Video transcript
Transcript from available captions; it may contain errors.
Hello, it's me. I'm Nuffield. Today, in this video, I would like to uh if you like to demystify some terms. Um so, the question would like to ask ourselves uh AI, agentic AI, uh some new terms AGI, and ASI four names for the same thing. Of course, they are not. And we would like to answer these questions in the next in this video. Make sure that we understand Um here is a road map. Um so, we would like to look into AI is there basically. Uh it powers many everyday systems. Agentic AI still is there as well. It's also a real thing. But it adds planning to use uh checking, kind of trying uh to perform multiple tasks, and has goals. And these actions are uh uh uh uh are there to achieve these goals. On the other hand, AGI, which is generalization of AI, is a research goal whose definition and achievement are debated. So, there's no single definition of AGI. And the other one is the ASI, which is an hypothetical level of capability beyond the human intellectual capability. So, they're surpassing the human uh capability. Think of AI uh is uh that stands for uh artificial intelligence. Think of it as an umbrella of a software that can look at patterns, understand, and generate language, interpret images, make predictions, support decisions, uh machine translation, recommendations, spam filters, image creation, modern conversation AI all fit under what's so called this umbrella. People sometimes describe today's AI as one task only. This is too simplistic. Uh because today's foundation uh can work uh across many types of tasks. Uh nonetheless, uh current systems have important limits. We need to understand that. Uh in generalization, autonomy, and understanding. Or uh would like to narrow AI as a shorthand for today's non-AGI system. So, the AI is nowadays or it's now a non-AGI system. Imagine asking a conventional chatbot, "What um shall I or how should I uh organize my trip?" It might produce an itinerary, but you still have to use the travel websites, check details, and make arrangements. An AI agent, on the other hand, starts with the outcome you want. Creates a plan, uses connected tools, checks what happened, and any adjustments depending on its permissions and also services. It might research, draft, update, record, and or arrange services. However, agents do not have unlimited independence. Human approval and intervention, tool permission, security controls, and verification remain essential, especially for spending money or making you know, large decisions that have a huge impact. Um Uh uh you know, uh uh Uh agentic AI uh basically uh there are multiple types of actions agentic AI can perform. I'm showing you some examples of these actions. A research um agent uh can gather um sources and organize uh uh a cited briefing. Um but you have to verify this as Uh a coding agent can edit project, uh run uh debug issues, uh uh of course within that environment that is permitted. Uh but software review still matters. A travel agent, for example, can compare options, prepare or update itinerary, um and some systems can take booking uh steps where uh they are given uh permission. Uh analytics agent, for example, can improve data, identify trends, and produce reports. Um These are examples of some agents. Uh some uh tools uh AI tools uh agentic AI tools uh recent ones have been developed by well-known uh AI um agents. Um things like chat uh chat GPT work, uh things like uh Claude, uh code copilot studio from Microsoft. Uh minus research topic uh build sites and presentations. There are multiple These are not a comprehensive list. Uh some advisory AI advisory team. So, these are still agent. Imagine you are a solo entrepreneur asking, "I have, say for example £5,000 and I have three business ideas. Which market should I investigate first?" A simple chatbot might provide you general options or your opinions. An agentic system can help you structure the question, gather current evidence, compare competitors, look for demand in the market, test assumptions, prepare a recommendation with uh practical plans. Chat GPT work, for example, is an example of a general-purpose agent that can research across resources and produce reports and presentation or presentations. Claude research follows multi-step research or searches, make decision about what uh to investigate next, and provide sources. So, these have different kind of fundamental uh models uh in the backbone. Google's Gemini deep research uh is built to plan and synthesize large research projects, including competitive intelligence and market analysis. Perplexity, for example, computer is another multi-step digital worker uh that can browse, research, create, and connect tools to finish um and the outcome. Their products and plans vary. Uh some are research-specific. Some can take actions in connected apps. None has the uh judgment, accountability, or access to your full circumstances of a qualified human consultant obviously. Demand verification, check citation, and ask it uh to show uncertainty, investment, uh legal, and or medical advice requires various specialist reviews, and shouldn't rely on these agents as yet, right? Um AI also, you can now assemble a digital creative team without starting each job from a blank page. So, bringing everything under one uh under uh one roof. Say you want to launch, for example, a cybersecurity training business. A design agent in Figma can help you draft a user interface layout, and provide feedback in actual design file. Canva, on the other hand, AI is a a conversational creative partner that can create and refine marketing graphics, social content, and presentations. Gamma's creation agent takes a brief, asks questions, plans, and builds a presentation, including a research, writing, design, and imagery. Um while letting you uh revise it Uh, lovable turns plain language requests into working website or applications and lets you refine them, uh, in conversation. Know the important distinction, not every AI-designed feature is a fully autonomous agent. An image generator that creates one picture from a prompt is AI-assisted creation. A system that explores the pre-plans and executes several coordinated steps is a closer is closer to the agentic button. A human should review uh, brand consistency, image rights, claim security, uh, citation resources, accessibility, and anything published to customers. So, you have to make sure before you publish it out, you shouldn't trust have a blind trust to these agents as yet. Now, the research AGI AGI means artificial general intelligence. The usual idea is AI, uh, with broad flexible capabilities comparable to those of people across many intellectual domains. Learning, reasoning, adapting, and transferring knowledge uh, to unfamiliar problems. Uh, researchers disagree about the exact threshold um, how to measure it and whether, uh, autonomy or uh, should be part of of that definition. We should therefore avoid claiming that, uh, particular benchmark uh, release or a demo proves, uh, it is an AGI. As of this video preparation, AGI has not yet been published by a broadly acceptable consensus and or standard. The important distinction is the generality of capability, not a glowing robot brain or human consciousness. Right. So, today's AIs uh, AI versus AGI, this is conceptual comparison, not a claim that every current AI model behaves identically. Today's system can be excellent at many tasks and can transfer knowledge across related activities, but their performance can, of course, can be different based on number one, who is using it, and the way they look into it. And they often depends on specialized training or external tools. AGI would imply, again, is not here yet. AGI would imply far more robust generalization, flexible learning, and reasoning across very different domains. We cannot test an actual AGI in production because there's no widely accepted demonstration that the threshold has been reached. Autonomy is different axis. A capable present today model can be wrapped in an agent, but that alone does not make it an AGI. Now, coming to the next one is the super one. So, ASI stands for artificial super intelligence. Imagine capability that does not merely match humans across broad intelligent intellectual tasks. But sub- substantially exceeds humanity's best abilities in science, strategic problem-solving, creativity, and other cognitive domains. This remains hypothetical. There is no established timeline, measured scale, or proven development path. The diagram is illustrative, not quantitative. So, there's no such a thing as yet, but people say we will reach there. Alongside potential benefits are questions about control, alignment, accountability, and how human can evaluate decisions made by system that is much more capable than they are all together. These issues are actively discussed in AI safety and governance. What where do we stand now? Here is a complete picture. AI is the overall field and the umbrella. Modern AI is really in daily usage. Authentic AI refers to systems that can pursue goals, achieve tasks, still need human intervention through multi-step actions using tools and feedback often with a human oversight. AGI is a goal of broad human comparable or human surpassing general intelligence or intellectual capability. Definitions differ and its achievement is unconfirmed. ASI, on the other hand, is hypothetical intelligence subsequent to beyond human capability. So, avoid presenting these four names as four successive software versions. Well, we see we do see this. They describe different concepts and agency and general intelligence can vary independently. So, what people get wrong? Let's correct four common misconceptions misconceptions. First, today's ChatGPT and similar systems are not established AGI. They are authentic AI like ChatGPT work but not they are not established. So, there's no single definition or agreed definition to AGI. But, it's also wrong to call them text-only. Many modern AI systems, indeed, are multimodal that can work with images, audios audio and tools. Second, AI, AGI, and ASI are not synonyms. Basically, they are different probably AGI Uh, means general human comparable capability where whereas ASI means going significantly beyond that. Super, like superhuman. And third, agentic AI is merely is not merely an improved chatbot. It coordinates actions and tools toward an objectives or an objective. Fourth, um, saying we are still early should not be treated as a timeline prediction. Much is changing, but no one can uh, fix a date for AGI or if uh, if ASI will ever come anyway. So, what to watch next? Uh, what should people watch over the next coming years? First, agents integrated with everyday productivity and business applications going to be a real a reality. Uh, especially the permissions and ability to complete tasks. Uh, secondly, stronger evaluation of reasoning robustness, generalization, and safety. Uh, demonstrations alone are not enough. Third, uh, the guardrails, oversight, regulations, and rules for high-impact usage. Fourthly, uh, your own opportunity. Experiment with well-defined tasks. Learn how to verify results and adapt your skills rather than chasing every headline. Ask what system can uh, can do to can today perform under which conditions and with what human uh, with what human checks. Um, so AI is a broad field of machines performing tasks associated with intelligence. Agentic AI is a design pattern where AI plans, uses acts, takes, and adapts towards a goal. AGI is not yet established level of flexibility of flexible general intelligence across domains. ASI is hypothetical. So, AGI still in research. ASI is hypothetical intelligence that greatly exceeds human intellectual performance collectively. If someone uses these terms interchangeably, remember the distinction field agency generality and superhuman uh pro Lastly, I would like to say thank you so much for watching this video. Uh and uh it's worth looking for resources to learn the basics of AI to start with. And uh um I've brought in the comments some of these uh references which I find really useful. Open AI, a practical guide for uh to building agents. That's uh And there's a Google DeepMind uh levels of AGI research paper. Uh IBM, uh they have also I cited they have also called what is artificial general intelligence AGI. And that was uh all these documents actually produced in 2026. I'll leave a link in the comments underneath the video. Thank you so much and I'll see you in future videos.
Assuntos: Agentes de IA, Assistente pessoal, AGI, ASI. Publicado no YouTube em 2026-11-07.
Comentários
Postar um comentário