Transcript 0:00 Something strange is happening. My LinkedIn feed has changed from hype about ChatGPT announcements and N810 workflows into endless posts about vibe coding and coding agents. 0:12 Claude Code and Cowork were getting the most attention. Then Claude Bot came out of nowhere and just completely blew up. 0:21 At Anthropic's request, it's changed its name to Mult Bot now, by the way, and despite massive FOMO, I haven't quite finished setting it up yet. 0:29 Instead, someone recommended Google Antigravity to me, and I gave that a go. It's a bit less technical than Mult Bot, as you don't need to set up in the terminal. I said I didn't wanna approve anything, and away it went. 0:42 Within about 15 minutes of watching it work, it had built exactly what I wanted. It works under my supervision, but quite independently, and it was certainly better than me at this multi-stage task. 0:54 It was a moment where I thought, "This is AGI." I've had this feeling before, actually. 1:01 In the summer, I built an SEO competitor analysis agent that was able to review long-form web articles and tell me how to improve them against Google Search Quality Raters Guidelines, which is a nearly 200-page PDF. 1:15 When I first reviewed the results, I thought, "This would be too expensive for a human to do manually and probably too hard to accurately cross-reference." 1:24 So the agent had surpassed human capability at a reduced cost, and this was GPT 4.1. It wasn't even the current models. And it's not just me thinking this. 1:35 Just recently, I posted a video on the case against AGI, which, for reasons I'll get onto, I think is still totally valid. 1:44 A few people in the comments of that video said they think we've already reached AGI, and I actually say something similar in the video. 1:52 But to add some heavyweights to this thesis, Pat Grady and Sonya Huang at investment fund Sequoia Capital recently published an essay titled "2026: This Is AGI." 2:04 Their opening remarks include the line, "While the definition is elusive, the reality is not. AGI is here now." 2:13 The definition of AGI is genuinely very fuzzy, which is something I'll get onto, and something the writers clearly acknowledge. 2:20 But the Sequoia partners have attempted to simplify it through defining it as the ability to figure things out. That's it. 2:28 They argue that three key ingredients have now converged: baseline knowledge from pre-training, reasoning ability from inference time compute, and iteration capability through long-horizon agents. 2:40 As they put it, "Generally intelligent people can work autonomously for hours at a time, making and fixing their mistakes, and figuring out what to do next without being told. 2:51 Generally intelligent agents can do the same thing. This is new." 2:56 The essay cites research from Meta showing exponential progress in AI's ability to complete long-horizon tasks, with performance doubling every seven months. 3:07 If this trajectory continues, agents could reliably complete tasks taking human experts a full day by 2028 and a full year by 2034. A full century, they suggest, could come by 2037. 3:20 That sounds like a technological singularity to me, something I've explored in another video. 3:26 Their closing point is that the kind of AI we're used to through prompting, like ChatGPT, they're great conversationalists, but their impact has been limited. Now, I think that's actually true. 3:39 Most people who use ChatGPT think it's an amazing tool, but it's kind of an advanced search engine or something to bounce ideas off. It doesn't do that much other than provide natural language responses. 3:50 The next couple of years will be about AI applications that can do multi-step tasks like Claude Code or Google's Antigravity, so long as you give them permission. 4:00 I also think these applications will become much more user-friendly, and you won't need a terminal. Claude Cowork, announced in January 2026 and created via Claude Code, is an example of this. 4:11 Now, Sequoia Capital has invested in some of the biggest AI companies in the world. They obviously have a vested interest in saying this, but plenty of others agree with them. 4:20 I'm a little bit skeptical about the statements of AI company CEOs because they have a big interest in promoting the AGI narrative to maximize shareholder value. 4:29 But Elon Musk, Dario Amodei, and Sam Altman all have provided similar timelines. In a podcast recorded at the end of 2025, Elon Musk says he believes that we will achieve AGI in 2026. 4:44 Within three years, robots will surpass surgeons. By 2030, the collective intelligence of AI will exceed that of all humanity. Dario Amodei, CEO of Anthropic, doesn't define it as AGI, rather powerful AI. 5:01 In his essay, "Machines of Loving Grace," he wrote that powerful AI could come as early as 2026, though there are also ways it could take much longer. 5:11 Anthropic also presented this timeline in their official recommendation to the White House in March 2025, stating, "We expect powerful AI systems will emerge in late 2026 or early 2027. 5:23 Powerful AI systems will have the intellectual capabilities of matching or exceeding that of Nobel Prize winners across most disciplines." Amodei describes this as a country of geniuses in a data center. 5:35 At the World Economic Forum in 2026, he also said "I have engineers within Anthropic who say, 'I don't write any code anymore. I just, I just let the model write the code. I edit it. I do the things around it.' 5:48 I think, I don't know, we might be six to 12 months away from when the model is doing most, maybe all, of what SWEs do end to end." 5:58 This pretty much matches the Sequoia view.If a coding agent can run end-to-end, then isn't that just AGI anyway? And it's becoming increasingly clear that such agents can. Sam Altman has typically been bullish. 6:12 He wrote on his blog in January 2025 that we are now confident we know how to build AGI as we have traditionally understood it. 6:20 We believe that in 2025, we may see the first AI agents join the workforce and materially change the output of companies. His second point was basically correct. 6:31 2025 was the year when AI agents joined the workforce, albeit not without frustration. But the difference between a GPT-4o agent versus a top-level model like GPT 5.2 is really quite vast. 6:45 Models like GPT 5 or Gemini 3 get more stuff more accurately correct more often. I've also accepted that it's just smarter than me generally. When I ask GPT 5 a question, I normally get a highly accurate answer. 7:00 Isn't this just AGI anyway? If you saw the vast improvements of now against the often hallucination-filled early years of ChatGPT, wouldn't you think it AGI? 7:11 One of the reasons why I think it isn't seen that way just yet is because it hasn't been properly let out of its box. 7:17 I imagine because you're watching this video that you're probably an AI enthusiast and use it regularly. But the average knowledge worker doesn't. 7:26 Regular paid usage of chatbots is still not even at a majority in the same way that Google Search absolutely is. Meanwhile, chatbots or just models themselves don't have a whole lot of agency. 7:39 It's only when they're placed in a multi-step workflow or agentic task with considerable permission that you see their greatest potential. 7:47 There's also a clear dismissal of AI by a lot of people, and I think some of this comes from using older versions without much success. It hallucinates or produces slop are the regular complaints. 7:58 But with enough tuning of system prompts, you can shape AI to making very high-quality outcomes in coding, writing, and image generation. 8:08 In my previous video, I argued that one of the biggest obstacles to declaring AGI was the lack of any agreed definition. 8:15 Indeed, in August 2025, even Sam Altman dialed back and called it not a super useful term because it means many different things. 8:24 He went on to say, "I think the point of all this is it doesn't really matter, and it's just this continuing exponential of model capability that we'll rely on for more and more things." 8:36 Intelligence is not very clearly defined in the first place. We struggle to define human intelligence with any precision. IQ tests capture something, but nobody believes they capture everything. 8:47 Emotional intelligence, creativity, practical wisdom, social reasoning, these are all important to how humans interpret people's intellect. 8:55 Yet we can't agree on how to measure them or even whether they're distinct capacities. 9:00 If we can't rigorously define human intelligence, perhaps expecting a rigorous definition of artificial general intelligence was always unrealistic. 9:10 The demand for a precise definition before acknowledging AGI might be a form of motivated reasoning, a way to keep moving the goalposts because we're uncomfortable with what the arrival of AGI actually means. 9:22 In June last year, Sam Altman wrote a blog post called The Gentle Singularity. It's kind of a strange article, but the point the title makes is that we perceive the singularity as a specific event. 9:34 Now, the term singularity is also interchangeable with a more violent-sounding intelligence explosion, a sudden and uncontrolled takeoff. But possibly, we're actually moving into the singularity already. 9:47 When people live in historic moments, it doesn't often seem quite as seismic as the future perspective on it might suggest. 9:54 That's because while life may be changing from an old way of doing things, there's still a huge array of human traits and behaviors that normalize that situation. 10:02 And while historians like to bracket events with dates so they can focus their work, the unfolding in reality doesn't really work like that. 10:09 If you take something massive like the Second World War, then when it started really depends on where you lived. In 1937 in the Far East, 1939 in Europe, and 1941 in the USA and the USSR. 10:22 Although the USSR was also part of the 1939 action. 10:25 Then also, while the war is clearly bracketed by hostilities starting, it was really part of a much longer 30-year cycle, which included World War I, and could stretch back even further to German unification in the 1870s. 10:40 Now, I digress a bit, but my point is that AGI development is likely to be a much more evolutionary process than a sudden event. I very much doubt we're going to suddenly wake up one day and feel like AGI has arrived. 10:53 It sort of feels like the generative AI era started with the arrival of ChatGPT in November 2023, but it didn't. Of course it didn't. 11:02 Generative AI applications have been around before this, and AI development stretches back to the 1950s, with it becoming much faster paced during the 2010s. Such shifts come more gradually than it may seem. 11:15 But Sequoia's straight-up functional approach sidesteps this problem entirely. They don't ask whether AI systems have achieved some abstract benchmark. They ask whether these systems can figure things out autonomously. 11:27 By that measure, the definition debate almost becomes irrelevant. The systems either work or they don't, and increasingly, they do work. 11:36 In my own experience with Google Antigravity, the system worked autonomously for 15 minutes, made decisions, and corrected its own errors, and produced exactly what I needed. 11:45 By Sequoia's functional definition, the ability to figure things out, that's AGI. There are also reports of Tesla achieving a US coast-to-coast fully self-driving trip of 2,700 miles. 11:58 That sounds like a very long horizon agent to me, doing a lot of things in order over the course of about 45 hours of driving time. Of course, I share some skepticism too.Otherwise, I wouldn't have made a video about it. 12:11 And I think some of the skeptics raise many valid points. Yann LeCun, formerly the chief AI scientist at Meta and Turing Award winner in 2019, remains critical. 12:22 In mid-2025, he said there was absolutely no way that we would reach human-level AI by scaling up LLMs. 12:29 We would be able to ask questions to an artificial intelligence with a gigantic memory, but this is quite different from having a Nobel Prize-winning equivalent that is able to create novel insights. 12:39 He doubled down on this view in December, saying, To 12:44 this, Google DeepMind CEO Demis Hassabis said LeCun was plain incorrect here. 13:05 He's confusing general intelligence with universal intelligence. 13:09 Now, Hassabis doesn't reject AGI, but he has some much more conservative timelines on it than, say, Sam Altman or Elon Musk, predicting that it's more likely to arrive in five years than the next two. 13:20 But what's interesting about LeCun and Hassabis is that they both work or have worked at two of the big tech corporations involved in the AI race and have much longer timeframes than those working on the smaller AI-focused companies like XAI, OpenAI, and Anthropic. 13:36 It correlates to me that hype from the newer companies is driven for investment reasons. 13:42 LeCun does cut quite a lonely voice amongst the top executives and scientists by saying large language models won't enable human-level machine intelligence. But there are still strong arguments for why they won't scale. 13:53 Tim Detmers from Carnegie Mellon argues that AGI and superintelligence are physically impossible fantasies born from treating intelligence as an abstract concept rather than a physical process. 14:06 Mainly this comes from a scaling problem and that we're near the limits of what we can do with current architecture, a view echoed by OpenAI's former chief scientist Ilya Sutskever. 14:16 According to Detmers, the future belongs not to those chasing mythical AGI or superintelligence, but to China's approach of widespread economic diffusion, integrating good enough AI everywhere for practical productivity gains rather than racing toward an impossible goal. 14:33 I also find the human-level AGI definition philosophically difficult. Going back to definitions, human intelligence is pretty broad. 14:41 Only a minority of people are really good at coding, for instance, and even fewer are Nobel Prize-winning scientists. 14:48 But if we were to create a country of geniuses in a data center, then it wouldn't take very long to bypass human-level intelligence and go into much faster learning and feedback loops. 14:58 Although scientific observation in biology may somewhat hold this process up, simulated scenarios could lead to faster breakthroughs. 15:07 In Superintelligence, Nick Bostrom writes that the train may not stop, pause, or even decelerate at Humanville Station. It's likely to swoosh right by. So powerful AI is not actually AGI anyway. It's superintelligence. 15:22 And it's one of the reasons why I'm prepared to go with the Sequoia definition of AGI of just being able to figure things out and accept that AGI is already here. 15:31 The thing with the AGI debate is it's impossible to be certain. We're talking about a fairly abstract concept, loosely defined, with varying times predicted of its arrival. 15:42 In my previous video, I presented the case against AGI and found it persuasive. The definition problem, the technical constraints, the history of mispredictions, all of that remains true. 15:53 But something has shifted in my direct experience. 15:56 Watching an AI system work autonomously for 15 minutes, making its own decisions, correcting its own errors, and producing exactly what I needed, that felt different from using a tool. 16:07 It felt like working with a colleague. The Sequoia thesis isn't that we've solved artificial consciousness or created digital minds. Their claim is simpler. AI systems can now figure things out. 16:20 They can work autonomously for extended periods, iterate on their own mistakes, and solve problems they weren't explicitly trained for. By that definition, AGI isn't coming. AGI is here. 16:34 Whether that's a cause of celebration or concern depends on what happens next. Sam Altman writes about a gentle singularity. Eliezer Yudkowsky warns that if anyone builds it, everyone dies. Both could be right. 16:48 What seems increasingly clear is that we're crossing some kind of threshold. The debate about when AGI arrives may have been settled while we were still arguing about what it means. 17:02 You've been watching Absolutely Agantic. We're a media consultancy and startup focused on AI, and we create weekly YouTube videos and a newsletter we send a couple of times a week. 17:13 In a world of AI hype, we're trying to strike a balance, looking at a range of opinions to cut through an increasingly complex world. 17:21 The best thing you can do to support us is to sign up to the newsletter in the description or subscribe to the channel. But also don't forget to like and leave us a friendly comment. See you next time. 17:32 It'll be Absolutely Agantic.
