Google Just Dropped The Singularity Bomb

追加: 2026-05-31

[00:00:02]There's a strange crack opening in the AI world right now. On one side, the benchmarks still say these systems are flawed, limited, and far from real human intelligence. On the other side, AI agents are already coding, researching, paying, planning, helping with science, solving math, and cutting real work from days to minutes. That contradiction is the whole story. Because maybe the singularity doesn't start when AI becomes perfect. Maybe it starts when imperfect AI becomes useful enough to speed up everything around it. Deis Hassabis, the CEO of Google DeepMind, stood up at a conference recently and said, "We're currently standing in the foothills of the singularity." Now, that's not just some casual tech bro hyperbole. This is the guy who literally won a Nobel Prize in chemistry for his work with Alphafold. And he's not exactly known for making dramatic statements. He's usually known for his restraint, and those words will likely end up in the history books, no matter how things play out. But here's the thing. Last June, he predicted AGI might happen between 2030 and 2035. And last week, he narrowed that window down to 202930.

[00:01:12]That's a massive acceleration in prediction, and it happened in less than a year. What makes this moment different is that Habibus isn't just making predictions about some distant future.

[00:01:22]He wanted to be authentic about what he's thinking with AGI. And for him, the singularity means the era we're in right now. He believes that when AGI happens, it'll be 10 times the industrial revolution at 10 times the speed. That's a staggering claim when you really think about what the industrial revolution did to humanity. But Habis isn't alone in this assessment. In January 2026, Elon Musk posted on X that we have entered the singularity and that 2026 is the year of singularity. And shortly after, Daario Amod, CEO of Anthropic, stated that we don't know if AI models are conscious. Patrick Collison from Stripe suggested that quarter 1 2026 could be looked back on as the first quarter of the singularity. OpenAI's VP of research, Aiden Clark, hinted in March that AGI may have in some form already arrived. And OpenAI President, Greg Brockman, said OpenAI has line of sight to AGI, while Mark Andre claimed it was reached roughly 3 months ago with the latest frontier models. Now, what's actually driving this shift? It's not just one thing. There's this concept called recursive self-improvement that's been theoretical for decades, but it's starting to show up in reality. Release cycle compression is the most visible evidence that recursive learning is already happening. While it used to take labs 6 to 12 months between major releases, it now takes weeks. And Frontier Labs are beginning to automate large fractions of their research operations. With their intelligent agent workforces potentially growing from thousands to hundreds of thousands of agents working together towards the single objective of making AI smarter within a year or two. And Hosabibus made it personal, describing how AI coding agents have collapsed timelines that once seemed fixed, saying he's been vibe coding things. Even little game prototypes in his spare 1 or two hours in the small hours of the morning, which would have taken 6 months before. This isn't some abstract demonstration. This is the CEO of DeepMind personally experiencing a massive acceleration in what's possible. The AI agents that everyone's been working on, they're not chatbots anymore. In 2026, AI agents are evolving from conversational systems into operational software that can operate across tools, systems, and workflows inside real business environments, handling complex tasks by planning actions, sequencing tasks, and executing workflows across multiple systems. They're moving into real working roles, becoming embedded in workflows rather than remaining limited to demos. And AWS added new payment capabilities for autonomous agents, enabling AI systems to complete transactions and take more direct action inside enterprise workflows. But let's talk about what's happening in science because that's where things get absolutely mindbending. There's this new thing called Axiomrover and it's quietly been making mathematical breakthroughs.

[00:04:20]Eight papers from Axiom Improver have appeared on Arcive since February with five already accepted at peer-reviewed journals. It proved that 100% of primes are partially regular and under certain conditions that Rammanujan's toao misses 100% of primes. We're talking about century old problems in mathematics being solved by machines. Then there's the Chan Zuckerberg Biohub which released what they're calling a world model of protein biology. It's built on ESMC, a language model trained on 2.8 8 billion sequences from across all of life, plus ESM fold 2 for atomic structures and an ESM atlas mapping 6.8 billion proteins. Multi-agent AI systems like co-scientist and Robin can autonomously generate hypotheses, design experiments, analyze data, and refine research questions, streamlining the scientific discovery process. And these systems demonstrated potential in identifying novel drug candidates and targets in biomedical research. The acceleration is showing up in really specific ways, too. SAP's sustainability AI agents in beta are delivering measurable results, including greater than 50% reduction in packaging, compliance, review hours, scenario, simulation time cut from a day to 20 minutes, up to 80% reduction in manual GHS classification effort, and over 20% fewer packaging compliance errors. These aren't incremental improvements. These are orderof of magnitude shifts in productivity. There's also this fascinating metric from a translation company called translated. According to translated, it takes a human translator roughly 1 second per word to edit another human translator's work. In 2014, editors needed about 3.5 seconds per word to fix a machine translated suggestion. By 2022, that had fallen to about 2 seconds. And if that curve kept moving, machine translation could reach human level editing effort by the end of the decade or possibly sooner. That's a concrete, measurable path toward a specific type of singularity. Now, not everyone agrees we're there yet. Yan Lun says current AI systems aren't genuinely intelligent. His argument is that real intelligence shows up when you solve new problems without any prior training, not in accumulated knowledge. Oriel Vinyals, co-lead of the Gemini program, splits the difference. Today's models are strong at code and math, and reasoning keeps getting more general. And if someone had shown him these models 7 years ago, he probably would have called them AGI. But the ability to learn from experience and produce real breakthroughs is still missing. But here's where things get complicated. ARK AGI 3 was introduced in March 2026 by the ARK Prize Foundation to test interactive experience-based reasoning.

[00:07:08]Can an agent explore, infer goals, build a world model, and keep learning over time? And humans solve 100% of their environments, while frontier AI systems, as of March 2026, score below 1%. So, by some measures, we're nowhere close, but by others, we're already there. What's undeniable is that the infrastructure supporting all this is evolving at a breakneck pace. Nvidia's upcoming Vera CPU based on ARM 64 posted the best performance ever seen on ARM, outscoring top Intel and AMD x8664 chips. Germany's Envision reported the first single molecule spin photon interface using a triplet groundstate carbine opening molecular cubits as a viable platform.

[00:07:54]And get this, CBN Nanotechnologies in Ottawa achieved the first simultaneous spatial and chemical control over mechanic carbon fabrication via an inverted mode STM. That's literally placing atoms on demand, dragging us closer to those diamondoid dreams people used to think were science fiction.

[00:08:13]Hosibus said he chose his words to provoke more urgency among governments, economists, and the broader public to prepare for increasingly powerful AI, using terms that were a little bit provocative and referring to a potential AI executive order that would mandate testing before new models are released.

[00:08:32]He said safety needs to be accelerated and this is a good moment to kind of strike while the iron is hot. There's also been this concerning situation with Anthropic's Mythos model, which apparently became what people are calling an alarming warning shot about how quickly AI systems are evolving and how unprepared companies and governments remain. The model was deemed too dangerous for public launch, according to reports from April 2026. And the money flowing into this space is absolutely staggering. In just a couple days, in early May 2026, over $5.5 billion in capital was specifically targeting the deployment gap in the enterprise sector. These aren't research projects anymore. This is about getting this stuff into production at massive scale. Hasibus said all the leading labs are quite focused on recursive self-improvement, noting there will be clear gains in terms of speed of research, but also risks with that type of system. And while we're not yet at the point where the systems are getting better on their own, the pace of development is clearly accelerating with what's being seen as soft self-improvement in the sense of coding agents making engineers much more productive. What's fascinating is watching how different parts of the tech ecosystem are responding. Robin Hood is now open to AI agents, letting customers hand trading and credit card decisions to AI. Bus Patrol, which installed AI cameras on tens of thousands of US school buses, plans to convert them into automatic license plate readers and hand the data to cops. YouTube is automatically tagging significant AI use. The world is actively reorganizing itself around these systems. Alphabet self-driving car division, Whimo, is testing AI models that would give autonomous vehicles a kind of imagination to react to unpredictable or dangerous situations. And Hosabus said texttovideo models could be the key to generalpurpose robotics and artificial general intelligence, noting that an AGI is going to have to understand the physical world. The political response has been chaotic. On May 21st, 2026, the president was supposed to sign an executive order creating a voluntary federal review process for Frontier AI systems. But hours before the ceremony, he pulled the plug. The concern was slowing America's AI industry at exactly the moment the United States believes it has the lead. Illinois did pass SB 315, requiring Frontier Labs to publish catastrophic risk plans alongside a first in the nation third-party AI safety audit mandate. What's happening in the scientific community is remarkable, too. New science skills in Google anti-gravity and three new experimental tools on Google labs are designed to help accelerate core steps of the scientific method built with co-scientist alpha evolve empirical research assistance and notebook LM.

[00:11:27]Scientists are literally using AI to build the tools that will create better AI. The debate about whether this is actually the singularity comes down to definitions. The idea of singularity describes the moment AI exceeds beyond human control and rapidly transforms society. And the tricky thing is that it's enormously difficult to predict where it begins and nearly impossible to know what's beyond this technological event horizon with no light that flips on and says here it too is. For Hassabis, the word captures something simpler and more immediate. The point beyond which meaningful prediction becomes impossible because the transformation will be so complete. And what makes 2026 significant in his view is not just the pace of model development, but the lived experience of agentic AI systems that plan, act, and deliver results across multi-step tasks with minimal human input. Whether you think we're at the singularity or just approaching it kind of depends on which metrics you look at and which expert you listen to. But what's undeniable is that something fundamental has shifted. The release cycles are compressing. The capabilities are exploding. The infrastructure is evolving at unprecedented speed. And the people building these systems, the ones who actually know what's under the hood, are using increasingly dramatic language to describe what's happening. So yeah, maybe we're in the foothills like Habis says. Maybe we've already crossed the threshold like Musk claims. Or maybe we're still years away like Lun argues.

[00:12:59]But one thing's for sure, the conversation has fundamentally changed and the technology is moving faster than almost anyone predicted just a year ago.

[00:13:07]All right, let me know your thoughts in the comments. Subscribe for more AI updates. Hit the like button if you enjoyed the video. Thanks for watching and I'll catch you in the next one.