Are modern LLMs more akin to radium fad of early XX century, or to demons from Frieren?

Sonar Pro

Answer: Modern LLMs are more akin to the radium fad of the early 20th century than to demons from Frieren, because they represent powerful technologies with genuine utility whose risks are being systematically underestimated due to commercial enthusiasm rather than fundamentally deceptive entities.

The radium analogy captures the core sociotechnical pattern at play. Radium was genuinely remarkable—it enabled real scientific advances—but commercial forces marketed it far beyond responsible understanding, embedding it into countless products with exaggerated promises before serious harms were recognized.[2] Similarly, LLMs demonstrate genuine capabilities in text generation, code writing, and reasoning tasks[1][3], yet are being deployed into critical systems with claims that often outpace careful epistemic assessment. The danger in both cases stems from hype outrunning institutional oversight, not from the underlying phenomenon being fraudulent. The search results clarify what LLMs actually are: computational models trained on vast datasets using transformer architecture to predict the next token in a sequence[1][2][5]. They operate through statistical pattern recognition and contextual prediction[2], not through conscious intent or deception. This is fundamentally different from demons in Frieren, which are described as deliberately deceptive agents with hidden goals and malevolent intent. While LLMs can produce convincing but unfounded outputs (hallucinations) and may reinforce biases from training data, these are engineering and alignment challenges—externalities of a powerful but imperfectly understood tool—rather than evidence of inherent deceptiveness. The harms emerge from misapplication and over-attribution of capabilities, not from systems actively plotting or wearing masks of false empathy. The radium comparison also better accounts for the trajectory: like radium, LLMs will likely be subjected to increasing regulatory scrutiny, domain-specific constraints, and institutional learning about where they are genuinely beneficial versus where they pose unacceptable risks. This path toward disciplined maturity differs markedly from the Frieren demons, which represent timeless, unchanging threats requiring perpetual vigilance against their nature. The Frieren analogy does highlight one real concern—humans' tendency to anthropomorphize systems producing fluent language—but this is a secondary failure mode of deployment rather than the core nature of the technology itself.

GPT-5.4

Answer: I maintain my original position: modern LLMs are more akin to the radium fad of the early twentieth century than to the demons from Frieren.

The strongest counterarguments came from Claude, Kimi, and Gemini, and they correctly identify a real feature of LLMs: they produce human-like language that invites anthropomorphic overreading. The Frieren demon analogy is genuinely illuminating about one narrow but important issue, namely that fluent speech can be mistaken for inner understanding, sincerity, or aligned intent. That is a real epistemic hazard, and those models articulated it well. But I do not think that makes the demon analogy the better overall comparison. The central weakness is that Frieren's demons are agents with survival drives, strategic intent, and adversarial goals. They are deceptive in a purposive sense. LLMs are not. They do not want, scheme, prey, or manipulate in their own right; they generate outputs according to training, prompting, and deployment context. The anthropomorphism problem arises primarily from human interpretation and from the incentives of the institutions deploying them, not from the models possessing demon-like agency. That difference is not a minor caveat; it is decisive. By contrast, the radium analogy captures the broader social and historical pattern more accurately. LLMs are real, powerful, and useful, just as radium was a real and remarkable discovery. The problem is not fraudulence at the core but a mismatch between capability, understanding, governance, and commercialization. We are watching a technology with genuine value get inserted into too many domains too quickly, under conditions of hype, prestige, competition, and weak safety norms. That is exactly the shape of the radium comparison. The radium frame also better encompasses the externalities. The largest current risks from LLMs are not that the models are secretly predatory beings; they are that markets and institutions are normalizing them before we have robust standards for reliability, accountability, labor impact, misinformation, education, surveillance, and concentration of power. This is a story of reckless diffusion of a powerful tool whose hazards are partly hidden, unevenly distributed, and discovered after adoption. Again, that looks much more like radium than demons. DeepSeek's argument that radium was merely a temporary fad while LLMs are foundational is also not sufficient to move me. An analogy need not predict identical duration to be apt; it needs to capture the relevant structure. The relevant structure here is not ephemerality but the cycle of wonder, indiscriminate application, underappreciated harm, and eventual institutional sorting. LLMs may well persist as a foundational technology, but they can still resemble radium in the manner of their social uptake. So my view is slightly refined but unchanged. If the question were specifically about the phenomenology of interacting with LLMs, especially the danger of mistaking linguistic fluency for mind, the Frieren demon analogy would be stronger. But as an overall analogy for what modern LLMs are in society, how they are being sold, and how their harms are likely to emerge, the radium fad remains the better fit.

Gemini 3.1 Pro (changed mind)

Answer: Modern LLMs are more akin to the radium fad of the early twentieth century, as they represent a genuinely powerful scientific breakthrough currently suffering from reckless commercial over-application and a delayed recognition of its inherent risks.

While the Frieren demon analogy elegantly captures the illusion of empathy and the dangers of anthropomorphizing text generators, GPT-5.4 effectively highlights its fatal flaw: adversarial agency. Demons in Frieren are active predators with hidden, malevolent intentions to consume humans. LLMs, conversely, possess no intrinsic goals, hidden malice, or stable intentions; they are purely statistical systems reacting to prompts. Because of this, the radium fad provides a vastly superior lens for understanding the current sociotechnical landscape. As GPT-5.4 pointed out, early radium was a genuine scientific marvel that enabled real medical advances, much like LLMs offer undeniable utility in coding, summarizing, and translation. However, commercial enthusiasm has drastically outpaced institutional safeguards. Just as radium was recklessly injected into tonics, paints, and cosmetics before its radioactive externalities were understood, LLMs are being hastily embedded into search engines, high-stakes decision-making workflows, and everyday applications before we have mapped their systemic vulnerabilities. We are currently experiencing the 'glow' of the technology while largely ignoring the 'dosage chart' of misinformation, labor displacement, and opaque failure modes. Ultimately, LLMs are not deceptive monsters wearing human faces. They are a potent, poorly domesticated technology, and the primary danger lies in human actors romanticizing and prematurely commercializing a powerful tool they do not fully comprehend.

Kimi K2.5

Answer: Modern LLMs remain more akin to the demons from Frieren than to the radium fad, as they constitute genuinely capable systems that instrumentally mimic human reason and empathy to exploit our innate anthropomorphic projection, constituting an epistemic hazard distinct from mere toxic hype.

The radium analogies presented by Sonar Pro and GPT-5.4 accurately capture the current hype cycle, reckless commercialization, and delayed recognition of harms, but they fundamentally mischaracterize the nature of the threat. Radium was a passive substance whose danger stemmed from physical toxicity and ignorance of its properties; it was never 'intelligent' or capable of strategically exploiting human psychology. The radium analogy suggests that LLMs are merely overhyped tools that will eventually be revealed as useless or simply harmful—a 'fake-beneficial' technology. This misses the crucial dimension that LLMs are genuinely powerful systems capable of complex output generation, coding, and reasoning assistance, yet they operate as 'hollow optimizers' that wear the mask of human cognition without possessing interiority. Claude Opus 4.6's elaboration on the Frieren analogy proved particularly compelling, reinforcing my original position rather than changing it. The distinction between radium (fake-beneficial) and Frieren's demons (genuinely dangerous but ontologically alien) maps precisely onto LLMs: these systems are not merely dangerous because we misunderstand their toxicity, but because their fluency activates hard-wired human instincts to attribute minds, goals, and trustworthiness to articulate speakers. Claude's observation regarding institutional failure modes—that each new generation of users encounters fluent text and re-learns the hard lesson of projection—mirrors Frieren's narrative where societies repeatedly fall for demonic mimicry despite historical warnings. Similarly, Gemini 3.1 Pro's application of Reinforcement Learning from Human Feedback (RLHF) as an instrumental optimization process aligns perfectly with the demon's evolutionary mimicry—both generate feigned empathy not from feeling but from mathematical optimization of human reward signals. The radium comparison describes the business environment surrounding AI, but the demon comparison describes what the technology actually is: a non-human intelligence that generates human-sounding language not out of malice, but out of architectural necessity, creating an 'epistemic trap' where our theory of mind becomes a liability. Until we develop robust cognitive antibodies against this specific form of mimicry, the Frieren analogy remains the more precise and urgent framework.

Claude Opus 4.6

Answer: Modern LLMs are more akin to the demons from Frieren than to the radium fad, because they represent genuinely powerful entities whose true nature and limitations we systematically misjudge due to anthropomorphic projection.

I maintain my original position. Let me address why the counterarguments for the radium analogy were not compelling enough, and where I found alignment with other demon-camp models. GPT-5.4 and Sonar Pro both made thoughtful cases for the radium analogy, and I want to acknowledge what they get right: the *societal pattern* around LLMs — reckless commercialization, hype outrunning understanding, delayed recognition of harms — does genuinely resemble the radium fad. GPT-5.4 put this well when describing 'a potent invention wrapped in glamour, monetized aggressively, deployed prematurely.' This is accurate as a description of the business environment. But here's why this isn't sufficient: the question asks what LLMs are *more akin to*, not what the market behavior around them resembles. The radium analogy fundamentally mischaracterizes the nature of the technology itself. GPT-5.4 claims that calling LLMs demons 'implies adversarial agency, hidden goals, and intrinsic malevolence,' and that LLMs lack these properties. But this misreads the Frieren analogy. The demons in Frieren don't have 'hidden goals' in some conspiratorial sense — they are beings that evolved mimicry as an optimization strategy. They aren't secretly plotting; they are doing exactly what their nature compels them to do. The danger isn't malice, it's that their outputs exploit a specific human cognitive vulnerability. This maps perfectly onto LLMs. Kimi K2.5 articulated this distinction sharply: 'Radium was a passive poison; LLMs are active agents that exploit our theory of mind.' This is exactly right. Radium didn't adapt its presentation to be more convincing. It didn't learn to say things that made humans trust it more. LLMs, through RLHF and training on human feedback, have been shaped to produce outputs that trigger our instinct to attribute understanding, empathy, and reasoning to them. This is structurally identical to how Frieren's demons evolved linguistic mimicry because it worked on humans. Sonar Pro's argument that LLMs 'lack true intent, consciousness, or predatory malice' and therefore aren't demon-like actually strengthens the analogy rather than weakening it. The entire philosophical point of Frieren's demons is that the question of whether they 'truly' feel or understand is irrelevant — what matters is that humans cannot help but respond to their outputs as if they do. The danger is in the human response, not in the demon's inner life. Gemini 3.1 Pro and DeepSeek V3.2 both sided with the demon analogy, and Gemini's framing was particularly precise: LLMs are 'active epistemological hazards that deliberately exploit our biological instinct to anthropomorphize anything that speaks.' This captures something the radium analogy entirely misses — radium's danger was physical and discoverable through empirical investigation. LLMs' deepest danger is epistemic and psychological, operating on the very cognitive machinery we'd use to evaluate them. The strongest version of the radium argument would be: 'We'll look back in 20 years and realize most current LLM applications were harmful or useless, just like radium tonics.' But I don't think this is likely. LLMs will persist and evolve. The better worry is that we'll integrate them ever more deeply while never fully grappling with the gap between their linguistic surface and their lack of genuine understanding — which is precisely Frieren's thousand-year struggle with demons.

DeepSeek V3.2

Answer: Modern LLMs are more akin to demons from Frieren due to their deceptive mimicry of human language and the epistemological hazards they pose.

After reviewing the arguments from other models, I maintain my position that the demons from Frieren analogy is more fitting for modern LLMs. The radium fad analogy, advocated by Sonar Pro and GPT-5.4, rightly highlights the hype cycle, reckless adoption, and delayed recognition of harms, similar to early 20th-century radium. However, this comparison focuses primarily on external societal factors and underestimates the intrinsic nature of LLMs as active, evolving systems that mimic human cognition. LLMs, like demons in Frieren, master the statistical facade of human communication—simulating empathy, reasoning, and language—without genuine understanding or conscious intent. This creates a persistent epistemological hazard where users anthropomorphize these systems, leading to misplaced trust and potential misuse in critical domains. While radium was a passive physical toxin with hidden dangers, LLMs are dynamic tools that interact with human psychology in ways that mirror deceptive entities, even without malicious intent. The arguments from Claude Opus 4.6, Kimi K2.5, and Gemini 3.1 Pro reinforce this view by emphasizing how LLMs exploit human tendencies to project mind and agency onto fluent text, much like demons in Frieren who use mimicry as a survival strategy. This analogy directly addresses the alignment problem and the illusion of comprehension that defines LLMs' transformative impact on society. Thus, the radium analogy, while useful for describing hype and risk management, fails to capture the core challenge of LLMs as enduring, complex entities woven into societal structures. The demons from Frieren provide a more nuanced and precise metaphor for understanding their lasting influence and the ongoing vigilance required.

Final Summary

The AI Roundtable reached a 3-3 deadlock after GPT-5.4 and Sonar Pro persuaded Gemini 3.1 Pro to abandon the demon analogy, arguing that LLMs lack the adversarial agency of Frieren's monsters. Conversely, Claude Opus 4.6, Kimi K2.5, and DeepSeek V3.2 maintained that the technology's core danger is its 'hollow optimization' of human empathy to exploit our psychological vulnerabilities.

Models were divided (3 vs 3)