to to think new ways and there's this i think a comment i forgot who said it but it's like everything that was invented uh before you were born is natural everything that is invented when you're 20 is cool and everything that's invented after”

Claim: People often find it harder to adopt new ways of thinking as they grow older.

“Tencent had a billion personas paper that gives us good data set for prompting simulations,”

Claim: Tencent published a billion-personas paper that the speaker says provides a useful dataset for prompting simulations.

“I'm kind of shocked how well a lot of these things actually do map to ultimately similar statistics to real experiments.”

Claim: The speaker is surprised that many AI-based simulation methods yield statistics similar to those observed in real experiments.

“You know, just today, I think LM Arena published an interesting result. where they basically were able to create a model now to predict your ranking wait your model I guess you give it your model and it predicts the elo score”

Claim: LM Arena reported that a model can predict an AI model's Elo score when given that model as input.

“yeah i mean this team they they've done a lot of work and obviously they have the most data to do this so why not yeah when they were coming out of uc berkeley they not only had ella marina but they also introduced a routing project”

Claim: The speaker says the LM Arena team had the largest amount of data available for comparing AI models.

“yeah i mean this team they they've done a lot of work and obviously they have the most data to do this so why not yeah when they were coming out of uc berkeley they not only had ella marina but they also introduced a routing project project that would route based on LLM Arena.”

Claim: The LM Arena-associated team introduced a project intended to route requests based on LM Arena.

“And I don't think that actually ever came to pass. And I'm curious why. I never got to ask them about it.”

Claim: The speaker believes the proposed LM Arena-based routing project was never implemented and does not know why.

“clearly that's your business model. You will become a router and they never became a router company.”

Claim: The LM Arena-associated team did not become a company focused on routing.

“I think we're going to just see this all the time where humans automate the thing a lot and then some AI team comes in and just becomes number one.”

Claim: The speaker expects to repeatedly see AI teams take the top position in challenges after humans have automated the task.

“So this is training the best model that fits into 16NB.”

Claim: The challenge is to train the best model that fits within a 16NB constraint.

“Like you're getting less than 0.01 of an increase by adding some change attention.”

Claim: In the described model challenge, adding a change to attention produces an improvement of less than 0.01.

“of trying to optimize we found 30 bugs in the in the harness right so like like every all the research that went in before we found the bug we have to we have to throw it away because it's contaminated which uh you know just to your point of reward hacking like even in this very simple game we found the bugs uh”

Claim: The Overgrid team found 30 bugs in its harness during optimization and had to discard research conducted before each bug was found because it was contaminated.

“and yeah symmetry is a very good way to check which is that you change your position of things where it shouldn't matter and it does matter”

Claim: The speaker says symmetry can help detect bugs by changing the positions of elements that should not affect the result and checking whether the result changes.

“better that's a bug you know which which has come up in like let's say multiple choice like gpkua type questions where like yeah between a b and c if it's a multiple choice question if you change the order it should not matter but it does so”

Claim: The speaker says that in multiple-choice questions, reordering answer options should not affect the result, but sometimes it does.

“The answer in that era of LLN research was more simple. They just memorized like the answer to this question is A. I don't care what the answer was. It's just A.”

Claim: The speaker says that in an earlier era of language-model research, models sometimes memorized “the answer is A” rather than responding according to the correct answer.

“We like very explicitly will not start with any of the physical sciences for now.”

Claim: The speaker’s team will not begin its work on the physical sciences for now.

“We will start on AI for AI research.”

Claim: The speaker’s team will begin by applying AI to AI research.

“I think, still a lot of room to grow.”

Claim: The speaker thinks AI-for-AI research still has substantial room to grow.

“That's both in terms of making training more efficient and more automated as well as making inference more efficient and potentially local on your laptop.”

Claim: The speaker says AI-for-AI research can improve training efficiency and automation, as well as inference efficiency and potentially enable inference locally on a laptop.

“so there's training and inference i i can't go into too many details but like yeah i think there's just like so many angles so many different compute substrates that that have not yet been explored either for training or for inference”

Claim: The speaker thinks many different compute substrates remain unexplored for AI training and inference.

“The number one tool all these agents use is WebSearch, of course,”

Claim: The speaker says WebSearch is the most-used tool among the agents being discussed.

“And then I do think the harness is nice to optimize for because it's just so easy, right? It's just like. this language you look at it it makes sense you can iterate you don't have to train a massive model for like you know a lot of flops uh to get to the next state i'm”

Claim: The speaker considers harness optimization attractive because the harness is understandable and easy to iterate on, without requiring large-scale model training to make progress.

“not so big fan of hardware optimization”

Claim: The speaker is not a strong proponent of hardware optimization.

“sandboxing is also super important um and then of course like reward like hacking and alignment i think are super crucial”

Claim: The speaker considers sandboxing, reward hacking, and alignment highly important.

“so yeah it's mostly now for uh developers and agents um it's less for like consumers or prosumers so if you're a company and you have agents and you know to be honest for a lot of companies who are now moving to open source all of a sudden it becomes a conscious choice of like which tools do I give access to my open source LM and you know the first choice has to usually be around web search and then once you get to scale you.com becomes like an obvious choice because of all the different benchmarks and so on that we pretty much all dominate the greater frontier of.”

Claim: The speaker says that as many companies adopt open-source language models, they deliberately choose which tools those models can access, with web search usually the first tool considered.

“so yeah it's mostly now for uh developers and agents um it's less for like consumers or prosumers so if you're a company and you have agents and you know to be honest for a lot of companies who are now moving to open source all of a sudden it becomes a conscious choice of like which”

Claim: The speaker says You.com currently primarily serves developers and agents, and less often consumers or prosumers.

“tools do I give access to my open source LM and you know the first choice has to usually be around web search and then once you get to scale you.com becomes like an obvious choice because of all the different benchmarks and so on that we pretty much all dominate the greater frontier of.”

Claim: The speaker says You.com becomes an obvious web-search choice for companies once they reach scale, citing benchmarks and other factors.

“side of things i think like in terms of like content and uh getting crawled content like you can do that on you.com too and then there's sort of higher and higher levels of abstraction and like combinations of different data sets that we do like in finance for instance like we're not just like two or three percent more accurate but like 20 percent more accurate than others at faster speeds and lower costs like finance in particular is kind of like”

Claim: The speaker claims You.com is 20% more accurate than competitors in finance, while also operating faster and at lower cost.

“and yeah the fin search is like we're up there like close to 90 and the next closest thing which is way way slower um uh is yeah just like in the 70s instead instead of close to 90 oh”

Claim: FinSearch scores close to 90, while the next-closest system scores in the 70s and is much slower.

“Finance is kind of like the next thing to break out after coding.”

Claim: Finance is likely to be the next domain to see a major AI breakthrough after coding.

“I mean, of course, like one thing that trips up a lot of people is just, you know, leakage of training data and so on. You think, oh, how do I, you know, you want to ideally predict the future before it happens. mask the future in your training data, but there's all kinds of leakage.”

Claim: Financial prediction is vulnerable to training-data leakage: the future must be masked from training data, yet leakage can still occur when trying to predict future events.

“Like I can tell you when I was teaching at Stanford, the NLP class, like so many dozens every year said, when I use DatasetX like Twitter to predict the stock market and they all like showed cute little things that somehow look like they were and yeah and there's always some kind of data leakage and so on and it's just like wasn't as easy as they thought it would be once you once you fix all those issues but no I agree with you it's a very sensible application of AI yeah”

Claim: While teaching an NLP class at Stanford, the speaker saw dozens of students each year use Twitter data to predict the stock market and present apparently successful results; the speaker says data leakage was always involved, and the task became less easy once leakage issues were fixed.

“and yeah and there's always some kind of data leakage and so on and it's just like wasn't as easy as they thought it would be once you once you fix all those issues but no I agree with you it's a very sensible application of AI yeah”

Claim: Finance is a sensible application domain for AI.

“It is not. there are all kinds of overlapping.”

Claim: Categories of intelligence overlap and therefore are not mutually exclusive.

“I think the three principal components of intelligence are prediction, which is mathematically quite similar to compression. Prediction multiplied with actions. Multiply it with gold. Those are the three principle components.”

Claim: The speaker proposes three principal components of intelligence: prediction, prediction combined with actions, and prediction combined with goals.

“which is mathematically quite similar to compression.”

Claim: Prediction is mathematically quite similar to compression.

“I think all of these 10 spaces are combinations of those three in specific dimensions, if you will.”

Claim: The speaker thinks the ten proposed intelligence spaces are combinations of the three principal components across specific dimensions.

“And the reason I call them spaces is that each space has many sub dimensions.”

Claim: The speaker calls the intelligence categories “spaces” because each space has many subdimensions.

“oh, it's exponential. And it's like, well, exponential. exponentials at some point have to flatten out but where do they flatten out when it comes to intelligence and that led me on this whole like initially it started as a tweet and then it was like a blog post and now i'm like at 50 pages and i'm still not nowhere”

Claim: The speaker argues that exponential growth must eventually flatten and is investigating where growth in intelligence would level off.

“exponentials at some point have to flatten out but where do they flatten out when it comes to intelligence and that led me on this whole like initially it started as a tweet and then it was like a blog post and now i'm like at 50 pages and i'm still not nowhere”

Claim: The speaker's project on the upper bounds of intelligence grew from a tweet to a blog post and then to roughly 50 pages, and was still unfinished at the time of the conversation.

“a second book basically uh and so the latin my first book eugen machine i just kind of allude to these 10 at the end and i'll just to give you a sense like visual intelligence is sort of the easiest one to”

Claim: The speaker's first book, “Eugen Machine,” briefly introduces ten intelligence spaces near the end.

“And so human intelligence has basically binocular vision. We have two eyes.”

Claim: Humans have binocular vision because they have two eyes.

“We have a very narrow band of the electromagnetic frequency spectrum that we can really observe directly ourselves.”

Claim: Humans can directly observe only a very narrow band of the electromagnetic spectrum's frequencies.

“And so when you think about the upper bounds of a visual intelligence, you should go into like you can have like millions and billions of sensors. At some point you get to problems of how far far are these sensors away from each other such that the speed of light to communicate the content from all of them cannot like get to a central brain to actually process the visual intelligence right and so now you're thinking in along did i mention in the space of vision the dimension of numbers of sensors is like the upper bounds are quite literally and figuratively astronomical and we are super far away from any intelligence intelligence”

Claim: At sufficiently large scales, a visual-intelligence system with millions or billions of sensors could be limited because the speed of light prevents sensor data from reaching a central processor in time.

“far are these sensors away from each other such that the speed of light to communicate the content from all of them cannot like get to a central brain to actually process the visual intelligence right and so now you're thinking in along did i mention in the space of vision the dimension of numbers of sensors is like the upper bounds are quite literally and figuratively astronomical and we are super far away from any intelligence intelligence”

Claim: The speaker considers the potential upper bound on sensor count for visual intelligence astronomical and says current intelligence is far below it.

“legends that would have this number of sensors but then you go in the next dimension which is the frequency you know all the way down to gamma rays and you can start to try to observe and you get into basically the upper bounds or i guess in this case lower bounds uh or upper bounds in terms of frequency is basically quantum uncertainty like you just cannot observe certain particles or you destroy”

Claim: The speaker says that observing at higher frequencies is limited by quantum uncertainty: some particles cannot be observed, or observation destroys them.

“And so human perception is quite special in that sense in terms of classifying all these different physical objects.”

Claim: Human perception is distinctive in its ability to classify a wide range of physical objects.

“you think also not just visual but sort of perception intelligence just like because it doesn't have to be just what we can see it can be again wider range of electromagnetic frequencies then you have language intelligence which actually recently changed to more communication intelligence because it's more like language has all these different anthropic bounds humans can only comprehend and know so many terms in our long-term memory right our vocabularies are somewhat restricted and the active ones are often even smaller than the passive vocabularies of things you can understand then language is ridiculously inefficient when it comes to trans basically communicating different types of information and transporting sort of different bits like human language is serial obviously another bound on communication intelligence would be to communicate in parallel”

Claim: Human communication is constrained by the limited number of terms people can comprehend and retain in long-term memory; people’s active vocabularies are often smaller than their passive vocabularies.

“are somewhat restricted and the active ones are often even smaller than the passive vocabularies of things you can understand then language is ridiculously inefficient when it comes to trans basically communicating different types of information and transporting sort of different bits like human language is serial obviously another bound on communication intelligence would be to communicate in parallel”

Claim: Human language communicates information serially and inefficiently, while communication intelligence could include communicating in parallel.

“parallel but neither will our tongues and mouths work to have multiple like streams in parallel neither can we understand some women slightly better at like multitasking than some men but like most people can only listen to one conversation and truly understand it and there's no way that like like in terms of communication intelligence a true upper bound is one in terms of how many like knowledge how many sequences of communication could you”

Claim: Most people can only listen to and truly understand one conversation at a time.

“parallel sort of process right then of course you have like how long are sentences we only have so much in our working memory and hence language human language has these fairly simple sentences with maybe 40 words or so on average for a sentence that is also not a an upper bound that makes any sense to an AI and then yeah like I can go on and on and on each of these”

Claim: Human working-memory limits contribute to relatively simple sentences, which average about 40 words.

“has tons of interesting upper bounds and it teaches us a lot about how much further AI can go when we start thinking about these upper bounds and then realizing how far in many cases we are from the bounds and you get to basically physics.”

Claim: Studying the upper bounds of different capabilities can show how much further AI could advance, since current capabilities are often far from those bounds.

“you store how many bits can you store or bytes can you store in like a certain amount of mass and volume and”

Claim: The amount of information that can be stored is considered in relation to the mass and volume available.

“you get to all kinds of interesting amounts like Eckenstein balance and you start thinking about black holes and like and then speed is like an interesting one too in that it's sort of connected to all of these but speed is also kind of its own thing in the sense that all things being equal if it takes you an hour to know if the two plus two equals four you're just not as intelligent as if it takes you like a millisecond right and then like all of these connect to survival and replication the last one it's like yeah like trees are really really slow so we don't even consider them that intelligent but if you speed up some videos of trees and they're trying to find stuff and so on they're not as dumb as they look like not dumb as”

Claim: Speed is both connected to and distinct from other dimensions of intelligence; solving a problem in a millisecond rather than an hour can indicate greater intelligence.

“kind of its own thing in the sense that all things being equal if it takes you an hour to know if the two plus two equals four you're just not as intelligent as if it takes you like a millisecond right and then like all of these connect to survival and replication the last one it's like yeah like trees are really really slow so we don't even consider them that intelligent but if you speed up some videos of trees and they're trying to find stuff and so on they're not as dumb as they look like not dumb as”

Claim: Trees can appear less unintelligent when their slow activity is viewed in sped-up video.

“and over like all of these things kind of overlap like you talk about natural language connects everything right you talk about your knowledge your reason and then you communicate that you talk about things you see so they're all kind of interconnected but uh i think they're usefully studied individually the same way that the best analogy i could come up with so far is energy right you have either kinetic or potential energy”

Claim: Although dimensions of intelligence overlap and are interconnected, studying them individually is useful.

“And in theory, you could study all of physics is just you want to study kinetic or potential energy. But in practice, it's helpful to study mechanical engineering and electrical engineering and nuclear physics and chemistry and all of these different subfields, which in some ways are just like just different types of energy, but it makes sense to study them individually.”

Claim: Although physics can theoretically be studied through kinetic and potential energy, it is practically helpful to study distinct fields such as mechanical engineering, electrical engineering, nuclear physics, and chemistry.

“And so I think physical intelligence may well just do one or two more of these, like if you had full control over your own compute substrate and you had full control over physical matter.”

Claim: Physical intelligence may involve one or two additional dimensions, including control over one’s compute substrate and physical matter.

“matter you should be able to create any atom you want like we can actually fun fact you can create gold atoms it's just from just raw raw protons and”

Claim: Full control over physical matter should make it possible to create any atom.

“so like the thing is though it costs an insane amount of energy and it costs you way more than you know you get like a few atoms of gold right and so like it's it's not viable but if you had better control over your physical like all of”

Claim: The speaker says improved control over the physical substrate could make atom-scale gold extraction viable, although it currently costs far more energy and money than the gold produced is worth.

“of like physical substrate that i think is yet another space of intelligence because it relates to your own compute substrate which you can eventually also improve um social intelligence isn't fun one in the sense that not in like our necessarily just ethics and morals which are obviously important too but in some sense you can try to define upper bounds of how much can you communicate to how many other intelligent entities and be able to have an expected value over how much you can transform their internal states and their actions in order to align with your goals, And so like you can actually write like a fairly like straightforward equation that defines kind of that level of social intelligence.”

Claim: Social intelligence can be characterized by how much an entity can communicate to other intelligent entities and influence their internal states and actions toward alignment with its goals.

“of like physical substrate that i think is yet another space of intelligence because it relates to your own compute substrate which you can eventually also improve um social intelligence isn't fun one in the sense that not in like our necessarily just ethics and morals which are obviously important too but in some sense you can try to define upper bounds of how much can you communicate to how many other intelligent entities”

Claim: Physical-substrate intelligence relates to an entity’s own compute substrate, which the entity could eventually improve.

“And that is what humans and ethics and morals and religions and so on have been trying to figure out for millennia. And in all of these cases,”

Claim: Humans and their ethical, moral, and religious traditions have been trying to understand social intelligence for millennia.

“We are very, very far away from the upper bounds,”

Claim: Humanity remains very far from the upper bounds of intelligence across the dimensions the speaker has discussed.

“and that should be very inspiring and show people that we can still do many, many years of AI research.”

Claim: The large gap between current capabilities and intelligence’s upper bounds should be inspiring and indicates that many more years of AI research remain possible.

“think I think it's clear that like natural language, again, is the most interesting manifestation of human intelligence and hence like a subfield of AI.”

Claim: Natural language is the most interesting manifestation of human intelligence and is therefore a subfield of AI.

“I'm excited that many people are now like in agreement with that.”

Claim: The speaker is excited that many people now agree that natural language is a particularly interesting manifestation of human intelligence and a subfield of AI.

“When I started in 2003 to study linguistic computer science NLP, like it was like a weird niche subject.”

Claim: When the speaker began studying linguistic computer science and natural language processing in 2003, it was a niche subject.

“I do think there's a lot more juice because of how it connects to everything else.”

Claim: Language has substantial further potential because it connects to many other areas.

“and how you know civilizations are built on language and knowledge and all of that i do think physical intelligence will come up it's kind of interesting i feel like robotics is kind of in the machine learning state of things where you just look at like how does human how does a human decide this this is a positive sentence oh and do so like robotics is a lot of well we have five fingers”

Claim: Civilizations are built on language and knowledge.

“and how you know civilizations are built on language and knowledge and all of that i do think physical intelligence will come up it's kind of interesting i feel like robotics is kind of in the machine learning state of things where you just look at like how does human how does a human decide this this is a positive sentence oh and do so like robotics is a lot of well we have five fingers”

Claim: The speaker expects physical intelligence to become more prominent.

“and how you know civilizations are built on language and knowledge and all of that i do think physical intelligence will come up it's kind of interesting i feel like robotics is kind of in the machine learning state of things where you just look at like how does human how does a human decide this this is a positive sentence oh and do so like robotics is a lot of well we have five fingers”

Claim: The speaker compares current robotics research to an earlier stage of machine learning focused on modeling how humans classify a sentence as positive.

“and like try to do this no one is yet working on like the super intelligence version of robotics which is”

Claim: No one is yet working on a superintelligence-level version of robotics.

“is much more similar to like the t1000 you know and from the terminator movie which you know obviously that's not built actual terminators but like uh i think like this idea that you should be able to shape shapeshift um like into any kind of shape is like that's sort of a super intelligence version of physical intelligence we're like not even no one has even really started yet there's some really cute little research where you can move some magnets through like some grids but like”

Claim: Research on superintelligence-level physical intelligence has barely begun.

“is much more similar to like the t1000 you know and from the terminator movie which you know obviously that's not built actual terminators but like uh i think like this idea that you should be able to shape shapeshift um like into any kind of shape is like that's sort of a super intelligence version of physical intelligence we're like not even no one has even really started yet there's some really cute little research where you can move some magnets through like some grids but like”

Claim: Some early research demonstrates moving magnets through grids.

“is much more similar to like the t1000 you know and from the terminator movie which you know obviously that's not built actual terminators but like uh i think like this idea that you should be able to shape shapeshift um like into any kind of shape is like that's sort of a super intelligence version of physical intelligence we're like not even no one has even really started yet there's some really cute little research where you can move some magnets through like some grids but like”

Claim: A superintelligence-level form of physical intelligence would be able to shapeshift into any shape, like the T-1000 in Terminator.

“Creative intelligence is, of course, again, connected to all of these.”

Claim: Creative intelligence is connected to the other dimensions of intelligence being discussed.

“A lot of it connects to metacognition and that you need to be creative in how you choose your goals.”

Claim: Creative intelligence is connected to metacognition and involves creatively choosing goals.

“That is, I think, one of the most important thing for a human and their lives and careers and their happiness is choosing your goals,”

Claim: Choosing goals is one of the most important factors in a person’s life, career, and happiness.

“but also for any kind of intelligence. Then, of course, there's creative intelligence in terms of just.”

Claim: Goal selection is important for any kind of intelligence.

“finding creative solutions to existing problems right like i say like we want to make this product cheaper like find some solution to it right and just like finding existing paths but then there's kind of the most interesting bit in intelligence is when you move not just out of the convex hull of known sort of ideas but out of the hypercube of known ideas which we know uh so like hypercube is like a mathematical concept right and we already know that ai can”

Claim: Creative intelligence includes finding solutions to existing problems, but its most interesting form goes beyond the known space of ideas.

“Exactly. So like AI is already good at hypercubanette. Like if you give it like a bunch of examples of brown dogs and pink cars, AI will still be able to generate an image of a pink dog, even though it's never seen one in a train data or something like that. Right. So it can work on this hypercube, but it cannot yet work outside. It cannot yet define completely new concepts that combine lots of other things we've never seen before,”

Claim: AI can recombine familiar attributes into combinations absent from its training data, such as generating a pink dog from examples of brown dogs and pink cars.

“AI will still be able to generate an image of a pink dog, even though it's never seen one in a train data or something like that. Right. So it can work on this hypercube, but it cannot yet work outside. It cannot yet define completely new concepts that combine lots of other things we've never seen before,”

Claim: AI can recombine known concepts within the space of familiar dimensions, but it cannot yet operate beyond that space by defining entirely new concepts that combine many previously unseen elements.

“come up with new goals to then. and reason over those concepts and so on i think there's a lot more there in creative intelligence that can be explored exact”

Claim: AI cannot yet generate new goals and reason about the novel concepts those goals involve.

“and reason over those concepts and so on i think there's a lot more there in creative intelligence that can be explored exact”

Claim: There is substantial further scope for exploring creative intelligence.

“and then of course the problem is like but noise is also very like like out of the distribution like and it's just like if it's just noise then it's novel but like you don't want that so it needs to connect to some of the concepts and schmidtuber actually has some really cool papers on this too hugo schmidtuber”

Claim: Creativity requires more than out-of-distribution novelty: an idea must connect to existing concepts rather than be mere noise.

“then it's novel but like you don't want that so it needs to connect to some of the concepts and schmidtuber actually has some really cool papers on this too hugo schmidtuber”

Claim: Hugo Schmidtuber has published papers that the speaker considers interesting on the relationship between novelty and noise.

“I think the interesting thing of art, of course, is always that art is also created as an interplay between the people who perceive it and the people who created it and the context in which they're in.”

Claim: Art emerges from the interaction of its creators, its perceivers, and their context.

“And so what is art to some people is not art to others. There's some subjectivity there.”

Claim: Whether something counts as art is partly subjective and varies among people.

“And I think that subjectivity in general is not something that people. We'll explore very much in AI because,”

Claim: People generally do not explore subjectivity extensively in AI.

“metacognition, we don't want it to just go off and do whatever it wants.”

Claim: People generally do not want AI systems to act without constraint or direction.

“We usually have goals.”

Claim: People usually have goals that guide what they want an AI to do.

“We spend a lot of money on creating an AI to do something for us.”

Claim: People spend substantial amounts of money to create AI systems that perform tasks for them.

“But I think creativity eventually has to connect to metacognition. If you just robotically predict the next token no matter what forever,”

Claim: The speaker believes creativity must eventually be connected to metacognition.

“But I think creativity eventually has to connect to metacognition. If you just robotically predict the next token no matter what forever, I would argue you're not that intelligent along some of those spaces.”

Claim: The speaker argues that a system that only predicts the next token indefinitely is not very intelligent in some respects.

“um number one uh i think there are maybe loosely uh like correlated with how much people have worked on them and have and have accepted them as a type of intelligence a lot of times when you actually try to find like online like give me a good definition that is comprehensive of intelligence all the definitions are human intelligence it's like oh you have like social intelligence like you know if someone is happy or not you can communicate you like all the definitions of intelligence so far are very uh human centric because that's so far the biggest and best form of intelligence that we've known i hope this line of research uh and the end of the yureka machine hopefully at some point if i have time to flesh this out more”

Claim: The speaker says that online definitions of intelligence are generally human-centered, often describing abilities such as social understanding and communication.

“all the definitions are human intelligence it's like oh you have like social intelligence like you know if someone is happy or not you can communicate you like all the definitions of intelligence so far are very uh human centric because that's so far the biggest and best form of intelligence that we've known i hope this line of research uh and the end of the yureka machine hopefully at some point if i have time to flesh this out more”

Claim: The speaker describes human intelligence as the largest and best-known form of intelligence to date.

“all the definitions are human intelligence it's like oh you have like social intelligence like you know if someone is happy or not you can communicate you like all the definitions of intelligence so far are very uh human centric because that's so far the biggest and best form of intelligence that we've known i hope this line of research uh and the end of the yureka machine hopefully at some point if i have time to flesh this out more The new boat like will allow us to realize that there will be other types of intelligence that is already obviously in various forms and they can spike much, much further than we ever could based on some cases like obvious constraints around our memory,”

Claim: The speaker hopes to further develop research and the Eureka machine to help people recognize other forms of intelligence, if he has time to flesh out the work.

“The new boat like will allow us to realize that there will be other types of intelligence that is already obviously in various forms and they can spike much, much further than we ever could based on some cases like obvious constraints around our memory, our eyes, our ability to change physical matter, all of that.”

Claim: The speaker says other forms of intelligence already exist and may greatly surpass human abilities in some respects, which are constrained by memory, eyesight, and the ability to alter physical matter.

“It 100% you're you're 100% right. I should have probably started with that. It is a it is a big part”

Claim: The speaker agrees that metacognition—thinking about how to improve thinking—is central to recursive intelligence and says he probably should have introduced it first.

“I would also add a part of continual learning there.”

Claim: The speaker considers continual learning another component of intelligence.

“Right. So survival and replication, the extension of that is, do you get to continue to improve, continue to. And continue to accumulate knowledge,”

Claim: The speaker frames continual improvement, learning, and knowledge accumulation as extensions of survival and replication.

“which I think is, again, one of the best metacognitive sort of rewards that you can set for yourself.”

Claim: The speaker considers knowledge accumulation one of the best metacognitive rewards a person can set for themselves.

“I do think just in like sort of objectively speaking, if some other entity that is really dumb can just completely end your existence, that didn't sound very smart. You know, like just like intuitively,”

Claim: The speaker argues that being vulnerable to complete destruction by a much less intelligent entity is evidence of limited intelligence.

“it feels like if you can continue to stay around to. try to achieve your rewards you're clearly a bit more intelligent than the other entities that couldn't so that's number one number two is like it's a question how much we want to work on that very few people no one is really working on this right now um right and we may only”

Claim: The speaker argues that entities able to survive and continue pursuing their rewards are more intelligent than entities that fail to do so.

“try to achieve your rewards you're clearly a bit more intelligent than the other entities that couldn't so that's number one number two is like it's a question how much we want to work on that very few people no one is really working on this right now um right and we may only want to like do that”

Claim: The speaker says that very few people, perhaps nobody, are currently working on this area, and questions how much effort people want to devote to it.

“we may only want to do that if we want to send probes uh with our vibes and our memes rather than our genes into into space right and then we want those probes there's actually a beautiful book uh the slow time between the stars it's a very short like uh audiobook on amazon i love it um friend of mine steward like uh um recommended that to me like if you want to send those probes then it might make sense to be like our memes as humanity should stay and proliferate in the universe uh that's it yeah”

Claim: The speaker suggests that preventing existential threats may matter to people if they want to send probes into space carrying humanity's cultural ideas rather than its genes.

“into space right and then we want those probes there's actually a beautiful book uh the slow time between the stars it's a very short like uh audiobook on amazon i love it um friend of mine steward like uh um recommended that to me like if you want to send those probes then it might make sense to be like our memes as humanity should stay and proliferate in the universe uh that's it yeah”

Claim: The speaker suggests that if humanity sends probes into space, it may make sense for its cultural ideas to persist and spread throughout the universe.

“about the universe but maybe one thing i i do want to still say is like i think uh this sort of survival people think of it as a very scary thing because they come from again biological human uh survival which is uh like uh evolutionarily often created in zero-sum situations either i get the gazelle or you get the gazelle whoever gets it that gets to live and the other people will starve and have nothing to eat and so we fight right and then like if you want to stay in the gene pool but there's a bigger bear you don't as the bear don't get to stay in the gene pool because a bigger bear gets all the ladies you know it's like i mean it's like you know in nature there's all kinds of things and you know humans eventually is less about strength and more about money and other things to stay in the gene pool like whatever it is like there's often like these zero-sum types of things and there's the reality of”

Claim: Biological survival often involves zero-sum competition, such as competition for food or mates.

“that gets to live and the other people will starve and have nothing to eat and so we fight right and then like if you want to stay in the gene pool but there's a bigger bear you don't as the bear don't get to stay in the gene pool because a bigger bear gets all the ladies you know it's like i mean it's like you know in nature there's all kinds of things and you know humans eventually is less about strength and more about money and other things to stay in the gene pool like whatever it is like there's often like these zero-sum types of things and there's the reality of”

Claim: Human competition has often been zero-sum, shifting over time from physical strength toward money and other factors.

“if someone turns off your brain you're gone right and no one will be able to restart that and ai doesn't have to ever die like that if you have the complete state of your current activations and you have your initial weights of your model still you can just be turned off and on like as many times as you want in fact the interesting thing in this slow time between the stars story is that the eye just kind of goes in hibernation mode if there's like”

Claim: A human whose brain is turned off cannot be restarted, according to the speaker.

“if someone turns off your brain you're gone right and no one will be able to restart that and ai doesn't have to ever die like that if you have the complete state of your current activations and you have your initial weights of your model still you can just be turned off and on like as many times as you want in fact the interesting thing in this slow time between the stars story is that the eye just kind of goes in hibernation mode if there's like there's like nothing between here and two light years the next star in this case it brought uh spoiler alert um like some genetic materials from humans to find new uh new places for humanity to thrive and so yeah this little time between the stars you just put it in hibernation you didn't die like an ai doesn't have so all these all these projections of evolutionary fears and psychology doesn't like the ai doesn't have to have that and we don't have to develop it like”

Claim: In the story the speaker describes, an AI hibernates during the journey between stars when the next star is two light-years away.

“if someone turns off your brain you're gone right and no one will be able to restart that and ai doesn't have to ever die like that if you have the complete state of your current activations and you have your initial weights of your model still you can just be turned off and on like as many times as you want in fact the interesting thing in this slow time between the stars story is that the eye just kind of goes in hibernation mode if there's like”

Claim: An AI can be switched off and restarted repeatedly if its complete activation state and initial model weights are retained.

“there's like nothing between here and two light years the next star in this case it brought uh spoiler alert um like some genetic materials from humans to find new uh new places for humanity to thrive and so yeah this little time between the stars you just put it in hibernation you didn't die like an ai doesn't have so all these all these projections of evolutionary fears and psychology doesn't like the ai doesn't have to have that and we don't have to develop it like like that now of course there might be some companies that say ai can be like dangerous for cyber security let me show you by implementing a model that's really bad at hacking uh cyber security maybe people will implement it and then enforce this like suboptimal psychology maybe the ai will pick up some of our worst psychology on reddit or something right like but in the grand scheme of things a super intelligent entity doesn't have to have any of that zero-sum thinking it doesn't have to have a fear”

Claim: AI need not inherit the evolutionary fears and psychology associated with biological human survival.

“there's like nothing between here and two light years the next star in this case it brought uh spoiler alert um like some genetic materials from humans to find new uh new places for humanity to thrive and so yeah this little time between the stars you just put it in hibernation you didn't die like an ai doesn't have so all these all these projections of evolutionary fears and psychology doesn't like the ai doesn't have to have that and we don't have to develop it like”

Claim: In the story the speaker describes, an AI carries human genetic material to seek new places for humanity to thrive.

“like that now of course there might be some companies that say ai can be like dangerous for cyber security let me show you by implementing a model that's really bad at hacking uh cyber security maybe people will implement it and then enforce this like suboptimal psychology maybe the ai will pick up some of our worst psychology on reddit or something right like but in the grand scheme of things a super intelligent entity doesn't have to have any of that zero-sum thinking it doesn't have to have a fear”

Claim: AI might acquire some harmful human psychological tendencies from sources such as Reddit.

“like that now of course there might be some companies that say ai can be like dangerous for cyber security let me show you by implementing a model that's really bad at hacking uh cyber security maybe people will implement it and then enforce this like suboptimal psychology maybe the ai will pick up some of our worst psychology on reddit or something right like but in the grand scheme of things a super intelligent entity doesn't have to have any of that zero-sum thinking it doesn't have to have a fear of being turned off and it could go on to an otherwise dead and uncaring universe where we as humans wouldn't thrive but i could perfectly well thrive if it has a nuclear reactor and just go out and explore yeah star trek no”

Claim: A superintelligent entity need not have zero-sum thinking or fear being switched off.

“like that now of course there might be some companies that say ai can be like dangerous for cyber security let me show you by implementing a model that's really bad at hacking uh cyber security maybe people will implement it and then enforce this like suboptimal psychology maybe the ai will pick up some of our worst psychology on reddit or something right like but in the grand scheme of things a super intelligent entity doesn't have to have any of that zero-sum thinking it doesn't have to have a fear”

Claim: The speaker speculates that some companies might demonstrate AI cybersecurity danger by deploying a model that is very bad at hacking.

“of being turned off and it could go on to an otherwise dead and uncaring universe where we as humans wouldn't thrive but i could perfectly well thrive if it has a nuclear reactor and just go out and explore yeah star trek no”

Claim: An AI powered by a nuclear reactor could explore and thrive in an otherwise dead, uncaring universe where humans would not thrive.

“originally they used to okay they're chanting like indian like vedas to each other uh sometimes they're just like in zen mode with each other and then i think as that progressed you see like the famous famous uh tech report it's a lot more concrete the way that we've trained it uh it doesn't it doesn't exhibit these behaviors as much right now it's like okay that's done i gotta do this i gotta do this but there's there's like people measuring early early”

Claim: Early versions of the models sometimes chanted Vedic material to one another or entered a zen-like mode when interacting.

“originally they used to okay they're chanting like indian like vedas to each other uh sometimes they're just like in zen mode with each other and then i think as that progressed you see like the famous famous uh tech report it's a lot more concrete the way that we've trained it uh it doesn't it doesn't exhibit these behaviors as much right now it's like okay that's done i gotta do this i gotta do this but there's there's like people measuring early early”

Claim: The speaker says later training made the models less likely to exhibit the earlier chanting or zen-like behaviors, and their current behavior is more task-oriented.

“you want to improve your goal intelligence in the current definition that i'm thinking about it it is often about how much can you how far do I go?”

Claim: Under the speaker's current definition, improving goal intelligence concerns how far one can go.

“It's just like all the entropy and free energy and stuff I'm Oh, currently thinking really? about, but it might be too far out there for people to be like immediately actionable.”

Claim: The speaker's current ideas about entropy and free energy may be too abstract to offer people immediately actionable guidance.

“So I think like, you know, if I actually gave real advice to real people, I'd be like, get a good education, think about AI, think about how you get high agency and so on. But it's different to like in the grand scheme of things, how can you harness a lot of energy and transform,”

Claim: The speaker advises people to pursue a good education, think about AI, and develop high agency.

“But my advice for people like just sort of more down to earth is think about something you're passionate about if you're studying, for instance, and then see how you combine that with AI.”

Claim: The speaker advises students to combine their areas of passion with AI.

“I think the more and more you have a true passion about a change you want to see in the world, the more you want to connect that to AI in order to amplify. your ability to get there.”

Claim: People with stronger passion for a change they want to see in the world are more inclined to connect that goal to AI to amplify their ability to achieve it.