AI vs AGI vs ASI: Has AGI Already Arrived?
Publicado no IA Dots ·
AGIAI SignalASIdot-agidot-artificial_intelligencedot-asidot-google_deepminddot-nvidiaidioma-enInglêsyt-_poI2TDCny4
Vídeo em Inglês, publicado pelo canal Cost of Intelligence.
About this content: The summary, key points and topics below are generated automatically from the video transcript. They reflect the content, opinions, claims and predictions presented in the video and do not represent a position, validation or fact-check by IA Dots.
Video summary
The video explains the different tiers of artificial intelligence—Narrow AI, Artificial General Intelligence (AGI), and Artificial Superintelligence (ASI)—and discusses how the lack of a universal definition leads to varying claims about whether AGI has already been achieved.
Key points
- AI represents the entire ladder, while Narrow AI, AGI, and ASI are specific rungs on that ladder.
- Different organizations and experts, such as OpenAI, Google DeepMind, and Nvidia, use vastly different definitions and metrics for AGI.
- Current AI models exhibit a jagged profile, performing exceptionally well on specific tasks like math exams while struggling with basic long-term tasks or general adaptability.
- Massive financial investments in AI dwarf the resources dedicated to creating standardized testing benchmarks.
- Government regulations and executive orders are beginning to adopt specific terminology for advanced intelligence, further complicating the definitions.
Topics
Artificial IntelligenceAGIArtificial SuperintelligenceAI BenchmarksTech InvestmentAI Regulation
Video transcript
Transcript from available captions; it may contain errors.
AGI is basically here. Nvidia's boss said so out loud. >> Define it first. Goldman Sachs forecasts $1 trillion of AI investment in 2026, and nobody agrees what the word means. >> Okay, but AI, AGI, ASI, aren't those just the same thing with extra letters? >> No, they're rungs on one ladder, and nobody agrees where the rungs are. Give me 6 minutes and one drawing. By the end, you'll know what each word means, who gets to define it, and who profits when nobody does. Here's the ladder. Rung one, narrow AI, software that does one job well, like a chess engine or a spam filter. Rung two, AGI, artificial general intelligence, one system that can learn and do a wide range of tasks roughly as well as a person. Rung three, ASI, artificial superintelligence, better than the best humans at practically everything. >> So, AI is the baby, AGI is the grown-up, and ASI is the superhero. >> Close, but AI is the whole ladder. Narrow, general, and super are the rungs. Notice what's missing, a number. Nothing says exactly where rung two starts. Here are real definitions, word for word. OpenAI's charter, highly autonomous systems that outperform humans at most economically valuable work. A 2025 paper by 33 researchers including Yoshua Bengio, matching the cognitive versatility and proficiency of a well-educated adult. Nvidia's Jensen Huang was asked if an AI could build a billion-dollar company. His answer, "I think it's now. I think we've achieved AGI. To be fair, he hedged. You said a billion and you didn't say forever." And for superintelligence, Nick Bostrom's classic, "An intellect that is much smarter than the best human brains in practically every field." >> Four definitions, four different bars. Has anyone actually put a chatbot on the ladder? >> Google DeepMind did. Its paper sorts AGI into five levels. One, emerging. Two, competent. [music] Three, expert. Four, exceptional. Five, superhuman. Chat GPT and Gemini sit at level one, emerging AGI. >> So, by Google's own ladder, AGI is already here. >> Level one of five. Fair, that point is yours. The paper lists levels two to five as not yet achieved. Now try the other rulers. DeepMind's ladder says level one of five. Huang's billion-dollar test says now. The 33 author test gave GPT-5 an AGI score of 57% and Sam Altman says OpenAI is not quite there yet. Same machine, four rulers, four answers. >> So, the answer depends on who's holding the ruler. >> Exactly. So, what can these systems really do? In July 2025, a Gemini model scored 35 of 42 points at the International Math Olympiad gold medal level. Yet researchers describe a jagged profile, ahead of most people at maths and recall, behind at learning from experience and long-term memory. One economist told Fortune a recent model moved his playlist, but he had to sit there babysitting it for 3 hours. >> So, brilliant at the exam, hopeless at the playlist. [music] >> A benchmark called ARC-AGI-3 puts the bar plainly. As long as there is a gap between AI and human learning, we do not have AGI. So, why does the fuzziness last? Follow the money. Remember that trillion? Goldman Sachs puts about $581 billion of it in the United States. Fortune reports that some critics say companies like the term precisely because it's ill-defined. It's easy to build hype by claiming big strides. >> So, AGI is a sales pitch? >> Sometimes. And sometimes it's a law. More on that in a minute. Now our math. $1 trillion of forecast AI investment in 2026. [music] And the prize pool Google DeepMind put up to build better AGI tests, $200,000. A trillion divided by 200,000 is 5 million. About $5 million of investment for every $1 of prize money for the ruler. Prize money only, and labs run private tests, too. Still, the building money dwarfs the ruler money. >> We're building the skyscraper and still arguing about what the floor is. >> Roughly, yes. If this made it click, subscribe. It's free and it helps more people find the channel. I draw one of these every Tuesday and Sunday. >> And ring the bell. Ring it twice. >> One more place the word matters, the law. On September 29th, 2026, Executive Order 14434 told federal agencies to write superintelligence instead of AI in their own documents. Its definition of superintelligence? The one US law already uses for AI. A machine-based system that makes predictions, recommendations, or decisions. >> So, by that order, superintelligence is already here? >> For that order's purposes, yes. By Bostrom's definition, no. Same machines, new label. The order also gives the White House science advisor 60 days, until the 28th of November, to propose legal wording for a federal definition. A definition written into law decides which systems the word covers. >> Wait. So, does that order change any rules? >> Not by itself. It says existing regulations, contracts, and grants don't have to change. It's a naming rule plus a deadline. So, when a headline says AGI is here, ask four things. Who is saying it? Which ruler are they using? What test would prove them wrong? And what do they sell? Ask the ruler question first. A score of 57% means nothing until you know what 100% means. >> And if they pass all four? >> Then I'll draw the next rung. The verdict: AI is the whole ladder, AGI and ASI are rungs with no agreed height. So, always ask who holds the ruler. >> Fine. AGI is basically here on one ruler. >> Next drawing: Huawei versus Nvidia. Can China build AI without Nvidia? Subscribe and find more doodles in the playlist. And tell me in the comments what's your test for AGI. >> [music] [music] [music]
Assistir ao vídeo original no YouTube
Assuntos: AGI, ASI. Publicado no YouTube em 2026-10-15.
Seleção automática por assunto e idioma. O vídeo pertence ao canal de origem.