Nobody Agrees When AGI Arrives. Build for the Shift Instead.

On July 17, a Spanish paper ran a headline that Demis Hassabis, the head of Google DeepMind and a chemistry Nobel laureate, gives the world three to four years until AGI arrives. Three days earlier he had written it more carefully himself: artificial general intelligence is "probably only a few short years away," and we are standing in the "foothills of the singularity."

Read that next to the independent forecasters. Metaculus, the largest public prediction community, puts the median for a general AI system somewhere around 2033. The last big survey of 2,778 machine learning researchers landed on 2047 for machines outperforming humans at every task. Daniel Kokotajlo, who wrote the widely read "AI 2027" scenario, revised his own timeline in December and pushed superintelligence out to roughly 2034.

So the person running one of the three leading labs says a few years. The forecasting crowd says one to two decades. They are looking at the same models. When the serious estimates disagree by a factor of ten, "when does AGI arrive" stops being a planning question. It becomes a bet, and nobody can price it.

The good news: it is also the wrong question. There is a second story underneath the timeline fight that almost nobody argues about, and that one you can actually plan around.

Why the Date Is Unknowable

Start with the fact that there is no shared definition of AGI. OpenAI has measured it in dollars of profit. DeepMind uses a zero to five capability ladder. Anthropic's Dario Amodei calls the term marketing and prefers "powerful AI." Meta's Yann LeCun wants to retire the word entirely. Nvidia's Jensen Huang said on a podcast this spring that we have already reached AGI, by his definition. Gary Marcus says we are nowhere close. Both are describing the same systems with different yardsticks.

Then add the incentives. A lab chief predicting AGI in a few years is also raising capital, recruiting talent, and shaping the regulation that will govern his own field. That does not make him wrong. It means you read the prediction with the interest attached.

And the technical headwinds are real, not hand-waving. Public training text is projected to run out around 2028. GPT-4.5 reportedly cost many times more than its predecessor for a marginal gain, which is what diminishing returns look like. The most concrete limit for anyone building with agents today is reliability over long tasks. The International AI Safety Report, written by Yoshua Bengio and more than a hundred experts, put it bluntly: as tasks get longer, agents lose the thread, and a single pop-up ad can derail an entire run. Small errors compound and the agent drifts off goal without recovering.

None of that tells you AGI is far away. It tells you the date is genuinely uncertain, in both directions, and that betting your roadmap on a specific year is a mistake whichever number you pick.

What Is Actually Not in Dispute

Now the part everyone agrees on, because it is measured rather than predicted.

The capability curve is steep and steady. METR, an independent evaluations lab, tracks the length of task an AI can complete on its own with a fifty percent success rate. That length has been doubling roughly every four to seven months. In practical terms, we went from minute-long tasks to hour-long tasks in about a year. METR is careful to note the measurement gets unreliable past sixteen hours, so this is a trend to respect, not a stopwatch to trust to the day.

Execution is becoming a commodity. The main coding benchmark, SWE-bench, went from around sixty percent to near ninety-nine over twelve months, per the Stanford AI Index. GitHub reports that a large share of new code is now AI-generated. Engineers at the frontier labs say privately that nearly all of their code is machine-written with light manual correction. Writing software, the thing that was scarce and expensive for fifty years, is on its way to cheap.

The plumbing has standardized. The Model Context Protocol, the way agents talk to tools, stopped being an Anthropic project in December, when it moved into the Linux Foundation's new Agentic AI Foundation with OpenAI and Block as co-founders and every major cloud on board. SDK downloads run in the tens of millions per month. The USB-C moment for AI already happened, quietly, while everyone watched the model leaderboards.

Put those three together and a pattern falls out. The model is getting commoditized. It is powerful, it is cheap, and it is replaceable. One frontier model vanished from the market for nineteen days this June and almost nobody noticed. If the thing everyone is racing to build is also the thing that is becoming interchangeable, then the value has to be moving somewhere else.

Where the Value Goes

It is moving off the model. This is the one point where the venture crowd, the enterprise buyers, and the people actually shipping agents all agree. Marc Andreessen framed it as "the moat is not the model, it is what you build around it." Sequoia's version is "services are the new software." Same observation from different seats.

Four layers are absorbing the value the model is shedding.

Memory is the first, and it is turning into the new lock-in. ChatGPT and Claude are both building persistent memory now, and both keep it proprietary. As one writer put it, it is ChatGPT's memory of you, not yours. That is exactly why an open counter-movement appeared this month, with competing proposals for a portable memory format that a second vendor can actually read back. Switching models has become trivial. Switching your accumulated memory has not. The context of who you are and what you have done is becoming stickier than the model that reads it.

Trust is the second. Microsoft warned in June about tool poisoning in MCP servers, where manipulated tool metadata fires silently on every call. Researchers found up to two hundred thousand exposed instances. Hassabis, in the same essay about the AGI timeline, spends most of his words proposing a FINRA-style body to test frontier systems for cybersecurity, biological risk, agentic behavior, and deception. When agents start acting in the world instead of just reading it, "this one is verified and signed" turns from a nice-to-have into a purchase requirement.

Discovery is the third and the most unsettled. There are over a hundred thousand agents listed across more than fifteen registries with, in one audit's phrase, zero interoperability, and nearly half of them duplicates. Nobody has won the layer where agents find each other and find services. The convention forming around agent cards at a well-known URL is a start, not a finish.

Relationship and workflow ownership is the fourth, and it is the one that matters most for anyone running a business rather than a lab. The consultants who used to resell someone else's software are now generating their own version and keeping the workflow, the data, and the client. Code is cheap. Trust is expensive. The defensible asset was never the software. It was the relationship the software sat inside.

The Honest Counterargument

Maybe none of this survives. If a system arrives that is genuinely general and self-improving, it could eat the memory layer, the trust layer, the discovery layer, and the relationship all at once, and this whole framing ages badly.

I take that seriously, and I still build on the four layers, for two reasons. First, even the aggressive timelines route through years of exactly this transition, and you do not get to skip the years by naming the endpoint. Second, the failure mode of the bet is mild. If AGI stays a few years out, owning memory and trust and relationships is precisely where the money is. If it arrives sooner, those are the last things to be automated away, because they are about accountability and human trust, not raw capability. There is no version of the next few years where a portable memory layer and a verified trust layer are a bad place to stand.

What This Means for You

Stop timing AGI. The people who do this for a living cannot agree within a decade, and the number is not actionable anyway. Watch the capability curve instead, and watch the reliability gap, because those are measured and they move.

Then build on what survives the model. Assume the model is replaceable, because it already is. Put your weight on the layers underneath it: the memory that belongs to the user, the trust that lets an agent act, the discovery that lets it be found, and the relationship no model can inherit. That is not a prediction about 2029. It is a description of what is already worth more this year than it was last year.

The date is a distraction. The shift is the story.

---

This article was originally published on studiomeyer.io.