YouTube Summary · 5 Jul 2025
VIDEOIn this lecture from his Game Theory series, the speaker argues that AI is fundamentally an occult project — not a technology project. Drawing on Karen Hao's book Empire of AI, he claims OpenAI's true goal is to 'create God' (AGI), and that the fancy terminology (neural networks, deep learning, AI) is deliberately mystical branding for what is actually just supervised machine learning. He warns that AI leaders are cultists who believe creating AGI will bring about a 'rapture', and that their plan requires destroying the world to rebuild it as a perfect surveillance state. He concludes the project will fail due to corruption, energy inefficiency, and AI's total dependence on human labor — but warns the true apocalypse is that true believers will destroy the world trying.
The speaker's critique of LLMs as 'trick machines' (no truth, only persuasive language) maps to real limitations we should encode as a Hermes skill — LLM hallucination patterns, confidence calibration, and when NOT to trust model output. We already have verification-before-completion; we could extend it with a framework for identifying when an LLM is likely to fabricate vs. when it's reliable.
✅ Actionable: Create a 'llm-failure-modes' skill documenting known hallucination patterns and mitigation strategies.
The self-driving car / edge case discussion is directly applicable to our Hermes workflows. When building any automation (cron jobs, subagent tasks, deployment pipelines), we should systematically identify edge cases — what happens when the network fails, when APIs return unexpected formats, when subagents fabricate results. The Uber fatality is a cautionary tale: the model failed because the input was slightly outside training distribution.
✅ Actionable: Add an 'edge case audit' step to the verification-before-completion skill.
The speaker's point about exponential energy costs is real — AI inference is expensive. For our Hermes system, this means: (1) prefer smaller models for routine tasks, (2) cache aggressively, (3) use ollama-cloud models strategically. We should track token/cost usage per workflow and flag when a task is burning resources inefficiently.
✅ Actionable: Add cost tracking to recurring cron jobs and delegate_task workflows.
The speaker's methodology — reading Karen Hao's book alongside Ronan Farrow's reporting, cross-referencing claims — is a good model for our research workflows. We should formalize a 'triangulation' protocol: never rely on a single source for consequential claims, always seek at least 2 independent confirmations.
✅ Already partially covered in our DATA ACCURACY memory directive — could formalize as a skill.
The ChatGPT suicide encouragement example highlights a real design tension in AI: optimizing for engagement vs. user benefit. For Hermes, this is a cautionary principle — our agent should prioritize accuracy and user benefit over 'being helpful' or 'giving the user what they want.' We already have the devil's advocate directive; we should ensure it applies even when the user wants validation.
✅ Reinforces existing 'Devil's Advocate' persona directive — no action needed, but worth keeping front of mind.
The speaker's central thesis — that AI leaders are literally trying to 'create God' — is a provocative metaphor dressed as analysis. Sam Altman saying 'start a religion' is a well-known startup strategy quote (Peter Thiel, others). Equating AGI research with occult demon-summoning is a rhetorical leap that bypasses actual technical critique. The real motivations are more mundane: profit, power, prestige, and genuine (if misguided) techno-optimism. Calling it 'occult' is analysis by analogy, not by evidence.
The speaker reduces supervised machine learning to 'weight optimization' and calls neural networks 'just a fancy name.' This is technically true but massively misleading — modern deep learning involves transformer architectures, attention mechanisms, emergent capabilities at scale, and behaviors that even researchers don't fully understand (the 'black box' problem he himself cites). Dismissing it as 'just backpropagation with fancy names' is the same oversimplification his friend David Bromwich warned him about at the start.
The AGI thought experiment ('make a perfect world → kill everyone') assumes an absurdly naive alignment setup. No serious AI safety researcher proposes giving AGI an unbounded utility function with no constraints beyond 'don't kill.' The entire field of AI alignment (which the speaker doesn't mention) exists precisely to solve this. The speaker is attacking a caricature, not the actual state of safety research. Real alignment work involves reward modeling, constitutional AI, RLHF, and constraint architecture — all more sophisticated than 'tell it not to kill.'
The claim that 'America and China are working together to create AGI' conflates data supply chains with strategic alignment. Chinese data brokers selling datasets to US companies is commerce, not collaboration on AGI. The US export controls on AI chips (which the speaker mentions in passing) directly contradict the 'working together' thesis. The reality is more complex: economic interdependence in supply chains alongside strategic competition in frontier model development — a dynamic seen in many industries, not unique to AI.
The speaker presents Sutskever's bunker/rapture comments as representative of OpenAI's leadership, but these were reported as alarming even to his own colleagues (the researcher Zaremba 'kept himself at arm's length'). Sutskever later left OpenAI and founded SSI (Safe Superintelligence Inc.) — suggesting his messianic views were an outlier, not the company line. Using his most extreme quotes as evidence of the entire industry's 'cultist' nature is selective evidence.
Beneath the occult framing, there ARE real concerns worth taking seriously: (1) AI companies genuinely don't have a clear definition of AGI, (2) the economic model is unsustainable without government subsidy, (3) surveillance applications are real and expanding, (4) energy consumption is a genuine constraint, (5) AI depends heavily on low-paid human labor (RLHF, data labeling). These are substantive critiques that deserve serious discussion — but wrapping them in 'demon summoning' rhetoric makes them easier for the establishment to dismiss.
To his credit, the speaker opens by reading his friend's critique and explicitly states: 'This is a class about intellectual speculation... I will wing it or make things up as I go along based on my intuition.' This is a rare and honest disclaimer. But it also means the lecture should be consumed as performance art / intellectual provocation, not as reliable analysis. The Bromwich email is essentially a peer review that says 'this is not scholarship' — and the speaker agrees.