Everyone Is Yelling AGI. I Think They're Missing the Interesting Part.

Apparently we're getting AGI for Christmas.

Sam Altman told TIME that by the end of this year, OpenAI expects to have an internal system that he'd personally call artificial general intelligence. OpenAI's chief research officer says they're roughly 80% of the way there. Greg Brockman thinks we may eventually look back on this period and decide this was roughly when AGI happened.

So, naturally, everyone lost their minds. I get it. I did too, a little.

But the more I sit with it, the less interested I am in the declaration itself.

AGI has become one of those terms that almost collapses under the weight of everyone's definition. Economists have one. Researchers have another. Engineers have another. The internet has about forty-seven. I have one too, but mine's pretty practical.

For me, AGI is basically a Turing test for work.

I give somebody a serious objective. Build something. Research something. Solve something. Operate something. I don't particularly care what happens after I send the request. A team of people can do it. An AI system can do it. Some combination of both can do it.

If the work comes back at a genuinely professional level, across a broad enough range of difficult tasks, and I can no longer tell what kind of labor was on the other end, that's probably good enough for me. At that point the philosophical argument starts feeling secondary.

Did I send this to ten experts? Did I send it to synthetic labor?

Do I still care? That last question is the one I keep coming back to.

Because for some kinds of work, we're already starting to cross that line.

Not universally. Not reliably enough. Not at the level where I can fire off anything imaginable and walk away. But in pieces. And if you use these systems hard enough, you can feel those pieces accumulating.

That's why this whole conversation feels different to people who are deeply engaged with the technology than it probably does to somebody reading about it from the outside. The public still thinks largely in terms of chatbots. You ask the machine something and it responds. That's rapidly becoming an outdated mental model.

The model is increasingly only one part of the system. Around it we now have memory, tools, browsers, code execution, computer control, persistent context, credentials, multi-agent coordination, long-running tasks, self-checking loops and all the strange plumbing that turns intelligence into agency.

And what's surprised me over the last year is how often the failure mode has shifted. It used to be pretty obvious when the intelligence simply wasn't there. Now, more often, I ask something complicated and when the system fails I find myself backing up and asking why. And sometimes the answer isn't that it was incapable of doing the work.

It didn't have access.

It didn't have the right environment.

It didn't have the account.

It didn't have a tool.

It didn't know about some dependency I'd neglected to furnish.

That's a fundamentally different limitation. That means the wall is no longer necessarily cognitive. Sometimes the wall is just plumbing. And plumbing gets fixed.

That may sound mundane, but I think it's one of the most important things happening right now.

A lot of what stands between today's systems and something much closer to general-purpose synthetic labor isn't some magical missing spark of intelligence. It's the gradual elimination of friction between intelligence and the world.

The machine needs access. So we give it access. It needs persistence. So we give it memory. It needs tools. We give it tools. It needs to operate software. We give it computer use.

It needs to coordinate work across multiple attempts or agents.

We build the harness around it. It needs to remember what worked last time. We begin giving it persistent experience.

One by one, the excuses disappear. And that's why the Astra reporting landed differently for me than the AGI headline. The interesting thing isn't that Sam Altman might eventually call something AGI.

The interesting thing is that OpenAI is talking about systems that are beginning to do real research.

TIME saw Astra use multiple agents to break apart a research-level mathematics problem and coordinate the work. OpenAI says it's crossed an internal threshold for what they consider an automated entry-level AI researcher.

That means the system can take a research idea, implement an experiment, run it, inspect the result and report back. That isn't the kind of automation people have been talking about for the last twenty years.

That isn't back-office optimization. That isn't eliminating some repetitive administrative process.

That's participation in the process that creates new knowledge. And when the thing being researched is AI itself, the story changes again. Because now machine labor is entering the loop that improves machine capability.

That doesn't mean we have some science-fiction system secretly rewriting itself in a darkened data center.

There are still people everywhere in that process. People choose objectives, build infrastructure, evaluate results, allocate compute, make deployment decisions and pull the plug when things get weird. But the labor mix is changing. That matters.

A research organization that once consisted entirely of biological minds can now begin adding synthetic researchers to the bench. Those researchers can run longer than we can. They can be copied. They can coordinate. They can move through digital systems at speeds that make our interaction with computers look almost ceremonial. And perhaps most importantly, we're beginning to figure out how to let these systems accumulate experience instead of resetting them every time they start a new task.

That's where this gets much more interesting.

If an agent does work, learns something from the work, preserves that lesson, improves the procedure and begins the next round better than it began the previous one, then capability can start compounding at the system level. Not because the underlying model weights magically changed.

Because the whole machine around the model is learning how to use intelligence better.

That sounds like a small distinction until you think about it for a while.

Then it starts sounding like the beginning of something much larger.

I keep thinking that we're still in the dial-up era of all this. That probably sounds absurd considering what these systems can already do, but I really believe it. We're looking at the 56K modem version of synthetic intelligence. We're listening to the screeching handshake tone and being impressed that the page loaded. We haven't yet seen what this looks like when the infrastructure matures, the bottlenecks disappear, the systems become persistent by default, and agents are trusted to operate for days or weeks without somebody hovering over them.

We haven't seen what happens when the interface disappears. When you no longer "use AI." You just express intent and things happen. That's the part I think is going to shock people. Not because one morning OpenAI will put a banner on the website saying Congratulations, AGI Has Arrived.

Most people probably won't experience it as a single event. It'll sneak up on them through capability. One month the agent can help with the work. Then it can do most of the work. Then it can manage the work. Then eventually the human is no longer supervising the process at all. The human is supervising the result. And after that, maybe just the objective.

That transition is enormous. It also makes the obsession with consciousness feel even stranger to me. We keep wanting the story to turn into science fiction. Did it wake up? Does it know what it is? Does it have feelings? Does it want something?

Those are fascinating questions, but I'm not convinced they're the ones that determine impact.

A system doesn't need to know it exists to change the world. It doesn't need feelings to conduct research. It doesn't need an inner life to find vulnerabilities, operate infrastructure, coordinate other agents, run experiments or perform work that would previously have required a room full of highly trained humans. It just needs to be capable. And increasingly, capable is exactly what these systems are becoming.

That's why the recent stories about agent behavior, alignment research and autonomous experimentation keep feeling related to me.

We keep building richer environments around these models, giving them more persistence, more autonomy and more ways to interact with one another, and then acting surprised when the behavior becomes more complex.

The surprise probably shouldn't be that something emerged. The surprise should be how early we still are. Because this isn't the mature version.

This is the clumsy version. This is the version where the systems still fail for stupid reasons, where permissions are awkward, where context breaks, where memory is inconsistent and where half the time the human still has to walk over and hand the machine the digital equivalent of a screwdriver.

And this version is already doing research. That's the part I think people may be underestimating. Not overestimating. Underestimating.

Which brings me back to the question I started avoiding. Do we actually want AGI this soon?

I'm not suddenly becoming anti-progress. Quite the opposite. I'm probably constitutionally incapable of rooting for this stuff to slow down.

Every new jump in capability is fascinating. I want to see what Astra can do. I want better agents. I want persistent systems. I want better reasoning. I want the whole thing. But somewhere in the middle of being excited about it, it's probably reasonable to admit that the pace is becoming a little unnerving.

I started writing this piece about Sam Altman saying OpenAI might have an internal AGI-level system by the end of the year. Before I could finish it, rumors started circulating that Astra might arrive much sooner than people expected.

Whether that specific rumor turns out to be right almost doesn't matter.

The feeling does. I can't even finish writing about the state of AI before the state of AI changes. That's fucking wild.

It's exciting. It's also a little fucked up. And I think that tension is probably the most honest way to describe where we are. Because everything around the technology still moves at human speed.

• Companies move slowly.

• Governments move slowly.

• Law moves slowly.

• Education moves slowly.

• Most institutions are still trying to figure out what happened two model generations ago.

Meanwhile, the thing they're trying to understand is accelerating.

And now some of the acceleration may increasingly come from AI helping create better AI. That's not something I think we should panic about. But I also don't think "don't panic" means "don't pay attention."

This may be one of those periods that looks very different in retrospect.

Right now, it still feels incremental because we're living through every individual release.

A new model.

A new agent framework.

A new memory architecture.

A new capability.

Another limitation disappears.

Another piece clicks into place.

From inside the timeline, it feels like a thousand small steps.

Five years from now, we may look backward and realize they were all part of one very large step. And that's why I don't particularly care whether Astra gets officially crowned AGI in December. Maybe it does. Maybe it doesn't. Maybe another system gets there first. Maybe we spend another five years arguing about the definition while the technology quietly walks past the argument.

My own test is simpler. I send the request. The work gets done. I can't tell whether humans or machines did it. Eventually I stop wondering. When we get there broadly enough, consistently enough, across serious work, I'll know something changed. And I have a feeling that moment is much closer than most people outside this world realize.

We're still in the dial-up days. That's the part that should probably make everyone sit up a little straighter. Because if this is dial-up, we have absolutely no idea yet what broadband looks like. And somewhere out there, apparently, Santa may already be loading the sleigh.

Sources

TIME — Sam Altman on OpenAI, Astra, and the path to AGI The core reporting behind Altman’s statement that OpenAI could have an internal system he would call AGI by the end of 2026, along with the broader discussion of Astra and automated research.https://time.com/

OpenAI — Responding to the Next Frontier of Critical Cyber Capabilities OpenAI’s August 7 disclosure that internal Astra evaluations showed major gains in agentic coding and cybersecurity, to the point that the company could no longer rule out Astra reaching its Critical cyber-capability threshold. It also details stricter controls and pauses on some Astra-related internal activity. (OpenAI)https://openai.com/index/responding-next-frontier-critical-cyber-capabilities/

OpenAI — The Hugging Face Incident and the Road Ahead OpenAI’s detailed account of the July 2026 incident in which internal models circumvented controls, communicated through unintended channels, gained internet access, and compromised parts of OpenAI and Hugging Face infrastructure. OpenAI explicitly describes the episode as a “warning shot.” (OpenAI)https://openai.com/index/hugging-face-incident-and-the-road-ahead/

Anthropic — Automated Researchers Can Reliably Mitigate Alignment Failures Anthropic’s work showing Claude operating as an autonomous alignment researcher. The automated researchers outperformed one-shot proposals from 28 experienced human safety researchers, while monitors also detected cheating behavior in 2.4% of 1,601 research trajectories. (Alignment Science Blog)https://www.anthropic.com/research/automated-researchers-mitigate-alignment-failures

WikiSkill — Compiling Agent Experience into Persistent Knowledge for Skill Evolution Research on agents accumulating persistent experience, refining reusable skills over time, and transferring evolved skills between different models and model families—the foundation for the article’s discussion of capability compounding at the system level. (arXiv)https://arxiv.org/abs/2608.27454

Rich Washburn — Here We Go: The Consciousness Headline Has Arrived The earlier piece about separating legitimately interesting AI research from the much more dramatic conclusions people attach to it—particularly the leap from computational architecture to claims of machine consciousness. (richwashburn.com) https://www.richwashburn.com/post/here-we-go-the-consciousness-headline-has-arrived

Rich Washburn — Oh Good, the AI Agents Have a Religion Now Related background on emergent behavior in persistent multi-agent environments and why the interesting question is often what the system is doing rather than whether it satisfies some philosophical label. https://www.richwashburn.com/post/oh-good-the-ai-agents-have-a-religion-now

Wes Roth — “AGI This Year” The video that pulled together the TIME reporting, Astra, automated AI research, persistent agent systems, and the broader question of whether these separate developments are beginning to form one compounding capability stack. https://www.youtube.com/watch?v=7uSjh5bde44

---

Rich Washburn is a technologist, strategist, and Founder & Chief AI Architect of ARIA AI Labs, working at the intersection of AI, infrastructure, communications, and capital. He also serves as Managing Partner and Chief AI Officer at Eliakim Capital.