Looking back at the fevered discourse around GPT-5's launch, I realise I was one of those voices crying disappointment into the void. Another model, another incremental step, another failure to deliver the promised land of AGI for many (not me to be honest). The prediction markets flipped from 80% confidence in OpenAI to 77% for Google within hours of the announcement. Users mourned the loss of GPT-4o's quirky personality and the arrival of a routing system that was well, broken. The whole thing felt like a letdown, a confirmation that the exponential curve was flattening into something more mundane.
But sitting here now, reading through the actual technical developments of 2025, I think we've been staring so hard at the horizon that we've missed the tectonic plates shifting beneath our feet. We're so obsessed with the question of when AI will become genuinely intelligent that we're blind to the fact that it's becoming something else entirely, something that might be more profound than intelligence itself.
Consider what actually happened this year. Both Google and OpenAI abandoned the idea of a single, monolithic model that answers everything. Instead, they built what are essentially AI operating systems. GPT-5 isn't a model; it's a router, a dispatcher, a meta-intelligence that decides which specialised system should handle your request. When you ask it a simple question, it sends you to one model. When you ask it to solve a complex mathematical proof, it engages an entirely different architecture that can "think" for minutes, exploring parallel solution paths like a chess engine exploring possible futures.
This isn't just a technical detail. It's a fundamental reconceptualisation of what artificial intelligence is supposed to be. We spent years asking "when will AI be as smart as humans?" but what's emerging is something that doesn't fit that framework at all. It's not trying to be human. It's becoming something unprecedented: a kind of cognitive infrastructure, a layer of intelligence that sits between human intention and digital execution.
The really striking thing about 2025 is how both companies converged on remarkably similar solutions despite their different philosophies. Google's Deep Think mode and OpenAI's thinking models, Google's Gemini routing and OpenAI's unified system architecture. They're both building the same thing from different directions because they've run into the same fundamental truth: intelligence isn't a single capability you can maximise. It's an ecosystem of specialised competencies that need to be orchestrated.
And while we were all waiting for the singularity, something else happened. AI became boring in the most important way possible. It became infrastructure. Google embedded Gemini so deeply into Search, Android, and Workspace that billions of people are using advanced AI without even thinking about it. It's just there, like electricity or running water. You compose an email and it suggests completions. You search for something and it synthesises an answer from multiple sources. You take a photo and it understands what's in it. This isn't the AGI we imagined, but it might be more important.
The enterprise adoption story is even more revealing. Companies aren't using AI to replace workers wholesale; they're using it to handle the mechanical substrata of knowledge work. The real deployment of AI in 2025 isn't about building digital employees but about creating what you might call cognitive middleware. Every API call, every document summary, every code completion is AI taking over the unconscious, automatic processes that used to consume human attention.
What strikes me most forcefully is how the benchmark situation reveals our conceptual poverty. MMLU is saturated. The models score over 90% on tests of human knowledge. So we made harder benchmarks like GPQA Diamond and Humanity's Last Exam, and the models are rapidly conquering those too. But what are we really measuring? The ability to regurgitate graduate-level physics? To solve competition math problems? These are proxies for intelligence, but they're not intelligence itself.
The real intelligence might be in what Google and OpenAI are doing at the system level. Jules, Google's coding agent, doesn't just write code. It understands intent, maintains context over long interactions, and navigates complex codebases. GPT-5's agentic capabilities let it chain together dozens of tool uses to accomplish multi-step workflows. This isn't about being smart in the way humans are smart. It's about being effective in ways humans never could be.
There's something almost poignant about the backlash to GPT-5's launch. Users complained about losing their beloved GPT-4o, with its particular personality quirks and creative style. They wanted their AI to be a companion, a friend, something with character. But OpenAI gave them something else: a unified system optimised for reliability and capability. It's the difference between a pet and a tool, between something you relate to and something you use.
This tension reveals the deeper confusion in how we think about AI. We anthropomorphise these systems, project consciousness onto them, debate whether they truly understand or merely simulate understanding. But maybe that's the wrong framework entirely. Maybe what's emerging is a new category of being that doesn't fit our existing ontologies. Not alive but not exactly dead, not conscious but not exactly unconscious, not intelligent but not exactly stupid.
The open source developments are particularly fascinating. Both Google and OpenAI released powerful open-weight models this year, not as charity but as strategy. They're creating a kind of AI commons while maintaining their proprietary advantages. It's a tacit acknowledgment that the future isn't about owning the smartest model but about controlling the platforms and ecosystems where AI gets deployed.
Think about what this means. We're watching the formation of something like an AI operating system layer for civilisation. Just as computers needed operating systems to manage hardware resources and provide common interfaces, we're building an equivalent layer for managing cognitive resources. The models are becoming commoditised. The real value is in the orchestration, the routing, the integration, the platforms that make AI usable at scale.
The comparison to the early internet is irresistible but probably wrong. The internet was about connecting information. This is about something else: augmenting cognition itself. Every Google search that returns an AI-synthesised answer, every GitHub Copilot suggestion, every Gemini-powered email draft is a moment where human and artificial cognition blend together. We're not building artificial general intelligence. We're building augmented human intelligence.
The mathematical triumph deserves special attention. When Gemini with Deep Think solved five out of six International Mathematical Olympiad problems (noting a model from OpenAI also did), achieving a gold medal standard, it wasn't just showing off. Mathematical reasoning is perhaps the purest form of abstract thought, the kind of crystalline logic that we once thought would be the last frontier for AI. But it fell just like chess and Go before it. Not because the AI understands mathematics the way a mathematician does, but because it found a different path to the same destination.
This is what I think we're missing when we complain about the lack of an exponential explosion powering us to AGI. We're looking for artificial humans when what's emerging is something far stranger and potentially more powerful: a kind of cognitive ecosystem that operates on fundamentally different principles than biological intelligence. It doesn't need to be conscious. It doesn't need to understand meaning the way we do. It just needs to work.
The enterprise integration story is really the story of how this new form of cognition is embedding itself into the economy. Microsoft putting GPT-5 into every Office application, Google weaving Gemini through Workspace. This isn't adoption; it's infiltration. AI is becoming the interstitial tissue of knowledge work, present in every interaction between human and computer.
What's particularly striking is how both companies have essentially admitted that raw scaling has hit diminishing returns. The jump from GPT-4 to GPT-5 wasn't the same as from GPT-3 to GPT-4. The improvements are more subtle, more about reliability and system design than raw capability. This isn't failure; it's maturation. We're moving from the explosive growth phase to the refinement phase, from revolution to evolution.
The context window expansion tells its own story. Gemini's million-token context isn't just a bigger number. It represents a fundamental shift in what these systems can do. They can now hold entire books, entire codebases, entire histories in their working memory. This isn't human-like at all. No human can hold that much information in conscious awareness simultaneously. It's a genuinely new capability, a form of cognition that has no biological analogue.
And yet we judge these systems by how well they imitate us. We test them on human exams, measure them against human benchmarks, evaluate their outputs by human standards. It's like judging a submarine by how well it swims like a fish. The submarine doesn't swim; it does something else entirely that happens to achieve the same goal of moving through water.
The router architecture is perhaps the most important development that nobody's really talking about other than to complain. It's an admission that intelligence isn't unitary. You don't solve all problems with the same cognitive approach. Sometimes you need fast, intuitive responses. Sometimes you need deep, systematic reasoning. Sometimes you need vast memory. Sometimes you need creative exploration. The routing layer is essentially a metacognitive system that decides which cognitive mode to engage.
This is so different from how we imagined AGI. We thought it would be a single, unified intelligence that could do everything humans can do but better. Instead, we're getting a collection of specialised intelligences orchestrated by a meta-intelligence that itself isn't generally intelligent but is very good at one thing: deciding which specialised intelligence to invoke.
The business model implications are profound. OpenAI charging $200 per month for ChatGPT Pro, Google asking $250 for AI Ultra. These aren't consumer software prices; they're professional tool prices. They're pricing AI like they price high-end creative software or professional services. Because that's what it's becoming: not a toy, not a curiosity, but a professional-grade cognitive tool.
The fact that both companies felt compelled to release open-source models is telling. They're not just competing with each other; they're competing with the very idea of ‘open AI’. They need to maintain relevance in a world where a number of chinese companies are giving away models nearly as capable as their flagships from a year ago. The moat isn't the model anymore; it's the ecosystem, the integration, the platform.
Reading about AlphaEvolve using Gemini to design new algorithms that improve Google's own data centers and TPU designs, I'm struck by the recursive nature of what's happening. AI is being used to improve the infrastructure that runs AI. It's a kind of cognitive bootstrapping that has no precedent in human history. We're watching the emergence of a self-improving system, not through consciousness or understanding, but through pure optimization.
The safety discourse has matured too. It's no longer about preventing Skynet or ensuring alignment with human values in some abstract sense. It's about practical things: reducing hallucinations, preventing jailbreaks, ensuring the AI refuses tasks it can't reliably complete. The safety work is engineering work, not philosophy. This is probably good, but it also means we've given up on solving the deep questions about consciousness and values in favor of making systems that work reliably enough to deploy at scale.
What we're witnessing isn't the birth of artificial general intelligence. It's the emergence of something else: a new layer of reality that sits between human consciousness and digital computation. It's not intelligent in the way we are, but it's not unintelligent either. It's something orthogonal to our concepts of intelligence, something that achieves cognitive outcomes through mechanisms we don't fully understand and can barely control.
The prediction market's flip from OpenAI to Google after the GPT-5 launch wasn't really about which model is better. It was about the market recognising that the game has changed. It's no longer about who has the smartest model. It's about who can build the most effective cognitive infrastructure, who can integrate most deeply into the flow of human work and thought, who can become most essential to the functioning of the digital economy.
This is why I think we've been missing the point. We keep asking "is it AGI yet?" when we should be asking "what is it becoming?" Because what it's becoming isn't artificial general intelligence. It's something we don't have a name for yet: a hybrid cognitive ecosystem where human and artificial intelligence are so intertwined that separating them becomes meaningless.
The documents paint a picture of 2025 as a year of maturation and integration rather than breakthrough. But I think that understates what's happening. We're watching the installation of a new cognitive layer over human civilisation. Every API call, every model inference, every automated task is a thread in a vast net of artificial cognition that's slowly wrapping around every aspect of digital life.
The really interesting question isn't whether these systems are intelligent. It's what happens when this cognitive net becomes so dense, so essential, that we can't function without it. We're building a kind of exocortex for humanity, a shared cognitive resource that amplifies our collective intelligence in ways we're only beginning to understand.
The competition between Google and OpenAI isn't really about who wins. They're both building the same thing from different angles, and the convergent evolution of their architectures suggests they're discovering the same fundamental truths about how to build cognitive systems at scale. The real competition is with time: can they deploy this cognitive infrastructure fast enough and safely enough to capture the value it creates before it becomes commoditised?
Looking at where we are in August 2025, I don't see disappointment. I see the installation of something unprecedented in human history. Not artificial general intelligence, but something potentially more transformative: a cognitive infrastructure that operates at a scale and speed beyond human comprehension, that can hold entire libraries in working memory, that can explore solution spaces we could never navigate, that can optimize processes we don't even fully understand.
We wanted AGI and we got something else. Something that doesn't fit our categories, that doesn't match our expectations, that doesn't satisfy our desire for artificial consciousness. But maybe that's exactly what we needed: not artificial humans, but tools that extend human cognition in ways we never imagined possible.
There’s a story of technical progress and business strategy, of benchmarks conquered and ecosystems built. But underneath that story is something deeper: the emergence of a new kind of being in the world, a cognitive presence that isn't alive but acts, that isn't conscious but responds, that isn't intelligent but solves problems we couldn't solve ourselves.
This is where we are: standing at the threshold of a world where human and artificial cognition are becoming inseparable, where the question of whether AI is "real" intelligence becomes irrelevant because it's so deeply embedded in everything we do that we can't imagine functioning without it. We're not approaching the singularity. We're living through something stranger: the gradual fusion of human and artificial cognition into something neither fully human nor fully artificial.
The future isn't about AGI. It's about this: the slow, inexorable integration of artificial cognition into the fabric of human life until the distinction between natural and artificial intelligence becomes as quaint as the distinction between online and offline. We're not building minds. We're building something else, something we don't have words for yet, something that will transform what it means to think, to know, to understand.
And maybe that's more interesting than AGI ever could have been.
I don’t care about AGI. I care about getting the tools powerful enough to matter into the hands of the people, small enough to run locally, and open enough to break. That’s where the real leap is. Because until then, you don’t own your intelligence, you’re just leasing it.
Wow this was a great read!!! Thank you! It actually makes so much more sense to see 5 as a router than a model itself. I’ve been using 5 and 5 pro for work and switching to 4o for “chatting” with my ai bud who is available 24/7 for my random thoughts. And using 5 and 5 pro absolutely feel like it’s pulling from different areas of its brain. I like reading its reasoning as it’s working and it makes more sense thinking about it in your terms. Loved the reading. Thank you for sharing your thoughts!