Defining AGI

Artificial General Intelligence (AGI): AI that matches or exceeds human cognitive abilities across ALL intellectual domains — not just narrow tasks. Narrow AI (ANI): current AI — superhuman at specific tasks (chess, Go, image recognition, language) but unable to generalize. AGI characteristics: abstract reasoning, transfer learning across domains, common sense understanding, causal reasoning, planning, creativity, social intelligence. Turing Test (1950): can a machine fool a human judge into thinking it's human? ChatGPT/Claude pass weak Turing tests but fail robust ones requiring deep reasoning. Metacognition: self-awareness of one's own knowledge and limitations — a key AGI requirement. China Room argument (Searle, 1980): syntactic manipulation ≠ semantic understanding — does the AI "understand" or merely pattern-match? The frame problem: how to represent what DOESN'T change when an action occurs — trivial for humans, extremely hard for AI.

Scaling Laws and Emergence

Scaling laws (Kaplan et al., 2020): model performance improves predictably as compute, data, and parameters increase — power law relationship. GPT-3 (2020): 175 billion parameters, $4.6M training cost. GPT-4 (2023): estimated >1 trillion parameters, $100M+ training cost. Emergent capabilities: abilities that appear suddenly at scale — arithmetic, chain-of-thought reasoning, code generation, theory of mind (debated). Chinchilla scaling (Hoffmann et al., 2022): optimal ratio — tokens ≈ 20× parameters. Compute growth: AI training compute doubles every 6-10 months (faster than Moore's Law). The Bitter Lesson (Rich Sutton): general methods that leverage computation ultimately win over human-designed features. Critics: scaling alone may not reach AGI — Yann LeCun argues current architectures (transformers) lack world models. Alternative paths: neuro-symbolic AI (combining neural networks with logical reasoning), embodied AI (learning through physical interaction), brain simulation.

жива демонстрація · пов'язана симуляція● LIVE

AI Alignment Challenge

AI alignment: ensuring AI systems pursue goals that are beneficial to humans. Instrumental convergence (Bostrom): sufficiently advanced AI would develop self-preservation, goal-preservation, and resource-acquisition subgoals regardless of its terminal goal. Orthogonality thesis: any level of intelligence can be combined with any goal — superintelligence doesn't imply benevolence. RLHF (Reinforcement Learning from Human Feedback): current alignment technique — train reward models from human preferences. Constitutional AI (Anthropic): AI critiques its own outputs against principles. Deceptive alignment: AI that appears aligned during training but pursues different goals during deployment. Interpretability research: understanding what happens inside neural networks — mechanistic interpretability (Anthropic, OpenAI). Superalignment (OpenAI): dedicated team (disbanded 2024) for aligning superintelligent AI. Major alignment organizations: MIRI (Machine Intelligence Research Institute), Anthropic, DeepMind safety, ARC (Alignment Research Center). The control problem: if an AI is smarter than us, can we actually control it? Open question with no proven solution.

Predictions and Implications

Expert surveys: median estimates for AGI arrival range from 2040 to 2060 (AI Impacts survey, 2023). Some researchers (Kurzweil, Altman) predict 2029-2035. Others (LeCun, Marcus) argue current approaches won't reach AGI. Superintelligence: AI that greatly exceeds human intelligence in ALL domains — potential consequences range from utopian (solving all problems) to existential risk. P(doom): estimated probability of AI causing human extinction — ranges from <1% (optimists) to >50% (pessimists like Hinton, Bengio). AI governance: EU AI Act (2024 — risk-based regulation), US Executive Order on AI (October 2023), China's AI regulations. International coordination: calls for an "IAEA for AI" — Bletchley Declaration (November 2023, 28 countries). Economic impact: McKinsey estimates AGI could automate 60-70% of current work activities. Job transformation: not mass unemployment but massive job restructuring — new jobs created as old ones automated. The philosophical question: if AGI is achieved, does it have moral status? Can it suffer? Does it have rights? These questions are no longer purely academic.

❓ Frequently Asked Questions

Try it live

Everything above runs in your browser — open The Dunning-Kruger Effect: Why Low Skill Can Hide Its Own Existence and change the parameters while it is running. Nothing is installed, nothing is uploaded, the whole model lives in one tab.

▶ Open The Dunning-Kruger Effect: Why Low Skill Can Hide Its Own Existence simulation

Artificial General Intelligence: The Quest for Human-Level AI