AI Fundamentals Deep Dive — Deterministic vs Stochastic, World Models, Semantic Reasoners, and More

The full unedited livestream. After the Grok $200K Morse-code hack stream raised more questions than answers for a lot of viewers, Goju hopped online to teach the fundamentals from first principles — and to push back on some of the most common framing errors in the current AI safety conversation, including Yann LeCun's. This is the long-form version with audience Q&A, tangents, the SLM training-cost ballpark walk-through with Clayton, the hybrid-vehicle analogy for DIF+LLM, Moxie's pushback on ownership ambiguity, JB's question on deterministic AI weights, the mid-stream delivery interruption, the moving-house schedule disruption, and the closing practical advice. **The lecture has three movements:** **Movement 1 — Fundamentals**: What it actually means for a system to be deterministic, non-deterministic, or stochastic. Most people use these words interchangeably — including AI papers — and they shouldn't. **Non-determinism in CS is a temporal property, not an outcome property**. P vs NP, NP-complete, NP-hard are all about *time*. Today's neural networks are stochastic — random initialization + stochastic gradient descent — not non-deterministic. That's not academic: stochasticity breaks iterative refinement, which is what makes the current AI build cycle structurally fragile. Deterministic systems compose. Stochastic systems don't. The hybrid-vehicle analogy lands hard: petroleum cars vs full-electric vs hybrid — DIF + a language model is the hybrid, where you get stochastic ideation on top of a deterministic substrate so you can switch modes when something actually matters. **Movement 2 — Semantic Reasoning vs World Models**: Once the substrate is clarified, the next question is what's missing. Goju pushes back on LeCun's "world models" framing as overkill for software-only AI. World models are the right target for robotics and embodied AI — physical-world grounding (objects, gravity, causality). For everything inside a computer — code, finance, web, agent automation — what's needed is a **semantic reasoner**: a system that understands the meaning of concepts like "giving" so it doesn't confuse "translate this Morse code message" with the wallet-transfer instruction the message carries. The Grok hack is the anchor example. The crypto industry's reaction — hard-code blocklists — fixes the wrong layer. The big reframe: **hacking a deterministic system is bounded; hacking a stochastic AI system is unbounded**. That's the categorical difference between traditional hacks and AI takeover. Plus the formal scaffolding: symbolic reasoners vs learned vs rules-based, SAT/SMT vs DIF vs world models, fully-formal (100% correct, narrow) vs semi-formal (scalable, 100%), why code is semi-formal in practice, why DIF is the right shape for the safety layer, isolation forests as another semi-formal example, the personification trap. **Movement 3 — The Practical AI-Takeover Playbook**: Two pieces of advice. **Sandbox your AIs** (Goju runs his own Claude Code inside sandboxes — correct shape for the threat model, not paranoia). **Minimize MCP usage** (OAuth chains and tool-use permissions are easy hijack vectors). The xz-utils contrast: traditional supply-chain attacks took years of patient social engineering; AI tools collapse the attack-prep cost to nothing because the AI's behavior space is unbounded — effectively infinite zero-days. Meredith Whittaker's framing on Signal closes it out: the AI shift is functionally an invitation to turn your brain off, and the cost of that shift is real. This is the full unedited version. The three sister topical VODs are the polished cuts. Watch this for the live community texture — chat exchanges, Clayton's question on Merly building its own LLM (and the $10–30M training cost answer), the Australia / Brisbane / Melbourne aside, the delivery interruption, all the back-and-forth. Watch the topical cuts for the high-density content version. 💬 Where do you actually land on this? Drop your honest take. 🔔 Subscribe for honest tech analysis: https://youtube.com/@gojutechtalk 📺 Related: The Problem With Today's AI (In Simple Language) https://youtu.be/Cl7x2OhbPwU 📺 Related: The Dangers of Over-Reliance on LLMs and AI https://youtu.be/U1Dhfij4Uy0 📺 Related: Good Software Development & the Future of AI for Code (In Plain Language) https://youtu.be/JR22KE6kLMA #DeterministicAI #StochasticAI #NonDeterministic #PvsNP #NPComplete #NPHard #AIFundamentals #SemanticReasoning #WorldModels #YannLeCun #DIF #DeterministicIntentFolding #FormalMethods #SemiFormalSolvers #SATSMT #SymbolicReasoners #LearnedReasoners #IsolationForests #ChainOfThought #AITakeover #StochasticHacking #UnboundedExploits #SandboxAI #MCP #OAuthHijack #VibeCoding #xzUtilsContrast #InfiniteZeroDays #MeredithWhittaker #Signal #MerlyDIF #GrokHack #LiveDiscussion #GojuTechTalk