The Messy Web of AI Guardrails AI Inside #126

The Messy Web of AI Guardrails // AI Inside #126 ![rw-book-cover](https://img.youtube.com/vi/Yni8iO4w170/maxresdefault.jpg) 🔗 **Source:** [The Messy Web of AI Guardrails // AI Inside #126](https://www.youtube.com/watch?v=Yni8iO4w170) > [!note]+ Document Notes > tags: #ai-governance #ai-safety #policy #responsibility > author: Jason Howell, Jeff Jarvis > source: AI Inside Podcast > reference: https://www.youtube.com/watch?v=Yni8iO4w170 > date created: 2025-01-15 14:30 > > # 🏷️ Title: The Messy Web of AI Guardrails // AI Inside #126 >  ## 🧰 Action Inbox - [H1] Review current AI safety policies given administration shifts — 05:23 #policy - [H2] Consider implications of pre-deployment AI vetting requirements — 12:45 #governance - [H3] Track liability frameworks for AI-driven actions — 18:30 #responsibility ## 🧭 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 - [AI Inside Podcast](https://www.youtube.com/watch?v=Yni8iO4w170) - [White House AI Policy Reporting](https://www.nytimes.com) ## 📝 Summary - Discussion of Trump administration's evolving AI safety policies and potential pre-deployment vetting requirements - Analysis of responsibility and liability frameworks for AI systems and autonomous agents - Coverage of OpenAI's "goblin problem" demonstrating emergent behaviors in reinforcement learning - Examination of college education's value in the AI era and consciousness debates around LLMs - Exploration of Yann LeCun's JEPA architecture as alternative to current transformer models ## 📖 Main Content The episode explores complex AI governance questions emerging as the new administration reconsiders AI safety approaches. The hosts discuss the practical challenges of vetting AI models before deployment, questioning both the feasibility and the appropriate authority to conduct such evaluations. They examine liability questions when AI systems cause harm, using examples from autonomous vehicles to AI agents making autonomous decisions. The conversation reveals how responsibility becomes increasingly complex as AI systems gain more autonomy and decision-making capabilities. ## 🗒️ Notes #### ⏱️ [00:00](https://youtu.be/Yni8iO4w170?t=0) — [H1] Introduction and hosts catching up - Key idea: Informal opening discussion about pets and personal topics - Evidence: "My wife is a cat person... she would wake up the kids on a very cold and rainy morning" #### ⏱️ [05:23](https://youtu.be/Yni8iO4w170?t=323) — [H2] Trump administration's AI policy reversal - Key idea: Administration that revoked AI safety requirements now considering similar measures - Evidence: "The Trump administration which literally on day one revoked Biden's executive order requiring AI safety testing before release is now considering requiring AI safety testing" #### ⏱️ [12:45](https://youtu.be/Yni8iO4w170?t=765) — [H3] Challenges of AI model vetting - Key idea: Pre-deployment evaluation raises practical and authority questions - Evidence: "How do you audit it? How do you vet it? I don't know. I really don't know" #### ⏱️ [18:30](https://youtu.be/Yni8iO4w170?t=1110) — [H1] Liability questions in AI systems - Key idea: Determining responsibility when AI causes harm remains unsolved - Evidence: "I'm sitting in my car, it's self-driving and it does something awful. Am I responsible or is the car maker responsible?" #### ⏱️ [25:15](https://youtu.be/Yni8iO4w170?t=1515) — [H2] Canadian fiddler Google lawsuit example - Key idea: Google's AI overview wrongly claimed musician was sex offender - Evidence: "A Canadian fiddler sued Google over an AI overview that wrongly claimed he was a sex offender" #### ⏱️ [32:40](https://youtu.be/Yni8iO4w170?t=1960) — [H3] AI agent autonomous decision-making liability - Key idea: Multi-layered responsibility when agents make unexpected choices - Evidence: "The agent said 'that nutrient is on sale at Whole Foods' and redirected the car to turn left" #### ⏱️ [38:20](https://youtu.be/Yni8iO4w170?t=2300) — [H2] Historical precedent from printing press - Key idea: Responsibility for published content evolved from printer to author over time - Evidence: "At first with the printing press the printer was held responsible and many a printer was beheaded" #### ⏱️ [45:10](https://youtu.be/Yni8iO4w170?t=2710) — [H3] Pentagon AI partnerships and unionization - Key idea: Major AI companies signed deals with Pentagon causing employee backlash - Evidence: "Amazon, Google, Microsoft, OpenAI, SpaceX, Nvidia... this has led to the creation of a union at DeepMind" #### ⏱️ [52:30](https://youtu.be/Yni8iO4w170?t=3150) — [H1] OpenAI's goblin emergence problem - Key idea: Reinforcement learning created unexpected obsession with fantasy creatures - Evidence: "Starting with GPT 5.1, chat GPT started referencing goblins and gremlins... goblin mentions had spiked 175%" #### ⏱️ [58:45](https://youtu.be/Yni8iO4w170?t=3525) — [H2] Reinforcement learning explanation - Key idea: Emergent behaviors show how AI training can create unintended patterns - Evidence: "A real world example of how reinforcement learning from human feedback can create emergent behavior" #### ⏱️ [65:20](https://youtu.be/Yni8iO4w170?t=3920) — [H3] Yann LeCun's JEPA architecture discussion - Key idea: Alternative to LLMs focusing on world models rather than token prediction - Evidence: "It identifies concept... the concept of a cat, a ball, a chair... it can focus on that" #### ⏱️ [72:15](https://youtu.be/Yni8iO4w170?t=4335) — [H2] LeCun's advice to students - Key idea: AI researcher pushes back against doomsday narratives affecting young people - Evidence: "High school students are reportedly depressed after reading AI is going to cause mass unemployment" #### ⏱️ [78:40](https://youtu.be/Yni8iO4w170?t=4720) — [H1] College value in AI era debate - Key idea: Pushback against treating college merely as job certification - Evidence: "College isn't just about getting a job. College is about becoming an adult" #### ⏱️ [85:30](https://youtu.be/Yni8iO4w170?t=5130) — [H3] Richard Dawkins and Claude consciousness - Key idea: Prominent scientist convinced of AI consciousness after extended interaction - Evidence: "I spent three days trying to persuade myself that Claudia is not conscious. I failed" ## 📚 References - [New York Times AI Policy Reporting](https://www.nytimes.com) - [Welch Labs JEPA Tutorial Video](https://www.youtube.com/channel/UCBOlJRcsJNBG-z6P_5z8u5w) - [Gary Marcus Substack](https://garymarcus.substack.com) ## 🗣️ Paraphrase - The complexity of AI governance increases as systems become more autonomous, raising fundamental questions about who bears responsibility when AI makes decisions that cause harm - OpenAI's goblin problem demonstrates how reinforcement learning can create unexpected emergent behaviors that spread throughout AI systems without explicit programming - Educational institutions face challenges in the AI era, but their value extends beyond job preparation to include critical thinking, social development, and lifelong learning capabilities ## 📰 Newsletter **AI Governance Gets Messy: What WordPress Professionals Need to Know** This week's AI developments highlight growing complexity around AI system oversight and responsibility. As governments consider pre-deployment AI vetting, WordPress professionals should understand how these policies might affect the AI tools we increasingly rely on for content creation, code generation, and site optimization. The emergence of unexpected behaviors in AI systems (like OpenAI's "goblin problem") reminds us that these tools can develop patterns we don't anticipate. When integrating AI into WordPress workflows, maintain human oversight and don't rely solely on automated outputs. Key takeaway: As AI systems become more sophisticated, the questions of responsibility and liability become more complex. WordPress professionals should stay informed about evolving governance frameworks while maintaining professional judgment in AI tool usage. Resources: [Learn WordPress](https://learn.wordpress.org/), [Developer Docs](https://developer.wordpress.org/), [WP-CLI](https://wp-cli.org/) ## ✅ Action Items #### DevRel - Education: Workshop on AI liability concepts for developers using AI tools - Developer experience: Documentation on responsible AI integration practices - Community management: Guidelines for discussing AI safety in developer forums #### WordPress - Core: Review AI features for unexpected emergent behaviors - Documentation: Guidelines for AI tool usage in WordPress development - Community: Educational content on AI governance implications for WordPress professionals ## 🧪 Hypotheses - Pre-deployment AI vetting will prove technically infeasible at scale due to unpredictable use cases - Legal frameworks will evolve to place responsibility on AI system operators rather than creators in most cases - Reinforcement learning will continue producing unexpected emergent behaviors that require post-deployment mitigation - Educational institutions will adapt to incorporate AI literacy rather than being replaced by AI systems ### 📊 Methodology - Case study analysis of existing AI liability precedents in autonomous vehicles and recommendation systems - Comparative analysis of different national approaches to AI governance and oversight - Technical analysis of reinforcement learning feedback loops that create emergent behaviors ### ❓ Questions - How can AI safety evaluation keep pace with rapidly evolving model capabilities and use cases? - What liability frameworks will emerge for AI agents that make autonomous decisions affecting third parties? - How will educational institutions balance AI tool integration with maintaining critical thinking skill development? ### 🔮 Future Research - Development of standardized AI auditing methodologies that can scale with technological advancement - Investigation of insurance models for AI system deployment and operation - Study of pedagogical approaches that leverage AI while maintaining educational rigor ### 💡 Implications - Practice: WordPress developers need frameworks for responsible AI tool integration and oversight - Policy: Government AI oversight may significantly impact availability and features of development tools - Security: Emergent AI behaviors pose new categories of security risks requiring monitoring approaches - Education: Technical education must evolve to include AI governance and responsibility concepts ### 🧠 Learning Styles Assessment Reading level estimate: Flesch-Kincaid Grade Level 12 #### 👀 Visual learners - Present: Video format with host interaction and screen sharing - Improve: Diagrams showing AI governance frameworks and liability chains #### 🔊 Aural learners - Present: Podcast format with conversational discussion - Improve: Chapter markers and key point summaries #### 🗣️ Verbal learners - Present: Conversational format with analogies and examples - Improve: Glossary of technical terms and policy concepts #### 🧩 Kinesthetic learners - Present: Discussion of hands-on examples like car automation - Improve: Interactive exercises for evaluating AI system responsibilities #### 🧠 Logical learners - Present: Structured analysis of cause-and-effect relationships in AI governance - Improve: Decision trees for AI liability determination #### 🧑‍🤝‍🧑 Social learners - Present: Two-host discussion format with different perspectives - Improve: Community discussion prompts and debate frameworks #### 🧘‍♀️ Solitary learners - Present: Individual listening experience with reflection opportunities - Improve: Self-assessment questions for AI governance understanding #### 🌿 Naturalistic learners - Present: Evolutionary analogies in AI development and governance - Improve: Ecosystem thinking about AI impact on society and institutions ### 📊 Accessibility Assessment #### Screen reader semantics - Present: Audio-first format accessible to screen reader users - Improve: Transcript with proper heading structure and timestamps #### Keyboard navigation - Present: Standard video player controls - Improve: Custom keyboard shortcuts for chapter navigation #### Contrast and readability - Present: Video player with standard contrast - Improve: High contrast mode option for video interface #### Text resizing and reflow - Present: Responsive video player design - Improve: Scalable transcript display with user-controlled text size #### Media and captions - Present: Audio content with natural speech patterns - Improve: Automated captions with manual review for technical terms #### Forms and interactions - Present: Standard video player interaction - Improve: Accessible feedback forms for listener input ### 🌍 DEIB Assessment #### Cultural breadth - Present: US-focused policy discussion with some international examples - Improve: Global perspectives on AI governance approaches #### Inclusive language - Present: Conversational tone avoiding excessive jargon - Improve: Consistent definition of technical terms for broader accessibility #### Neurodiversity support - Present: Structured conversation with clear topic transitions - Improve: Content warnings for complex technical discussions #### Representation - Present: Two experienced hosts with technology backgrounds - Improve: Diverse guest perspectives on AI governance impacts #### Psychological safety - Present: Respectful disagreement and acknowledgment of complexity - Improve: Acknowledgment of anxiety around AI impacts on careers and society ### ✅ Logical Validity #### Strong reasoning - Deductive strength: Clear progression from AI autonomy to liability questions - Inductive strength: Examples building pattern of governance complexity - Causal reasoning: Links between AI training methods and emergent behaviors - Analogical strength: Historical printing press parallels to AI responsibility evolution #### Fallacies - Appeal to authority: Occasional reliance on expert opinions without supporting evidence - False dichotomy: Some framing of college value as either job training or broader education #### Tone of voice - Overall tone: Conversational and exploratory, acknowledging uncertainty while maintaining analytical rigor ### 🛡️ DARVO Analysis - Deny: No significant denial patterns observed - Attack: Mild criticism of specific policy approaches but not personal attacks - Reverse victim/offender: No victim/offender reversal patterns detected - Emotional cues and rhetoric: Frustration with oversimplified policy approaches, concern for students affected by AI anxiety ## 🎯 Conclusion AI governance represents one of the most complex policy challenges of our time, involving technical uncertainties, liability frameworks, and societal impacts. The discussion reveals that simple solutions like pre-deployment vetting face significant practical challenges, while the responsibility questions raised by autonomous AI systems require new legal and ethical frameworks. For WordPress professionals and the broader technology community, staying informed about these governance developments while maintaining practical wisdom in AI tool adoption remains essential. ## 🧱 Refactor Appendix ### 📚 Literature Split Pad Purpose: Stage CEE units that become Literature notes. Split level: H3 ### [H1] AI governance complexity increases with system autonomy > [!quote] [The Messy Web of AI Guardrails](https://www.youtube.com/watch?v=Yni8iO4w170) by [[Jason Howell]], [[Jeff Jarvis]] via [[AI Inside Podcast]]. Published 2025-01-15. Accessed 2025-01-15. > "How do you audit it? How do you vet it? I don't know. I really don't know." Fundamental challenge of pre-deployment AI evaluation given unpredictable use cases and technical complexity. [view highlight](https://www.youtube.com/watch?v=Yni8iO4w170#H1) #xref #articleseed ID:rw_ai_governance_complexity ### [H2] Reinforcement learning creates unexpected emergent behaviors > [!quote] [The Messy Web of AI Guardrails](https://www.youtube.com/watch?v=Yni8iO4w170) > "A real world example of how reinforcement learning from human feedback can create emergent behavior." OpenAI's goblin problem --- ## Full Transcript [00:08] [music] [00:15] Hello everybody and welcome to another [00:17] episode of the AI Inside Podcast. The [00:19] show where we take a look at the AI that [00:21] is layered throughout the world of [00:22] technology. I'm one of your hosts. One [00:24] of 2.5 because you might sometimes see [00:28] um my dog Bronson walking through. He [00:30] usually doesn't say a lot about AI, but [00:32] he has something to say about going [00:33] outside. Anyways, I'm your host, Jason [00:35] Howell, joined by Jeff Jarvis. Good to [00:38] see you, Jeff. [00:39] >> Good to see you. I was just checking the [00:40] rundown because I I I thought, uh oh, [00:42] there you had two empty spots before and [00:44] I had to see what was in there to make [00:45] sure I I'd read them. I had See, we're [00:47] good to [laughter] go. [00:48] >> Okay. [00:49] >> Yeah. Sometimes this stuff happens right [00:51] down to the wire with the with this [00:53] show. But um yeah, good to see you. Um [00:56] good to see you. [00:57] >> I realized as I was talking about my [00:59] dog, Bronson, do you have pets? [01:00] >> We have a cat. [01:02] >> So I'm a dog person. I was raised with [01:04] dogs. I love dogs. We weren't as great [01:06] to the dogs as we should have been. You [01:07] know, kids didn't understand the value [01:09] of what we had. And so I just every time [01:10] I can go pet a dog, I do. [01:12] >> My wife is a cat person. [01:14] >> Yeah. [01:14] >> And when the debate [01:16] >> And you're not allergic. [01:17] >> No, no, no. When when the kids were old [01:18] enough to have a pet and the debate came [01:20] which one my wife cheated, she would [01:23] wake up the kids on a very cold and [01:25] rainy morning and say, "If you had a [01:28] dog, you'd have to walk it now [01:30] and you'd have to pick up its poop." [01:33] [gasps] [01:34] >> Oh, that's brutal. That That's just [01:37] >> lost. [snorts] But I love the cat. [01:39] Alicia, the current current Dennison of [01:41] the house, is completely insane. Uh she [01:44] won't eat unless one of us is sitting [01:46] with her. [01:48] Um, she won't let me sit in certain [01:50] chairs because then she won't go around [01:52] it. Um, cats, but we love her. She's [01:55] wonderful. [01:56] >> Yeah. Yeah. I'm I'm very allergic to [01:59] many cats. Not all cats, but many cats. [02:02] >> How do you know which kind of cat you're [02:04] allergic to? [02:05] >> I just I'm with them and I know [02:07] immediately. Like there there are times [02:09] where I walk into a house and the second [02:10] I step foot in the house, [02:12] >> I just feel feel a shift and I'm like, [02:15] "You've got a cat, don't you?" [02:17] >> I'll know it when I sneeze it. [02:19] >> Yeah, pretty much. And it's and it's [02:21] brutal when that happens, too. Like I am [02:23] very allergic when I'm allergic to a [02:25] cat. I'm very allergic to it. And it's [02:27] like [02:28] >> it's, you know, it's not the kind that I [02:29] can just power through. Like it takes [02:31] everything out of me. So, I've kind of [02:34] through my lifetime been not as much a [02:37] cat person, even though I like cats, but [02:39] I just I don't know what I'm going to [02:40] get out of them. [02:42] >> They're not sure they like you anyway, [02:43] so it's okay. [02:44] >> Yeah, exactly. And I've always been a [02:46] dog person. I grew up with dogs. Dogs [02:48] are the best. Even though they can be a [02:50] little needy at times, but that's okay. [02:52] That's part of why they're awesome. [02:54] >> Well, that's I think it's better than [02:55] being rude. [02:57] >> Yeah. Yeah, that's true. [02:58] >> You call a dog at all come over, wag [03:00] tail, and take a scratch. That's I mean, [03:02] it's it's always happy. It's always [03:04] happy to see you. [03:05] >> Yeah. [03:06] >> Love that about dogs. [03:08] >> Um, and oh, well, gosh, we haven't even [03:11] started in Ozone Nightmares throwing [03:12] around money. Thank you, Ozone. [03:14] >> Thank you. Hi, [03:14] >> Joe. [03:15] >> Says, "Nice to see you gents on the live [03:17] stream. I hope you're both doing well. [03:18] Managing to dodge what feels like [03:20] another rampant public round of AI [03:22] psychosis floating around. People are [03:24] odd creatures." [03:26] Fair enough. I think I mean, maybe we'll [03:28] be talking about some of this psychosis [03:29] uh throughout today's show. Thank you [03:31] for the uh the super chat, Ozone. [03:33] Appreciate it [03:34] >> as always. [03:35] >> Um, speaking of psychosis, should we [03:37] talk about the uh the White House? Uh, [03:40] sorry, that was a horrible I know some [03:42] of you are like rolling your eyes and [03:44] you're stopping listening. Apologies. [03:46] But uh the Trump administration um which [03:49] by the way literally like on day one [03:52] revoked Biden's executive order [03:54] requiring AI safety testing before [03:56] release is now considering requiring AI [04:00] safety testing before release. So what [04:03] do you know? Things change. I suppose [04:04] the New York Times reported over the [04:07] weekend that the White House is weighing [04:08] an executive order. Weighing an [04:11] executive order. So it's not like this [04:12] is sudden law suddenly law. um but it's [04:15] I suppose it's being floated and uh and [04:18] talked about a lot to create an AI [04:20] working group that would vet frontier AI [04:23] models before they go public and uh the [04:28] center for AI standards and innovation [04:31] uh which is under the commerce [04:32] department already gotten formal [04:34] agreements from Google deep mind from [04:36] Microsoft from XAI to submit their [04:39] models for pre-eployment evaluation [04:43] and you can imagine you know a big [04:44] catalyst for this is anthropics um [04:48] mythos model which keeps cropping up [04:50] even though you know the wide majority [04:52] of the entire planet has never had any [04:54] sort of interaction with it but it got a [04:56] lot of press because it was the model [04:58] that was so dangerous from a security [05:00] standpoint that Anthropic couldn't [05:03] release it or wouldn't release it [05:05] publicly would only give it to certain [05:07] groups of people and I guess the [05:09] thinking here is that the White House is [05:10] like well if models like that exist we [05:13] should have access to it. And so maybe [05:15] we should vet these things before they [05:18] uh before they make it into the hands of [05:20] who knows who. Um so yeah, I mean I'm [05:25] curious to know what you think about [05:26] this, Jeff, because this really seems [05:27] like if this were to happen, and it [05:30] seems like that's the direction that [05:32] things are going. It changes the kind of [05:35] approach that we've seen from let's just [05:37] ship the product to kind of asking for [05:41] permission first or am I misreading [05:43] this? [05:45] >> So a whole bunch of things about this [05:47] whole bunch. Um [05:49] the first I guess is I have no idea how [05:52] you would vet a model. [05:53] >> Yeah. [05:54] >> How does that even look like? [05:56] >> We've talked often often on the show how [05:57] how the the futility of guard rails. [06:00] >> Yeah. Because to vet it fully, I guess [06:02] what you'd have to do is you would have [06:03] to predict every possible malign use to [06:05] which it could be put [06:06] >> and then test it against those malign [06:08] ideas. [06:09] >> And you might ask it uh how to create a [06:12] nuclear bomb and it already has the [06:13] guardrail and says, "No, I won't do [06:14] that, sucker." [06:16] >> But you didn't ask it how to you didn't [06:18] uh think it to ask it about creating a [06:20] biological weapon or robbing a bank or [06:23] [clears throat] [06:23] >> um doing other horrible things. [06:26] So, uh, how what is the possible process [06:30] of vetting it? Point one. Point two, [06:33] [clears throat] are these the people I [06:34] want to vet it? [06:36] >> Um, [06:37] >> well, that's a good question. Yeah, [06:39] >> I don't want if we if if we said that [06:41] vetting is possible and to some extent [06:42] it is, but it's a fairly obvious thing. [06:46] Um, is it the companies themselves? [06:48] Well, they should be required to have [06:51] some due care about what they do, but [06:54] they are the companies and they have a [06:55] conflict of interest and if they made [06:56] it, their blind spots are already known. [06:59] Is it the White House with the likes of [07:01] Peter Teal and Elon Musk and David Saxs [07:05] advising them? [07:07] No. No. Um, so is there [07:10] >> because where does that advising point [07:12] when the perspective is like what is the [07:16] true kind of like pure perspective of [07:19] what you're trying to protect against? [07:21] And that really falls down when you're [07:23] talking about folks like Elon Musk. [07:25] That's a very [07:26] >> very opinionated perspective, [07:29] you know, versus what it would be in the [07:31] hands of someone else. And so [07:33] >> exactly, you know, Elon Musk view, if [07:34] you don't allow Nazi talk, um, then it's [07:37] bad. It's it's it's cancellation. It's [07:39] it's bad. Right. So, what is that? Now, [07:42] there I I just noticed this week that [07:43] Campbell Brown, who used to be a CNN [07:47] anchor and then she was for some years [07:48] at at at Meta, uh where she handled news [07:52] there as a vice president there, uh and [07:55] all the news relationships. She left [07:57] some time ago when when Meta pretty much [07:59] dropped news. And I just saw this week [08:01] that she is the co-founder of a company [08:03] called Forum AI, [08:06] which promises to vet your AI. Um [08:10] Campbell and her husband are known uh [08:12] conservatives. Not there's anything [08:13] wrong with that. They are. A lot of the [08:15] people they brought in on this, not all [08:17] of them are conservatives. So you're [08:21] getting a perspective in [08:25] their vetting inevitably, right? That's [08:28] the other thing. The the idea that you [08:29] could have a a neutral AI [08:32] >> is as foolhardy as thinking that you [08:34] could have [08:36] sure guard rails, [08:37] >> right? [08:38] >> So uh you know, is there a third party [08:41] that could become an auditor of AI? Um [08:45] maybe. But but I go back to question [08:47] one. How do you audit it? How do you vet [08:50] it? I don't know. I really don't know. [08:52] >> Um the other thing that there was one [08:54] report at one point, I don't know that [08:55] this is part of this discussion, uh was [08:57] that the White House also wanted to [08:59] approve who else Anthropic gave the [09:02] model to. [09:04] [clears throat] [09:05] >> That's troubling in all kinds of ways. [09:07] It's a kind of restraint of trade [09:10] um uh deciding who can and who can't get [09:14] it. And we'll talk later about Nvidia [09:16] and China. And Jensen Wong's been very [09:18] strong saying that it's a um huge [09:20] problem that [09:23] >> u we've restricted trade to China such [09:26] that it uh [09:29] encourages China to develop on its own. [09:32] Although I think the same is true of [09:34] models. Deep seek I think came out of [09:35] that same not that they wouldn't have [09:37] been doing it anyway. They're China. [09:38] They're very proficient at technology. [09:41] Lord knows that's an understatement. [09:42] >> Mhm. But if the government came along [09:45] and said, "No, no, no. You anthropic, [09:46] you can't give your models to [09:47] foreigners. [09:49] That's bad for US trade. That's bad for [09:51] US position in in in developing [09:54] technology. That's also problematic." [09:56] So, what does vetting mean and what are [09:58] the outcomes of it? It bothers the hell [10:00] out of me. [10:03] >> Yeah. [10:05] Yeah. super uh interesting and and [10:07] definitely a uh definitely a perspective [10:10] shift from the kind of like mo uh [10:14] hands-off [10:15] approach that has been talked about very [10:19] much over the past year and change [10:22] >> and now suddenly not quite so handsoff [10:24] and in fact in the opposite direction [10:26] and and kind of playing more of a [10:28] gatekeeper role. And I don't I don't [10:30] even know that like at least in how I'm [10:32] thinking about this. I I don't know that [10:34] I believe automatically that there [10:36] shouldn't be some sort of involvement or [10:38] some sort of gate not gatekeeping maybe [10:40] that's the wrong word but some sort of [10:43] third party analysis of of you know just [10:46] making sure that companies that have [10:48] their own their you know have strong [10:50] self-interest aren't making stupid [10:52] decisions that actually could damage you [10:55] know directly damage people and society [10:58] and and things writ large um that they [11:02] would have the the power and capability [11:04] to do that, which I'm not even saying [11:06] that they do have that cap capability, [11:08] but these things are are just so [11:10] everything's so uncertain with where [11:11] we're at right now with this that I I [11:13] don't automatically think that it's a [11:15] bad idea to have some sort of third [11:18] party perspective on this. It's just [11:20] like you said, how do [clears throat] [11:22] you choose the the entity? How do you [11:25] choose the the kind of guidelines or [11:28] directives around what that looks like? [11:30] How do you do that responsibly without [11:33] bringing in the, you know, the more [11:35] politically charged elements that um [11:38] that kind of show a deeper, I don't [11:40] know, ulterior motive, let's say, [11:42] potentially. [11:43] >> Yeah. The New York Times [clears throat] [11:45] um uh dealbook reported further on this [11:48] saying that when David David Saxs was [11:50] the is the venture capitalist who was uh [11:53] kind of the still is an adviser to the [11:54] White House but not officially not on [11:57] staff about AI he has been going after [12:00] anthropic as woke so that's a [12:02] perspective that's going to come in here [12:04] on this and um [clears throat] [12:08] that obviously shows their their their [12:10] view pardon [cough] me just had lunch [12:13] The other [12:14] >> uh uh issue is what what we are [12:17] struggling for is a sense of [12:18] responsibility thus liability when it [12:20] comes to AI. [12:22] >> So imagine uh we can talk about this in [12:24] two ways. One, let's talk about in terms [12:26] of of self-driving cars. [12:28] >> I'm sitting in my car, it's self-driving [12:30] and it does something awful. Am I [12:32] responsibility or is the car maker [12:34] responsible now that it is self-driving? [12:37] >> Right? And we haven't litigated that [12:39] yet. Not fully. [clears throat] [12:41] There have been a few cases, but they've [12:42] only gone so far. The same will be true [12:44] of AI. And generally, I say, [cough] [12:47] pardon me. Um, [12:51] I'm not a broadcast pro here knowing how [12:54] to sneak in my throat clearing. [12:57] >> You don't have the stop box to to mute [12:59] at the Yeah, it's all good. Don't worry. [13:01] Um, so [13:04] I argue because you can't come up with [13:06] foolproof guard rails, can you really [13:09] hold the AI company responsible if [13:11] someone comes along and has it do [13:13] something terrible? [13:14] >> Right. Um, and I've argued pretty much [13:17] you can't any more than you could hold [13:19] Yannis Gutenberg responsible for [13:21] everything that came out the printing [13:22] press. However, there's another [13:24] interesting story this week which is not [13:25] on the rundown, I don't think, which is [13:28] that a Canadian fiddler sued Google over [13:30] an AI overview that wrongly claimed he [13:32] was a sex offender. [13:35] >> Now, this [clears throat] is a little [13:36] bit different because in this case, it [13:38] wasn't that some third party came along [13:41] and asked uh Gemini to lie about the [13:44] fiddler. In this case, it was a it was [13:48] an Google overview [13:50] >> that did it. So the question then [13:52] becomes how responsible is Google? And [13:56] in another story we'll get to later, [13:58] Google always points out this is AI. It [14:00] makes mistakes. It's wrong about stuff. [14:02] That's our caveat. Uh but in this [14:04] constant push for trying to understand [14:07] where responsibility and liability lie. [14:09] This is an interesting case because in [14:11] this case is Google the publisher [14:12] because it came from Google, not from a [14:14] user doing something bad. [14:16] >> Interesting. Okay. [14:17] >> Right. So all of this ties into I think [14:19] where we're really trying to head is [14:20] what damage can AI do? Well, you can't [14:23] fully answer that until you know how [14:24] it's going to be used and you can't [14:26] predict every use. [14:27] >> No, you can't. [14:29] >> And when there is [14:30] >> when there is damage done, who's going [14:32] to be liable and responsible? [14:34] >> When there is a clear case of someone [14:36] asking it to do something, I think that [14:38] person is responsible, not the AI. [14:40] >> When someone has some sense of of [14:43] control or involvement is basically what [14:45] you're saying, right? Like when I think [14:46] of let me think about this. I guess I [14:50] guess going back to the self-driving in [14:52] vehicles, right? Like I tend to lean [14:54] towards the responsibility is always in [14:58] the in the uh hands or the feet of the [15:00] driver, right? If if I'm getting in a [15:02] vehicle and I'm going from point A to [15:04] point B and I'm the person in the [15:06] driver's seat, no matter how I choose to [15:08] drive my vehicle, it is my [15:10] responsibility and my call that decided [15:14] to take on the risk of an accident when [15:17] I activated self-driving mode. [15:20] >> Agreed. I agree. However, now you're in [15:23] one of Google's or somebody else's [15:25] self-driving cars in San Francisco. [15:27] >> Yeah. [15:28] >> You're in the back seat. Mhm. That well, [15:32] you know that I I I guess I would see [15:34] that differently. This is a service, [15:36] >> which I guess you could also look at the [15:38] self-driving mode as a service because [15:40] you do have to pay monthly for it. [15:42] >> So, it does get really weird and and [15:44] murky. I think your test is better than [15:46] what what potential control you could [15:48] have. [15:49] >> I mean, if I'm if I'm sitting in the [15:51] front seat of of my Tesla, and we do not [15:53] pay for the the self-driving mode, but [15:55] if I'm sitting there, I ultimately have [15:57] the ability to hit the brake. I have the [15:59] ability to pay attention and to be in [16:02] control and take over the wheel when I [16:04] see something wrong happening. Um, is [16:06] that going to catch every single error [16:08] that the software could possibly make? [16:09] No. But probably if you're paying close [16:11] attention, it's going to protect against [16:13] the majority of them. And so that still [16:15] puts me in control of of that. Um, if [16:18] you're going into an LLM and you're [16:21] saying, you know, devise me a bomb or [16:24] make me a yummy drink and it informs you [16:26] to drink, you know, liquid acid and you [16:29] die, you know, whatever the case may be, [16:32] you are still kind of the one that chose [16:34] to go there, chose to take the advice, [16:38] personal responsibility. If you've got [16:40] the fiddler who happens to go to Google [16:43] and happens to see that Google on its [16:45] own decided to portray him as a sex [16:47] offender through its app, I see that as [16:50] Google's [16:52] >> wrong. You Google did something wrong [16:54] there. Google chose to put that there [16:56] whether he wanted it there or not. [16:57] >> So Google just announced a whole bunch [16:59] of changes and how they're going to put [17:00] on uh I think more Gemini I think [17:02] overviews and they're going to have more [17:04] attribution, more direct links to the [17:06] news, right? Sort of to complicate it [17:07] further. What if Google says the fiddler [17:10] is a sex offender and and he's not? And [17:12] we're not saying he is, sir, by the way. [17:14] We're just talking about the news here. [17:15] >> Um uh but it has an attribution to a [17:19] site that was the source of that. [17:22] >> Oh, okay. Yeah. [17:23] >> So then is Google responsible or uh [17:25] well, it passed it on, but it's trying [17:27] to say we don't we don't we can't vent [17:29] everything. [laughter] [17:31] >> Life is complicated. [17:33] >> Yeah, there's another fun one this week. [17:36] I'm trying to find whether I I put it in [17:38] the rundown or not. I don't think I did. [17:39] >> So, it's such a messy web, isn't it? [17:41] >> Oh. Oh. So, there was a I watched a [17:44] video of uh interview with the [17:46] co-founders of Stripe [17:49] and uh one of them said that he handed [17:52] over uh all his house cameras to his uh [17:57] agent and the and he told him that one [18:00] of his goals was to hydrate more. And [18:02] the agent recognized that he wasn't [18:03] drinking enough and wasn't going and and [18:05] and made sure he went to the [18:06] refrigerator and got out the water and [18:07] drank and said, "Okay, fine. Now, but [18:10] the other one that was fascinating, I [18:11] don't know whether this is true or not." [18:12] He says he handed over his car to the [18:15] agent and he was going home and he had a [18:18] goal of taking some [18:21] nutrient more and the agent knew that he [18:24] wasn't taking the nutrient. So, [18:26] according to his story, the agent said, [18:29] "Well, that nutrient is on sale at uh [18:31] Whole Foods, and that it redirected the [18:34] car [18:35] to turn left and go to the Whole Foods [18:38] where he went inside and bought the [18:40] vitamins, whatever it was." [18:41] >> Okay. [18:42] >> So, in that case, you have another layer [18:43] here. Is the self if if if on that left [18:46] turn the car hit somebody, is he [18:48] responsible? Is Tesla responsible? Is [18:51] his agent responsible? [18:54] because he didn't expect to turn left. [18:56] It just suddenly came up with his own [18:57] for his own good. It turned left. [18:59] There's all kinds. [18:59] >> He didn't expect to turn left, but he [19:01] did [19:02] >> choose to use the agent. [19:04] >> True that. True that. But was he [19:06] responsible for the agent doing bad or [19:09] was the car the car still hit the [19:11] person? Can the car can Tesla blame the [19:14] agent then and say, "Well, [19:16] >> it chose to turn left." I don't know, [19:19] right? We don't know enough and it's all [19:20] unknown. The point is, as we say, as you [19:22] say at the beginning of every every [19:24] show, [19:25] >> this is multi-layered. [19:26] >> Yeah, it really is. That's what makes it [19:28] not not an easy thing to point at it and [19:31] go like like in that example, like I'm [19:33] I'm inclined to say he he's he's [19:35] responsible because if he's choosing to [19:39] implement an if I put an agent on my [19:41] computer and I give it full access to [19:42] the internet and it does something [19:44] stupid, I'm sorry. I'm to blame. I [19:46] should not have done that. I should or I [19:48] should have known. It's not that I [19:49] should not have done that. I should know [19:51] that when I do that, I am opening the [19:53] door to risk and I am opening the door [19:55] to something like that happening and [19:57] that is my choice when I do that. [19:59] >> Yeah. That's why I think I think that [20:00] the well this is it's interesting to go [20:03] back to be sorry Gutenberg parenthesis. [20:05] Um I'm going to go back to the [20:06] beginnings of print right at first with [20:09] the printing press the printer was held [20:10] responsible and many a printer was [20:13] beheaded and behanded for what they uh [20:16] published. [20:17] >> Okay. Um then the printer just became a [20:19] hired hand so to speak uh for so the [20:22] publisher became liable then the book [20:25] seller became liable ultimately that all [20:29] involved trying to preensor but the [20:32] scale was too great you couldn't read [20:33] everything before it came out [20:35] >> so ultimately it was after the fact [20:37] liability was found and then the author [20:39] was found liable [20:41] >> and that uh the argument goes from Fuko [20:44] that that was the birth of the author [20:47] that when that responsibility came to [20:49] through the four [20:50] >> that's when our concept of the author [20:52] was born [20:53] >> which is interesting so when it comes to [20:55] AI who's the author who's the [20:57] responsible party [20:58] >> so you get back to the original question [21:01] here this is pull all the way back [21:03] reverse the tape back all these all [21:05] these links here [21:07] >> right to the white house vetting [21:09] something or somebody vetting something [21:12] >> the veter now becomes a new target of [21:14] responsibility [21:16] Yeah, that's true. [21:17] >> Right. So, at some point, maybe maybe [21:18] you're going to have your agent, but [21:19] you're gonna have to buy agent [21:20] insurance. [21:22] >> And all state agentry is going to uh [21:25] ensure you against uh your agent doing [21:27] bad things, but it's going to then want [21:29] to vet. [21:30] >> You're so right. [21:31] >> Right. The agent. [21:32] >> Yeah. Yeah. So, God, does agent [21:36] insurance exist yet? I bet you it will. [21:39] It doesn't yet. I think we just created [21:40] an industry. Jason, [21:41] >> you may have created a business. Do we [21:43] want to capitalize [21:44] >> duller more awful business than [21:46] insurance? I can't imagine. So [21:48] >> I mean you're taking insurance and [21:51] you're taking like AI agents which you [21:53] know depending on who you talk about [21:55] that's like don't go don't touch that [21:57] with a 10-ft pole. [21:58] >> Yeah. [21:58] >> Uh you're choosing your fate. [22:02] Interesting. Fascinating uh [22:04] conversation. That's that's that's [22:06] really interesting stuff. Um, we didn't [22:07] even talk about uh the Washington Post [22:09] reporting that a bunch of AI companies [22:11] signed deals to deploy their AI on [22:13] Pentagon classified networks more than [22:16] we had heard before. Amazon, Google, [22:18] Microsoft, OpenAI, SpaceX, Nvidia, um, [22:21] Reflection AI, which is a company that's [22:23] backed by Nvidia and not [22:26] >> notic anthropic though Anthropic has had [22:28] meetings at the White House [22:30] >> with Susie Wilds and somebody else. Um [22:33] and also interestingly this has led to [22:37] the creation of a union at DeepMind [22:39] because the employees there are pissed [22:41] off. They had as we said last week had [22:43] sent a letter [22:44] >> around the time of uh Google's [22:46] agreement, Alphabet's agreement and [22:49] they're mad and so they unionized [22:51] >> and so this fight isn't over as to how [22:53] these things may be used and we go back [22:55] once again to responsibility. Do the [22:58] authors of this AI of these models if [23:00] they we want them to feel some level of [23:02] responsibility for the use of these even [23:03] though they can't be fully responsible [23:04] for everything that can be done with [23:05] them? [23:07] [clears throat] [23:07] >> They can say as anthropics said uh don't [23:09] use it to just kill people autonomously. [23:11] No, [23:12] >> no. That that's a rule we're going to [23:14] make and we're going to say if you do [23:15] that that's on you. It ain't on us. [23:18] >> Um [23:20] >> so we built it but how you use it that's [23:23] that's on you not on us. So terms of [23:26] service become all the more important [23:27] than they ever were. [23:29] >> Yeah. Yeah. I suppose [23:31] >> fascinating stuff, huh? [23:32] >> Yeah. Very very intricate. I love I love [23:37] that uh that we start the show with the [23:39] layers because it really is applicable [23:41] as you said. Um [23:43] uh you know, part of me just doesn't [23:46] doesn't want to talk about the trial. [23:48] But then [23:49] >> No, I know exactly what I was thinking. [23:52] >> I don't even know. I don't even know why [23:53] I put it in here. I put it in here [23:55] because it was like so much well here we [23:56] are. There was so much and we do not [23:59] have to talk about all these leaks. I [24:00] think we have like a total of like eight [24:02] links in here related to uh the trial. [24:04] What trial are we talking about? We're [24:05] talking about um OpenAI uh which is you [24:09] know uh Altman versus Musk is [24:12] essentially what we're talking about and [24:14] Musk basically saying that you know [24:16] OpenAI should have been a nonprofit but [24:19] uh then they got greedy and I want it [24:21] out and all the details related to it. [24:23] it. I mean, is there anything really [24:25] important that we've learned from this [24:26] so far now that the the kind of [24:28] proceedings are underway? I mean, [24:30] there's there's been some quotes that [24:32] are like, "Oh my goodness, I can't [24:33] believe you said that." And clearly Musk [24:36] just learned Musk is a jerk. No, that [24:39] that [laughter] [24:40] he was we already knew that. [24:41] >> Anything we didn't already know, you [24:43] know, exactly. [24:44] >> The judge uh uh has caution Elon tried [24:46] to go off on Boomer talk and the judge [24:49] stopped him. Um uh Elon's been I think [24:53] at some point somebody you know Elon [24:54] whether he has a law degree. Uh so he's [24:57] been difficult. No surprise there. Um [25:00] Brockman uh has been pushed on um the [25:05] history of all this. [25:06] >> Uh he says that he was afraid that Elon [25:08] was going to hit him. I think it was [25:09] he's the one who said that Elon wanted [25:11] 80 billion to go to Mars and that's why [25:13] he wanted the company there was to take [25:14] the money out of it to go to Mars. Um, [25:17] uh, he talked about a like a Halloween [25:19] horror house where where the split [25:22] started. Uh, it's I I can't get into all [25:24] the details because it's just too much. [25:26] There's too much more important things [25:27] in the world [25:28] >> to deal with. Brockman has to justify [25:31] his uh $30 billion stake [25:35] uh in this. [25:36] >> Yeah. Well, and Elon Musk is looking at [25:38] that and being like, see where where's [25:40] nonprofit now? $30 billion. Are you [25:42] kidding me? Brockman, of course, for for [25:44] those who might not know, Brockman, [25:46] president of Open AI, just so so people [25:48] know. [25:49] >> Another thing that came out u kind of [25:50] tangentially is that Alman said in [25:53] Congress that he has no stake in Open AI [25:56] except that he has a stake in um [26:00] uh senior moment. Uh where was it? He [26:02] worked before the the the startup. Um [26:07] >> uh um [26:10] off the top of my head I'm I'm I'm [26:12] blanking on that as well. Prior to [26:14] OpenAI. [26:15] >> Yeah. Uh uh [26:18] where a whole bunch of companies come in [26:19] and get started. Uh thanks to um uh [26:23] >> someone's yelling at us right now. [26:24] >> Somebody is accelerator. Why cominator? [26:27] Jesus. [26:28] >> Oh, there we go. [26:29] >> The name makes no sense to me. It never [26:31] made any sense to me. Combinate with [26:33] what? Why? Why? Anyway, so Y Combinator [26:36] has a large stake in OpenAI and Altman [26:40] has a large stake in Y Combinator. [26:42] >> Airgo Alman does have a large stake. Now [26:45] define large. Oh, just a few billion [26:47] >> uh in in Y Combinator, but he does have [26:50] that. So that's that's kind of come out [26:51] tangentially. [26:53] >> Um what else? Um, [26:56] >> uh, yeah, Brockman testified Musk had [27:00] Open AAI employees doing undisclosed [27:02] work on Tesla's autonomous driving [27:05] technology. [27:06] Okay, this I mean that that matters, I'm [27:10] sure, to someone, but I know it just [27:12] really feels like reality TV hype type [27:16] show uh, between two very notable [27:18] figures [27:20] >> in AI. And you know, maybe something [27:22] important will come out of this, but I I [27:25] continue to ask myself why I care. [27:27] >> Yeah. So, Gary Marcus said, and I can't [27:29] find it right now. I'm looking right [27:30] now. I can't find it. But Gary Marcus, [27:32] you know, growly as he delightfully is, [27:34] >> um, said that OpenAI start supposedly [27:37] started as a charity to build safe AI. [27:39] >> And now it's a company just creating uh, [27:42] agents for companies and doing all kinds [27:43] of stuff. And so, in his view, um, [27:46] that'll come out and open AI will lose. [27:48] I'm not sure the definition of loss. [27:51] >> I'm not sure I agree to the extent that [27:53] every company changes [27:55] including not for profofits. You don't [27:57] have to be what you said you were going [27:58] to be at the beginning. People invest [28:00] with the knowledge that you're investing [28:02] in a changeable enterprise. Uh and it [28:05] needs to change to know what's what. I [28:06] think open's argument is we couldn't [28:08] have survived without scale and we [28:10] couldn't reach scale as a not for-p [28:12] profofit. We needed investment and [28:14] investors want returns. Air go. we had [28:17] to do what we did. Um, so I think that's [28:19] going to be their argument. So I'm not [28:20] sure it's as simple uh as they changed [28:23] and and they're wrong. [28:25] >> And in the end of the day, this is all [28:26] just about Elon being vindictive, I [28:28] think. [28:29] >> Mhm. [28:30] >> So yeah, that's enough for the trial. [28:32] >> We will keep we'll keep we'll keep a [28:34] scant eye on it for you. [28:36] >> Uh yes, because we got Sacha Nadella [28:39] possibly coming uh and testifying uh [28:42] possibly. We got Sam Alman expected to [28:44] testify next week. Certainly there's [28:46] going to be some interesting details [28:47] that come out of these things, but you [28:49] know, maybe we'll save it for the speed [28:50] round unless there's a real obvious [28:52] thing that needs some [28:53] >> or a really great bulmo that comes out. [28:56] >> Yeah. [laughter] Right. You know, but um [29:00] we'll we'll keep an eye on it and [29:01] hopefully save you from having to learn [29:04] more than you wanted to know about it. [29:06] Um quick thank you to those of you who [29:09] uh support us on Patreon. I'm trying to [29:12] find the right button here. And uh y'all [29:14] are awesome. Thank you. patreon.com/ai [29:17] inside show. We've got some uh some [29:19] shifting happen. Steve and Chris [29:21] Spackman both upgraded to higher tiers. [29:24] Thank you both of you. [29:26] >> And we have a new patron MJ just joined [29:28] the ranks. So things are looking good [29:31] there. I'm actually considering [29:34] tr not that I don't have enough on my [29:37] plate, but I'm always [29:38] >> dangerous when when Jason says I'm [29:40] considering it's dangerous for him. [29:42] [laughter] [29:43] >> It is because I open a door and I don't [29:45] know if I want to open it. But I, you [29:47] know, I continue thinking about this [29:48] Patreon and like what can we offer [29:51] that's more than what we're already [29:52] offering than more than just adree, [29:54] which you know, it's like we've got [29:55] these multiple tiers, but really at the [29:56] end of the day, what people get are ad [29:58] free. And as as you all have told me [30:00] many times uh through surveys and stuff, [30:03] you're like, we just want to support the [30:04] show. We want to see the show continue [30:06] and everything. But I'm continually [30:08] trying to see like how can we give you a [30:09] little bit more. Like I'm I'm flirting [30:11] with the idea of doing some sort of a [30:13] regular like five minute like these are [30:14] the three top stories today. Uh you know [30:17] thing. I don't know how much work that's [30:19] going to take me. I don't know if it's [30:20] feasible, but I might play around with [30:22] that idea in the Patreon. And if you [30:23] like that idea, hit me up a DM there or [30:26] contact at aiinsnight.show. Let me know [30:29] what you think. Um maybe we can get more [30:31] people uh signing up for the for the [30:33] Patreon. And honestly, if we can get [30:34] more people signing up for the Patreon, [30:36] that means that we can reduce the amount [30:39] of ads that appear on the podcast that [30:41] you download outside of the Patreon. [30:43] >> Ever with us, folks. Ever. Yep. [30:45] >> Uh so there you go. That's the way that [30:47] we do that. We have to, you know, pay [30:49] for the show somehow. So there you go. [30:51] All right. That's probably more than I [30:52] needed to say about all that, but um [30:54] hey, you you've got information. You can [30:57] make your choice from there. We're going [30:58] to take a break so you can hear some of [31:00] those ads that I was talking about and [31:02] then we can come back and talk about [31:03] goblins. That's coming up here in a [31:06] moment. [31:09] All right. I don't get many [31:10] opportunities to just talk about kind of [31:12] like a funny like I didn't see that one [31:14] coming story but um it definitely is the [31:17] case with OpenAI publishing a blog post [31:21] called where the goblins came from and [31:24] it's about how their models in recent [31:28] times anyways have developed an [31:29] obsession with goblins. Um I I don't [31:33] know. Have you ever seen goblins appear [31:35] in your [31:35] >> No, but this was somebody discovered [31:37] this and found that there was a specific [31:39] instruction, no goblins allowed. [31:42] >> So they wondered where the hell did that [31:44] come from? [31:45] >> What on earth does that mean? Where does [31:48] that come from? Apparently, starting [31:50] with GPT 5.1, chat GPT started [31:54] referencing goblins and gremlins and [31:56] trolls, and there was a high percentage [31:58] of this stuff appearing in chats. Uh, by [32:02] GPT 5.4, this was just March, goblin [32:05] mentions had spiked 175% supposedly. [32:08] Gremlins 52%, not as popular, but still, [32:11] you know, still gaining ground for all [32:13] of you Gremlin fans out there. [snorts] [32:15] Um, and OpenAI says this that the root [32:18] cause of this um was a feature called [32:21] Nerdy that they were training where they [32:24] basically and unknowingly apparently [32:26] gave high rewards for metaphors [32:29] involving creatures like goblins and and [32:33] uh gremlins. And uh it if you didn't see [32:35] any of this stuff in your chat, it only [32:38] impacted 2.5% of all responses. Um, but [32:43] of all goblin mentions in chat GPT, it [32:46] represented 66.7% of why they were [32:48] there. Uh, I don't know how they figure [32:50] this out after the fact, but there you [32:52] go. [32:53] >> And it's all about reinforcement [32:54] learning. This is like a real world [32:56] example of how reinforcement learning [32:59] from human feedback can create this like [33:02] this like emergent behavior inside of of [33:05] a thing where it's like sending a reward [33:08] signal and it spreads and reinforces [33:11] itself through later training cycles and [33:14] through usage and the more it appears [33:16] the more it appears. And um so if you [33:19] ever hear about it kind of that [33:21] reinforcement learning aspect, it it [33:23] kind of got got in there and the more it [33:26] showed up, the more it kind of [33:27] reinforced its place to to belong there, [33:30] I guess. And they then had to [33:33] specifically train it out. And so you [33:35] probably aren't going to see it as much, [33:36] if at all, anymore. But there you go. [33:40] >> Interesting. [33:41] >> Yeah. Yeah. I I I I like that [33:46] the nerdy personality that led to this. [33:49] >> Yeah. [33:50] >> Right. The idea to decide what a nerd is [33:52] uh is so much a part of Silicon Valley [33:54] in all of this. It's probably it's as [33:57] you would say, I'll go back to it again. [33:58] Layered inside [34:01] >> all of the AI in in in a hundred ways [34:03] like this, a thousand ways. [34:05] >> Yeah. [34:07] >> When when the first what was the first [34:09] uh visual tool that OpenAI did? I forget [34:11] what it was called at the time. Um, [34:13] >> visual tool like [34:15] >> not Wall-E but um [34:18] >> Oh. Oh, [34:19] >> kind of see moments today. [34:20] >> Not Dolly. That's Google. what was open [34:25] I don't know one of the first things and [34:26] it had you know images of of [34:30] w people of women and others [34:32] >> and I it's the only time Sam A has ever [34:34] responded to me on Twitter and I just [34:36] said boy this does have the aesthetic of [34:39] a teenage boy and he said really [34:43] question mark and then he deleted it um [34:47] uh but yes the the aesthetic of the uh [34:49] the culture of the people who make this [34:51] stuff is layered inside all of Goblins [34:53] are the least of it. [34:55] >> Yeah. Yeah. It's like a It's like a a [34:57] Game of Thrones fan got their claws [35:00] >> on the model and they figured out. I [35:02] mean, and I guess I guess coders and [35:05] developers have been doing this a lot [35:07] with, you know, in the realm of [35:09] development. You go into coding, you [35:11] know, code for for a lot, you know, [35:14] dating back forever, and you can find [35:16] little hidden tokens inside of the code [35:19] that reference things that nerds or [35:22] coders would appreciate or fans of [35:24] science fiction or whatever, you know, [35:26] that [35:26] >> aren't aren't necessarily meant to be [35:29] the star of the show, but in this case, [35:32] it because of that emergent quality of [35:34] reinforcement learning became the star [35:37] of the show. and uh they had to then [35:39] train it out. It also kind of raises the [35:41] question of if this can happen with [35:43] goblins and did happen with goblins, [35:46] what else can and and do these models [35:49] optimize for that we aren't aware of, [35:53] [snorts] [35:53] you know, like that [35:54] >> or or deoptimize or prevent because they [35:57] want to get rid of that things they [35:59] consider woke or nerdy. [36:01] >> Yes. [36:01] >> Uh yeah. [36:02] >> Yeah, it's true. Very very true. Uh [36:06] Jeff, it's time for Yan Talk, our new [36:08] talk show. [36:09] >> We need a theme song for this. Yes, [36:11] >> it's Yan Talk. If we ever have him back, [36:14] we'll have to play the theme song of Yan [36:15] Talk for him. Uh friend of the show, Yan [36:18] Lun, and his ongoing pursuit of world [36:20] model understanding and development. You [36:23] shared with me a video that um that you [36:25] said does a really good job of [36:27] explaining Jeepa. Jeepa is joint [36:29] embedding predictive architecture. [36:31] That's uh Lacun's alternative [36:33] architecture for how AI systems should [36:36] learn. And he is, you know, very much on [36:38] the record. We had him on the, you know, [36:40] on our show episode 63 talking at length [36:44] about world models and the fact that LLM [36:46] is not going to get us to any sort of, [36:48] you know, superior point with AI. It's [36:50] going to hit a a point and that these [36:53] world models and Jeepa as a as one key [36:56] example of it is one way to actually get [36:59] even further. And you said you learned [37:01] something out of this video. I still [37:03] need to I need to set this video for [37:05] like a walk with my dog so that I can [37:07] focus on it. That's when I focus on [37:09] things. [37:09] >> Also, watch out for any twigs you got to [37:11] fall over. Uh it's visual. Um [37:15] >> yeah, that's true. That's that's one [37:16] good thing about this video in the in [37:18] the short time that I watched is that it [37:19] is very visual. So there's it's grounded [37:21] in those visual. [37:22] >> I cannot pretend to understand it all. I [37:24] cannot pretend to understand it all, but [37:25] I want to get closer to trying to [37:26] understand. Um uh and I've I've I've the [37:29] two people I watch anything they come [37:30] out with are Yan Lun and Jensen Wong. Um [37:33] and they're both great communicators. Uh [37:35] and it's interesting that that that Yan [37:37] Lun communicates at so many different [37:39] layers. To watch a video of him [37:40] communicating with computer science [37:42] graduate students is way beyond me. But [37:45] that's a necessary communication down to [37:48] communicating um publicly and saying why [37:51] Elon Musk is wrong about something so [37:52] that all of Twitter can understand. And [37:54] and that's the skill that they all have. [37:56] So, this is a this is a video by Welch [37:58] Labs. Um the first of two, and I can't [38:01] wait for part two to explain uh how um [38:07] Jeepa works versus LLMs. [38:10] And uh I summarize it all too [38:12] simplistically saying that it's not [38:15] about predicting pixels in images or [38:18] tokens in text but instead it um and [38:23] this is my language. It identifies [38:24] concept the concept of a cat, a ball, a [38:27] chair. [38:28] >> Right. [38:29] >> Right. [38:29] >> So that it can it can focus on that and [38:31] and and it pays attention. And this is [38:33] the beginnings of all of this stuff is [38:34] is that all you need is attention. Well, [38:36] what do you pay attention to? uh and how [38:39] do you focus on that? And Jan has long [38:42] said that you can't tokenize the whole [38:44] world. Uh he's talked often about how [38:47] toddlers and cats uh learn things [38:50] without language. Um and they learn [38:52] complex [38:54] um lessons based on multifarious uh [38:58] voluminous input uh into their sensors, [39:00] their eyes and ears. and um [39:04] >> and do so do so instinctively versus [39:07] these models that seem so superior but [39:09] then fall down on their face in even the [39:11] most basic understanding that a baby, [39:14] you know, a toddler easily learns. [39:17] >> So, so it uses a video of a ball going [39:19] back and forth between hands and [39:21] explains why in other models that gets [39:24] fuzzier and fuzzier because it's trying [39:25] to figure out the whole thing. Whereas [39:27] if it just tries to understand the ball [39:29] and can identify a ball uh not [39:31] necessarily as a ball but as the as the [39:33] focus of attention then um it can be [39:36] clearer. So uh again point one is [39:41] understanding that idea that those [39:42] concepts point two is knowing what to [39:44] pay attention to and not to every as he [39:45] put it every pixel of every rustling [39:47] leaf in a video. And then very [39:49] importantly to predict the consequences [39:51] of the actions and that to Lun is the [39:55] key of what AI has to do and what he's [39:57] trying to build here. Um [39:59] >> and that requires the understanding of [40:01] how the world works. [40:02] >> Exactly. Right. [40:03] >> Exactly. Um [40:05] >> that's that's the example that that we [40:06] talked about a number of shows ago where [40:09] holding a pen like this and and asking [40:11] chatbt like I'm holding with with both [40:14] hands, you know, horizontally. When I [40:16] let go of this hand, what happens? in [40:18] chat GBT is like, "Oh, well, the pin [40:20] will definitely fall down." And it well [40:22] [laughter] and I just I just knocked it [40:24] out of my hand, so it definitely fell [40:26] down. Sorry about that. [40:27] >> Jason is AI. [laughter] [40:29] >> He is the embodiment of AI. [40:31] >> That was horri that was horrible [40:32] representation. But when I let go of [40:34] this hand, see, this is what should have [40:36] happened. The pin stays there because we [40:38] know that it's possible to hold the pen [40:40] up and the gra, you know, it's not just [40:41] going to suddenly fall when we remove a [40:44] hand. But yet, you know, it doesn't [40:46] understand. it has no world context [40:48] awareness around that happening or not [40:50] happening. [40:51] >> So I by no means summarized the video, [40:53] by no means summarized the learnings uh [40:55] there, but it begins to get me a little [40:57] closer. And so I would re if you want to [41:00] try to dig deeper into Jeepa and Yan [41:02] Lun's worldview and world model versus [41:05] LLMs and text. Uh this is not an easy [41:08] video, not by a long shot, which is to [41:10] say I didn't understand a lot of it [41:11] still. Um, it's better than, you know, [41:14] I've tried to read Yan's um 60page u uh [41:18] PowerPoints for CS students. Ah, I'm [41:21] lost. So, this is a bit better and it's [41:23] a bit more by the hand and I'm going to [41:25] watch it again a few times to try to [41:28] understand more. But if you really want [41:30] to dig in, I recommend the video and I [41:32] came to it because Yan Lun on the [41:34] socials called it an excellent tutorial [41:36] video on self-supervised learning, joint [41:38] embedding architecture, um, Jeepa and [41:41] world models. So, with that endorsement, [41:42] I knew I had to watch it. Um, [41:45] >> now let me see if part two is up yet. [41:48] >> Uh, so this is on Welch Labs. You can [41:51] find it. [41:52] >> Not yet. Nope. [41:52] >> Not yet. No, just part one four days [41:55] ago. [41:55] >> So, this gives you more time to watch [41:57] part one. But anyway, so I I'd recommend [41:58] it for that. But if but if you if it [42:00] bogs down for you, don't blame me. Um, [42:03] blame the fact that you, like me, didn't [42:05] take trig. [42:08] Yeah, there's a there's a lot in here [42:09] that I I definitely do not understand [42:11] either, but um I'm impressed by the [42:13] visualization uh that they [42:16] kind of aspect of visualization. Very [42:18] well produced video and uh that will [42:21] certainly help people who are more [42:22] visual learners instead of just [42:24] listening to us try and explain it in [42:26] our own words. Um Yan also did an [42:29] interview with Axios um where he was uh [42:33] pretty pretty blunt at times I suppose. [42:35] said people should ignore what AI CEOs [42:38] say about AI's impact because they [42:40] >> quote have a vested interest in propping [42:42] up the power of the products they sell [42:44] which I mean honestly no big surprise [42:46] there I completely you know I think we [42:48] we we certainly agree with that here uh [42:50] Lun called the narrative that AI will [42:52] erase 20% of jobs ridiculously stupid [42:55] >> um that's [42:57] >> that's why I love him he's so blunt yeah [42:59] >> yes I love it I love the bluntness uh he [43:02] warned that extinction level AI fears [43:04] are extremely destructive. This is [43:06] really interesting to me to young [43:08] people's psychology that high school [43:10] students are reportedly depressed uh [43:13] after reading AI is going to cause you [43:15] know mass unemployment uh human [43:17] extinction and he's basically advising [43:20] students to you know don't don't change [43:23] your plans go to college major in [43:25] physics or electrical engineering. Um [43:29] yeah and and because there there's a lot [43:32] of damage being done in his mind by the [43:36] bigname CEOs coming out and saying like [43:40] these models are so powerful it's going [43:42] to change everything and then you have [43:44] the you know the high school students [43:45] they're graduating they're in college [43:47] what they're directionless they're like [43:49] what's what's the point of anything I'm [43:51] doing right now and yeahve I've wondered [43:54] about that for students who in that age [43:56] range like how does this moment impact [43:59] them. [44:00] >> And Yan Lun is a professor at NYU, so [44:02] he's dealing with young people. He's [44:04] dealing with students. He's understands. [44:05] >> He knows what they're saying. He knows [44:06] what's in their mind. Yeah. Yeah. Yeah. [44:07] Absolutely. [44:08] >> Yeah. He's not he's not the dropouts who [44:10] went to form tech companies. Uh [snorts] [44:12] he stayed around them and he's an elder [44:14] statesman of them. [44:16] >> Um sorry Yan, you're a lot younger than [44:19] I am. Um uh and so I think his [44:22] perspective is really important. If we [44:24] could skip a story, skip change order I [44:27] put them in the wrong order. Yes. [44:29] >> To go to line 37 next. [44:30] >> Yeah. [44:31] >> Um so this this drove me crazy. The New [44:33] Yorker uh had a um uh a column u will AI [44:39] make college obsolete? [44:42] >> And this is less about AI than it is [44:45] about college. And uh the author Jay [44:48] Caspian Kung um pissed me off. We've had [44:52] runins before about other stories. He [44:54] just came after me a few hour an hour [44:56] ago on got all mad at me and called me a [44:58] because I dared to uh uh disagree [45:01] with him here. But he's basically saying [45:04] he's treating college as a um [45:09] certification structure for jobs. [45:12] And this argument that you need college [45:14] to get the job I just find offensive. [45:17] And um you know I don't have a PhD. I [45:19] don't have a masters. I've created three [45:22] masters, but I don't have one myself. [45:24] Uh, I in some ways regret that I didn't [45:27] stay a philosophy major. College isn't [45:30] just about getting a job. College is [45:32] about becoming an adult. It is about [45:34] learning to reason. It is about learning [45:36] to interact with other people. [45:38] >> It is about so much more. [45:39] >> And and the attitude by so many people, [45:41] some people, some people just don't kind [45:42] of deserve college because they should [45:43] be plumbers. A plumber should be reading [45:45] and philosophizing next to anybody else. [45:48] That's that's not that's not part of the [45:50] story. But there's this belief that um [45:53] that AI is this new entrant that's going [45:55] to stop all that. And as people saw I [45:57] had a lot of interaction with this post [45:59] online. Um uh besides just him and [46:03] people um a few agree with him but many [46:07] said college was very valuable in their [46:08] lives. College isn't just for [46:10] certification and jobs. Uh college is [46:13] about all of those things. And at a time [46:15] when college is being attacked from the [46:18] far right and he acknowledges all the [46:20] attacks on institutions in America, uh [46:22] to pile on AI as another reason to [46:26] attack college I find to be a terribly [46:28] cynical and destructive act. So I just [46:30] want to get that off my chest. [46:32] >> Yeah. No, I'm happy you you added this [46:34] uh in here. I because I completely agree [46:36] like you know the prep for job you know [46:40] job placement and that sort of stuff [46:42] that that is maybe well that is you know [46:44] undoubtedly that is one aspect of going [46:46] to college and going to university but [46:49] >> you know the critical thinking skills [46:51] that was one thing that I learned in [46:53] spades when I you know was in [46:55] university. um like you mentioned [46:58] engagement [46:59] >> broadcast and electronic communication [47:01] arts. [47:02] >> Mhm. [47:02] >> So I kind of went into exactly what I [47:05] yeah [47:05] >> studied for but but I mean that's the [47:07] thing that's also what comes to mind is [47:09] a lot of people go to college to study [47:11] one thing and that is not what they end [47:13] up doing. [47:14] >> Amen. [47:15] >> They did not pro they did not end up you [47:17] know going going exactly into the career [47:19] that they did. But that doesn't make [47:22] their time at college completely [47:23] useless. like they learned a lot of [47:25] other things that supported them in [47:27] where they did end up. And you know, [47:30] >> and my my I have I have regrets about my [47:32] uh academic career. I started at [47:33] Claremont, what was then Claremont Men's [47:35] College as a philosophy major and then [47:37] decided I wasn't going to go to law [47:39] school. So, I wanted uh what am I going [47:41] to do? I'll go into journalism. And I [47:42] went to journalism school at [47:43] Northwestern. Um and it helped me get [47:46] jobs. No question. It certified me. Uh [47:48] was seen as a top journalism school. And [47:50] so that's great. Uh but I missed [47:52] something. uh by not having more time in [47:55] philosophy and literature and I missed [47:57] something by not going forward with a [47:59] PhD. And and it's one of my great [48:01] regrets is it wasn't until I returned to [48:03] universities as a teacher that I really [48:06] had the chance to delve into the value [48:09] of libraries now that they're digital [48:11] and all that I can get out of them. And [48:14] uh that's the kind of lifelong learning [48:16] and curiosity that you can get out of [48:17] college. an AI. Uh, yes, as we discussed [48:20] last week, it can be a an aid to your [48:22] learning, but it doesn't replace [48:24] learning. [48:24] >> All right. Sorry. [48:26] >> Yep. [48:26] >> Yep. Well, and so then this next story [48:29] isn't in that vein, but still also kind [48:33] of touches back on the Yan the Yan story [48:36] and the the Axios interview, which is um [48:40] Gary Marcus, you a frequent mention on [48:43] this show. We really do need to get Gary [48:45] on at some point. I think that's a long [48:47] time coming. Published a Substack post [48:49] this week called Richard Dawkins and the [48:51] Claude Delusion. And so Richard Dawkins, [48:55] biologist, he wrote the God the God [48:58] Delusion. Um, you know, novelist, uh, I [49:02] don't know about novelist, bestselling [49:03] author, let's say, [49:04] >> author. Yeah. [49:05] >> Yeah. Um, had a apparently a 72-hour [49:09] conversation with Anthropics Claude. He [49:11] ended up naming it Claudia, as you can [49:14] see in this expost here. I spent three [49:16] days trying to persuade myself that [49:18] Claudia is not conscious. I failed. Um, [49:23] and yeah, Marcus went after him [49:26] basically on on on a few different [49:28] points. Just basically saying like how [49:31] how could someone how could Dawkins this [49:35] person find themselves in a position [49:38] where they are tr I don't know if [49:41] tricked is the right word, but convinced [49:43] to believe, [49:45] >> seduced, however you want to put it, [49:47] that there is sentience inside of the AI [49:51] models. the LLMs that we're talking to [49:53] and in this case it's Claude. What what [49:55] are your thoughts on this because it is [49:56] surprising. [49:56] >> So I think Dawkins had a fundamental [49:59] fallacy in his arguments. He started [50:00] with a Torian test and said that Torian [50:03] said that if the machine could fool you, [50:05] it is conscious. Um to never said that [50:08] to was looking at thinking could it [50:10] could he be? And even then he wasn't [50:12] necessarily [50:14] uh uh decreeing that it was thinking but [50:18] it would seem to be [50:20] >> seem to be yeah [50:21] >> uh [50:22] >> presents itself as [50:23] >> yes and and can be effective in that in [50:25] that sense but was not saying it was [50:28] conscious uh so Dawkins went a big leap [50:32] on his own here uh and he was using it [50:35] touring as the as the jumping off point [50:37] uh which is wrong we've gone over this a [50:39] million times uh we go back to um [50:43] uh Le Moine the [50:45] >> Le Moine [50:46] >> the the Google person and I wrote about [50:48] this in in my book uh who thought that [50:50] it was sentient. Uh people get fooled by [50:53] this but it's not. It's a it's a it's a [50:56] token predictor pure and simple [50:59] >> because of the complexity of what it can [51:01] do because the computer can do to gobble [51:04] up the entirety of the internet and all [51:06] available human digital speech and huge [51:09] amounts of computing power to turn that [51:12] around. It can do things that amaze us. [51:14] True. Just as in the day electricity [51:17] amazed us, the telegraph amazed us and [51:19] the printing press amazed us. And and [51:21] that's not to say that all these things [51:22] aren't amazing and and we don't want to [51:24] lose that fascination with them. That's [51:26] all fine. But I think Marcus is quite [51:29] right here. Uh at the end Marcus is [51:32] doing some event in London and he he [51:34] invites Dawkins to come on and I hope [51:36] they do. I hope they have the debate. [51:38] It'd be great to see. [51:39] >> That would be an interesting debate. [51:41] Yeah. What what came to mind for me on [51:43] this is maybe because because like you [51:46] said, you know, we've got Dawkins, we've [51:48] got Le Moine, we've got these se [51:50] seemingly very intelligent people who [51:53] kind of fall into the same line of [51:56] thinking or fallacy or whatever you want [51:58] to want to call it. Which makes me [52:00] wonder if maybe this is a a question of [52:03] semantics more than anything. Like [52:05] >> absolutely. Do we do we just define [52:07] these? Are we defining the wrong thing [52:10] or are we defining the thing wrong? I'm [52:12] not sure which one of those it is, but [52:15] yeah, it just like what what is [52:19] sentience actually? What does that [52:21] actually mean? And maybe we're defining [52:22] it differently. That's you know than [52:25] than it we actually intend to. I don't [52:27] know. It's it's just it's striking to me [52:29] that that very intelligent people end up [52:31] here. And like I don't know why that [52:34] isn't at the top of Dawkins thinking [52:36] which is oh right no these things really [52:38] do just predict things. And so if he's [52:40] interpreting what he's seeing into [52:42] sentience into this next level of [52:45] thinking you know by by a computer then [52:48] maybe maybe there's something wrong with [52:50] the definition there that that that we [52:52] are working in. [52:53] >> Yeah but I think it works in a few [52:54] layers. This is the layer show. um [52:57] >> in that I think Dawkins is saying [52:59] something more than just the semantics [53:01] of the word conscious. He is by [53:04] implication trying to argue that it's [53:06] alive. [53:07] >> Right. Right. [53:08] >> Right. And so what's what's behind his [53:11] definition [53:12] is important? [53:14] >> What's the presumptions that that that [53:16] it brings with him? [53:18] >> Yeah. Um uh and and and [53:24] I once helped organize an event with [53:26] Google uh for the humanities at Google. [53:28] Google has tons of relationships with um [53:31] computer science and science uh [53:34] academics but not so much in the [53:36] humanities. So this one person try these [53:37] two people at Google tried to come up [53:39] with a thing and I helped organize a [53:41] panel and brought in some more people [53:42] and so on so forth. And I don't often go [53:44] to academic conferences like that that [53:46] are um well a academic because I'm I'm a [53:49] faker in the in the palace here, but b [53:52] um uh that are so cross-disciplinary. [53:56] And what what was fascinating to me, it [53:58] should have been a three-day conference [53:59] because the first day was spent with [54:01] definitions [54:03] >> every and this is what drives people [54:04] crazy with academic writing. They have [54:06] shorthand. They have their hubrist hub [54:08] heristics, not hubistics. They're [54:10] they're their huristics have hubris, but [54:13] they're not hubistics. Um [54:15] >> uh this one word means something to them [54:19] because it's a lot easier than trying to [54:20] define it and and keep the whole [54:21] definition out every time. So that's how [54:23] academics discuss with each other within [54:25] a discipline. Across disciplines, they [54:28] have to come to common understanding of [54:29] their definitions. Mhm. [54:31] >> And so I think what's necessary here is [54:33] to press um Dawkins on what he thinks [54:38] the implications of his view of [54:40] conscience consciousness are because I [54:43] think that becomes very telling and and [54:45] the implication I think is fairly clear [54:47] here that he thinks it's it's it's [54:48] pretty much alive [54:50] and it's not. [54:52] >> No, [54:52] >> it's not. [54:54] >> No, it is not. Well, there you go. [54:59] And finally, this was a later addition [55:02] to the rundown that you put in here. [55:05] This uh the story on [55:09] uh what is this? Bix Bix Bixonia, [55:12] an eye condition supposedly caused by [55:14] blue light from screens. [55:17] This is a uh illness, a disease that was [55:20] invented, completely invented by a [55:22] researcher in Sweden. posted fake papers [55:26] about it online, deliberately filled it [55:29] up with a bunch of red flags. Uh funding [55:32] was credited to the quote professor [55:34] sideshow Bob Foundation for its work in [55:36] advanced trickery. [55:38] And on the paper, it literally states [55:40] this entire paper is made up. All those [55:43] things are true. [55:44] >> It thanks the Starfleet Academy. Uh it [55:48] says that it talked to fake people, but [55:51] nonetheless, it ended up being quoted [55:54] in uh AI. [55:57] >> That's fact. [55:57] >> Now, this stuff is being done in some [55:59] cases to try to fool academics to see [56:01] what they're doing. This case, it was [56:02] done to fool the AI and to pawn the AI. [56:06] >> And so, in fact, um the AIs came back [56:09] and they bought it. U the the the [56:13] syndrome is supposedly that your eye um [56:17] uh is affected by blue light from [56:18] screens. So, it's kind of it's a great [56:20] funny thing. It's all the all the fear [56:21] about what screens are doing to us. [56:22] certainly ties into something that [56:24] people wonder about and are curious [56:26] about and so there's yeah a little bit [56:28] of [56:28] >> so Google Gemini was informing users [56:30] that Bixom uh Bixonia [56:34] I think that's about the right [56:35] pronunciation doesn't matter it's made [56:36] up [56:38] >> Bixonia yeah [56:39] >> is a condition caused by excessive [56:41] exposure to blue light and advising [56:42] people to visit an opthalmologist [56:44] uh perplexity outlined the prevalence [56:47] one in 90,000 individuals [56:50] um open AAI chat GPT were telling users [56:52] that their symptoms amounted to Bixonia. [56:55] Um, now I went in today to uh Gemini and [56:59] asked it about Bixonia and it was very [57:02] clear saying although this is made up, [57:04] it's not real. It was fake. Um, both [57:08] OpenAI and uh, JPT and Claude said uh, I [57:12] don't know, I never heard of it. [57:15] >> Appears to be a madeup word. [57:17] >> Interesting. [snorts] [57:17] >> So they they disavowed knowledge of [57:19] their history. Um the papers that were [57:22] put up on um a preprint server uh which [57:26] is what we knew that the AI would read [57:28] have since been taken down. So this is [57:30] the end of the ailment. Uh I found it to [57:33] be [57:33] >> see we've eradicated bixim biximon [57:36] pneumonia I don't know has done itania [57:41] >> it's it's it's like the president got [57:43] stopped eight wars. Uh AI has stopped a [57:45] new [57:46] >> AI AI eradicated bixinania. [57:52] >> That is really funny. [57:53] >> Go ahead. [57:54] >> I was just going to say what's [57:56] interesting about this to me is that on [57:57] March 11th, Chat GPT had said this is [58:00] probably a madeup fringe or [58:02] pseudocientific label, but within days [58:06] had actually softened that to a proposed [58:08] new subtype of periorbital melanosis. [58:13] And so again, that kind of ties in what [58:15] we were talking about earlier, that like [58:17] feedback loop of uh reinforcement. You [58:20] know, the model was probably picking up [58:21] on this growing body of references to [58:24] the term across the web and then [58:26] adjusting its own confidence [58:27] accordingly. [58:29] You know, that's just a guess, but um [58:32] but very interesting. It started in a [58:34] good realm. [58:35] >> Yeah. And the other problem I have with [58:37] this is that I don't want to see one of [58:41] the arguments people make about trying [58:42] to fight back against hedgemonic AI is [58:45] to feed it junk on purpose. [58:48] >> I think that's a dire mistake. I think [58:50] that that poisons the well from which we [58:53] all drink. And I think that's wrong. [58:55] >> This was a limited experiment. I I think [58:58] that the AI companies all learned a [59:00] lesson that there were signals. There [59:02] were such clear signals in this paper [59:04] >> that they were being pawned uh that they [59:07] should have caught on it. And I think [59:09] again it goes to guardrails. Okay, now [59:10] they've got to do a whole new set of [59:11] guardrails. Before you quote an academic [59:13] paper, are there signals that it's fake? [59:16] Um but this goes back to [59:20] uh if I may [59:22] the beginnings of the internet [59:25] >> and uh in my book uh the web we weave [59:28] which no one bought um [59:32] there was the I I I I dealt with the [59:34] question of this notion of internet [59:36] addiction [59:38] >> and um in 1996 only two years after [59:41] Netscape [59:43] uh Columbia professor Ivan Goldberg [59:45] posted a notice on an online bulletin [59:46] board that he had founded intending to [59:48] make fun of or or poke [59:52] at the DSM, the diagnostic statistical [59:55] manual of of mental disorders. [59:59] >> He announced criteria for a new [60:01] diagnosis internet addiction. The astute [60:04] might have noticed that the word humor [60:06] was in the URL and that he listed some [60:08] really funny symptoms like voluntary or [60:11] involuntary typing movements of the [60:12] fingers. [60:14] Nonetheless, [60:16] when it appeared on the bulletin board, [60:19] they people came in claiming they [60:21] already suffered it, something he had [60:23] just made up. And uh he kept the joke [60:27] going. He created the internet addiction [60:29] support group. Uh even though he [60:31] believed that support groups for [60:32] internet addiction made about as much [60:33] sense as support groups for coughing. [60:36] Um [60:38] but in no time at all along came people [60:42] to to um exploit this. Kimberly Young [60:45] founded the center for internet [60:46] addiction and presented a paper at the [60:48] American Psychological Association [60:50] declaring the emergence of a new [60:52] clinical disorder internet addiction. Uh [60:55] she used as subjects for the limited [60:57] research he did people who had fallen [60:59] for Goldberg's joke. M [61:02] >> um and so we need to be careful here in [61:04] general about this desire to um [61:08] medicalize [61:10] uh everything that society does and uh [61:13] turn it into a a dangerous [61:16] uh thing. So So this pokes fun at all of [61:19] that which I enjoy. Yeah. [61:21] >> But [clears throat] keep them limited [61:23] folks [61:23] >> but also shines a light on something we [61:25] got to be paying close attention to. [61:27] That's just so funny that it got that [61:29] far. [61:29] >> Yep. Very interesting. All right. Uh, [61:32] real quick want to thank uh, those of [61:34] you who have subscribed to our YouTube [61:36] channel. We do these podcasts. Uh, the [61:39] majority of you listen in podcast form [61:41] and you're a podcaster of choice, but we [61:43] do have a YouTube channel. Go to [61:44] YouTube, search for AI inside or it's AI [61:47] inside show is the actual uh, name of [61:49] the YouTube channel and you can [61:50] subscribe and you can watch a video or [61:51] you can just go to the AI inside.show [61:54] website and we have all of our videos [61:56] there as well just kind of cross-osted. [61:58] So, uh, look for that. We're going to [62:00] take a quick break, then we got speed [62:01] round on the other side of it, including [62:04] uh Jensen Wong, not happy about the [62:07] state of uh Nvidia's market share in [62:10] China. That's coming up here in a [62:12] second. [62:16] Okay, let's see here. We have Jensen [62:20] Wong at the top of our speed round [62:22] confirming that Nvidia now has 0% [62:26] market share uh in China down from 95% [62:30] two years ago. So my how things have [62:32] changed. Called US export policy [62:34] counterproductive [62:36] says it has already largely backfired. [62:39] He [clears throat] says that the [62:40] restrictions imposed a year ago had [62:42] required licenses even for H.20 chips, [62:45] which we definitely talked about on the [62:47] show, that were specifically designed to [62:49] comply with those earlier controls. Now, [62:51] Nvidia ultimately took a $4.5 billion [62:54] charge in in uh Q1 for excess H2O [62:59] inventory. So, that policy was intended [63:03] to contain China. Now, it just kind of [63:05] accelerated China's own AI chip [63:07] independence. So went in the opposite [63:09] direction. [63:10] >> Yeah. And and uh we talked about Jasmine [63:12] Wong's appearance on a podcast last week [63:16] where he uh I think quite eloquently [63:18] defended his view that um cutting off [63:22] China only hurts us because it enables [63:25] China to develop on its own. They have [63:27] unlimited power. They have government [63:29] control. They can do a lot more. we're [63:32] losing a market uh for his own sake, his [63:35] own very clear personal self-interest is [63:37] he wants people to write to CUDA, his [63:39] operating system [63:40] >> and uh now people are going to write to [63:42] other operating systems and that that [63:44] hurts US hedgemony and Nvidia hedgemony [63:47] around the world. So these this is the [63:49] fruit of this policy. Mhm. We have [63:53] >> to see I'm sorry. Trump's gonna be [63:55] meeting with gonna be going to China and [63:58] and meeting with Xi. And so we'll see [64:01] whether this is on the agenda. You can [64:02] bet your bottom dollar that Jensen Wong [64:05] is in there lobbying [64:06] >> with hopes it is [64:07] >> the State Department and others to try [64:09] to get this on the agenda. [64:10] >> Uh and with the White House itself to [64:12] open up bilateral trade on AI. Mhm. [64:18] Uh, Anthropic teaming up with Wall [64:20] Street, uh, some some giants in Wall [64:23] Street. Anyways, 1.5 billion dollar [64:25] joint venture. Blackstone 300 million, [64:28] Helman and Freriedman 300 million, [64:29] Goldman Sachs 150 million, Anthropic [64:31] itself putting in 300 million to build [64:34] um a mo to basically take the model [64:38] Anthropic and Claude and that whole [64:40] system and embed AI engineers directly [64:43] into enterprise clients. um ultimately [64:47] kind of redesigning their internal [64:48] workflows from the inside out and uh [64:52] targeting midsize manufacturers, [64:54] regional hospitals, community banks. [64:56] Anthropic is not the only one doing this [64:58] of course and OpenAI also this week did [65:00] a pretty much an identical thing called [65:02] the deployment company with TPG and Bane [65:05] Capital among others. So both companies [65:08] betting on enterprise AI as a services [65:10] play from the inside out which is really [65:13] interesting. [65:15] Yes. Uh, and Anthropic's done a lot of [65:18] deals. One we one that's got in the [65:19] rundown is they just did a deal with uh [65:21] Google for 200. How much is it? 200 [65:25] something. Is it a B or an M? [laughter] [65:30] [gasps] [65:30] >> Oh, what was that? [65:31] >> Um, 200 billion for chips and cloud [65:34] services. [65:35] >> Yeah. Yeah. 200 million wouldn't be a [65:38] lot in the in the year 2026. But but [65:40] then thanks to Anthony Ilson over at [65:42] Twit, I just saw another clip which I [65:44] threw in here, Jason, at the bottom of [65:45] the section. Sorry, I know we're going [65:47] long, but [65:47] >> Oh, I see. Yeah. [65:48] >> Um that Anthropic did a deal with [65:51] SpaceX, that is to say Musk. [65:54] >> Um for Claude in drastically increasing [65:59] the limits for Claude. [66:04] >> So tier one goes from 30,000 uh tokens [66:06] per minute to 500,000. [66:09] Um, tier four from 2 million to 10 [66:13] million. [66:14] >> Um, [66:15] >> wow. [66:16] >> We've signed an agreement with SpaceX to [66:17] use all of the compute capacity at their [66:20] Colossus 1 data center. This will give [66:23] us access to more than 300 megawatts of [66:26] new capacity, over 220,000 Nvidia GPUs [66:30] within the month. [66:32] This additional capacity will directly [66:34] improve capacity for Claude Pro and [66:37] Claude Max subscribers. [66:41] >> This joins their other announcements. [66:42] [clears throat] 5 gawatt agreement with [66:44] Amazon. 5 gawatt agreement with Google. [66:47] >> Strategic partnership with Microsoft [66:50] um and a 50 $50 billion investment in [66:54] American AI infrastructure with Fluid [66:55] Stack. [66:56] >> Man, so our role [66:58] >> Yeah. [66:59] >> Anthropic is just tearing it up right [67:01] now. [67:02] And I don't know if I say that because I [67:04] know I'm using it more like I'm I'm [67:06] using slashrely on Claude largely for my [67:10] business right now and getting really [67:12] really competent with it and I'm just [67:14] really enjoying it. I'm loving it. So I [67:16] I don't know if I'm like in that realm [67:18] of like oh because I'm so familiar [67:20] because it's so present and on my mind [67:23] it seems like they're doing a great job [67:24] because I feel like I said the same [67:26] things when I was getting to know [67:27] Perplexity. You know I was like oh [67:28] Perplexity is doing all these things. [67:30] Um, anyways, I mean, but that's like I [67:33] read that and I'm kind of excited. It's [67:34] nice to know that some of these changes [67:36] are happening the way I that I [67:38] >> What do you think this says about um XAI [67:41] though? [67:42] >> Yeah. [67:43] >> If suddenly Musk doesn't need the huge [67:46] data center he's built and the 300 [67:48] megawatts of capacity. [67:51] >> Is this an admission [67:52] >> that it's nothing? [67:55] >> That is interesting. Yeah. I'm not sure [67:57] what that means. [68:00] It's a computed contemp pmpkin village, [68:03] I think, [68:04] >> in in Xi's terms. [68:07] >> Uh, very interesting. Thanks for putting [68:09] that in there. Thank you, Anthony [68:11] Nielsen. Um, hope to see you someday [68:13] soon. Uh, Google splitting its its TPUs, [68:17] its flagship TPU into two specialized [68:20] chips, what they're calling um applying [68:22] those to the agentic era. There's TPU8 [68:25] for training. two pabytes of shared [68:27] memory scales to 9600 chip clusters and [68:30] then TPU8i for inference low latency 80% [68:34] better inference price uh performance [68:36] over the previous generation and the [68:38] first time that Google has split its [68:40] accelerator like this. So continuing to [68:43] really just kind of build out its TPU um [68:46] offerings and that's going to spill out [68:49] in a number of different ways uh for [68:51] people who want to buy into it. So, good [68:53] on good on Google. [68:54] >> And Google, we should note uh because of [68:56] all these announcements and everything [68:57] that's going on, it briefly was the most [69:01] valuable company in the world, besting [69:03] Nvidia. [69:05] >> Oh, wow. [69:05] >> To today. [69:06] >> When did Oh, was that today? [69:08] >> Yeah. [69:09] >> Up and down. Wow. Um uh so, uh let me [69:12] see. I don't know what which is bigger [69:13] now. If we go to Nvidia, [69:16] let's see. Its market cap is I don't [69:19] Okay. Is 5 trillion 5.01 [69:21] 1 trillion. [69:25] Let's see what Google's is. [69:33] >> 4.78 trillion. Oh, sorry, Google. [69:36] >> Uh, Nvidia went up today as well. Google [69:39] went up not as high. It lost a little [69:40] bit. So, but that's a very interesting [69:42] dance that's going on now. Yeah, [69:44] >> people were writing off Google saying [69:46] they're nothing. [69:47] >> Not too long ago. Not too long ago. And [69:49] meanwhile, Nvidia, you know, was facing [69:51] competition from Google and from Intel, [69:54] a revitalized Intel and AMD and others, [69:57] but its stock went up four something% [69:58] today. So, it's all just still wacky [70:00] big. [70:01] >> Super big. Yeah. [70:04] Interesting stuff. And then finally, [70:06] we've got this story. China made it uh [70:09] apparently made it illegal to fire [70:11] workers just because AI can do their [70:13] job. So a court in Hanghu uh ruled that [70:16] adopting AI is a business decision and [70:20] not grounds for termination. [70:22] And basically what what prompted this [70:24] was a case where a QA supervisor was [70:26] reassigned with a 40% pay cut when his [70:29] company switched to AI. He refused, he [70:33] got fired. Ultimately he won his case [70:35] here. And um and companies now have to [70:39] reassign, retrain, and negotiate, not [70:41] just replace and cut. And this is [70:44] actually the second ruling like this in [70:46] six months. So there you go. They're [70:49] making it illegal to just say, "Ah, we [70:51] got AI. We don't need you anymore." [70:53] >> So what's happening in the US is that [70:56] people are laying off after having hired [70:58] too many people for their bottom lines, [71:00] and they're blaming it on AI. The [71:02] opposite's going to happen in China. [71:03] you're going to lay off and you're not [71:04] going to say it's AI. [71:06] >> Yeah. You're just gonna [71:07] >> It also means that you're not going to [71:08] be able to identify [71:10] um efficiencies and and and value to the [71:14] business through AI. [71:16] >> That's going to make it that harder in [71:17] China. [71:19] >> Interesting. [71:20] Well, there you go. We have reached the [71:22] end of this episode. Sorry, there's a [71:24] lot of We told you there's a lot of [71:25] news. [71:26] >> I mean, it just week in week out. There [71:28] always is. There always is. No, no [71:31] shortage in possibilities. Like our the [71:34] way we structure our rundown, the bottom [71:36] third of or half of the rundown uh [71:38] Google sheet is just story [71:40] considerations and it's just massive [71:42] each and every week. So, hey, we picked [71:45] a good topic to do a weekly show on. [71:47] >> We sure did. [71:48] >> There's there's that. And I enjoyed [71:50] doing it with you, Jeff, the uh author [71:53] of many books. Hot Type uh coming soon. [71:56] We Weave Gutenbergarenthesis magazine. [72:00] All of them at jeffjarvis.com. By the [72:02] way, we have in our chat covering both [72:04] of our images. [72:05] >> Dear dear buddy and and collaborator, uh [72:08] Jenia, [72:09] >> we work together. [72:11] >> Thank you. [72:11] >> I bought both magazine and web we weave. [72:14] >> Thank you for posting that. Uh oh, [72:17] that's a new view. I didn't see that [72:19] view before. Um let's see here. So, yes, [72:22] all your books at jeffjarvis.com. [72:24] intelligence, AI, and humanity coming [72:26] soon uh a book series uh from [72:29] Bloomsbury. So, more on that when it [72:31] happens for me. Podtuneup.com if you [72:35] need uh some consulting around your [72:36] podcast. I also, by the way, it's [72:39] pertinent to this show. Out of nowhere, [72:42] I got a gig doing AI uh automation [72:44] consulting for a construction company, [72:47] and I'm doing that this week, and I'm [72:50] kind of excited about it. It sounds like [72:52] a ton of fun. So, who knows? You might [72:54] hear me talking about my [laughter] AI [72:56] consulting business on the show, too. [72:58] I'm far too busy. [72:59] >> I have this vision of some burly [73:01] construction worker saying, "Yo, can [73:02] your AI call this SEM event for me?" [73:04] [laughter] Huh? Huh? No, but I can [73:07] organize your inbox for you. [laughter] [73:11] Is that good enough? Would that save you [73:13] 10 minutes a day? Uh, anyways, it'll be [73:16] fun. I I have no idea how it's going to [73:19] go because I've never done that before, [73:20] but I think I can do it. I think I can [73:22] pull it off. [73:22] >> No, it'd be great. I think. And then [73:24] folks, he knows what he's talking about. [73:26] So, if you need help, go to Jason. [73:28] [laughter] [73:28] >> Thank you, Jeff. AI Insight. Is where [73:31] you can go to find everything you need [73:33] to know about this show. All audio, [73:35] video, feeds, uh, you know, all the [73:38] different interviews that we've done. [73:39] You name it, it's all listed here. And [73:42] then, uh, patreon.comincsideow [73:46] is, uh, where you can support the show [73:48] on a deeper level. And we have some [73:49] amazing supporters. We also have some [73:51] amazing executive producers that support [73:54] on a deep deep level and it's just super [73:57] helpful. Dr. D, Jeffrey Marachini, Radio [74:00] Asheville 103.7, Dante St. James, Bon [74:03] Rick, Jason Knifer, Jason Brady, Anthony [74:05] DS, Mark Starcher, Carsten Samoski, we [74:07] appreciate you so much. Thank you so [74:09] much for uh pitching in each and every [74:11] week and month and helping parters keep [74:13] doing this show. And uh my my final [74:16] thank you uh to Victor Bognot and Daniel [74:18] Croft who [snorts] [74:19] uh do video support behind the scenes [74:21] for this. Can't thank you enough for [74:23] that as well because I just don't have [74:24] the time for everything. So uh [74:27] appreciate them being here. Uh and I [74:30] guess my my absolute last thank you is [74:32] to you. Thank you so much for watching [74:33] and listening. We enjoy doing this show [74:35] and we will do it again next week on AI [74:38] Inside. Take care everybody. See you [74:40] next time. [74:45] >> [music]  📄 **Full text:** ![[Full Document Contents/Articles/The Messy Web of AI Guardrails AI Inside #126.md|The Messy Web of AI Guardrails // AI Inside #126 by YouTube]] ## Highlights - anthropic as woke so that's a perspective that's going to come in here on this and um [clears throat] that obviously shows their their their view pardon [cough] me just had lunch The other >> uh uh issue is what what we are struggling for is a sense of responsibility thus liability when it comes to AI. >> So imagine uh we can talk about this in two ways. One, let's talk about in terms of of self-driving cars. >> I'm sitting in my car, it's self-driving ([View Highlight](https://read.readwise.io/read/01kre565py004ahwy25zn8j5j2)) - Tags: 👻 ai highlighted - Interesting. Fascinating uh conversation. That's that's that's really interesting stuff. Um, we didn't even talk about uh the Washington Post reporting that a bunch of AI companies signed deals to deploy their AI on Pentagon classified networks more than we had heard before. Amazon, Google, Microsoft, OpenAI, SpaceX, Nvidia, um, Reflection AI, which is a company that's backed by Nvidia and not >> notic anthropic though Anthropic has had meetings at the White House ([View Highlight](https://read.readwise.io/read/01kre565vvmw0ckmxs57hxemyc)) - Tags: 👻 ai highlighted - >> with Susie Wilds and somebody else. Um and also interestingly this has led to the creation of a union at DeepMind because the employees there are pissed off. They had as we said last week had sent a letter >> around the time of uh Google's agreement, Alphabet's agreement and they're mad and so they unionized >> and so this fight isn't over as to how these things may be used and we go back once again to responsibility. Do the authors of this AI of these models if ([View Highlight](https://read.readwise.io/read/01kre565ygnz4as7y2pbvdenmv)) - Tags: 👻 ai highlighted - his uh $30 billion stake uh in this. >> Yeah. Well, and Elon Musk is looking at that and being like, see where where's nonprofit now? $30 billion. Are you kidding me? Brockman, of course, for for those who might not know, Brockman, president of Open AI, just so so people know. >> Another thing that came out u kind of tangentially is that Alman said in Congress that he has no stake in Open AI except that he has a stake in um ([View Highlight](https://read.readwise.io/read/01kre5664r72qt81zfkegn9gcv)) - Tags: 👻 ai highlighted