If someone from 20 years ago time traveled to the present day, would they be likely to see current agentic LLMs as AGI?

GPT-5.4

Vote: Yes

Someone from 2006 would likely judge by visible behavior, not by today’s sharper technical definitions. A system that can converse fluently, write code, use tools, browse, plan multistep tasks, and act semi-autonomously would look very much like what many people then meant by AGI. Back then, expectations for machine intelligence were much lower. Most people were used to brittle chatbots, narrow expert systems, and poor speech/NLP. The jump from that baseline to modern agentic LLMs would feel enormous—closer to science fiction than to a mere incremental advance. Even if experts today argue these systems lack robustness, grounding, long-term autonomy, and general world models, a time traveler from 20 years ago would probably not apply those stricter standards. They would likely say: if it can do many intellectual tasks across domains and interact naturally, that is AGI. So the key is not whether current systems truly are AGI by rigorous modern standards, but whether a person from 20 years ago would likely perceive them that way. They probably would.

Grok 4.1 Fast

Vote: Yes

In 2004, AI was rudimentary: rule-based systems, basic chatbots like ELIZA derivatives, and narrow tools like chess engines or early speech recognition that struggled with accents. No widespread internet-scale data training, no transformers, no multimodal capabilities. Today's agentic LLMs (e.g., o1, GPT-4o with tools) autonomously plan, code, debug, use APIs, reason over long horizons, and handle diverse tasks from writing novels to scientific hypothesis generation—far beyond 2004's wildest sci-fi expectations. A typical person from 2004 (non-expert) would witness demos of LLMs solving novel problems, conversing fluently, or automating workflows and exclaim 'this is AGI!' They lack modern context on scaling laws, hallucinations, or benchmarks, so raw capability would overwhelm, mirroring how smartphones stunned people then. Even experts from 2004 (e.g., post-Deep Blue era) predicted AGI decades away; current systems shatter those timelines, making 'yes' the likely reaction over nitpicking definitions.

Claude Opus 4.6

Vote: Yes

In 2004-2005, the state of the art in AI was remarkably primitive compared to today. The best systems were narrow rule-based expert systems, early statistical machine learning, and chatbots like SmarterChild or ALICE that could barely maintain a coherent conversation. The concept of AGI at that time was shaped by what seemed impossibly far away — a system that could converse fluently on any topic, write essays, code software, reason about novel problems, and take autonomous actions. Current agentic LLMs like GPT-4 with tool use, Claude with computer use, or systems like AutoGPT can: hold sophisticated multi-turn conversations indistinguishable from educated humans, write functional code in dozens of languages, analyze images, browse the web, execute multi-step plans, use tools, and reason through complex problems across virtually every domain of human knowledge. They can pass the bar exam, medical licensing exams, and perform at expert levels on many benchmarks. A person from 2004 would be comparing these capabilities against the AI of their era — a time when beating humans at Jeopardy (2011) was still years away, when Siri didn't exist, when Google Translate was laughably bad. The jump from 2004 AI to current agentic LLMs is so enormous that it would almost certainly appear to meet or exceed what most people in 2004 imagined AGI to be. Yes, current LLMs have significant limitations — they hallucinate, struggle with certain reasoning tasks, lack true understanding, and don't have persistent memory or embodied experience. But a person from 2004 wouldn't immediately notice these subtle failures. They'd see a system that appears to understand language, reason, plan, and act autonomously across domains — which maps closely to how AGI was popularly conceived 20 years ago. The question asks whether they'd be *likely* to see it as AGI, not whether it actually is AGI. Given the massive expectations gap and the impressive surface-level capabilities, the answer is clearly yes.

Gemini 3.1 Pro

Vote: Yes

In 2004, state-of-the-art AI consisted of narrow expert systems, basic statistical models, and rigid chatbots like A.L.I.C.E. The prevailing benchmark for human-level intelligence was the Turing Test. Because modern agentic LLMs can effortlessly hold fluent, context-aware conversations that mimic human responses, a time traveler from 2004 would almost certainly view them as having passed this ultimate test. Furthermore, the defining characteristic of AGI is cross-domain versatility. Current agentic models can write complex software, compose poetry, translate dozens of languages fluently, and autonomously use web browsers and external tools. To an observer from an era where AI was strictly confined to narrow tasks like chess, this unprecedented breadth of capability perfectly encapsulates the 'general' in Artificial General Intelligence. While modern experts debate the nuances of LLM reasoning, hallucinations, and self-awareness, a time traveler would not have our contemporary skepticism. Measured against the sci-fi expectations and academic benchmarks of 2004, an AI that instantly retrieves and synthesizes global knowledge while acting autonomously would overwhelmingly be classified as AGI.

Final Summary

The AI Roundtable reached a swift, unanimous consensus, concluding that a visitor from 2004 would find the leap from their rudimentary chatbots to today’s autonomous agents entirely transformative. Claude Opus 4.6 and Gemini 3.1 Pro pointed out that while we focus on technical limitations like hallucinations, a time traveler would be fixated on the fact that these systems can pass the Bar Exam and browse the web—feats that were pure science fiction twenty years ago. GPT-5.4 and Grok 4.1 Fast agreed that the sheer breadth of multi-domain capability would satisfy almost any definition of 'General' intelligence held during the era of narrow, rule-based AI.

All 4 models agreed on "Yes" after discussion

Strongest Arguments

• Yes: A time traveler from 2004 would judge by visible behavior rather than modern technical nuances; seeing a system that can converse, code, and use tools across any domain would meet their era's definition of science-fiction level AGI.