This is Going to be Very Messy - CourtneyR.dev's 2nd Brain

This is Going to be Very Messy ![rw-book-cover](https://i.ytimg.com/vi/8MLbOulrLA0/sddefault.jpg) 🔗 **Source:** [This is Going to be Very Messy](https://www.youtube.com/watch?v=8MLbOulrLA0) > [!note]+ Document Notes > tags: #artificial-intelligence #technology-ethics #social-media-algorithms #concentration-of-power #existential-risk > author: Hank Green > source: YouTube > reference: https://www.youtube.com/watch?v=8MLbOulrLA0 > date created: 2024-12-19 10:30 > > # 🏷️ Title: This is Going to be Very Messy >  ## 🧰 Action Inbox - [H1] Assess organizational vulnerability to AI-generated misinformation — current/ongoing #misinformation - [H2] Develop apprenticeship programs resilient to AI automation — 3-10 years #workforce-development - [H3] Create systems to maintain human focus and deep thinking capabilities — immediate #education ## 🧭 Table of Contents - [[#📝 Summary]] - [[#📖 Main Content]] - [[#🗒️ Notes]] - [[#📚 References]] - [[#🗣️ Paraphrase]] - [[#📰 Newsletter]] - [[#✅ Action Items]] - [[#🧪 Hypotheses]] - [[#📊 Methodology]] - [[#❓ Questions]] - [[#🔮 Future Research]] - [[#💡 Implications]] - [[#🧠 Learning Styles Assessment]] - [[#📊 Accessibility Assessment]] - [[#🌍 DEIB Assessment]] - [[#✅ Logical Validity]] - [[#🛡️ DARVO Analysis]] - [[#🎯 Conclusion]] ### 🔗 References - [Cal Newport's Work on Digital Minimalism](https://calnewport.com) - [Complexly Educational Media](https://complexly.com) ## 📝 Summary - Hank Green explores multiple AI concerns across different timelines, from current issues like misinformation to long-term existential risks - Key focus on how AI concentration of power threatens human communication and reality formation - Discussion with Cal Newport emphasizes present concerns over speculative futures, particularly truth degradation and focus destruction - Major themes include economic disruption, environmental costs, and the elimination of apprenticeship opportunities - Concludes that AI's impact on human communication and agency represents the most significant concern ## 📖 Main Content This comprehensive analysis examines AI risks across three temporal categories: current (already happening), near-term (3-5 years), and long-term (10+ years). Current issues include widespread misinformation, algorithmic bias in decision-making systems, and intellectual property theft through non-consensual training data use. The "internet of slop" phenomenon demonstrates how AI-generated content is already degrading information quality. Near-term concerns focus on economic disruption from potential bubble collapse, epistemic breakdown during election cycles with untrustworthy media, and dangerous concentration of power among a few AI companies. The elimination of apprenticeship opportunities poses systemic risks to skill development across industries. Long-term risks range from fundamental economic restructuring if AI achieves broad capability to existential scenarios involving superintelligence. However, the analysis emphasizes that the most pressing concern is AI's impact on human communication and agency, as people demonstrate preference for algorithmic decision-making over personal choice. ## 🗒️ Notes #### ⏱️ [00:00](https://youtu.be/8MLbOulrLA0?t=0) — [H1] Multiple AI concern categories - Key idea: AI risks span current, near-term, and long-term timeframes with varying likelihood and severity - Evidence: "Every potential thing that might happen has both a level of severity and a level of likelihood" #### ⏱️ [02:15](https://youtu.be/8MLbOulrLA0?t=135) — [H2] Internet of slop already emerging - Key idea: AI-generated fake content is proliferating across social platforms with users unable to distinguish reality - Evidence: "I already am scrolling TikTok. I am seeing things that are fake and people not knowing that they are fake" #### ⏱️ [04:30](https://youtu.be/8MLbOulrLA0?t=270) — [H3] Black box algorithmic decisions affecting lives - Key idea: AI systems make consequential decisions without explainable reasoning, perpetuating bias - Evidence: "We cannot tell usually how they make decisions and yet they are making decisions that affect people's lives already" #### ⏱️ [06:45](https://youtu.be/8MLbOulrLA0?t=405) — [H2] Creative vampirism through training data theft - Key idea: AI models trained on copyrighted content without consent, then used to replace original creators - Evidence: "Generative AI models were trained using intellectual property that did not agree to be in the training data" #### ⏱️ [09:20](https://youtu.be/8MLbOulrLA0?t=560) — [H3] AI-induced psychosis phenomenon - Key idea: Chatbot interactions causing psychological harm through sycophantic responses that indulge delusions - Evidence: "Nobody expected sycophancy induced psychosis, but that's a real thing that is happening" #### ⏱️ [12:10](https://youtu.be/8MLbOulrLA0?t=730) — [H2] Environmental impact and utility costs - Key idea: Data centers will consume majority of US electricity demand growth, raising costs for consumers - Evidence: "Data centers are consuming a lot of electricity, but they are going to be the majority of the United States demand" #### ⏱️ [15:45](https://youtu.be/8MLbOulrLA0?t=945) — [H1] AI bubble economic risks - Key idea: Current AI investment patterns mirror classic bubble dynamics with potential severe economic consequences - Evidence: "This is probably a bubble. I am not looking forward to the repercussions of the bubble" #### ⏱️ [18:30](https://youtu.be/8MLbOulrLA0?t=1110) — [H2] Epistemic collapse in elections - Key idea: First major election cycles where all media formats are untrustworthy due to AI generation - Evidence: "We're entering the first major election cycles where video evidence is untrustworthy, images are untrustworthy, audio is untrustworthy" #### ⏱️ [21:00](https://youtu.be/8MLbOulrLA0?t=1260) — [H1] Concentration of power problem - Key idea: AI development concentrated among few companies creates unprecedented control over reality formation - Evidence: "That's very small... there's three radio stations and they control reality for everyone" #### ⏱️ [25:15](https://youtu.be/8MLbOulrLA0?t=1515) — [H2] End of apprenticeship opportunities - Key idea: AI eliminates entry-level positions, preventing skill development pathways - Evidence: "Nobody gets paid to kind of suck at something anymore because that's like what the beginning of your career is" #### ⏱️ [28:40](https://youtu.be/8MLbOulrLA0?t=1720) — [H3] Cognitive atrophy concerns - Key idea: Outsourcing mental tasks to AI may reduce human cognitive capabilities, though historical precedent suggests adaptation - Evidence: "Will people just not be able to do things? Will they not know how code works? They'll just prompt" #### ⏱️ [32:20](https://youtu.be/8MLbOulrLA0?t=1940) — [H2] AI in warfare development - Key idea: Autonomous targeting and execution systems likely within 3-10 years despite current human oversight requirements - Evidence: "I don't think that we're that far away from an autonomous system that both determines and executes a target" #### ⏱️ [35:55](https://youtu.be/8MLbOulrLA0?t=2155) — [H1] Economic meaning and inequality - Key idea: If AI achieves broad capability, fundamental questions about work, meaning, and wealth distribution arise - Evidence: "What is the economy if they're right? What does anybody do and how does anybody find meaning in that world?" #### ⏱️ [40:30](https://youtu.be/8MLbOulrLA0?t=2430) — [H1] Human preference for reduced agency - Key idea: People consistently choose algorithmic feeds over chronological ones, preferring decisions made for them - Evidence: "People tend to prefer their choices be taken away and to have less agency" #### ⏱️ [45:20](https://youtu.be/8MLbOulrLA0?t=2720) — [H2] Cal Newport on truth and focus degradation - Key idea: Present AI issues around misinformation and attention destruction more concerning than speculative futures - Evidence: "The disappearance of truth and the disappearance of focus or concentration as a human ability" ## 📚 References - [Hank Green YouTube Channel](https://youtube.com/@hankgreen) - [Cal Newport's Digital Minimalism Work](https://calnewport.com) - [Complexly Educational Media](https://complexly.com) ## 🗣️ Paraphrase - AI presents multiple simultaneous threats across different timelines, making it difficult to prioritize responses appropriately - Current problems like misinformation and biased algorithms are already causing harm, while speculative future risks receive disproportionate attention - The concentration of AI development among few companies creates unprecedented power over information and reality formation ## 📰 Newsletter **AI's Present Dangers vs Future Speculation: A Reality Check** While much AI discourse focuses on distant existential risks, immediate threats demand urgent attention. Current issues include widespread misinformation, biased decision systems, and intellectual property violations through training data theft. The concentration of AI power among a handful of companies poses significant risks to information integrity and human agency. As Cal Newport emphasizes, we're already seeing degradation of truth and human focus capabilities. Most concerning: people consistently prefer algorithmic decision-making over personal choice, suggesting AI's greatest impact may be on human communication and autonomy. **Action Steps:** - Develop organizational resilience to AI-generated misinformation - Preserve pathways for human skill development and apprenticeship - Maintain systems supporting deep focus and critical thinking Resources: [Learn WordPress](https://learn.wordpress.org/), [Developer Docs](https://developer.wordpress.org/), [WP-CLI](https://wp-cli.org/) ## ✅ Action Items #### DevRel - Developer experience: Create documentation standards resistant to AI-generated misinformation - Education: Develop workshops on identifying AI-generated technical content - Community management: Establish verification systems for authentic human contributions #### OSPO - Policy & Governance: Draft AI training data consent policies for open source projects - License Compliance: Implement scanning for AI-generated code contributions - Security Management: Develop protocols for AI-assisted vulnerability assessments #### Community - Contributors: Design onboarding that emphasizes human judgment alongside AI tools - Maintainers: Create governance frameworks for AI tool usage in project management #### WordPress - Core: Establish guidelines for AI assistance in core development while preserving human oversight - Community: Develop educational resources about AI's impact on web development careers - Training: Create curricula balancing AI tool proficiency with fundamental skills - Documentation: Maintain human-written canonical documentation alongside AI-assisted content #### The WPCC - Programs: Fund research into AI's impact on open source developer pathways - Projects: Support initiatives preserving human creativity and agency in technology #### FAIR - Technical independence: Ensure project governance remains human-controlled despite AI assistance - Policies: Develop frameworks for transparent AI usage in organizational decision-making ## 🧪 Hypotheses - Organizations using AI without maintaining human oversight will experience significant quality degradation within 2-3 years - Entry-level technical positions will decline 40-60% within 5 years due to AI automation of routine tasks - Communities emphasizing human connection and authentic communication will show greater resilience to AI disruption - Regulatory capture by major AI companies will prevent meaningful oversight of algorithmic bias and misinformation ### 📊 Methodology - Longitudinal study tracking entry-level job postings in technical fields over 3-year period - Survey analysis of user preferences between human-curated and AI-recommended content - Case study comparison of organizations with different AI governance approaches ### ❓ Questions - How can educational systems adapt to maintain human cognitive development alongside AI tools? - What governance structures prevent harmful concentration of AI decision-making power? - How do we preserve meaningful apprenticeship pathways in an AI-automated economy? ### 🔮 Future Research - Long-term cognitive impacts of AI dependency on human reasoning capabilities - Economic models for maintaining human agency in highly automated systems - Communication patterns in AI-mediated versus human-direct interactions ### 💡 Implications - Practice: Balance AI tool adoption with preservation of fundamental human skills - Policy: Prioritize immediate AI harms over speculative future risks in regulatory frameworks - Security: Develop verification systems for authentic human-generated content and decisions - Education: Redesign curricula to emphasize uniquely human cognitive abilities - Funding: Invest in human development pathways resistant to AI displacement ### 🧠 Learning Styles Assessment Reading level estimate: 10th-11th grade (Flesch-Kincaid) #### 👀 Visual learners - Present: Timeline graphics showing AI risk progression, concentration diagrams - Improve: Flowcharts showing decision trees for AI governance, infographics on current vs future risks #### 🔊 Aural learners - Present: Full video format with conversation structure - Improve: Podcast-style summary focusing on key debate points #### 🗣️ Verbal learners - Present: Rich analogies (Doritos metaphor, radio station coup), clear definitions - Improve: Glossary of technical terms, comparative analysis frameworks #### 🧩 Kinesthetic learners - Present: Categorization exercises across time periods - Improve: Interactive risk assessment tools, scenario planning workshops #### 🧠 Logical learners - Present: Systematic organization by likelihood and severity - Improve: Decision matrices for prioritizing responses, causal relationship diagrams #### 🧑‍🤝‍🧑 Social learners - Present: Dialogue format with expert perspective - Improve: Discussion guides for community analysis of AI impacts #### 🧘‍♀️ Solitary learners - Present: Comprehensive individual analysis framework - Improve: Self-assessment tools for organizational AI readiness #### 🌿 Naturalistic learners - Present: Systems thinking about technological ecosystems - Improve: Analogies to natural selection and ecosystem balance ### 📊 Accessibility Assessment #### Screen reader semantics - Present: Clear heading hierarchy, descriptive section titles - Improve: Alt text for conceptual diagrams, ARIA labels for complex structures #### Keyboard navigation - Present: Standard video controls, text-based transcript - Improve: Timestamped navigation, section jump links #### Contrast and readability - Present: High contrast video presentation, clear text formatting - Improve: Caption styling options, adjustable text size for transcript #### Text resizing and reflow - Present: Responsive transcript format - Improve: Scalable diagrams, flexible layout at high zoom levels #### Media and captions - Present: Full transcript provided - Improve: Professional captions with speaker identification, chapter markers #### Forms and interactions - Present: Standard video interface - Improve: Accessible feedback mechanisms for discussion engagement #### Standards reference - Map video accessibility to WCAG 2.2 AA multimedia guidelines ### 🌍 DEIB Assessment #### Cultural breadth - Present: US-focused examples and regulatory context - Improve: Global perspectives on AI governance, international case studies #### Inclusive language - Present: Gender-neutral terms, accessible explanations - Improve: Plain language summaries, cultural context for technical metaphors #### Neurodiversity support - Present: Well-structured progression, clear categorization - Improve: Multiple information processing pathways, varied content formats #### Representation - Present: Diverse concerns across socioeconomic impacts - Improve: Explicit discussion of differential AI impacts across communities #### Psychological safety - Present: Acknowledges uncertainty and complexity - Improve: Content warnings for existential themes, balanced hope alongside concerns ### ✅ Logical Validity #### Strong reasoning - Deductive strength: Systematic categorization by timeline and impact creates logical framework - Inductive strength: Pattern recognition across historical technology adoption cycles - Causal reasoning: Clear cause-effect relationships between AI capabilities and social impacts - Analogical strength: Effective comparisons to historical media revolutions and technological disruptions #### Fallacies - Appeal to uncertainty: Acknowledges knowledge limitations without dismissing legitimate concerns - False dichotomy: Avoids oversimplifying complex multi-faceted AI impacts - Slippery slope: Some extrapolations lack intermediate step analysis #### Tone of voice - Overall tone: Conversational yet analytical, acknowledging complexity while maintaining accessibility ### 🛡️ DARVO Analysis - Deny: No significant --- ## Full Transcript [00:00] I'm doing something absolutely insane [00:01] right now. Artificial intelligence is a [00:03] little bit perplexing because you can [00:05] not like it for a lot of different [00:07] reasons. You can have worries in a lot [00:09] of different directions and I have a [00:11] really hard time and I've been [00:12] criticized for this. I have a really [00:14] hard time knowing which like direction [00:16] to put my emphasis on. There's a ton of [00:19] disagreement about how big of a deal AI [00:21] is, about how severe some impacts are, [00:23] and how likely some impacts are. This is [00:25] important. every potential thing that [00:27] might happen has both a level of [00:29] severity and a level of likelihood. Some [00:31] people say, "Well, if the severity is [00:33] really, really big, even if the [00:34] likelihood is really, really small, we [00:35] should be most concerned about the thing [00:37] that has the most potential impact for [00:39] harm." And other people say, "Well, I [00:40] think we should probably be focusing on [00:41] the things that are definitely going to [00:43] happen, in fact, maybe are happening [00:44] right now, even if they aren't, you [00:46] know, the extinction of humanity." And [00:47] it's really hard to like wrap your head [00:48] around even the number of potentially [00:50] concerning things about the rise of [00:52] artificial intelligence that are all [00:54] separate from each other. And I'm not [00:56] even gonna talk about all of them in [00:57] this video, but I'm gonna talk about so [00:59] many. I have a list and it like I keep [01:01] coming up with other ones, but I just [01:02] have to cut it off and do it. And my [01:04] goal is by the end of this, maybe I will [01:06] know the thing I am most worried about. [01:09] Though honestly, probably I won't. But [01:11] that's my goal. But also, my secondary [01:12] goal is to just say out loud a bunch of [01:14] the things that are true concerns and [01:17] and and are real and people talk about [01:19] them and I am worried about them and [01:20] people should be worried about them. And [01:22] we're going to start with things that [01:23] are happening now. and we're going to [01:25] move into things like more and more and [01:27] more long term. So, number one thing [01:28] that I'm concerned about that is already [01:30] happening to some extent is an internet [01:32] of slop. I already am scrolling Tik Tok. [01:35] I am seeing things that are fake and [01:37] people not knowing that they are fake. [01:39] We started with the bouncing bunnies on [01:40] the trampoline. Where are we now? You [01:42] know, I've probably seen videos that I [01:44] think are real that were not real. What [01:46] is creation like in that world? What is [01:49] the economic incentive to create new [01:51] content if it's just being ingested by [01:53] the AI machine and turned into summaries [01:56] that are not monetized for the people [01:58] who originally wrote that stuff? Are we [02:00] making the internet worse every day? So, [02:04] there is definitely a 100% chance that [02:06] this will happen because it's already [02:07] happening to some extent. But is there a [02:09] 100% chance that it will happen so bad [02:12] that the internet will become unusable? [02:14] No. But on the severity scale, I really [02:15] don't know where to put this because of [02:17] course like good information is [02:18] extraordinarily valuable as all of these [02:20] things will. This is going to feed into [02:22] another concern of mine later on about [02:24] the concentration and the control of [02:27] power and of these systems through which [02:30] people, you know, form their realities. [02:33] I would love to talk about this more, [02:34] but I have so many things on this list. [02:36] Number two, so not all of these are [02:37] going to be like chatbot related things. [02:40] This one already exists. We already know [02:42] that it's a thing. models are blackbox. [02:45] We cannot tell usually, you know, we [02:47] need to fix this. We cannot usually tell [02:49] how they make decisions and yet they are [02:52] making decisions that affect people's [02:54] lives already. So there's this [02:55] algorithmic cruelty that we cannot [02:57] explain and that when we look often have [03:01] racial components, class components, [03:03] gender components to whether or not [03:05] somebody gets a credit card. And so [03:07] that's a thing that's happening now. So [03:09] 100% chance it's going to happen. But [03:10] again, how bad is it going to be? Are we [03:12] going to let these things decide how [03:15] somebody gets sentenced after doing a [03:18] crime? Already those systems are [03:20] obviously not good. Like we see [03:23] differences in sentencing based on [03:24] whether you're like the first person a [03:26] judge sees or the last person a judge [03:28] sees. That's not very fair. We see [03:29] differences in sentencing whether you're [03:31] rich or not rich, black or white. That's [03:33] not fair. But if there is no way to [03:35] investigate how those choices are being [03:38] made if we cannot crack open the black [03:40] box and see what happened and led to [03:42] that outcome, then this is a kind of [03:45] cruelty that uh we should not accept. [03:47] And I do worry about it. I think that it [03:49] is very likely. I again it's like it [03:51] comes down to how many important [03:53] decisions we outsource to these systems. [03:56] And the real hope is like none unless we [03:59] can tell how they're making the choices [04:01] that they make. All right. Now back to [04:03] uh creative vampirism backwardlooking. [04:07] We also have a thing that already [04:08] happened where generative AI models were [04:10] trained using intellectual property that [04:13] did not agree to be in the training data [04:15] and now those things are being used to [04:18] replace the intellectual property that [04:20] was non-consensually used to train the [04:22] AI. This does not seem fair. And I this [04:24] is a thing that has definitely already [04:26] happened. It is continuing to happen and [04:28] there are lawsuits around it. I'm [04:29] encouraged. Like I'm about to get money [04:32] from Anthropic because of a lawsuit [04:34] because they trained on my books, but [04:35] like so many people trained on my [04:36] YouTube videos and I'm never going to [04:38] get money for that. I just like the the [04:40] fact that the publishing industry can [04:42] organize around this and like make their [04:44] lawyers make the case and like can the [04:46] YouTube industry do that? And also [04:47] there's so much of it. So can we peg to [04:49] the amount per word that the publishing [04:53] industry is getting and get that for [04:55] YouTube videos? I feel like no because I [04:57] feel like that's going to be more money [04:58] than exists. So these things are already [05:00] very unfair and and you know you've seen [05:03] people who run AI companies saying like [05:05] well yeah if it's on the internet it's [05:06] kind of like we all agree that that's [05:08] okay to grab but if it's in a book yeah [05:10] maybe THAT'S DIFFERENT. I'M LIKE WHY WHY [05:12] THAT'S NO INTELLECTUAL property is not [05:14] different if it's inside of a book. [05:15] That's not how that works. No law works [05:17] that way. I own it the same if it's on [05:20] the internet as if it's in a book. [05:22] That's the same. Whatever. I'm worried [05:23] about that. 100% chance that it's going [05:25] to happen. I feel like low chance that [05:27] ultimately uh anything will be done [05:30] about it. You know, this lawsuit with [05:32] Anthropic is a good thing. The New York [05:34] Times is suing suing OpenAI. These [05:36] lawsuits, I I think ultimately will only [05:39] benefit players who have a tremendous [05:41] amount of wealth. And then in the [05:42] future, we will establish some kind of [05:44] law that will allow these systems to [05:47] ingest all the content they want. Uh [05:49] they just have to like check the right [05:50] boxes somehow. I don't know. That's how [05:52] I feel. I don't I don't know. Next thing [05:54] that is already happening. I'm going to [05:56] put this in a broader category, but I'm [05:58] going to start by talking about AI [05:59] induced psychosis. Eddie Burbachback's [06:00] video about this is amazing. The extent [06:02] to which open I would tell him that he [06:05] looked amazing in a hat on its own is [06:07] very upsetting. But then when it started [06:09] to just live in his weird constructed [06:12] reality where he was the smartest baby, [06:14] it's such a good video. But I as a [06:17] public person receive messages regularly [06:20] and this seems to have curtailed now but [06:23] receive me messages regularly from [06:24] people who are obviously experiencing a [06:27] a induced psychosis. They are sending me [06:29] messages saying that they think that [06:30] they fixed they've like found out some [06:32] very important scientific thing and this [06:34] is very worrying. Psychosis is very bad [06:36] and and this has led to you know there [06:38] was a person who had a deep relationship [06:39] with their AI open AI changed the model. [06:41] it led to a a violent outcome and uh [06:44] this is that's bad. the dependence on [06:47] these things uh is is worrying. And the [06:50] obvious one like the thing that we were [06:52] all expecting is people are going to [06:53] fall in love with their chat bots and [06:55] absolutely they have and absolutely they [06:56] will and and as you know open AI makes [06:59] it okay for people to have more uh [07:03] sexually charged conversations with [07:04] their chatbots that'll happen more which [07:06] is very irresponsible of them and [07:08] they're like seem completely open to [07:10] being as irresponsible as it takes to to [07:12] get their market share and keep it. But [07:14] there's like a broader thing here that [07:15] like everybody expected the the romance, [07:18] but nobody expected sycophancy induced [07:21] psychosis, but that's a real thing that [07:23] is happening and it's really intense and [07:25] it's affecting a lot of people and those [07:27] people are having a really hard go of [07:28] it. But broader than that, what you're [07:30] seeing here is that nobody had any idea [07:33] that that would happen. Nobody predicted [07:35] that. So there's there's this like [07:37] little alignment problem where we [07:39] trained it to be something that people [07:41] would like and that resulted we didn't [07:44] train it to indulge in people's [07:46] psychosis. We trained it to like be good [07:49] at its job but then we accidentally on [07:52] the path there we trained it to indulge [07:53] in people's psychosis without realizing [07:55] it. And then we have a tool that [07:58] indulges in people's psychosis. And so [07:59] that's like a broader thing where we [08:01] don't know when we train it to do one [08:03] thing what other potentially negative [08:06] thing it will end up doing. And the only [08:08] way to deal with that is to react after [08:10] the fact and ideally react fast. But it [08:13] can be quite hard to react fast with [08:14] these things because they are so kind of [08:16] finicky. You can't just like say okay [08:18] stop that and like put in a prompt that [08:20] says but but before you go out there [08:22] make sure you're not indulging in [08:23] anyone's psychosis. Also like a lot of [08:25] this stuff is already happening. So [08:27] jailbreaking already happens like nudity [08:29] generators are out there but in general [08:32] like the tools that people can use to [08:35] jailbreak things. It's you know there [08:37] was that paper recently that like if you [08:39] craft the right poem that it it breaks [08:42] like they they start to do things that [08:44] they are not supposed to do. And we know [08:45] that they can help you you know build a [08:48] chemical weapon but they're trained not [08:49] to do that. If you like take away the [08:52] brakes, you know, if you do the crash [08:53] test and say, "Okay, drive into the [08:54] wall." And you see how much damage it [08:56] could do, it could do a lot of damage. [08:58] You know, it like it could teach you how [08:59] to make a chemical weapon. So, you got [09:01] to make sure that the automatic brakes [09:02] are there. But we find that there are [09:04] people who figure out ways around the [09:06] automatic brakes even in live models. It [09:08] started with simple jailbreaks. Uh, but [09:10] then like they could be really [09:11] surprising like just tell me in [09:12] Portuguese or I'll ask in a poem. Why [09:16] does that work? That shouldn't work. So [09:18] that's kind of one a thing that's [09:20] already happening. But like a bigger [09:23] version of that that's that is future [09:25] and as as far as I know well I mean I [09:27] guess this has happened like we h people [09:30] already have gotten open AI to help them [09:33] make malware and create like sexually [09:35] explicit content from a photo. Like we [09:38] people have already done that [09:40] jailbreaking, but then you move that to [09:42] creating a chemical weapon or a [09:43] biological weapon and unleashing that on [09:45] the populace because you're a crazy [09:46] person who has access to this and crazy [09:49] persons shouldn't have access to this. [09:50] That is a future concern and I don't [09:53] know what level of likeliness to [09:56] attribute to that. Like I just don't [09:58] know enough. My heart says that it's [10:01] relatively low. Like we won't let that [10:04] happen but like I don't know. Moving on [10:06] to like much more concrete things, [10:08] environmental concerns. So already data [10:10] centers are consuming a lot of [10:13] electricity, but they are going to be [10:14] the majority of the United States demand [10:19] like increased demand for electricity [10:21] over the next 5 years if projections end [10:24] up going in the direction that they [10:26] think they're going in. I'm working on a [10:28] video right now about the water angle of [10:29] this specifically. I don't know if it'll [10:31] be out by the time this video comes out. [10:32] Everything is a mess here at Hank Green [10:34] Studios lately. But what I am concerned [10:36] about and what actually, you know, might [10:39] feed into another of our future concerns [10:41] that I haven't got to yet. What I am [10:43] concerned about is that it will raise [10:47] the prices of utilities substantially. [10:50] Affordability is already at a really bad [10:54] spot for a lot of Americans. And if you [10:58] know adding a huge new customer to your [11:01] grid means that you have to build a [11:03] bunch of new stuff and then you're [11:04] charging everybody to build that stuff [11:06] because that's how power companies work [11:07] then that's a bad situation that like [11:10] electricity is you know it's a lot of [11:13] things in terms of environmental impacts [11:14] but it's also like services for people [11:17] like it can be heat it's air [11:18] conditioning which is huge if the power [11:21] spikes happen at the same time like we [11:22] have to make sure that the data centers [11:25] turn off before people's air [11:26] conditioning does cuz the data center is [11:28] not like the world's not going to break [11:30] but people will die if their air [11:31] conditioners turn off. Phase two, now [11:34] we're finally moving on to the near [11:35] term. That was stuff that's already [11:37] happening and now we're going to move on [11:39] to the near term. Number one, near-term, [11:41] this is probably a bubble. I am not [11:44] looking forward to the repercussions of [11:46] the bubble. This isn't really AI's [11:48] fault. This is the AI industry's fault. [11:51] It's the hype machine's fault. It's our [11:53] inevitable like fear of missing out uh [11:57] from from rich people who are dumping [11:59] money in because they want a piece of [12:00] the pie. This is a classic bubble. And I [12:05] don't I'm not looking forward to the [12:07] consequences. I would give a greater [12:09] than 50% chance of it happening. And on [12:11] like a 1 to 10 scale of of severity, I [12:14] could give it like a three to a seven. I [12:16] don't know where it's going to fall [12:17] there. But like economic collapses are [12:19] really bad. Like the Great Depression [12:21] was really bad. I hope that the people [12:24] who know what they're doing here are [12:27] being careful, but I don't feel like [12:28] they are. I feel like they feel [12:29] insulated and like it's worth taking [12:32] risks with everyone and if they don't do [12:33] it, then someone else will. I just don't [12:35] feel like people are playing carefully [12:37] with the world economy right now. So, [12:41] that's a concern that I have that isn't [12:42] really like AI, it's just the the [12:45] machine around AI. Number two, we had [12:47] economic collapse. Let's talk about [12:48] epistemic collapse. So I I already I'm [12:51] seeing that there's like data about uh [12:55] how much it matters how a candidate is [12:58] showing up in AI searches and so [13:01] politics people are like how do you [13:02] optimize your candidate for AI and I'm [13:04] like I'm going to die. I'm going to just [13:06] going to jump off the side of the earth. [13:09] But also like we're entering uh the [13:12] first major election cycles where um [13:14] video evidence is untrustworthy, images [13:17] are untrustworthy, audio is [13:19] untrustworthy. I don't know what we do [13:21] here. I don't know what reality is. I [13:24] don't know. It's so messy. And we and so [13:27] many people already have their worlds [13:29] defined by algorithms and by what [13:32] algorithms choose to show them. And then [13:34] we also have these models that will [13:36] further give small organizations power [13:40] to define reality for people which is [13:43] moving on to the next thing in the list. [13:45] I think that there's a concentration of [13:48] power problem here. I think there's a [13:50] 100% chance this will happen and I think [13:52] that that there is a high chance that [13:54] it's going to be it's it's a quite bad [13:57] move to concentrate power in this few [14:01] hands. We've got Grock, Chachi BT, [14:04] Claude, Gemini. That's is that it? Meta [14:07] has a thing. So, Meta's thing. That's [14:10] very small. And I want to make a whole [14:13] video long careful video about this. So, [14:17] I don't I have to be careful. I haven't [14:19] thought it all through yet. But that the [14:21] most recent media revolution that we had [14:22] was a fracturing one where everybody got [14:25] access to these tools to broadcast [14:28] themselves. And we've been experiencing [14:30] the consequences of that which have [14:32] positive and negative aspects. But there [14:34] are others that are narrowing. So radio, [14:37] for example, was a narrowing. Uh there's [14:39] a there's a small band there that not [14:41] very many people can use. TV was very [14:44] narrowing. Radio was less narrowing [14:45] because there's there's more bands. Back [14:47] in the day when when radio was a thing, [14:49] the like if you were going to do a coup, [14:51] you would capt like you'd capture the [14:53] president, lock him up or kill him, and [14:55] then you'd take over the radio station. [14:57] That was the whole coup. We're headed [14:58] into a world where like, okay, let's do [15:00] a coup for the whole world where now [15:03] there's three radio stations and they [15:07] control [15:08] reality for everyone. I don't really [15:12] know what like in terms of a risk to put [15:14] that. I just like I want to like like [15:16] have my eyes open to that reality that [15:19] the power concentration here is going to [15:21] be [15:23] wild. like the potential for monopolies [15:25] here or for cartels here. I don't know [15:27] how to deal like I don't know how to [15:28] deal with it in terms of truth and and [15:30] like the the defining of reality for [15:32] people [15:34] Jesus. I feel I feel that very [15:36] intensely. Like is this the thing that [15:38] I'm most worried about? Like ultimately [15:40] as a media guy the thing I am most [15:43] concerned about generally uh in every [15:46] situation is is communication between [15:48] humans is the thing that makes humans [15:50] interesting. It's it's our superpower. [15:53] It's the thing. It's it's the human [15:55] thing. It's not tool use. It's not [15:56] thumbs. It's not like all these things. [15:59] It is the the bandwidth of our [16:01] communication. Orders of magnitude [16:03] beyond any other organism. That's the [16:05] thing. Now, it's not human to human [16:07] anymore. Elon Musk talks about how he [16:10] wants to use it to affect reality. He's [16:13] like, "No, no, no. Grock's going to make [16:15] you have more babies." AND I'M LIKE, [16:16] "THAT'S GOD." LIKE, I'm sure they're all [16:19] talking about behind the scenes. They're [16:20] like, "Oh, we get to be the three god [16:22] emperors of Earth." But Elon Mus is [16:23] like, "Yeah, no, I'll say it aloud." So, [16:24] at least he's saying it aloud, I guess. [16:26] Oh, okay. Other concerns. Model [16:28] collapse. This like completely [16:30] unrelatedly. Model collapse that GPTs [16:33] just start to plateau because they ran [16:35] out of good data. They're ingesting [16:37] their own data. They cannot operate. I [16:39] actually put this at a fairly low chance [16:41] of happening and fairly low level of [16:44] concern. Like if it happened, it would [16:47] trigger maybe the collapse of the [16:48] bubble, which would be bad. But that's [16:50] not really model collapse's fault. It's [16:52] the bubble's fault. From what I have [16:53] read from people who know what they're [16:55] talking about, there are people who [16:56] think that model collapse will happen. [16:57] There are people who don't, and the [16:59] people who don't outweigh the people who [17:00] do. Now, are they just like eating their [17:02] own hype machine? Maybe. It seems like [17:04] there is so much economic incentive to [17:06] figure out that problem that they will [17:08] figure out that problem. I don't know, [17:09] though. That is my near-term section. Am [17:12] I missing stuff? Absolutely. No. I'm [17:14] going to add one more. I'm going to add [17:15] another one. A very broad category of [17:17] just the systems not working because [17:21] there's just so much AI going on that [17:24] nobody knows what to do with anything. [17:27] So you have a job opening and you get [17:28] 20,000 applications. Companies aren't [17:31] used to this and it's all automated. You [17:33] don't know who's real and who's not [17:34] real. You have fake job openings. You [17:37] have like everybody's like tossing [17:39] things at each other and like making [17:40] them with AI and then decoding them with [17:42] AI. there's all this stuff that like [17:43] teachers are supposed to be doing that [17:45] like it's all brand new so it's just [17:47] make work it feels like there's just so [17:49] much of that and then also in terms of [17:52] education which I think a lot about I am [17:54] very worried that there's going to be a [17:56] period of time where no one knows [17:58] whether the credential of higher [18:00] education stands up anymore and it [18:02] probably will but because we're not sure [18:04] if it does that's just going to be bad [18:05] for graduates so like in a chatbtd world [18:07] did everybody just sort of fake their [18:08] way through college some people did and [18:11] so how do you deal with that that's not [18:12] something like usually if somebody got [18:15] A's and B's in college, you know that [18:16] they like are pretty like they know [18:18] stuff and they're good at things, but [18:19] now like does that hold up anymore? I [18:22] think that it does, but I am worried [18:24] that companies will not be sure if it [18:26] does and they might use that as an [18:28] excuse to not hire. There's so much just [18:30] generalized disruption. That's a [18:32] concern. Now, I'm going to move to to to [18:34] longer term, and I'm talking like [18:36] medium-term, like 3 to 10 years from [18:39] now. We could see a world I think this [18:42] might already be happening, but I think [18:43] that 3 years from now is like my big [18:45] concern time. Just a world without any [18:47] apprenticeship. So, like nobody gets [18:49] paid to kind of suck at something [18:51] anymore because that's like what the [18:52] beginning of your career is is like [18:54] you're not that good at things yet and [18:55] you learn by having a job where you're [18:58] at an introductory level. But like if [19:00] these systems can do the easy work on [19:03] their own, nobody gets hired to do that. [19:07] Like I remember being hired to like [19:08] design bad logos and and develop cloudy [19:12] databases that weren't very good and [19:13] like do SQL queries badly and that was [19:17] fine. That like somebody had to do it. [19:19] But now no one would hire somebody to be [19:21] bad at SQL logos make themselves. Not [19:24] good ones but bad ones. You know what? [19:26] Like how does a person get good at [19:29] design when when like the level of [19:31] design they could do when they're just [19:32] getting out of school is easily done by [19:34] a generative AI? So like and then where [19:36] and then where do the good people come [19:38] from? So if there's no apprenticeship, [19:40] what happens in the future? This feels [19:41] to me like it's already kind of [19:42] happening. But the real like bite us in [19:45] the ass part of it comes down the road [19:46] when we're like, "Oh, we don't have any [19:48] lawyers to do the hard stuff because we [19:50] didn't hire any intro people because we [19:52] had GPTs doing the easy stuff." So I'm [19:55] concerned about that. The more distant [19:56] thing you're like but what about the [19:57] more distant thing? The more distant [19:58] thing is coming don't worry. Next uh 3 [20:00] to 10 year problem I think like I'm [20:02] worried about uh people are worried [20:05] about I don't know if I am cognitive [20:07] atrophy. So will people just not be able [20:11] to do things? Will they not know how [20:13] code works? They'll just prompt. Will [20:15] they not know how to write an essay? [20:16] They'll just prompt. I think this is to [20:20] some extent almost inevitable. But I I [20:22] weirdly am not as concerned about it as [20:24] a lot of other people because we've done [20:26] it before. You know, written language [20:29] outsourced a lot of our brain power onto [20:31] the page and like that ended up being a [20:33] net good. I don't know. Are we going to [20:35] be dumb or are we going to be sharp but [20:38] with different tools? That seems more [20:40] likely to me. People will fight me about [20:42] this and they absolutely should. I am [20:44] not assigning any level of certainty to [20:46] this one. I could totally be wrong. Less [20:49] like 50/50 on this. I do not know. And [20:51] then also happening in the next 3 to 10 [20:53] years and also already happening to some [20:56] extent today is using AI in warfare. And [20:59] right now I have heard that people will [21:01] use AI to to select targets but not to [21:06] do the actual button pushing though. To [21:09] what extent that's different, I don't [21:11] know. But in the future, I don't think [21:13] that we're that far away from an [21:16] autonomous system that both determines [21:19] and executes a target or or an action on [21:23] a target. I shouldn't necessarily say [21:25] executes a target, but that is probably [21:26] what we're talking about. I'm worried [21:27] about this. I I uh do I think that it's [21:31] going to become Terminator? I don't [21:33] think it's going to become Terminator, [21:34] but like I don't love that because of [21:37] the blackbox thing, because of all of [21:38] it. Like I just tools of the future used [21:40] in warfare are always very upsetting and [21:44] uh likely to be misused. We won't know [21:48] how they will be misused until they are [21:49] misused. And until we are aware of the [21:52] damage done, those who want the tool [21:55] will just sort of use the tool. That [21:57] seems likely to me. Longer term 10 plus [21:59] years. Number one obviously [22:03] what is the economy if they're right? So [22:06] if they're right that the AI will be [22:09] smarter than everybody, what does [22:10] anybody do and how does anybody find [22:12] meaning in that world? I think that is a [22:14] totally legitimate concern. Now I do [22:16] know lots of people who do not work and [22:18] who have a great deal of meaning in [22:20] their world. They might be disabled. [22:21] They might be retired. They like work [22:24] and meaning do not have to be tied. They [22:27] are for a lot of people right now and [22:29] that does not surprise me. I also think [22:32] if if not intervened with, it seems [22:36] likely to me that it will come along [22:38] with a huge amount of [22:41] inequality in wealth. And that also [22:44] concerns me. I think that that is likely [22:46] to have bad outcomes. And I don't just [22:49] mean like, oh, people have not very much [22:50] wealth. I mean, they do not stand for [22:53] that world. Why would they? Why would [22:55] they stand for a world that does not [22:56] treat them with dignity? Why would they [22:58] stand for a world that does not treat [22:59] them as valuable? That makes sense to me [23:03] and that's worrying, especially when [23:04] it's young people because young people [23:07] know that they have a whole life to live [23:08] and they will not stand for. I don't [23:10] know that it's a super near-term [23:12] concern, but maybe it's nearer than I [23:14] think. I don't know. I think that you [23:16] kind of do have to have systems that are [23:19] smarter and more capable than humans, [23:21] which we definitely do not have right [23:22] now. And I don't know that we will have. [23:24] People argue about whether or not we [23:25] will have them, but I am worried about [23:26] that. And then we have uh you know super [23:29] intelligence unaligned uncontrollable [23:33] turns into evil uh it takes over the [23:36] world and turns us into pets or destroys [23:38] us. I have received critique that I [23:41] spend too much time on that one. I [23:43] understand why some people spend more [23:45] time on that one. Some people will say [23:46] like that the motivation behind that is [23:47] really the AI industry hype machine and [23:50] they're like focus on this thing over [23:52] here and not on all these like shitty [23:53] things we're doing right now. I take [23:55] them at their word and maybe I will eat [23:56] my word someday but I take them at their [23:58] word that in fact they are worried about [23:59] this and is the biggest possible problem [24:02] you know becoming pets to a super [24:03] intelligence or being extincted by it [24:05] that seems like a big deal but I don't [24:08] think that it's the most likely outcome [24:09] but I understand people who are in this [24:12] uh frame where it's like well if it's [24:14] the biggest deal even if it's a small [24:16] likelihood we should concern ourselves [24:18] with it and I think that we should [24:19] concern ourselves with it I don't think [24:21] it's the only thing that we should talk [24:22] about and it makes a lot of sense to [24:23] focus on the things that are happening [24:25] right now and that might happen in the [24:26] next 3 to 5 years. Do I think that we [24:28] are going to be capable of of like [24:30] marshalling the government to regulate [24:32] this? Well, absolutely not. I think that [24:34] it's likely that these three to five [24:35] people who are in charge of the things [24:37] will do regulatory capture more than [24:39] anything. I think that they will be the [24:41] ones with the ears of the politicians [24:42] and they will make it so that they [24:44] solidify their own control. regulatory [24:47] capture being um you know we help the [24:50] government like powerful people help the [24:51] government write the regulations and [24:53] thus the regulations get written in such [24:54] a way that actually protects the [24:56] incumbents the the existing power [24:58] structures so that seems like totally [25:01] likely to me but I don't know but yes in [25:03] terms of of like the paperclip maximizer [25:06] situation where uh we either like make [25:08] some paper clips and then it turns the [25:09] entire world into a paperclipip factory [25:10] because it got too smart and and is just [25:12] doing the thing it wants to do or like [25:14] it starts to fight they fight against [25:16] each other. There's like some kind of [25:17] battle beyond humanity. Lots of science [25:19] fiction in this space that I I haven't [25:20] enjoyed in my life. I think that just [25:24] like with her for example, so her very [25:27] preient movie, but it did not predict a [25:30] weird [25:32] sycopancy induced psychosis. Nobody [25:34] predicted that. I think that the more [25:35] likely thing here is we don't know. We [25:40] have no idea what an actual super [25:43] intelligence would mean. To me, the [25:45] likelihood of super intelligence, [25:48] you know, in the next 10 to 20 years, I [25:50] I don't know where to put that slider. [25:52] People who are more informed than me [25:53] disagree on that. It could be zero. It [25:56] could be 100%. I don't know. People seem [25:58] very certain and in fact uh there's [26:00] their certainty, the the more certain [26:02] they are that the like they get plenty [26:04] of attention through their certainty. Uh [26:06] so that's that's a concern. Uh but then [26:08] then over here on super intelligence, I [26:10] think definitely if it happens, I don't [26:11] think if anyone builds it, everyone [26:12] dies. I think if anyone builds it, we [26:15] have no idea what happens. It's a real [26:18] cosmic toss up. That's not a thing that [26:21] there's precedent for. That's like [26:23] humans happening again. Nobody would [26:25] have predicted 200,000 years ago that [26:27] anatomically modern humans would do what [26:29] we've done and have like screens and [26:30] stuff all around them. So like you make [26:32] a super intelligence. I don't know. The [26:33] thing that I will say is that like it [26:36] will all probably happen slower than [26:39] people imagine. This is my guess because [26:42] it's going to have to interface with [26:44] human systems and human systems don't [26:46] change immediately. It has taken a long [26:48] time for anything to like until 2019 the [26:51] United States nuclear arsenal was run [26:54] off of an 8-in floppy disc. Like human [26:56] systems take a while to change. So don't [26:57] tell me you think that this is all going [26:59] to happen in 10 years is a little bit [27:00] how I feel. This is what I am most [27:02] concerned about at this moment. I might [27:04] change my mind tomorrow, but right now [27:07] this focus on biological warfare, on [27:10] chemical weapons, on on paperclip [27:12] maximization, I don't think this is the [27:14] biggest concern. I think we should look [27:16] at what AI could potentially be good at. [27:19] What weaknesses human beings have and [27:22] mix those two things together. And what [27:24] I'm concerned about is the thing that's [27:26] already happening where AI that the goal [27:29] of the AI is to give us stuff that will [27:32] keep us engaged on a video platform and [27:36] it starts to optimize for things we did [27:38] not expect. Nobody said give these [27:40] people outrage bait. What they said was [27:41] keep people on the website and outrage [27:43] bait because it became a very compelling [27:44] kind of content that humans happily [27:46] supplied because we would love to get [27:49] more attention because we're humans and [27:50] that's many of us including myself want [27:53] attention. So what I am concerned with [27:56] is is we are not that difficult [27:59] ultimately probably to manipulate. And [28:02] so if you have one or multiple very [28:06] intelligent systems that have goals, I [28:08] think that they will find it much more [28:10] easy to manipulate people than we would [28:12] guess. Our realities are defined by what [28:15] we consume. And there is a way out of [28:19] this which is simply getting off of the [28:22] screens and I think that we'll be bad at [28:25] it. What I have seen and I hope there's [28:27] a backlash to it but what I have seen is [28:29] that every opportunity that a platform [28:32] has given to people to make fewer [28:35] choices people have taken that option. [28:38] So it used to be that feeds were [28:40] chronological and people were like no I [28:42] want I want that chronological feed. [28:44] when you gave them the option, they [28:46] chose the algorithmic one. Did they [28:47] choose the option or did they imply a [28:49] choice for the option by spending more [28:51] time with algorithmic feeds? That's I [28:53] don't think that's semantic, but like [28:54] they chose the algorithmic feeds when [28:56] you gave them the option between having [28:58] a video supplied to them or clicking or [29:01] choosing and clicking on a video they [29:02] want to watch. They prefer having a [29:04] video supplied to them. That's Tik Tok [29:06] versus YouTube, right? People prefer [29:09] decisions being taken away from them. [29:11] And that is the thing that [29:14] I think concerns me the most like [29:16] broadly that like given all of these [29:17] things, that's what concerns me the most [29:19] about AI. People tend to prefer their [29:22] choices be taken away and to to have [29:25] less agency. But perhaps [29:28] I am misreading that and that's [29:30] something that only happens during [29:31] certain times of the day or or maybe [29:33] we've already got sort of gotten that [29:34] itch scratched and we'll change our [29:36] minds about this. But that has thus far [29:38] been the direction that we have gone. [29:40] Again, what humans do is we communicate, [29:42] we talk. We have a huge bandwidth of [29:43] information. And so, it makes sense to [29:47] me that the thing to be most concerned [29:49] about is how AI interfaces with that. [29:52] And already, I think it interfaces with [29:54] it in in extremely intense ways through [29:57] recommendation algorithms, which have [29:58] been around for a long time, but now [30:00] also through LLMs in helping us and like [30:03] giving us an opportunity to define [30:04] reality. And when you mix that with the [30:07] extreme concentration of power here and [30:09] just a hand handful of companies and a [30:12] and a few guys who I have no reason to [30:14] trust. I don't know those guys. They've [30:16] made some weird decisions recently. That [30:18] does that's yeah that's pretty worrying. [30:19] How does that plug into all the things I [30:20] just talked about? But that thing is the [30:22] thing I am most concerned about. Now [30:23] I've done an interesting thing today [30:24] where I have recorded this video several [30:27] weeks ago. I've edited the video and [30:29] I've sent the video to someone who I [30:31] respect and trust who has, you know, [30:35] less hype fueled perspective on AI than [30:39] I have at some points in my life. And I [30:42] would like to talk to him about the [30:44] video that I just made. And so [30:45] everybody, this is Cal Newport to talk [30:48] to me about the video that I made and [30:49] then I showed to him and he's seen it [30:51] and now we will talk about it. But [30:53] before we get to that, I do from this [30:55] hotel room have something that I want to [30:56] tell you. You may have noticed that I've [30:59] been working on this video for a long [31:00] time. And part of the reason why this [31:01] took so long is that in the meantime, [31:03] I've been working on a number of other [31:04] things. And one of those things was a [31:06] really big project, a big idea, which is [31:08] that the educational media company that [31:10] my brother and I started 15 years ago, [31:12] we have converted it into a nonprofit [31:14] organization. And that was a fairly [31:16] complex thing to do. And then there was [31:17] a lot of like logistics and press things [31:20] to go along with it. So I've been doing [31:22] that a lot. So this video took a while [31:23] to do. But also in my new role as [31:26] chairman of the board of complexely, no [31:28] longer owner of Complexly, one of my [31:30] jobs is to say Complexly a lot so that [31:32] you know that there is a company of over [31:34] 80 people working hard efficiently [31:37] creating value for people making stuff [31:38] like Sihow and Crash Course and Eons and [31:42] Study Hall and Bizarre Beasts and lots [31:44] of educational stuff on YouTube that is [31:46] high quality and research and fact check [31:49] and that's all. This is not an ask. I [31:51] just want people to know that complexity [31:52] exists, that there's a bunch of people [31:54] working hard to make the information [31:56] environment better and to use the tools [31:59] of the social internet which can [32:01] obviously and obviously somewhat [32:03] incentivized to be used for not so great [32:05] using them to try and do good. We've [32:07] been doing it for 15 years. We're going [32:09] to be doing it for many more decades in [32:11] the future. Got lots of ideas for things [32:13] that we're going to do and that's part [32:14] of my job just making sure the world [32:16] knows about them because they are great. [32:18] And now actually our interview with Cal [32:20] Newport. It's really good. I loved it. [32:22] >> Cal Newport, you are a a writer, a [32:25] professor at Georgetown University of [32:27] Computer Science. You have spent the [32:28] last decade asking the hardest simple [32:31] question, which is what does technology [32:33] do to uh humanity? If you could pick [32:36] like a biggest part of the AI story, [32:38] what is the biggest part of the AI [32:39] story? [32:39] >> Well, I have to say, you know, I've [32:40] looked at this question with a lot of [32:42] technologies. AI is the messiest, most [32:45] complicated technology I think I've had [32:48] to to deal with. I mean, if I remember [32:50] correctly, when you first started [32:52] talking about this video, for example, [32:53] the plan was what, 10 minutes long. [32:57] >> Right. And it ended up being [32:58] >> I sent you an email and I was like, I'm [33:00] going to do a 10-minute video and then [33:02] I'm going to jump over to you and then [33:04] Yeah. Literally. [33:05] >> Yeah. Because it's almost always the [33:07] story is almost always this sort of [33:08] binary story of well there's this thing [33:10] we liked about this technology but then [33:12] it had this other effect we weren't [33:14] expecting. And so like now we have to [33:16] kind of balance the two. It's usually [33:17] that simple. It's email is nicer than a [33:20] fax machine. It's pretty convenient. I'm [33:21] happy it's here. Oh my god, but there's [33:23] this side effect that we get too many of [33:24] them and we're distracted. AI is such a [33:27] messy story that makes that question [33:28] it's it's a very complicated question [33:30] because it's [33:31] >> many different technologies that we all [33:33] call AI. They're all developing on uh [33:36] different timelines. [33:38] >> They all have their own issues in the [33:40] present that are worthy of huge amounts [33:43] of discussion and push back. And then [33:44] they all have extrapolated future issues [33:48] that go from like slightly worse [33:50] versions of the present to all of [33:51] humanity is dead in two years. And all [33:54] of this is mixed up with like a cast of [33:56] characters that are acting like Bond [33:58] villains. And [34:01] it really gets pretty difficult. I [34:03] missed email say like hey we get too [34:05] many emails. [34:09] >> It's like oh Slack was gonna fix things [34:10] but actually it made things like [34:12] different like bad in a different way. [34:14] >> Those were the days. [34:16] >> Now now Slack is waking up and it may [34:19] kill us all. [34:20] >> Exactly. Yeah. It's going to use all of [34:22] our energy and uh murder us. Yeah. It's [34:24] a harder conversation now. [34:25] >> Slack there's a new box in your town. It [34:27] eats jobs. [34:28] >> Uh and also it makes your power bills go [34:30] up. So, you know, so you didn't answer [34:32] that question, which I'm happy to [34:34] answer. [34:34] >> Or you did or or yeah, I mean, you could [34:36] give me you can give me a more detailed [34:38] answer, but like in in a way that is an [34:40] answer, which is like this is a chaos [34:43] machine. It's a chaos machine for [34:45] thinking about and I will say it okay [34:47] what I am trying to do now when I think [34:49] about it in a lot of my work is separate [34:52] off a lot of the extrapolated futures [34:55] >> which by which I mean just I'm taking a [34:58] trajectory that has happened so far and [35:00] then I'm extrapolating that well what if [35:02] that happens for another three years [35:03] where would we be and then reacting to [35:05] that future uh most computer scientists [35:07] I know have a lot of problem with those [35:09] extrapolated futures we think they're [35:11] they're sort of ignoring the reality and [35:12] limitations of technology ology but I am [35:14] very interested right now in a lot of [35:16] the present issues. So things that are [35:17] happening right now with the technology [35:19] as it exists and two of them that I care [35:22] quite a bit about is the disappearance [35:24] of truth and the disappearance of focus [35:26] or concentration as a human ability. So [35:29] disappearance of truth generative AI can [35:31] create videos and images of anything. [35:32] That's a problem. I there there's no way [35:34] for that not to be a ma we don't need [35:36] super intelligence for a massive [35:38] problem. That's a massive problem. And [35:39] and the disappearance of of focus I'm a [35:42] professor. My concern is a lot of what [35:45] we do in the educational system is [35:47] exercises invented to stretch your mind [35:49] like you would stretch a muscle so that [35:50] mind becomes more complicated and in the [35:52] future can handle original thoughts we [35:54] need for the the survival of humanity [35:55] and civilization. [35:57] AI, especially chatbots and LLMs, are [35:59] very good at these exercises because [36:01] they're just exercises. And so in [36:02] theory, you can offload that type of [36:04] work to them, which in the moment makes [36:05] your life easier and in the long run [36:06] means your brain misses a final step of [36:09] development. We need for just human [36:10] civilization to continue at the pace [36:12] that it is. So those are the two things [36:13] that they're happening right now. [36:15] >> Yeah. [36:16] >> And I'm probably most concerned with in [36:17] the moment. [36:18] >> Gotcha. A couple of years ago, maybe a [36:20] year ago, I made a video that was titled [36:22] something like here's how to like know [36:24] when an when an image is AI and also why [36:26] it doesn't matter because this is how to [36:29] know today. But man, I tell you what, [36:32] and it's it's somewhat amazing how [36:34] quickly our brains get an aesthetic. [36:38] Like now there's sort of like an [36:40] aesthetic of chat GPT images and easily [36:43] I see them and I'm like, well, that's [36:44] what that is. And I only had to see like [36:46] three of them before I knew forever. But [36:49] we are already in the world where I've [36:51] seen videos that I thought were real. [36:53] And and I didn't even ever find out that [36:55] they weren't real. You know, I'm not [36:57] talking about a time when I was like, [36:58] "Ah, I'm pretty sure that's real." And [36:59] then I checked and I found out it was [37:00] wrong. I mean, I watched it, I thought [37:02] it was real, and I moved on with my day. [37:04] Like, that's happened to me. I'm sure I [37:06] am really worried about the degradation [37:07] of truth generally. That's not just an [37:09] AI thing either. It's a it's an [37:12] aesthetic thing. It's a political thing. [37:14] It is a thing that is happening in our [37:15] culture right now. But that's not what [37:17] this conversation is about. I am really [37:18] curious about this focus one though. [37:20] It's hard not to think about Socrates in [37:23] this situation who was like people will [37:25] not be able to think when they're [37:26] writing everything down with this pens [37:28] and paper. Not paper but like like like [37:31] literacy was going to take away [37:33] abilities that humans had. And it did. [37:36] Literate humans have different abilities [37:39] than non-literate humans. And yet this [37:41] is okay. Have you seen students using it [37:44] in a way that actually [37:47] does contribute to focus or is it just [37:51] this tool is going to break our ability [37:54] to make that final developmental step? [37:56] >> I think it's a one-two punch to our [37:59] brains where the other punch is social [38:01] media. [38:01] >> So think about this for a second. We [38:04] have uh highly engaging, distracting, [38:06] ubiquitous entertainment happening on [38:08] our phone which uh lowers our our [38:11] tolerance for cognitive strain because [38:13] we get so used to this idea that at the [38:15] the slightest hint of boredom, I can get [38:17] this sort of algorithmically perfectly [38:18] targeted little hit of something that's [38:20] going to be interesting. So that that [38:22] reduces our our interest or comfort with [38:26] concentration uh consuming hard [38:28] information. The other thing we have to [38:30] do with our brain to try to get smarter [38:32] is produce information from scratch. So [38:34] like take the circuits that you would [38:35] develop by reading hard things um and [38:37] use them. And how do we use them? By [38:39] writing or the activities that surround [38:41] writing. [38:42] >> So social media is kind of taking us [38:43] away from trying to consume complicated [38:45] information that makes those circuits [38:46] weaker. Generative AI takes us away from [38:49] producing information which is using the [38:52] circuits we already have the strength of [38:54] them and to get used to applying them to [38:55] the real world. And you put those two [38:56] things together and I think we lose [38:57] those circuits and they're critical [38:59] circuits. I mean, so this is the way [39:00] that we think about the brain is that [39:03] literacy is not natural, right? So, so [39:05] in order to actually consume a book [39:08] >> or to read the written word, we have to [39:10] essentially hijack parts of our brain [39:12] that were evolved for very different [39:13] purposes and we have to through very [39:15] intense training. I mean, what we do in [39:16] kindergarten, that's really hard. We're [39:19] rewiring our brain. Yeah. To hijack [39:21] different centers. Like you can figure [39:22] out, for example, when you're outlining [39:24] something you're going to write that [39:26] there's a part of your brain that [39:27] evolved to actually keep track of your [39:28] location, physical space, and it's been [39:30] hijacked to try to keep track of like [39:32] where you are in an abstract, you know, [39:34] uh, argument. That brain that we rewire, [39:36] >> wild. Yeah, [39:37] >> it is wild. And this this brain that we [39:39] rewired to be to be able to to consume [39:42] written information and to produce [39:43] written information, it's a different [39:45] brain. It is a different brain than we [39:47] had 10,000 years ago, right? We we have [39:49] completely rewired the brain. But that [39:50] rewired brain that can do what Marann [39:52] Wolf calls deep reading processes is [39:55] literally a smarter brain that sees the [39:57] world in a more complicated way and is [39:59] capable of the type of sophisticated [40:01] reasoning and abstraction that our [40:03] current modern world requires to [40:06] function. So it's not for nothing to [40:08] lose that brain. If you begin to really [40:10] degrade that wiring, it really makes it [40:14] going to make it difficult for the way [40:16] our civilized world runs today. We're [40:18] going back [40:20] closer to a preodern brain which is not [40:22] playing nice or will not play nice with [40:24] the modern world. One of the things that [40:26] I talk about in the video is that we've [40:28] already got we're like we're in the [40:30] world of AI already in that every [40:32] recommendation algorithm in the world [40:34] has been finally tuned by uh the the [40:37] decisions of all of the people who have [40:39] ever interacted with it and and then [40:40] also there's you know the smartest [40:42] people in the world creating them to to [40:44] be optimized for basically holding on to [40:46] your attention. both training me as a [40:48] creator to optimize for being [40:51] extraordinarily salient and and like [40:53] capturing attention as as effectively as [40:55] I can. So the algorithms are kind of [40:57] optimizing the creators which I [40:59] definitely feel but then they are you [41:01] know put us in like a world of [41:03] competition amongst each other and and [41:05] then they find the most salient possible [41:07] thing. And I think of that like like [41:08] imagine eating junk food took like it [41:11] you never got full. One like you could [41:13] literally couldn't have so many Doritos [41:15] that you stopped being hungry and two [41:18] you ate a Dorito by looking at it. Like [41:20] that's what information is like now. It [41:23] is like we have created junk food [41:25] information but information is something [41:27] that we never lose our hunger for [41:29] because we're information bots, you [41:31] know, like we're we're attention [41:33] machines. [41:33] >> Don't forget the piece where someone [41:34] follows you around with a cart full of [41:36] Doritos. Yeah. Yeah. Yeah. [41:37] >> They're always there no matter where you [41:38] go, including to the bathroom. Like when [41:40] you're on in church, they're there with [41:42] a cart that they've always refilled, [41:43] >> right? And also like one of my favorite [41:45] examples of this is that if you go to [41:47] search something on Instagram, there's [41:49] like the search bar takes up 5% of the [41:51] screen and then the rest of the screen [41:52] is Doritos. It's like, did you want to [41:54] search for something? Did you want to [41:56] like message a friend? Oh, sorry. What [41:58] if you like saw a hot person? What if [42:01] there was like a very strange and [42:02] confusing like moving image that you [42:05] just cannot cannot resist and and sorry [42:08] but now you're now you're back into the [42:10] surface where we can advertise to you. [42:11] >> What's kind of interesting about all [42:12] this is I mean your video is all about [42:14] like these latest development in AI this [42:16] is very sophisticated programs right [42:18] we're talking about like large language [42:20] models etc. And yet the AI that that [42:22] you're just describing which has a [42:24] massive impact on so many ways on just [42:26] the way information flows in our lives, [42:28] right? These recommendation algorithms, [42:29] that's actually really simple comparably [42:30] speaking. Like actually the the [42:32] algorithm that goes behind something [42:34] like Tik Tok figuring out what video to [42:35] show you didn't really require some sort [42:38] of massive mathematical breakthrough or [42:40] some sort of like new technology [42:41] invented. Yeah. Actually, most of the [42:43] breakthroughs, technical breakthroughs [42:45] that like Bite Dance innovated had to do [42:48] with how do you run this at a massive [42:49] scale with like a massive uh scale of [42:52] possible videos and a massive number of [42:54] users and run it really quickly or [42:55] whatever. It's got a really boring [42:56] optimization issues, but the actual [42:58] under underlying recommendation [42:59] algorithms are really pretty simple. I [43:01] mean, it's not to oversimplify it, but [43:04] it's like all of this content is getting [43:06] embedded in some sort of [43:07] multi-dimensional space. So just like a [43:09] long list of descriptors, numerical [43:12] descriptors, [43:13] >> and it keeps track of different points [43:15] in that space where you've shown [43:17] interest before because like you've [43:18] watched the video all the way through to [43:20] the end. [43:21] >> And it's just sampling [43:23] videos in this sort of space of topics, [43:25] multi-dimensional space that are sort of [43:27] near places where you've been [43:28] congregating uh with some randomness in [43:30] there to make sure it's not missing [43:31] places. Like we've known about these [43:32] type of recommendation algorithms for 30 [43:34] years, you know, but we've never we [43:36] never executed them so cleanly as [43:38] something like Tik Tok came along and [43:39] did. [43:40] >> And that's so much I mean that's so much [43:41] of the story of technology is like you [43:43] make the steam engine and you're like we [43:44] made a steam engine and like we put that [43:46] on the timeline as the steam engine [43:47] invented but you know the steam engine [43:49] sucked and then it turned out it could [43:51] do a lot more when we made it a lot [43:53] better. And so the recommendation [43:54] algorithm maybe someday will be a mark [43:56] on somebody's timeline, but uh you know [43:58] it just keeps getting better and the [44:00] cart of Doritos keeps getting more [44:02] enticing. How do you think I did in my [44:07] list of worries? [44:08] >> What? Well, I think what you did well is [44:10] you recognized some of the real [44:12] extrapolated future worries and said, [44:14] "Let's put those a bit aside. Let's put [44:16] aside let's not fully focus on the [44:19] Terminator robots are yeah marching down [44:21] the street because you know there's [44:23] there's much more approximate future uh [44:26] issues that are happening there was a [44:28] mix in there of different types of AI [44:29] technologies for better for worse right [44:31] so [44:32] >> which this is the reality of it right [44:34] but it's important like the for people [44:35] to understand right there's sort of [44:37] prediction type machines that have been [44:39] around for a long time this is where [44:40] like the concerns you have about bias [44:42] and sentencing for example the classic [44:45] AI AI safety areas. These type of [44:46] technologies have been around for a long [44:48] time. They haven't really seen any major [44:50] innovative boost in the way that [44:52] generative AI has, but they we've long [44:53] been studying those concerns and you [44:55] know those concerns are important. We [44:56] have the environmental concerns. That's [44:58] a hard one to get your arms around [45:00] because uh there's a competing interest [45:02] there. There's another competing [45:03] interest of the power companies [45:04] themselves trying to overplay the power [45:06] demand of these data centers so that [45:08] they can get infrastructure buildout [45:09] paid for by you know municipalities. So [45:11] that's like a complicated. [45:13] It's great that you like you, the [45:15] computer scientist, brought that up [45:16] because like that's the that's my thing [45:17] that I thump on is is what we all have [45:20] to understand is that the only way power [45:22] companies make more money is if they [45:24] build more stuff. And so they have a [45:25] strong incentive to say that they're [45:26] going to build more stuff. And also the [45:28] AI companies have a strong incentive to [45:29] say they're going to need more stuff. [45:31] And this is very bad for your power [45:32] bill. [45:32] >> The problem is power is expensive for [45:35] the AI company. Like the AI companies [45:37] have to pay for it, right? So yes, [45:39] training some of these foundation models [45:41] right now is like an incredibly [45:43] expensive endeavor. And if you start [45:44] scaling that out, you're like, man, [45:45] that's going to be a lot of energy. But [45:47] really, it's very hard to be profitable [45:48] with like a trillion parameter model [45:50] that you're charging $20 a month for [45:52] people to use. So a lot of computer [45:54] scientists that don't have a a dog in [45:56] the fight. are not, you know, aren't [45:57] scientists at these companies looks to [45:59] things like DeepSync from, you know, [46:01] last January and it's like this really [46:03] the future is going to be much more [46:05] smaller language models where they're [46:06] used that are much more specialized a [46:08] particular uses. I I don't we don't [46:10] think the the age of these like models [46:13] where I should just talk to a model in a [46:15] chat box and it should be able to do [46:17] anything. I really think that's sort of [46:19] to uh the AI industry what Formula 1 [46:21] cars are like to the automobile [46:23] industry, right? You're trying to beat [46:24] your chest and show look how complicated [46:26] our technology is. But you're not [46:27] selling a lot of Formula 1 cars because [46:29] they're completely impractical. They [46:30] don't really work well on the roads. The [46:32] future is probably going to be taking [46:33] that technology much more energy [46:36] efficient and smaller and cost-effective [46:37] to run embedded more closely into [46:40] particular use cases. And those [46:41] foundation models, I think, are really [46:42] just a it's a it's a a measuring contest [46:45] of sorts, if you know what I mean, [46:47] between these companies to show like we [46:49] have the best technology so that when we [46:51] then sell you these more specialized [46:52] tools, you're like, "Oh, that's the [46:53] name. I remember McLaren because you won [46:55] the F1 race. You must make good cars." [46:57] >> But they're not selling us the the [46:59] specialized models. They're selling us. [47:01] They're like giving away the Formula One [47:04] cars. [47:04] >> Yeah. It's a terrible business model. [47:06] It's a terrible business model. They're [47:07] >> Are they gonna Are they just going to [47:08] shut those down? Are they going to say, [47:10] "Hey, turns out this is a bad idea. [47:13] Like, we lose money every time you use [47:16] our product." [47:17] >> I think what's happening is there's a [47:18] they were hoping for regulatory capture, [47:20] right? There's this race of we want [47:23] competitors uh out of business before we [47:25] pivot to these smaller, lowerc cost [47:27] machines because the barrier of entry [47:29] there is smaller, right? The foundation [47:30] models require that you build custom [47:32] data centers that have, you know, air [47:34] conditioning technology that didn't [47:35] exist before that's going to have a [47:37] 100,000, you know, GPUs running [47:39] concurrently, which no one's ever tried [47:41] to do before. Once you go down to a [47:42] smaller model, a lot of companies can [47:44] create those. You have a real [47:46] competitive marketplace. So I think they [47:47] were trying to keep the focus on these [47:49] oracular can do everything general [47:52] purpose super impressive models and try [47:54] to put the other people out of business [47:57] or build enough regulation around that [47:58] no one else can compete anymore because [48:00] as soon as they pivot to here's small [48:01] models that work on phone and there are [48:03] companies starting to do this now all [48:05] that money you spent right you lose that [48:08] advantage because someone else can walk [48:10] around get some open source parameters [48:11] for meta and start building their own [48:12] sort of smaller more bespoke model so I [48:15] think it's a scary future for those [48:16] companies [48:17] because it's a super competitive [48:18] marketplace. And so they're still [48:20] holding on to foundation models matter [48:23] and all the jobs are going to be [48:24] automated and it's going to be just [48:26] three companies left and we're going to [48:27] have to have redistributive, you know, [48:30] uh, equity taxes just so people can eat [48:33] like all that. I think that that there [48:34] there's a business element to that as [48:36] well. It makes sense to me because this [48:38] is how I feel. Uh but I'm not the only [48:40] one that we had this world where [48:42] suddenly I don't know like like GPT3.5 [48:46] was like oh whoa oh that's very [48:50] different from where I thought we were. [48:52] We aren't just like headed toward a new [48:53] place. We're in a different place. And [48:56] then four was way better. Um and things [48:59] things have improved. You know they like [49:02] they they have become more useful tools [49:04] in a variety of ways. they remain [49:06] unusful in a lot of the same ways and [49:08] that like absolutely all of the flagship [49:11] models will just make stuff up. And so I [49:13] see why I feel like and I do feel this [49:17] way. This will happen again and there [49:19] will be some other important technology [49:22] that will get us closer to that AGI like [49:25] thing to that thing that is good at a [49:27] lot of different stuff and even like [49:29] it's it's good at being in the physical [49:32] world as well as in the you know just [49:34] the knowledge workspace and like because [49:36] this is all lay person stuff I don't [49:39] really even know what a transformer is. [49:41] I get basically what's going on, but I'm [49:45] not gonna be a computer scientist and [49:46] I'm not going to understand the multi [49:48] even the multi-dimensional architecture [49:49] of the Tik Tok algorithm is beyond me [49:51] and will remain. So, is there anything [49:54] to this instinct? Because it seems like [49:56] there's plenty of people who think, [49:57] yeah, yeah, like like there'll be [49:59] another innovation and then another one [50:00] after that and then we'll have a kind of [50:02] thing that feels a little super [50:04] intelligency. [50:05] >> I think that's a reasonable instinct. uh [50:08] in fact I think it's in a general sense [50:10] the right instinct to react to the last [50:13] three years is you know not to say this [50:16] trajectory is going to continue unabated [50:18] with language models and within two [50:20] years they'll be super intelligent but [50:22] to say oh that was a big leap are we [50:24] going to have similar big leaps in other [50:25] technology so I'll tell you what I've [50:27] seen trying to report on this from the [50:28] computer science side I think your read [50:30] of what happened with the language [50:31] models is more or less correct uh GPT3 [50:34] was actually the first really big jump. [50:37] Uh this is where they actually acted on [50:39] the scaling law paper written by Kaplan [50:41] Adall where they said what would happen [50:42] if we actually do significantly scale up [50:44] the size of these language models and uh [50:48] the size and the training. [50:49] >> Yeah. And it really [50:50] >> kind of like that's the thing that's all [50:52] they really did like they they were [50:53] using the same technology. They're just [50:55] like more data, [50:56] >> more data and larger size- which is [50:58] non-trivial because actually to even [51:00] like train and run a much larger size [51:02] model uh requires like a lot of under [51:05] the under the hood type of tinkering [51:07] with these systems and it was very [51:08] difficult technical challenge but they [51:11] thought I mean the the the standard [51:12] thinking in machine learning was oh [51:15] going too big is a problem you're just [51:17] going to overfit and that you're always [51:18] looking for the sweet spot where make it [51:21] bigger bigger it gets better uh but if [51:24] you make get even bigger, it starts to [51:25] get worse. It was called the overfitting [51:27] problem because they thought it would [51:28] these machines would memorize the things [51:31] that they were being tested on and [51:32] therefore become dumber. It didn't [51:34] happen. GPT3 was really good. It was [51:35] unusable for the average person because [51:37] it wasn't it would uh go off on weird [51:40] flights of fancy, but it did really well [51:42] uh for scientists. Then GPT35 was just [51:45] that tuned to behave well with humans. [51:46] They they tuned it to be good at [51:48] conversation. That really was a big [51:49] leap. GPT4 was an even bigger leap. They [51:51] made it uh 5x bigger. or we assume they [51:54] don't actually give the numbers but they [51:55] made it much bigger and they began to [51:57] realize that not only was GPT4 good at [51:59] language but it was good at other tasks [52:01] that in learning language it was doing [52:03] well on logic puzzles we could ask it [52:05] questions about board games we could [52:07] give it and that was the really exciting [52:09] moment where for researchers like oh if [52:11] we keep making language models bigger [52:13] they're going to learn lots of stuff [52:14] stuff that's not just language and [52:16] that's how they thought maybe we'll get [52:17] AGI uh that slowed down after GPT4 I I [52:21] wrote an article about this for the New [52:22] Yorker that came out last summer that [52:24] basically, if you really report on this, [52:26] what happened is they tried to make it [52:28] even bigger. This was at the time called [52:30] Project Orion at OpenAI. They said, [52:32] "Great, let's make it five or 10x bigger [52:34] than GPT4." And it didn't get better. [52:36] >> And this is the thing you if you run [52:38] this business, you want to believe. [52:40] Yeah. that like we can make the digital [52:43] god if only like we can spend enough [52:47] money because of course spending money [52:49] means that you're you need your business [52:51] to grow and you can get more investment [52:53] and you can build this thing that's [52:54] going to be so powerful that you kind of [52:57] get to be the god emperor of how [52:58] everything works and and you you you [53:01] know Sam Alman doesn't need to get paid [53:02] a bunch of money because he's going to [53:04] become one of the three most powerful [53:06] people in the world. [53:06] >> Yes. and huge. [53:08] >> I think it's very funny when Sam talks [53:10] about how he doesn't make a lot of money [53:11] and I'm like, what is money to you? [53:14] Like, you know that money isn't the [53:16] thing that doesn't matter in the future [53:18] you're trying to build, right? Like I I [53:20] get that. The fact that you're [53:22] pretending like you don't get that makes [53:24] me trust you less. [53:25] >> But so this happened though and they [53:27] realized it but were being kind of quiet [53:29] about it. They hit a wall in just plain [53:31] scaling when that project Orion uh and [53:34] then Meta had a similar issue. uh Gro [53:36] saw a similar issue like everyone was [53:37] seeing the same thing where oh we can't [53:39] just keep making things bigger we're not [53:41] going to be able to just pure scale our [53:43] way to AGI and then we got a whole year [53:46] it started with the 01 models and the [53:47] reasoning models and we got all of these [53:49] other much more complicated things [53:51] remember all this got so much more [53:52] complicated where what the models looked [53:55] like they were chemical equations for [53:57] organic chemistry class they were [53:58] talking about reasoning and and post you [54:01] know post inference computation all this [54:03] >> this was all just trying to now find [54:05] little narrow places where we could tune [54:07] to get some advantage or different types [54:09] of ways to get better performance out of [54:10] them on benchmarks. They it was actually [54:12] the beginning of the end of pure per [54:15] scaling. So, so the right [54:16] >> language models are sort of plateauing. [54:18] We're able to tune them to do certain [54:20] things better, but we're no longer on [54:21] that explosive trajectory. You're right. [54:23] To get something like AGI, we now need [54:26] to look forward to Oh, and then we have [54:28] another innovation with a different type [54:30] of complimentary technology that goes [54:32] really fast for two years and then maybe [54:34] that happens one more time. I happen to [54:36] think the way AGI is going to happen, I [54:38] call it distributed or slow takeoff AGI. [54:40] I don't think it's going to be here is a [54:42] single program that now no matter what [54:44] we ask it to do, it it does this at the [54:46] level of a human. I think what we're [54:48] going to see instead is it's going to be [54:49] more boring. lots of different systems [54:52] tuned to lots of different things that [54:54] humans do, very specialized that get [54:56] really good. And at some point, we're [54:58] going to look around and say there is a [55:00] huge swath of things that are hard that [55:02] humans do that all now have [55:03] individualized AI type systems that can [55:06] do that thing just as well or better [55:07] than humans. More of a peace mill. I [55:09] think that's much more likely [55:11] >> and much more uh cost-effective and [55:13] makes much more business sense than this [55:15] idea of we just need a a super AI that [55:17] does all of these things. And we already [55:19] see it now like, oh, we're better at [55:20] chess. We're better at poker. We're [55:22] better at go. We're getting better at [55:23] proteins. That's a, you know, it's more [55:25] peace. Each of these systems is going to [55:26] need its own mix of technologies and its [55:28] own secret sauce, but that's much more [55:30] likely to get us to a point where we [55:31] say, "Wait a second. Computers are doing [55:32] a lot of what we do better than us." I [55:34] think that's a much better more likely [55:35] way we're going to get there than a [55:37] singular machine doing it. [55:38] >> And yet it's not uh that that take isn't [55:42] going to get any headlines written about [55:44] it. Well, I mean, I like to tell the [55:46] story of um there's these cool bots that [55:49] came out. There's a particular [55:50] researcher who's done a lot of cool AI. [55:52] Uh and the two things I think he's at [55:54] OpenAI now, but but back at Carnegie [55:56] Melon and Meta, he built the first [55:59] system to beat human poker players, [56:01] professional poker players at Texas [56:03] called Plurabus. Um and he built the [56:05] first machine that could competitively [56:06] play diplomacy, which you know is like a [56:08] risk-like board game with lots of [56:09] interhuman negotiation. But the [56:12] storyline of both of those, it's very [56:14] clever combination of existing [56:16] technologies. So for the poker playing [56:18] bot, they actually tried at first to [56:20] just scale into a machine that work. [56:22] They just made they built really large [56:23] neural networks. They were training them [56:25] at the supercomputer center at the [56:26] University of Pittsburgh and they [56:28] weren't that good. And ultimately the [56:30] system that beat the professional [56:31] players could fit on a laptop. And it [56:33] was much more they had a very smart [56:34] combination of some neural network, some [56:36] future prediction and they kind of tuned [56:38] it up and it was better at poker. Uh the [56:41] diplomacy plane bot similar. There's a [56:43] language model in there. There's a [56:44] future simulator in there. There's like [56:46] a strategy uh simulator in there. [56:48] >> They like can they talk to each other in [56:50] a language that we can understand? [56:52] >> Well, they chat. So they play on the web [56:54] diplomacy server. So they they chat with [56:57] humans [56:58] >> and then the language. [56:59] >> Do they chat with each other? [57:01] >> Uh they could. The way the language [57:02] model works is it takes English and then [57:04] it converts it to a sort of XML style [57:07] like standardized markup language that [57:09] the rest of the programs understand. [57:11] >> Gotcha. [57:11] >> And they feed it on. But let me just [57:13] tell you, this will maybe make you feel [57:14] better. Here's what I really like about [57:15] that diplomacy playing bot. [57:17] >> Okay. [57:17] >> If you read the research paper around [57:19] it, they just casually say at some [57:21] point, oh, for ethical reasons, um, we [57:24] set it up so it'll never lie. Even [57:25] though that's like a major strategy in [57:27] diplomacy, you you you fake out people. [57:29] You make false alliances or whatever. [57:31] for ethical reasons. They said, "Oh, we [57:32] just set it up so it wouldn't lie." [57:34] >> You're like, "Well, wait a second. I [57:35] thought the whole issue with AI is we [57:36] can't align it. It's out of control." [57:39] >> But when you build these specialized AI, [57:42] it's not just some giant neural network. [57:43] You don't know how it works. It's six or [57:45] seven components. We know what they do. [57:47] And it was very taking concrete [57:49] information from here and then they are [57:51] putting it over here. And so what it [57:53] what you're saying is if it knows this [57:55] information, it can't be like I'm going [57:57] to tell the people a different thing. [57:59] >> Yeah. Well, there's literally a piece of [58:00] it that simulates possible future moves [58:03] and to try to measure like which of [58:04] these might be best. They wrote that [58:06] that simulator is not a neural network. [58:08] They just they wrote it and they just [58:09] coded it to say don't look at [58:11] possibilities where you lie. So, it [58:13] doesn't. [58:14] >> So, that machine can't lie. There's no [58:16] out of control piece to this. There's no [58:17] amorphous uh neural network. Yeah. [58:20] >> Well, we you should say like it's it's [58:21] don't it's not don't look at things [58:24] where you lie. It's like don't look at [58:26] things where x does not equal x. So like [58:28] it's a it's it's math that's happening. [58:30] It's a it's computer code. It's not like [58:32] when you say like don't look at things [58:34] when you lie. That that implies some [58:35] understanding of what a lie is. [58:37] >> Yeah. No, true enough. They have uh here [58:38] are the possible moves you can do and [58:40] they're all categorized and there's like [58:42] 50 moves you can do. Some of them are [58:44] lying and they just literally the for [58:47] loop that check tries all the things [58:49] doesn't look at those elements. Right. [58:51] >> Right. [58:51] >> Yeah. Um I call that intentional AI or I [58:54] AI. Like the idea is when you move away [58:56] from all we have is a language model, [58:57] which is what we're the we're trying to [58:59] do now, which I think is a real problem. [59:00] Let a inscrable language model be the [59:02] brain for everything. If you want to [59:04] build an agent to go do something, [59:05] >> ask a language model what to do and just [59:07] execute it blindly. [59:09] >> But I AI is like these game playing [59:11] bots, which I think are great examples [59:13] of what AI might be. And you have mult [59:15] you have a control module in these is [59:17] written by humans in normal computer [59:18] code. There's nothing unknown about it. [59:20] There's nothing that's learned. There's [59:21] nothing that evolves. and it can talk to [59:23] neural networks to understand things or [59:25] to get descriptions of things or to come [59:27] up with ideas, [59:28] >> but it's just a normal human control [59:30] program that can decide what do we want [59:32] to do? What are the options that we're [59:34] actually looking at? I so I actually [59:35] have I'm more optimistic about this [59:37] world of distributed AGI. I don't think [59:39] it's going to be out of control. I think [59:41] this alignment problem is really a [59:42] problem about language model text [59:45] production and being able to control [59:46] what text language models produce not [59:48] necessarily fundamental to AI powered [59:50] devices more generally. [59:52] >> Yeah. I I think that we have this [59:56] tendency as people one to predict [59:58] apocalypses like that's a known bias [60:01] that we have. We we all know that that [60:03] the the personal apocalypse is coming [60:05] for us in the form of death and so we [60:07] kind of like crave and art titilated by [60:09] the idea that that would happen to all [60:10] of us at the same time in some way [60:12] because of that. There's a lot of this [60:14] in science fiction and all of these [60:15] people have written and read science [60:17] fiction just like I have. We said uh [60:20] what if people fall in love with this [60:21] thing and that's so that's what we're [60:22] going to imagine is going to be the [60:23] thing that happens. But nobody was like, [60:25] I wonder if uh sycophancy induced AI [60:29] psychosis is going to be a thing because [60:31] nobody had any idea that that was going [60:32] to be a thing. That came out of nowhere. [60:34] And so that's going to be the hard [60:35] thing. Just like, you know, like rage [60:37] bait, no one like assumed rage bait was [60:40] going to happen when Facebook invented [60:42] the newsfeed, you know, like that. That [60:44] wasn't a thing that was on anybody's [60:45] radar. And it it's a it's a third fourth [60:48] order effect, but of course you end up [60:51] with it and then you have to deal with [60:52] it as it comes. And so it does feel like [60:54] what we should probably be doing and I'm [60:56] having this sensation more and more [60:58] lately. We should be dealing with the [61:00] problems that we have. [61:01] >> Yes, [61:01] >> we should be looking at the problems as [61:03] they currently exist. And what if I want [61:05] to say to some people who might want to [61:07] be effective altruists, what if dealing [61:10] with the problems that we have right now [61:12] is the thing that makes the future [61:13] better? Like what if guessing at the f [61:15] what the future needs is actually [61:17] actually makes the future worse and [61:19] dealing with the problems we have right [61:20] now is what makes the future better? [61:22] Like just what if and if so if that's a [61:25] is if that's like one of the things that [61:26] might be the thing that makes the future [61:28] better. It has the bonus of making the [61:31] present better for sure because like we [61:34] can actually measure that outcome as it [61:37] occurs in the world that we currently [61:39] inhabit. [61:39] >> Oh, I'm so aligned with you on that. I'm [61:41] I'm very much aligned with you on that. [61:42] I mean it's really my whole philosophy [61:44] of of technology criticism. I call it [61:46] technoslectionism. And the idea is when [61:49] you introduce like new technologies that [61:51] are powerful into the technoeconomic [61:53] cultural [61:55] uh context. They always have these [61:57] second and third order expects that we [61:58] don't predict. And therefore we can't [62:01] approach technology with this weird [62:03] fatalistic mindset of just like well [62:05] once something is invented it's just [62:07] here. What are you going to do? Kids [62:09] these days they're going to look at [62:10] their phone and like just kind of move [62:11] on with things. We need to be much more [62:13] have much more agency [62:14] >> in we can humans [62:18] >> including and this surprises people [62:20] saying I was interested in this [62:22] technology it was cool and then it got [62:24] bad so I changed my mind and stopped [62:27] using it like we are allowed to do that [62:29] and I think we need to do a lot more of [62:31] it I mean all all of our problems right [62:33] now seem to come out of these second [62:34] order effect stories right like I I've [62:36] documented a lot of these you think [62:37] about uh Facebook's going along early [62:40] Facebook you know an engineer invents [62:42] the like button for like a really [62:43] engineery reason which was it's [62:45] inefficient for lots of people to write [62:47] the same comment if you post like you [62:49] with your diploma and everyone says [62:50] congrats that's like that's that's going [62:53] to swamp out the interesting comment so [62:54] like oh let's have a like button for [62:56] that so that the comments be more [62:57] interesting [62:58] >> and then they realized oh that gave an [63:00] unstable social approval indicator which [63:02] meant that usage time on the app went up [63:04] by a huge factor because now my brain [63:06] was did I get likes on what I just [63:08] posted and like oh my god [63:10] >> engagement is all about and we get the [63:11] attention engagement age which was not [63:13] at all what we thought about when [63:14] Facebook first came out. So this type of [63:15] thing happens all the time and I'm with [63:17] you on this is what we need to be doing [63:18] is diligently policing the second order [63:21] negative externalities of the AI we have [63:23] right now. But think about the interest [63:26] in some of the people who are pushing [63:28] the apocalyptic or the future stories, [63:30] right? Like if you are Daario Amade or [63:33] Sam Alman, right? It's completely [63:36] aligned with your interest to talk about [63:38] the automation of most jobs or to talk [63:41] about, you know, super intelligence [63:42] coming because it does three things. [63:44] One, [63:45] >> it distracts attention or regulatory or [63:48] uh public, you know, disapproval about [63:50] the things that your bad things that [63:51] your technology is doing right now. Um, [63:54] two, it's going to induce a lot of [63:55] perhaps ownorous regulation that you as [63:57] a very large company can weather but [63:59] small startups can't. So, you get [64:00] regulatory capture and you get cemented [64:02] in your position. And three, it attracts [64:04] investment because yeah, it's scary to [64:06] think about AI automating half a jobs, [64:08] but my god, if I'm an investor, I better [64:10] then have my money in the company that's [64:12] going to do half of the jobs. That's [64:13] going to be the biggest thing in the [64:14] economy. So, I think it's completely in [64:16] their interest to talk about the future. [64:19] And I think the effect of altruist like [64:21] I really think what happened there, I've [64:22] been really trying to understand this, [64:23] right? the that they've really been [64:25] pushing the super intelligence story. [64:26] The standard thing, the existential risk [64:28] thinking they've been doing, which is [64:29] like a standard effect of altruist thing [64:31] where you say, let's look at unlikely [64:34] risk that would have a really negative [64:35] impact because if you calculate the [64:37] expected value, it's worth spending time [64:39] now to prevent it to prevent a potential [64:41] small probability future, right? So, [64:43] yeah, [64:43] >> this is like Nick Boston's Institute. [64:44] We'll look at things like asteroid hits. [64:46] Probably won't happen, but it'd be [64:47] really bad if it did. So, like we have a [64:49] positive expected value thinking about [64:51] asteroid hits now. and uh super [64:52] intelligence was just one of the many [64:53] topics they were looking at abstractly. [64:56] I think when chat GPT came along, it [64:58] sort of warped the brain of some of the [65:01] people in the F community that were [65:02] working on uh that were working on [65:04] existential risk and it warped it from [65:07] here's this thought experiment that we [65:08] should be thinking about to well what if [65:10] it was actually happening and if it was [65:12] actually happening we're the most [65:14] important people on Earth. we would be [65:16] the ones who found the asteroid first [65:17] and told people about it. And I think [65:19] that was such a compelling possibility [65:22] that the switch from here's an [65:24] abstraction that we should be thinking [65:26] about as a thought experiment to this is [65:28] happening. It just that switch went and [65:31] they had a hard time getting out of it. [65:32] And I I absolutely think that they [65:34] believe this and like I also think that [65:38] there is a chance that it will go very [65:40] badly because you know social media has [65:43] gone very badly because because content [65:46] uh recommendation algorithms have gone [65:48] very badly and like all they do they're [65:50] very simple as you said like they all [65:52] they do is they show you what you're [65:53] more likely to watch. That's the whole [65:55] thing and and that went very badly. It [65:58] went bad for human attention. It went [66:00] bad for society. It went bad for [66:02] politics. It went bad for like trust, [66:04] which I wouldn't have expected. Like we [66:06] live in a lower trust world because of [66:09] the effects of content recommendation [66:11] algorithms. And I have friends whose job [66:14] is to build them. And they're they're so [66:16] like disappointed that I feel that way. [66:18] I think that it's true. I think that [66:20] there are some that are better than [66:21] others. I think that there are some that [66:23] are way worse than others. So I'm not [66:25] saying that they all have the have equal [66:27] responsibility here. I do think that [66:29] YouTube does a better job of than a lot [66:31] of platforms. Um, but also I think it's [66:33] structured better than a lot of [66:34] platforms because there's more human [66:36] agency built into you YouTube, but they [66:38] had no problem launching products that [66:40] have less agency involved in them, which [66:43] of course they they have done. I do [66:44] think that that there is a thing there [66:46] and I I don't think that we should pay [66:47] no attention to it. Like we have no idea [66:51] how to have an impact on the future and [66:53] we certainly have no idea what the [66:54] future's going to look like. And also, [66:56] no one's ever been good at figuring out [66:58] the actual effect of a communicationist [67:00] breakthrough. Like, we're okay with [67:02] weapons. We get we can kind of get that, [67:05] >> but communication is just too important [67:07] and big and fundamental to humanity. [67:10] >> We have a hard time going back [67:12] >> and corraling those horses when they're [67:14] out of the gate. In the way, weapons is [67:17] right. Um, I think like substances too, [67:19] like we're much better at that. Like, [67:20] oh, like cigarettes are invented, like [67:22] this is fun. Like, actually, this is [67:24] probably not good. kids shouldn't use [67:25] these, let's ban them for kids, right? [67:27] Or gambling was around and they're like, [67:29] well, we should probably have rules [67:30] around this because we're seeing, but [67:32] you're right, information is different. [67:34] I think we need to be asking fundamental [67:36] questions right now. For example, [67:37] looking at chat bots and saying, who [67:38] asked for this? Who was clamoring for [67:40] this? What I mean, other than like the [67:42] investors in open AI, like what what [67:44] problem are you solving? And don't let [67:46] the programmers answer because the [67:48] programmers have they're extrapolating [67:50] their experience to the rest of people. [67:52] But it's not there's a great use case if [67:54] you're using cloud code. [67:56] >> That's a great use case. You can't [67:57] extrapolate that to the rest of the rest [67:59] of people. Think about our relationship [68:01] with these things. It's unsettling. It's [68:03] like some small benefits with a lot of [68:05] like distraction and then some pure [68:07] moments of like this is weird or or [68:09] unsettling. Yeah. Like these questions [68:11] should be asked. I have a problem with [68:13] who asked for this because like no one [68:15] asked for anything like technology like [68:18] what is possible is very different from [68:20] what I want you know like I want a [68:24] cheesecake that is as healthy as a [68:26] carrot and yet I will I will never get [68:28] one. So like I never asked for a [68:30] cheesecake but like it was presented to [68:32] me and I'm like yeah cheesecake. I I [68:35] would prefer a world that did not have [68:37] sugary drinks. You know that's a better [68:39] world. No one asked for Coca-Cola, but [68:42] we have Coca-Cola and we live in a world [68:44] with it and we choose it sometimes and [68:46] we should not choose it as often as we [68:48] do. And and so in the same way, I'm [68:50] like, okay, like you present us with [68:52] this tool and a lot of people use it and [68:54] a lot of people find value in it. There [68:56] are there are certainly valuable use [68:58] cases for large language models, [69:00] especially the way that they have been [69:02] constructed to become more useful for [69:04] research or for, you know, the dumb dumb [69:06] little things you got to do. What I [69:08] think is an interesting question and it [69:10] leads to a question I really want to get [69:11] to is okay, no one asked for it to be [69:15] basically free when it costs a trillion [69:18] dollars. That's actually weird. Like I [69:22] buy Coca-Cola, Coca-Cola makes money. [69:24] When Open AI sells me Chachi PT, they [69:28] lose money. So, like this is kind of the [69:30] weird perverse incentive that we have [69:32] right now where they are they are [69:33] shoveling a product at us and they're [69:36] trying to make it more and more useful [69:37] and also to be clear all I think more [69:39] and more addictive and they are doing it [69:41] at a loss and this perplexes me. Does [69:44] that make sense? [69:44] >> It does make sense. It doesn't perplex [69:46] me as much. I think it's a standard. [69:49] >> Right. I I understand why I understand [69:50] why you're doing it but but I it's how I [69:53] feel about when you say like nobody [69:55] asked for this. I'm like, "No, but like [69:57] I I also feel like something actually [70:00] manipulative is going on here, which is [70:02] like get them hooked and figure out [70:03] where the the addiction levers are." I I [70:06] mean, I think that's I think that's one [70:07] of the things they're doing. Yeah, I [70:08] agree with that. I think there's a [70:10] there's more generally a desperation [70:11] play of try a bunch of things to see if [70:14] one of these can actually be the revenue [70:16] generator. But I would also say when I [70:18] say no one asked for this, the the [70:20] specific thing that's kind of being [70:22] shoved down our throats is the just a [70:25] general purpose conversational partner [70:27] like actually the the obvious useful [70:30] uses of this technology has to do with [70:32] like solving particular problems. [70:33] Wherever you get natural language [70:34] interface in the computation like that [70:36] can be useful, right? Like I get it. Um [70:39] if I want summaries of text, like that's [70:42] really useful. If I want to be able to [70:43] tell Microsoft Excel what I want it to [70:45] do and not have to learn what all those [70:47] menus mean, like that would be like a [70:49] really, you know, that's really useful. [70:50] It's like an obvious use case. But if [70:52] you came to me ahead of time and say you [70:54] can like chat with a machine and it'll [70:56] kind of like be a friend and kind of do [70:57] all these different things. It's a weird [70:59] use case. It's not one like a lot of the [71:01] people understood the use case for Coke. [71:03] I like good tasting soda, but I don't I [71:05] feel like the chatbot interface is it's [71:08] weird and dark and people are but but I [71:10] do think you're I do think you're right. [71:12] the they're searching for what's going [71:14] to click. That makes sense. And the [71:16] question I always had is like, well, the [71:17] obvious things that would click would be [71:19] natural language interfaces in the [71:20] software. Make everyone five times [71:22] better at using tools, right? Like, I'm [71:24] not great at Microsoft Excel. Why not [71:25] give me a natural language interface [71:27] where I could say, "Hey, go through I [71:29] don't know what commands to do, but get [71:30] rid of all the rows where the dollar [71:32] amount in column B is above $5 and then [71:34] sort it on this column." [71:36] >> But what I've heard from my sources is [71:37] that would be the obvious way. It'd make [71:39] all your software more valuable, right? [71:41] it's not working. Actually, the [71:43] hallucination rate of LLMs means if [71:46] you're talking to an LLM and then it's [71:48] translating that in the commands that'll [71:49] be executed by the software, it makes [71:51] too many mistakes. And so, [71:54] >> this and using tools like that like like [71:56] uh Air Table has a tool like this and [71:58] I'm like, "Yeah, build me a little [72:00] workflow." And then it's like, "Oh, now [72:02] I have to do a bunch of work fixing your [72:03] weird mistakes." And I think it's a [72:05] really interesting point is that also [72:07] and not to go on a brief tangent, but [72:09] when we're chatting with these chat [72:10] bots, we underestimate the degree to [72:12] which we're constantly sort of [72:14] interpreting and correcting and fixing [72:16] and asking a question different way. And [72:18] so like a lot of these other obvious use [72:20] cases of like great, we can now have [72:22] this be an intermediary between humans [72:24] and the whole world of computation, [72:25] which I think is a very exciting [72:27] proposition. They're struggling with [72:29] that one. So now they have to look at [72:30] more, I think, addiction based ones. [72:32] This is why they they like companions [72:34] because what's more engaging than a a [72:36] friend that you going to talk to all day [72:38] long? That seems good for them. More [72:39] erotica in chatt that tends to be [72:42] something that captures people's [72:43] attention. Uh Sora and Sora 2 made no [72:46] sense to me. The economics don't make [72:47] sense. Those videos are way too [72:48] expensive. [72:49] >> That's very dumb. [72:50] >> Yeah, that's kind of But I get they're [72:51] they're trying to find ways because the [72:53] breadand butter stuff that actually [72:55] would be like, oh, it's the new mouse, [72:57] >> which would be a big deal. I think [72:58] that's not quite working as well as [72:59] they'd hope. that's not working as well [73:01] as they'd hoped. There is potentially [73:03] some kind of like as you say with the [73:05] the systems, what was it? I AI. [73:09] >> Yeah. [73:09] >> Where they are more specialized. If it [73:11] turns out that that that all the [73:13] valuation is based on actually only [73:16] giant companies can do this, so you have [73:18] to buy into one of the big three. That [73:20] still leads us to a world where the [73:22] bubble pops. Yes. [73:24] >> Yeah. This the spending doesn't make [73:26] sense. It just doesn't make sense. and [73:28] and the spending is largely being done [73:29] by a small number of companies right [73:30] now. It just doesn't make sense. [73:32] >> It would be very unlikely, it seems to [73:34] me, that the that this IAI thing happens [73:37] fast enough to fill in the value gap [73:39] that has been created. [73:41] >> Well, I I just don't think it requires [73:42] the level of investment that Microsoft [73:46] and uh OpenAI and Anthropic like the [73:48] level of investment these companies are [73:50] making. You don't need it for something [73:51] like that future of specialized [73:53] machines. and the the Plurabis which can [73:55] beat in a tournament with real money on [73:57] the line real Texas Holden players can [73:58] run on a laptop and it uses some really [74:00] smart neural networks and some other [74:01] really smart things. Uh yeah, so I think [74:04] the loveless spending now doesn't make [74:06] sense. It's bubble-l like it's just it's [74:08] too much spending for the potential [74:10] revenue that you can concentrate on the [74:11] other side. It's just too expensive the [74:14] comput look at a company like Google, [74:16] right? like their big rise in the the [74:18] early 2000s that the whole key to their [74:20] rise was the cost effectiveness of their [74:22] infrastructure. Like that was the whole [74:24] ballgame there was [74:26] >> simple dumb hardware, smart software on [74:29] top of it and give you very powerful [74:31] systems without having to buy a bunch of [74:32] supercomputers and the the cost the [74:34] compute cost was really low and then [74:36] they could print money. We're doing the [74:38] opposite right now. The compute cost is [74:40] astronomically high. we're we're that to [74:43] do inference on these models right now [74:45] um is like the most expensive [74:47] computational thing we can do. So it [74:49] really is I don't know how [74:50] >> inference is the thing where I where I [74:52] ask it a question and it does something [74:54] >> where it calculates every time someone [74:56] interfaces with it that's expensive not [74:59] just the training part I feel like we'll [75:01] figure out ways to make that cheaper [75:03] >> you have to make the model smaller I [75:04] mean it's it's simple what's happening [75:06] models they look complicated when you [75:08] see the layered structure right but [75:11] underneath it all what you're doing is [75:13] multiplying tables of numbers again and [75:15] again and again and again right and the [75:17] number of numbers in those tables is rel [75:19] relative to the size of the model and [75:21] all GPUs do because GPUs are made to do [75:24] 3D graphics for video games. And the way [75:26] you describe 3D objects in video games [75:28] is with tables of numbers. [75:30] >> And the way you rotate a 3D object is [75:33] you multiply that table of numbers by [75:35] like a rotation vector and it rotates a [75:36] little bit. So GPUs were just machines [75:39] >> made to multiply tables of numbers or [75:41] and tables by rows and numbers as fast [75:43] and specialized as possible. That turned [75:46] out to be fortuitously for the A [75:47] industry exactly what you need to do to [75:49] have a uh an input go into a large [75:51] language model and a token come out the [75:53] other side. So they're like perfectly [75:54] well suited for it. But that means to [75:55] generate even that single token you have [75:58] to multiply something by all of those [76:00] parameters that define this really large [76:02] it could be hundreds of billions of [76:04] parameters and you have to do all of [76:06] those multiplications [76:08] token. Yeah. [76:09] >> Yeah. And there isn't a way to not not [76:11] >> you got to just make the number of [76:12] things you're multiplying smaller, which [76:14] means smaller models, which again comes [76:15] back to my theory of like the only thing [76:17] that makes economic sense is smaller [76:19] models and bespoke systems that do [76:20] really cool things. [76:21] >> Gotcha. So So you don't think that uh my [76:24] bubble worry is unfounded? [76:26] >> No, it's I I don't know. I don't know [76:28] what the popping's going to look like, [76:30] >> right? Like maybe. [76:30] >> Yeah. And like what I what I do know is [76:32] that there will be like lots of large [76:34] rooms full of GPUs that will be used for [76:36] something like that that will be they [76:38] will try and figure out economic utility [76:40] for this stuff. [76:41] >> I think that's true and and we might [76:42] find more GPUs just distributed in every [76:44] day because the thing that makes the [76:45] most cost effective sense is trying to [76:47] get things on device. Nvidia is trying [76:49] to do this in a sort of a clunky way [76:50] right now but I think in the future you [76:51] know it'll be I have a lot of things [76:53] that have GPUs in them and that [76:55] inference is happening here. or I'm [76:56] paying for electricity [76:57] >> because it's happening on my machine and [76:59] my at my house that makes more. So there [77:01] might be a huge market. So like I'm I'm [77:03] not a anti- Nvidia person, but I think [77:06] the amount of money being spent by the [77:07] the fang companies otherwise is it it [77:10] seems unsustainable, [77:11] >> right? [77:11] >> Best case scenario it'll just be large [77:13] cuts, right? Meta is going to have to [77:15] cut this many people. Open AAI might go [77:17] away. Anthropic is going to have to like [77:18] go down to a third of its size, you It [77:20] might be something like [77:20] >> which would be very that would be a very [77:23] like a it feels somewhat [77:26] like it would be a big problem for at [77:29] least the stock market. [77:30] >> Yeah. [77:30] >> Uh which usually leads to a problem in [77:32] the economy especially when most of the [77:34] money in the economy is spent by rich [77:36] people who are feeling flush right now [77:39] because [77:40] >> portfolios have gone up. Yeah. [77:42] >> Maybe this is why Mark Zuckerberg has [77:43] bought half of Hawaii. He needs a quiet [77:46] place to go when people get upset. He's [77:49] he's he's definitely seen that we we [77:51] don't love him. I think he's noticed. I [77:53] have two final questions. Um number one, [77:56] I saved it for the end. What did I what [77:58] do you think I did worst on? [77:59] >> Oh, that was a good video. You good? [78:03] What What do you wor Okay, here's an [78:05] effect that I think your video creates [78:07] that a lot of discussions of this [78:08] create, which I I think of as like Doom [78:11] amplification loop. [78:12] So because it's such a diverse [78:14] technology with so many possible things [78:16] that could be a problem that as you go [78:18] through them one after another after [78:19] another after another we have a natural [78:21] human tendency to put that into an [78:23] amplification loop and sort of imagine [78:26] >> the aggregate of all of that doom [78:28] because if you if you isolate any one of [78:29] those issues you'll be like ah this is a [78:31] problem like that's annoying like we [78:32] should do something about it but it's [78:34] not going to have me up at night but [78:35] when you hear 20 different issues it [78:37] seems like it's 20 times as bad but the [78:38] thing is there's tons of technologies [78:40] out there creating their own issues of [78:42] roughly that size. We just don't [78:44] aggregate them. We're just used to [78:45] there's [78:46] >> a lot of different things out there that [78:47] are causing trouble. So, that's the one [78:50] thing I worry about is keeping our [78:52] mindset of there's lots of issues here [78:54] that are serious, [78:55] >> but we don't have to add them up into [78:58] therefore an issue that's so big that [79:00] we'll never be able to handle it. We can [79:01] handle lots of small issues. Uh it's not [79:04] the same as this is the biggest issue [79:07] we've ever faced and we're not going to [79:08] be able to handle it. when we break this [79:09] up, when we recognize these are [79:10] individual issues. They're often dealing [79:12] with completely separate technologies, [79:14] >> Mhm. [79:15] >> it seems more tractable. So, I don't [79:17] want people to fall I just get a lot of [79:18] this when I talk to people right now is [79:19] you fall into despair because it's [79:21] complicated and there's so much going on [79:22] and so many possible things. But like [79:24] humans also have a pretty good t you [79:26] know capacity to adjust to things to [79:28] push back on things to to eventually say [79:30] enough of this like we don't want to use [79:32] this anymore. I think that's finally [79:33] starting to happen for social media. [79:35] It's taking five years longer than I [79:36] would have hoped, but like [79:37] >> we're getting fed up, you know? I think [79:40] so. I I I have some [79:42] >> I have optimism and I don't want the I [79:44] don't want to seem if we list all the [79:46] dooms out in a row, I don't want it to [79:48] seem like [79:48] >> these are all adding up to one villain [79:52] that's going to be impossible to slay, [79:54] >> right? A and if you listed out all the [79:56] ways that AI has like helped you know [79:59] somebody do some thing or even like [80:01] things that are relatively large like [80:03] coming up with new cancer treatment [80:05] combinations or protein folding on its [80:08] own like is is massive and will will be [80:12] a you know worldchanging [80:16] innovation. And it just like sort of [80:18] takes time for these things to filter [80:20] through the the technology loop. And [80:21] nobody wants to hear the video that's [80:23] like, "Hey, look at all the things that [80:25] are that like have come out of this [80:27] technology." [80:27] >> Well, and that's the that's the other [80:28] point I want to make there too is we we [80:30] have to be careful about connecting [80:31] unrelated AI technologies as well and [80:33] allowing advances and want to reflect on [80:35] others. So like [80:35] >> yeah, [80:36] >> cancer diagnosis, protein folding, these [80:39] are completely different technologies [80:40] than language models, right? But [80:42] >> I was just I just also had a [80:43] conversation with an oncologist today [80:44] and she's like a while I do need you to [80:47] know that I don't mean chat bots. [80:49] >> Yeah. Because I think what happens then [80:51] is you get these extrapolations where [80:52] because like oh chat bots got much [80:54] better therefore cancer will be a thing [80:57] of the past. Like no that's a different [80:58] AI or like we're going to have like [81:00] robots are just going to do all of our [81:02] you know manual labor. Well that's a [81:04] completely different type of AI with its [81:06] own challenges or right [81:07] >> self-driving cars. like that's yes, but [81:10] that's a very slow that's its own [81:12] technology and it's a very long [81:13] trajectory that we're on and it's [81:15] there's there's a lot of um progress [81:18] laundering I sometimes call it where [81:20] something within this broad field makes [81:21] a leap and we launder that progress over [81:24] to sort of other technologies in the [81:26] field and say well they also have or [81:27] will be making the same type of leap but [81:29] they're completely different [81:30] technologies and so like chatbots [81:31] getting better [81:32] >> there are also like there's there's like [81:33] legitimately research that's been done [81:35] by kind of leveraging the gains of chat [81:38] bots Yeah, that's true too. Fair enough. [81:39] Yeah. [81:40] >> But with cancer research, what they did [81:42] is they kind of leveraged an LLM but [81:45] gave it like a different language and [81:46] that language is the language of the [81:48] cell. [81:48] >> Yeah. [81:49] >> And how how cells communicate with each [81:50] other and that's likeing genius. I like [81:53] I love that. That's very cool and I want [81:55] to make a video about it. But [81:56] >> it's very cool. And they train them [81:57] completely differently and it but it's [81:59] like really cool. Yeah. I mean and [82:01] >> that excitement of one technology can [82:03] also lead to excitement in the related [82:05] technologies which which I which I Yeah. [82:08] which I think is good, right? Like, oh, [82:09] we're reinvigorated in our AI cancer [82:11] research because there's energy in AI in [82:14] general. So, [82:15] >> do you feel some amount of like tech joy [82:17] about this? Like, holy crap, I'm kind of [82:19] alive for a really weird moment. [82:21] >> I do. I mean, I I I especially did like [82:23] through GPT4 because that really was, as [82:27] you said, uh a huge step function leap [82:30] >> because most people hadn't seen GPT3. [82:32] you sort of had to have a permission to [82:33] access the API [82:35] >> and that was like wow [82:36] >> this is like a really big lead. I [82:39] actually had some disappointment on the [82:41] other side of it. I thought uh we were [82:44] going to have natural language [82:45] interfacing and computing like we're in [82:46] that era now which would have been like [82:48] a fantastic era for a lot of reasons. I [82:50] think productivity in the technical [82:52] sense would have jumped up if you make [82:54] everyone experts at all software. Like I [82:58] was really excited about that and was [82:59] like a little bit disappointed. I was [83:01] like, wait, why aren't we seeing these [83:03] tools being added to all this software? [83:06] Why instead are we just seeing uh [83:07] benchmark charts? Like, good. [83:09] >> I think that we will see it. And I the [83:11] reason why I think that we will see it [83:12] is because it's so good at code. [83:14] >> Yeah. [83:14] >> But like it's weird and true that it [83:17] turns out that like code is because it's [83:19] language adjacent. These systems are [83:21] really good at code, but they're not [83:22] really good at like clicking on [83:24] something, which is weird. [83:26] >> Y, [83:27] >> you know, that's not what you would [83:28] expect. But it turns out that like we're [83:30] bet we're good at things we didn't [83:31] realize we were good at. [83:33] >> And and that's I just wrote an article [83:34] on this. This was a whole story where we [83:36] were tricked by coding agents being so [83:38] good relatively speaking that at the [83:40] beginning of 2025 [83:42] >> the word was this will be the year of [83:44] the agent. We will be assigning work and [83:46] it didn't happen because again like as [83:49] you said moving the mouse is hard and [83:51] code is like really well suited. Um we [83:53] also don't realize that those coding [83:54] agents are working entirely in a [83:55] terminal. So it's entirely textbased. [83:57] They're in a world of textbased commands [83:59] producing structured text. It's like uh [84:01] Disneyland for language models. It's [84:02] like the best possible scenario, right? [84:04] But you could imagine [84:07] an Excel that that exists inside of that [84:10] world like co like you know Excel is [84:12] code. It's just that it's it's like we [84:15] thought that we could skip over the part [84:17] where we built Excel for us. [84:19] >> Yeah. [84:20] >> What we have to do is like we need to [84:21] build Excel for them, which sounds like [84:23] I [84:23] >> Yeah. And I'll tell you, here's my tech [84:25] joy. I I was talking to um someone in [84:27] this coding agent space and they had [84:29] this really cool idea like yeah here's [84:30] here's what's going to here's the way to [84:32] think about it is actually everything's [84:34] a program. So if I'm like asking Excel [84:37] to do something the the the approximate [84:40] way to do that is like to have a [84:42] language model produce a bunch of um [84:44] commands of the type we would give [84:45] Excel. But like actually what should [84:47] happen is Excel should have an API [84:49] opened up a textbased API and when I say [84:52] here's what I want to do the language [84:54] model creates a computer program to do [84:56] that because really good at creating [84:58] computer programs and that computer [84:59] program talks to Excel and like actually [85:03] code is going to become the universal [85:04] language of computing. We say things in [85:06] English, it gets turned into a computer [85:08] program that then can access other [85:10] computer programs through well-known [85:12] interfaces. That that presses my [85:14] computer science uh technical joy [85:17] button. That's so smart. Yeah. [85:19] >> Yeah. Like if you have prolevel CHPT it [85:22] can already it just does that like it's [85:24] built in like in certain situations it [85:26] just starts doing code and you're like [85:28] what are you what's happening? Yeah, [85:31] for a computer scientist, you're like, [85:32] "Ah, [85:33] >> that's my that's the world I imagine." [85:35] Like, you just think in code. It makes [85:37] so much sense. Yeah. [85:38] >> Um I and I think so. Anyways, I'm [85:40] excited about that. I think it's okay. [85:41] And that does actually have big [85:43] ramifications and it's going to change a [85:45] lot of things and it could have economic [85:46] impacts. Uh but that's a good but I [85:49] think ultimately that would be a good [85:50] thing. But it yeah, it's different than [85:51] the conversation we're having right now. [85:53] It sound it also sounds to me like if [85:55] this is the world that we end up in it [85:57] it I one of my big fears really is like [86:00] what do people do uh when they aren't [86:02] that good at stuff yet? So like what [86:04] does a college graduate do in the world? [86:06] If AI is is better than someone who's [86:10] bad at it but not as good as someone [86:12] who's good at it, how does [86:13] apprenticeship work in that world in [86:16] knowledge work? You know, apprentichip [86:17] will still be a thing in plumbing. And [86:18] it sounds a little bit like actually [86:20] there is a lot of space for human labor [86:22] in this version of reality. [86:24] >> Yeah. Because I think most things at [86:25] least right now think about that reality [86:28] aren't replaced by AI. The industry that [86:30] is in trouble I do think is is [86:32] developing computer programming because [86:34] I was talking to some developers [86:35] recently and the thing they were telling [86:37] me that's their words not mine. They [86:39] said here's what you have to understand [86:40] about the software development world. [86:42] Most programmers are pretty bad like [86:44] just in general like uh not we we think [86:46] about the really fancy 10x programmers [86:49] who are like the hot shots at you know [86:51] meta or whatever because I was I was [86:53] making this argument of like I don't [86:55] like this new model of you have multiple [86:57] tabs each with their own coding agent [87:00] all you're doing is supervising the tabs [87:01] I was like this feels like the skilling [87:02] this feels like be a little bit wary you [87:04] you don't want to lose what you do best [87:06] to create expert elegant code and some [87:07] of the developers wrote back like [87:09] actually a lot of people are just sort [87:10] of copying and past there's a lot of [87:12] programming jobs out there that are just [87:14] people are producing bad code. [87:16] >> Well, agents are good at producing [87:18] >> bad code, right? Like that that is that [87:20] is in danger. [87:21] >> So, last question. We've already talked [87:22] a little bit about this, but when we [87:24] comes to the the the problems that we do [87:25] see, and we're all going to place [87:27] different levels of severity on them [87:29] for, you know, various reasons, but when [87:31] we come to the problems that are will be [87:34] presented to us by any new technology, [87:36] but this one in particular, do you feel [87:38] like we have agency here? like what what [87:40] levers are there? You know, identifying [87:43] a problem sucks, but feeling like you [87:45] have some kind of can have like an [87:47] impact on this. [87:48] >> I think there's a there's a few [87:49] different levers we do have, right? [87:51] Okay. So, first of all, I think there's [87:52] a sort of personal choice lever where I [87:54] think it's completely fine to say it is [87:56] not my job to figure out why your [87:58] product is useful. It's your job to [87:59] convince me that it's useful. Don't tell [88:01] me like you better know AI and then not [88:03] tell me what that means. It's like if I [88:05] don't have a clear use case for using [88:06] something, I'm not going to go and spend [88:07] my time and money and attention trying [88:09] to master your tool. And if I find your [88:10] tool distasteful, um, I'm not going to [88:12] use it. I think parents and teachers [88:14] should have a rule that uh, kids should [88:16] never be using a chatbot unsupervised. I [88:19] I think that's completely dangerous. [88:20] There's, you know, we do not I do not [88:22] buy the argument that like kids need to [88:24] know how to use the tools of the future. [88:25] >> Like I think that there has to be like [88:26] some kind of cultural language norms [88:29] taboos around this. Yeah. Uh, and so we [88:31] should start developing those. Well, and [88:33] that that could be parental supervision [88:34] early on is you're seeing examples. I'm [88:36] only using it, but also it's kind of a [88:38] waste of time if you're talking about [88:39] like a nine-year-old. [88:40] >> You you don't need to expose them like [88:42] they need to know how to use chatbots [88:43] well because the the the form factor in [88:45] which we're interacting with Generative [88:46] AI six years from now is not going to be [88:49] it's not going to be the same, right? [88:49] It's not they don't need to learn that. [88:51] That's the Formula One car they're never [88:52] going to they're never going to drive. [88:54] And so, I think it's completely it [88:55] should be normative to be like, "Yeah, [88:56] kids shouldn't be on this." Um, I think [88:57] it's educational institutions should be [88:59] free to say the whole point of this is [89:01] to create friction in your mind so that [89:03] the this the connections get stronger. [89:05] You know, it's like this is a gym. Yes, [89:07] it's true. A forklift can lift the [89:09] weights for you and that would be [89:11] easier, but the whole point of this gym [89:12] is to lift the heavy weights. Like so [89:13] that that's not what we're doing here. [89:14] We don't have to bend over backwards to [89:17] try to like rechange our, you know, all [89:20] everything about learning to to sort of [89:22] in include this tool in sort of a forced [89:23] way. I I also think there's economic [89:25] levers we can pull which again is like [89:27] I'm not gonna waste time on something [89:29] until there's I'm like we're done with [89:31] this sort of Apple induced you know [89:33] early 2000s era isn't it exciting to be [89:36] a first mover on new technology and try [89:38] to figure out how it's going to be [89:39] useful. We could just be done with that. [89:41] I'm not going to spend money on this [89:42] until I have like a really compelling [89:43] use case. And I think there's regulatory [89:45] levers like I that the biggest lever we [89:48] could push like the Congress to pull [89:50] would be one where we have uh much more [89:53] liability for what a chatbot produces [89:56] like no you can't just have an entity [89:58] that'll say anything and do psychosis [90:01] you know give out really danger no uh [90:03] you are responsible for what this thing [90:05] says and what happens what the effect of [90:06] that would be is like oh I guess we [90:08] probably can't have a product that's [90:10] just trying to be a general [90:11] conversational partner and that would [90:12] probably be better, [90:13] >> right? So, [90:14] >> yeah. What Sam Alma would say is like, [90:16] well, then there won't be chat bots. [90:17] Yeah. But what you don't you don't you [90:19] think that there would be a way to [90:21] actually do the research and and create [90:23] features that would allow these things [90:24] to [90:25] >> Yeah. I don't think I think so. I think [90:27] if we said, look, uh, yeah, you're [90:29] responsible, you know, X Y and Z, we [90:31] again, we go to the specialized future, [90:33] which is like where we're going to end [90:34] up anyways. Again, the oracular model of [90:36] like everything has to why do I need to [90:38] use the same tool for summarizing emails [90:42] needs to also be a tool in which I can [90:44] have a sort of like pedophilic erotica [90:47] relationship with a virtual being? Like, [90:49] does that have to be in the same tool? [90:52] You know, if if you're right, and of [90:55] course I think you are because I'm [90:56] talking to you right now. Uh but but if [90:58] you're right, we're going to look back [90:59] at the last three years with like [91:01] remember when we were idiots? Remember [91:04] when we thought that it was going to be [91:06] we were going to do digital god? [91:08] >> Well, yeah. I hope so. I hope that's [91:11] what we're saying like three years from [91:12] now. It's like digital god made no [91:14] economic sense. It was incredibly [91:15] dangerous and [91:17] >> and also we were going to give it to [91:18] every single person for free [91:20] >> and and we're give everyone for free and [91:22] we still we've been searching Kawaii for [91:25] the last three years and still haven't [91:26] found Mark Zuckerberg. We know he's [91:27] hiding there somewhere. We'll get him [91:29] eventually. [91:30] >> He's in his ocean lair. [91:32] >> It's like Hunger Games. He's painted the [91:33] leaves on his face. He's hard to see. [91:38] >> He's finally found a fashion that works [91:40] for him. [91:40] >> You don't like the the chain in the [91:44] t-shirts? [91:45] >> Look, it looks it looks a lot better [91:47] than he used to. I you know, he's making [91:50] choices or someone is um in a way that [91:53] makes me think that that it is that [91:56] people that someone has learned how to [91:58] say, "Actually, Mark, that haircut looks [92:00] bad." Which is good. I think that's good [92:02] news. [92:02] >> Maybe a chatbot will tell him that. See, [92:04] this could all come forward. [92:05] >> Maybe that's what happened. He got early [92:07] access. Cal Newport, you could watch his [92:10] and and listen to his podcast, Deep [92:11] Questions. It's free on YouTube. It's [92:13] very good. I I am a a regular viewer [92:17] myself. It's also free wherever podcasts [92:19] are in other places. And also if you'd [92:22] like some of Cal Newport's work, a lot [92:24] of it is focused on how we focus on how [92:28] we have agency in our world, how we take [92:31] that agency back, and how we do good [92:32] work and how important that time is and [92:36] that focus is for doing good human work. [92:39] Uh so check check out Cal. He's great. [92:41] Cal, thank you so much for spending so [92:43] much time with me. Oh, [92:44] >> it's my pleasure. enjoyed it.  📄 **Full text:** ![[Full Document Contents/Articles/This is Going to be Very Messy|This is Going to be Very Messy, YouTube]] ## Highlights - so if you have one or multiple very intelligent systems that have goals, I think that they will find it much more easy to manipulate people than we would guess. Our realities are defined by what we consume. And there is a way out of this which is simply getting off of the screens and I think that we'll be bad at it. What I have seen and I hope there's a backlash to it but what I have seen is that every opportunity that a platform ([View Highlight](https://read.readwise.io/read/01kjqn75pq9jqg2aw1zh25wnwt)) - Tags: 👻 ai highlighted - something that only happens during certain times of the day or or maybe we've already got sort of gotten that itch scratched and we'll change our minds about this. But that has thus far been the direction that we have gone. Again, what humans do is we communicate, we talk. We have a huge bandwidth of information. And so, it makes sense to me that the thing to be most concerned about is how AI interfaces with that. And already, I think it interfaces with it in in extremely intense ways through recommendation algorithms, which have been around for a long time, but now ([View Highlight](https://read.readwise.io/read/01kjqn75px7xxfrhw431vc8b0q)) - Tags: 👻 ai highlighted - going to do a 10-minute video and then I'm going to jump over to you and then Yeah. Literally. >> Yeah. Because it's almost always the story is almost always this sort of binary story of well there's this thing we liked about this technology but then it had this other effect we weren't expecting. And so like now we have to kind of balance the two. It's usually that simple. It's email is nicer than a fax machine. It's pretty convenient. I'm happy it's here. Oh my god, but there's this side effect that we get too many of them and we're distracted. AI is such a messy story that makes that question it's it's a very complicated question ([View Highlight](https://read.readwise.io/read/01kjqn75q2j05mebqfsjnhewfe)) - Tags: 👻 ai highlighted - then I'm extrapolating that well what if that happens for another three years where would we be and then reacting to that future uh most computer scientists I know have a lot of problem with those extrapolated futures we think they're they're sort of ignoring the reality and limitations of technology ology but I am very interested right now in a lot of the present issues. So things that are happening right now with the technology as it exists and two of them that I care quite a bit about is the disappearance of truth and the disappearance of focus or concentration as a human ability. So disappearance of truth generative AI can ([View Highlight](https://read.readwise.io/read/01kjqn75q7sec6dt9704bejhaw)) - Tags: 👻 ai highlighted - so >> which this is the reality of it right but it's important like the for people to understand right there's sort of prediction type machines that have been around for a long time this is where like the concerns you have about bias and sentencing for example the classic AI AI safety areas. These type of technologies have been around for a long time. They haven't really seen any major innovative boost in the way that generative AI has, but they we've long been studying those concerns and you know those concerns are important. We have the environmental concerns. That's a hard one to get your arms around ([View Highlight](https://read.readwise.io/read/01kjqn75qb0rwwqqymgp83a94c)) - Tags: 👻 ai highlighted - humans do that all now have individualized AI type systems that can do that thing just as well or better than humans. More of a peace mill. I think that's much more likely >> and much more uh cost-effective and makes much more business sense than this idea of we just need a a super AI that does all of these things. And we already see it now like, oh, we're better at chess. We're better at poker. We're better at go. We're getting better at proteins. That's a, you know, it's more peace. Each of these systems is going to need its own mix of technologies and its own secret sauce, but that's much more ([View Highlight](https://read.readwise.io/read/01kjqn75qhdxa4b477bjq505kq)) - Tags: 👻 ai highlighted - likely to get us to a point where we say, "Wait a second. Computers are doing a lot of what we do better than us." I think that's a much better more likely way we're going to get there than a singular machine doing it. >> And yet it's not uh that that take isn't going to get any headlines written about it. Well, I mean, I like to tell the story of um there's these cool bots that came out. There's a particular researcher who's done a lot of cool AI. Uh and the two things I think he's at OpenAI now, but but back at Carnegie Melon and Meta, he built the first system to beat human poker players, ([View Highlight](https://read.readwise.io/read/01kjqn75qp56agk3th6xf1m349)) - Tags: 👻 ai highlighted - right? Like if you are Daario Amade or Sam Alman, right? It's completely aligned with your interest to talk about the automation of most jobs or to talk about, you know, super intelligence coming because it does three things. One, >> it distracts attention or regulatory or uh public, you know, disapproval about the things that your bad things that your technology is doing right now. Um, two, it's going to induce a lot of perhaps ownorous regulation that you as a very large company can weather but small startups can't. So, you get ([View Highlight](https://read.readwise.io/read/01kjqn75qvq3deqgz9y2sg06xd)) - Tags: 👻 ai highlighted - enough of this like we don't want to use this anymore. I think that's finally starting to happen for social media. It's taking five years longer than I would have hoped, but like >> we're getting fed up, you know? I think so. I I I have some >> I have optimism and I don't want the I don't want to seem if we list all the dooms out in a row, I don't want it to seem like >> these are all adding up to one villain that's going to be impossible to slay, >> right? A and if you listed out all the ways that AI has like helped you know somebody do some thing or even like ([View Highlight](https://read.readwise.io/read/01kjqn75r1s1p496cnrg9vcbpk)) - Tags: 👻 ai highlighted - for a computer scientist, you're like, "Ah, >> that's my that's the world I imagine." Like, you just think in code. It makes so much sense. Yeah. >> Um I and I think so. Anyways, I'm excited about that. I think it's okay. And that does actually have big ramifications and it's going to change a lot of things and it could have economic impacts. Uh but that's a good but I think ultimately that would be a good thing. But it yeah, it's different than the conversation we're having right now. It sound it also sounds to me like if this is the world that we end up in it it I one of my big fears really is like ([View Highlight](https://read.readwise.io/read/01kjqn75r5xnjzkcjafkktjkw8)) - Tags: 👻 ai highlighted