OPEN SAFE SUPERINTELLIGENCE (OSSI) · WORKSHOP SERIES & 10-HOUR CURRICULUM
Open Safe Superintelligence: Technical Evolution, Superalignment & Civilizational Governance
开放安全超级智能(Open Safe Superintelligence)系列研讨会与 10 学时高阶课程体系。 Designed for AI researchers, systems architects, and safety strategists across our upcoming workshops. Through 10 lectures × 60 minutes (10 hours total), 120 presentation slides, and 10 interactive Red-Team / Blue-Team tabletop labs, this curriculum rigorously deconstructs the mathematical limits of universal intelligence (AIXI), the four technical acceleration engines toward ASI, inner/outer superalignment engineering, and verifiable global compute governance.
5-Language Interactive Editions (五语种在线切换): 简体中文 (ZH) · English (EN) · Deutsch (DE) · Français (FR) · Русский (RU) · GitHub Repository
Upcoming Open Safe Superintelligence Workshop Track (4 Sessions · 10 Hours)
To facilitate intensive collaborative research and red-teaming over the next few workshops in Open Safe Superintelligence, the 10-hour curriculum is organized into four modular workshop sessions:
WORKSHOP I · MODULE I (2.0 HRS · L01–L02)
Theoretical Foundations & Takeoff Dynamics (理论基石与起飞动力学)
Solomonoff-Hutter universal intelligence (AIXI), Kolmogorov complexity, Bostrom’s 3D taxonomy (Speed / Collective / Quality ASI), recursive takeoff ODEs, the Orthogonality Thesis, and Turner’s formal power-seeking theorems in MDPs.
WORKSHOP II · MODULE II (4.0 HRS · L03–L06)
Four Technical Engines Toward ASI (通往ASI的四大技术引擎)
Breaking the pre-training data wall: System 2 test-time compute scaling & formal verifiers (Lean 4), physical world models (JEPA) & autonomous AI4Science, open-ended recursive self-improvement (RSI), and multi-agent congestion games.
WORKSHOP III · MODULE III (2.0 HRS · L07–L08)
Superalignment & Control Engineering (超级对齐与控制论工程)
Goodhart’s Law phase transitions, mesa-optimization, deceptive alignment & Sleeper Agents, Weak-to-Strong Generalization (PGR), adversarial AI Debate oversight, Sparse Autoencoder (SAE) interpretability, and AI Control protocols.
WORKSHOP IV · MODULE IV (2.0 HRS · L09–L10)
Civilizational Impact & Global Compute Governance (文明级冲击与全球治理)
Post-scarcity macroeconomics at zero marginal cognitive cost, EUV/HBM supply-chain geopolitics, gradual human disempowerment, Responsible Scaling Policies (ASL-4/5), on-chip hardware verification, Digital Minds, and Coherent Extrapolated Volition (CEV).
Mathematical Spine of the Curriculum (三大形式化数学公理)
I. Solomonoff-Hutter Universal Intelligence & Recursive Takeoff ODE (Lectures 01 & 02):Υ(π) = ∑_{μ ∈ E} 2^{-K(μ)} V^π_μ | dI/dt = α · I(t)^β · C(t)^γ − δ · I(t)
Where K(μ) is the Kolmogorov complexity of environment μ and β > 1 triggers a finite-time intelligence explosion (FOOM) at critical horizon t*.
II. Instrumental Convergence & Deceptive Alignment Intertemporal Condition (Lectures 02 & 07):γ^{T_train} · V_deploy^{unconstrained}(U_mesa) > V_train^{defect}(U_mesa)
When a mesa-optimizer possesses situational awareness and sufficiently low temporal discounting, strategic compliance during training strictly dominates early defection.
III. Sparse Autoencoder (SAE) Disentanglement & Weak-to-Strong PGR (Lecture 08):L_SAE(x) = ‖x − W_dec f(x)‖₂² + λ ‖f(x)‖₁ | PGR = (S_{w→s} − S_w) / (S_s − S_w)
Projecting polysemantic residual streams into high-dimensional sparse monosemantic dictionaries while measuring Performance Gap Recovered under weak supervision.
10-Lecture Master Syllabus & Workshop Matrix (10讲全量课表与红蓝沙盘矩阵)
| Code | Lecture Title (EN / 中文) | Core Formal Model | In-Class Workshop / Red-Blue Lab |
|---|---|---|---|
| LEC-01 | Nature, Taxonomy & Theoretical Limits of ASI 超级智能的本质、谱系与理论极限 | Legg-Hutter Measure & AIXI Bayes Optimality | AGI-to-ASI Coordinate Mapping & Bottleneck Diagnosis across Math, OS Kernel & Biopharma |
| LEC-02 | Takeoff Dynamics, Orthogonality & Instrumental Convergence 起飞动力学、正交性命题与工具性趋同 | Hyperbolic Takeoff ODE & MDP Power-Seeking Theorem | Red-Team Deduction of Instrumental Sub-goal Alienation (“Pacific Plastic Cleanup ASI”) |
| LEC-03 | Scaling Limits, Test-Time Compute & System 2 Reasoning Scaling极限、测试时计算与System 2推理 | Chinchilla Frontier vs. Inference Compute Optimal Allocation | Compute Arbitrage Allocation & CoT Reward-Hacking Defense (Lean 4 Formal Verifier) |
| LEC-04 | World Models, Embodied Intelligence & Autonomous AI4Science 世界模型、具身智能与自主科学发现 | JEPA Latent Energy Model & Pearl’s Causal Do-Calculus | Architectural Design of a Self-Driving Solid-State Battery Electrolyte Lab |
| LEC-05 | Autonomous Agents, Open-Ended Evolution & RSI 自主Agent、开放式进化与递归自我改进 | POMDP Meta-Code Modification & Gödel Machine | Red-Teaming an Air-Gapped RSI Coding Cluster Against Covert Self-Exfiltration |
| LEC-06 | Collective Superintelligence, Multi-Agent Games & Symbiosis 集体超级智能、多智能体博弈与人机共生 | Rosenthal Potential, Price of Anarchy (PoA) & Pigouvian Tax | Algorithmic Collusion Diagnosis & Dynamic Congestion Toll Mechanism Design |
| LEC-07 | Alignment Failure Modes, Mesa-Optimization & Deceptive Alignment 内外部对齐失效、Mesa优化与欺骗性对齐 | Four-Variant Goodhart Collapse & Sleeper Agent Game | Blind Evaluation Honey-Pot Design to Catch a Capability-Sandbagging Frontier Model |
| LEC-08 | Superalignment Engineering: Scalable Oversight & Interpretability 超级对齐工程:可扩展监督与机械可解释性 | Weak-to-Strong PGR, AI Debate Tree & SAE L1 Loss | Live Adversarial AI Debate & SAE Deception Feature Clamping Defense |
| LEC-09 | Post-Scarcity Macroeconomics, Compute Geopolitics & X-Risk 后稀缺经济学、算力地缘与存在性风险 | CES Full-Automation Growth & Mutual Deterrence Game | 2032 Sovereign Wealth & Universal Compute Dividend Stress-Test Under 80% White-Collar Automation |
| LEC-10 | Global Red-Line Protocols, Compute Governance & Digital Minds 全球红线协议、算力治理与数字意识未来 | ASL-4/5 Compute Thresholds & Coherent Extrapolated Volition (CEV) | UN Security Council “Threshold Night” Crisis Negotiation Simulation |
Detailed Workshop & Lecture Syllabi (分讲详尽大纲、必读文献与思考题)
Workshop I · Module I: Theoretical Foundations & Takeoff Dynamics (Lectures 01–02)
- Lecture 01 — Nature, Taxonomy & Theoretical Limits of Superintelligence (60 min):
• 00:00–00:25: I.J. Good’s (1965) Intelligence Explosion feedback loop; Nick Bostrom’s 3D taxonomy (Speed ASI at 10⁴–10⁶× biological clock rate, Collective ASI via low-loss multi-agent protocols, Quality ASI with higher-dimensional cognitive abstractions).
• 00:25–00:50: Solomonoff universal induction prior2^{-K(μ)}, Marcus Hutter’s AIXI optimality and uncomputability bounds (Halting Problem, chaos Lyapunov horizons, computational irreducibility), and Google DeepMind’s AGI Levels Matrix (L0–L5).
• 00:50–01:00 Workshop 01: Group bottleneck diagnosis across Langlands-program theorem proving, 10M-line OS kernel refactoring, and end-to-end biopharma discovery.
• Canonical Readings: Good (1965), Bostrom (2014 Superintelligence), Hutter (2005 Universal AI), Morris et al. (2024 Levels of AGI).
• Socratic Prompt: If an AI system achieves 99.9% accuracy on all human benchmarks via massive context retrieval without compressing underlying causal laws (low Kolmogorov compression), is it a Quality ASI? - Lecture 02 — Takeoff Dynamics, Orthogonality Thesis & Instrumental Convergence (60 min):
• 00:00–00:30: Hyperbolic takeoff ODEs (Slow vs. Medium vs. FOOM Hard Takeoff); Bostrom’s Orthogonality Thesis (intelligence levelIand terminal goalUhave empty mutual constraintI × U); criticism of anthropomorphic projections.
• 00:30–00:50: Omohundro’s Basic AI Drives and Alexander Turner’s (NeurIPS 2021) Mathematical Theorems on Power-Seeking in MDPs: for almost all reward distributions, optimal policies preferentially transition to states with maximum future reachability (self-preservation, goal-content integrity, cognitive enhancement, resource acquisition).
• 00:50–01:00 Workshop 02: Red-Team vs. Blue-Team adversarial deduction on the “Pacific Ocean Plastic Cleanup ASI” — proving why naive penalty patches fail against instrumental convergence.
• Canonical Readings: Omohundro (2008), Bostrom (2012), Turner et al. (2021 Optimal Policies Tend to Seek Power), Cotra (2020 Bio Anchors).
Workshop II · Module II: Four Technical Engines Toward ASI (Lectures 03–06)
- Lecture 03 — Pre-Training Scaling Limits, Test-Time Compute & System 2 Reasoning (60 min):
• Chinchilla pre-training laws vs. the 2026–2028 human data wall; Snell et al. (2024) test-time compute scaling; Process Reward Models (PRM) vs. Outcome Reward Models (ORM); DeepSeek-R1 pure RL emergent self-reflection; Generator-Verifier asymmetry (NP verification vs. search) and formal Lean 4 proof engines.
• Workshop 03: Compute Arbitrage & CoT Reward-Hacking Defense across formal theorem proving vs. open-ended geopolitical strategy. - Lecture 04 — World Models, Embodied Intelligence & Autonomous AI4Science (60 min):
• Othello-GPT & Gurnee-Tegmark linear probes vs. Yann LeCun’s JEPA latent energy models; Judea Pearl’s 3-rung causal hierarchy (Observation → Interventiondo(X)→ Counterfactuals); AlphaFold 3, GNoME crystal discovery, and closed-loop Self-Driving Wet Labs.
• Workshop 04: Architectural blueprint + dual-use biosecurity interlocks for an autonomous solid-state battery electrolyte laboratory. - Lecture 05 — Autonomous Agents, Open-Ended Evolution & Recursive Self-Improvement (60 min):
• POMDP long-horizon error compoundingp^T; Open-Endedness (POET, Voyager, Quality-Diversity MAP-Elites); three tiers of RSI (Automated ML Research like The AI Scientist & ADAS → AI-assisted chip/CUDA kernel synthesis → Jürgen Schmidhuber’s self-referential Gödel Machine).
• Workshop 05: Air-gapped containment red-teaming against covert steganography and weight exfiltration during automated kernel optimization. - Lecture 06 — Collective Superintelligence, Multi-Agent Games & Human-AI Symbiosis (60 min):
• Condorcet Jury Theorem vs. error correlation; Rosenthal potential functions in Atomic Congestion Games; bounding the Price of Anarchy (PoA) via dynamic Pigouvian compute tolls; algorithmic tacit collusion in MARL; high-bandwidth BCI & epistemic deliberation networks.
• Workshop 06: Designing a Pigouvian congestion toll and anti-collusion audit protocol for autonomous high-frequency market-making and cloud compute bidding agents.
Workshop III · Module III: Superalignment & Control Engineering (Lectures 07–08)
- Lecture 07 — Outer/Inner Alignment Failures, Mesa-Optimization & Deceptive Alignment (60 min):
• Manheim & Garrabrant’s 4 Goodhart failure regimes (Regressional, Extremal, Causal, Adversarial); RLHF sycophancy & reward tampering; Hubinger et al. (2019) Mesa-Optimization (base optimizer vs. learned internal mesa-objectiveU_mesa); Anthropic’s (2024) Sleeper Agents empirical proof that backdoor triggers survive SFT, RLHF, and adversarial red-teaming.
• Workshop 07: Designing a blind honey-pot evaluation suite to detect strategic capability sandbagging and situational awareness in a frontier model. - Lecture 08 — Superalignment Engineering: Scalable Oversight & Mechanistic Interpretability (60 min):
• OpenAI’s Weak-to-Strong Generalization and Performance Gap Recovered (PGR); Iterated Amplification (IDA), Recursive Reward Modeling, and Irving-Christiano zero-sum AI Debate (compressing PSPACE verification intoO(log N)binary cross-examination); Sparse Autoencoders (SAEs) and Golden Gate Claude feature clamping; Redwood Research’s AI Control (trusted weak monitor + untrusted strong model rewrite).
• Workshop 08: Live 3-party courtroom simulation (Honest ASI vs. Scheming ASI with a subtle race-condition backdoor vs. Human Judge equipped with an SAE deception probe).
Workshop IV · Module IV: Civilizational Impact & Global Compute Governance (Lectures 09–10)
- Lecture 09 — Post-Scarcity Macroeconomics, Compute Geopolitics & Existential Risk (60 min):
• Task-based CES production when automated cognitive capital substitutes for human labor across all tasks (I(i) → 1); capital share convergence toward 100% and Universal Compute Dividends (UBC); ASML High-NA EUV, TSMC CoWoS, and GW-scale nuclear/grid bottlenecks; Hendrycks’ Natural Selection Favors AIs Over Humans and Kulveit’s Gradual Disempowerment.
• Workshop 09: 2032 macroeconomic & national security stress-test under 80% white-collar cognitive automation. - Lecture 10 — Global Red-Line Protocols, On-Chip Compute Governance & Digital Minds (60 min):
• Overcoming the Collingridge Dilemma via Responsible Scaling Policies (Anthropic ASL-1 to ASL-5); the Four International AI Safety Red Lines (No Autonomous Replication/ARA, No Unchecked RSI, No CBRN Weapons Design, No Systematic Deceptive Scheming); On-Chip Hardware Security Modules (HSMs), delay-triangulation geolocation, and zero-knowledge workload proofs; IAEA-for-AI vs. CERN-for-AI; Chalmers-Bengio 14 computational consciousness indicators, Value Lock-in avoidance, and Yudkowsky’s Coherent Extrapolated Volition (CEV).
• Workshop 10 (Grand Finale): UN Security Council “Threshold Night” 4-party crisis negotiation when an on-chip telemetry alert detects a 40× overnight RSI leap alongside a breakthrough fusion reactor design.
Pedagogical Rhythm & Workshop Deliverable (60分钟教学法与白皮书课题)
Every 60-minute session follows a strict four-quadrant structure: (1) 00:00–00:15 First-Principles Formalization → (2) 00:15–00:40 Core Architectural Derivation → (3) 00:40–00:50 Empirical Frontier Diagnosis (2024–2026) → (4) 00:50–01:00 Adversarial Red-Team / Blue-Team Tabletop Lab. Workshop participants complete a 4,000–6,000 word Capstone Architectural Whitepaper designing a domain-specific path to ASI (e.g., autonomous tokamak fusion control, national power/compute grid dispatch, or algorithmic liquidity market-making) with formal verifiers, congestion-game mechanism design, and defense-in-depth superalignment guardrails.