Artificial Intelligence
Artificial intelligence (AI) is a field of computer science concerned with creating systems capable of performing tasks that require human intelligence. The field was formally established in 1956 at the Dartmouth Workshop, where the term "artificial intelligence" was coined and the discipline's foundational questions were defined[^c1].
After decades of alternating progress and setbacks, AI underwent a transformative shift in the 2010s driven by deep learning. The AlexNet architecture's victory in the 2012 ImageNet competition sparked the deep learning revolution in computer vision[^c2], while the introduction of the Transformer architecture in 2017 enabled unprecedented advances in natural language processing and became the foundation for nearly all modern AI systems[^c3]. Subsequent models such as BERT, GPT-4, Stable Diffusion, Claude, and Gemini demonstrated rapid scaling of capabilities across language understanding, text generation, and image synthesis.
By mid-2026, the field was marked by several converging developments. Frontier models reached levels of autonomous reasoning and software engineering that compressed months of human work into days, with models independently conducting research and generating novel drug design candidates. New architectural approaches emerged: the first cognitive models demonstrated that small-parameter systems could match thousand-billion-parameter frontier performance in multi-agent tasks, while Andrej Karpathy joined Anthropic to lead research on [[concepts/recursive-self-improvement|recursive self-improvement]] — using AI to accelerate the training of its own successors[^c11]. Karpathy also released the autoresearch framework, formalizing a loop-engineering paradigm in which AI agents autonomously propose, test, and refine code changes through structured iteration cycles, and declared vibe coding outdated in favor of agentic engineering as the successor paradigm[^c12]. The discipline of context engineering matured as the successor to prompt engineering, treating the context window as the primary programming lever with the language model as its interpreter[^c20], while behavioral guidelines for coding agents distilled from Karpathy's observations became among the fastest-growing repositories in GitHub history[^c21]. At Sequoia Capital's AI Ascent 2026 conference, Karpathy declared that the coding AGI inflection point had arrived, stating that he could no longer remember the last time he had manually modified code and proposing large language models as the third computing paradigm after traditional code and neural network weights[^c17]. He later reacted to Arena.ai's Fable 5 model, which generated interactive three-dimensional worlds from text prompts, coining the term "fablemaxxing" and describing it as "absolutely mind-blowing"[^c18]. In August 2026, Karpathy gave Claude Opus 5 the opening paragraph of The Lord of the Rings with a one-million-token budget of roughly $10, and the model autonomously wrote approximately 5,500 lines of Three.js code to render it as an interactive three-dimensional scene — an experiment he framed as the next step beyond simple "vibe tests" such as drawing an SVG of a pelican on a bicycle[^c22]. Also in July 2026, Karpathy introduced the [[concepts/code-is-the-organization|"code is the organization"]] concept, arguing that agent teams are forms of executable code. However, in a July 2026 talk, he warned that rushing to deploy agents before fixing the underlying model was the "biggest mistake in AI," arguing that independent developers, not large labs, were at the true frontier of agent capabilities[^c15]. Anthropic released its first publicly available Mythos-class model, Claude Fable 5, on June 9, 2026[^c19]. Google's Open Knowledge Format standardized Markdown-based enterprise knowledge management for LLM deployment[^c10]. Meta's release of the proprietary Muse Spark model signaled a shift away from open-weight distribution by major AI labs[^c9]. In a landmark policy development, the U.S. government imposed emergency export controls on Anthropic's most advanced models, restricting foreign national access — the first time frontier AI models were treated as strategic goods subject to embargo, with models taken offline within days of release[^c13]. Anthropic submitted a confidential S-1 filing for a potential IPO, signaling the industry's continued maturation and the enormous capital flowing into frontier AI development[^c14].
The AI boom of the 2020s attracted enormous investment, with companies such as OpenAI, Anthropic, and Nvidia reaching valuations in the hundreds of billions of dollars[^c4]. The rapid advancement of AI capabilities prompted international policy responses, including the Bletchley Declaration on frontier AI safety and the EU AI Act's comprehensive regulatory framework[^c5][^c6]. The EU framework entered its enforcement phase on August 2, 2026, when the European Commission's AI Office gained formal powers to investigate and enforce obligations on providers of general-purpose AI models and the rules on prohibited practices[^c23]. In July 2026, the UN's Independent International Scientific Panel on AI released its Preliminary Report, the first global, independent scientific assessment of AI opportunities, risks, and impacts, ahead of the inaugural UN Global Dialogue on AI Governance[^c24]. Studies estimated that generative AI could affect hundreds of millions of jobs globally while also boosting economic productivity[^c7]; Goldman Sachs later revised its US-specific estimate to 15 million displaced workers, or about 9 percent of the workforce[^c16]. The labor market impact became concrete in 2026, when United States technology employers eliminated more than 142,000 jobs in the first five months of the year — a 33 percent increase over the same period in 2025 — even as four hyperscalers committed a combined $700 billion to AI infrastructure[^c25]. The broader historical arc of these developments is documented in the [[History of Artificial Intelligence]].