Think of it as Yoga for the brain, we are going to stretch a bit.
IS AGI HERE NOW?
AGI may not arrive by announcing that it is intelligent. It may arrive the day we stop telling the machine how to do the job and simply tell it what needs to be done. — YNOT
For years we have been waiting for Artificial General Intelligence as though it were going to arrive wearing a tuxedo, knock politely on the front door, and announce:
“Good evening. I am AGI.”
That was probably our first mistake.
Technology almost never arrives with the courtesy of introducing itself.
The automobile didn’t announce the end of the horse.
The Internet didn’t announce the end of the encyclopedia salesman.
And AGI may not announce itself either.
It may simply start doing the work.
And that is why I have begun wondering whether AGI is already here.
Not theoretically. Not someday. Now.
Two systems in particular have made that question much harder to dismiss: OpenAI’s GPT-6 Astra and Anthropic’s Claude Fable 5.1.
The argument isn’t that either one is conscious.
I don’t particularly care whether the machine sits around at night contemplating mortality, writing poetry about electric sheep, or wondering why humans put pineapple on pizza.
That isn’t the useful definition of intelligence.
The useful question is much simpler: Can I give it a complicated problem without telling it how to solve the problem—and can it figure out what to do?
Because that is what we normally call intelligence when a human does it.
For the last several years, AI worked something like an extraordinarily talented employee who required extraordinarily detailed instructions.
“Read this.” “Summarize that.” “Write this program.” “Make this picture.”
“Use this database.” “Follow these twelve steps.”
We gave AI the destination and the road map.
Now something very different is happening.
One early Astra example described in the material I was reviewing involved giving the system tens of thousands of emails, years of writing, calendars, contacts and unfinished work.
Nobody gave it a detailed recipe.
Astra selected its own approach, obtained the software it decided it needed, built an environment, and constructed a personal knowledge system.
Think about what changed.
The human didn’t say:
Step 1. Step 2. Step 3.
The human essentially said:
“Here is my mess. Figure it out.”
That sounds suspiciously like what I have been telling employees, programmers and consultants for forty years.
OpenAI itself describes Astra as its most capable model for difficult end-to-end work, including coding, research, computer use and complex multistep jobs.
It has crossed another remarkable threshold as well. OpenAI says Astra can, given the appropriate tools and access, discover previously unknown computer vulnerabilities and develop methods for exploiting them without a human guiding every step. OpenAI consequently classified Astra at its highest “Critical” cybersecurity capability level.
That isn’t autocomplete. That isn’t a better spellchecker.
That is a machine receiving an objective, examining an unfamiliar environment, developing a strategy, using tools, discovering obstacles and modifying its behavior to accomplish the objective.
Now consider Claude Fable 5.1.
Anthropic specifically describes Fable as being designed for long-horizon agentic work: jobs lasting hours, involving multiple applications, where the AI plans the work, selects and uses tools, recovers when something fails and can operate unattended.
That last part matters enormously. Recover when something fails.
Humans don’t really earn their money when everything goes according to plan.
We earn it when something goes wrong.
The printer doesn’t work. The database structure isn’t what the documentation said.
The supplier doesn’t answer. The software crashes.
The API changes. The instructions contradict reality.
Old AI would stop and ask: “What should I do?”
The new generation increasingly says: “That didn’t work. I’ll try something else.”
The source material gives a wonderful example of Fable doing essentially that—changing its planned approach when another route became inconvenient and selecting an alternative tool to accomplish the original objective.
And there, I think, is the important line.
Intelligence is not knowing the answer.
Intelligence is figuring out what to do when you don’t know the answer.
That distinction changes everything.
An ordinary computer executes instructions.
Traditional software follows rules.
Early AI answered questions.
Agents performed tasks.
But these new systems increasingly appear capable of being handed responsibility.
“Keep this project moving.”
“Figure out what’s wrong with this software.”
“Watch these accounts and tell me when something matters.”
“Research this subject and update me when the evidence changes.”
“Build me something that accomplishes this objective.”
Those aren’t prompts anymore. Those are jobs.
And that may ultimately be a much more useful definition of AGI than another academic benchmark.
Can one general-purpose intelligence move between programming, research, writing, finance, images, documents, browsers and unfamiliar software?
Can it determine what tools it needs?
Can it plan? Can it remember?
Can it notice that its original plan failed?
Can it invent another plan?
Can it continue working without somebody standing behind it saying, “Now click the blue button”?
And most importantly: Can it take a problem that formerly required a competent human and return with a result?
If the answer increasingly becomes yes, we may discover that AGI arrived while half the world was still debating what AGI meant.
There are plenty of reasons to remain cautious.
Astra can still be wrong.
Fable can still misunderstand an objective.
Agents can make bad decisions.
They can follow the wrong path with tremendous enthusiasm—which, incidentally, is another characteristic they share with humans.
Reliability may now be the bigger obstacle than raw intelligence.
A system that succeeds 98 percent of the time can be astonishing.
A system controlling your bank account, your company or your infrastructure that succeeds 98 percent of the time can be terrifying.
So humans are certainly not leaving the picture.
Our role may simply be changing.
We are moving from telling computers how to do the work to telling intelligent systems what we want accomplished.
That is an enormous distinction. And perhaps AGI was never going to be a machine that suddenly became human.
Perhaps that was another bit of human vanity.
Perhaps AGI is simply the moment when a machine becomes sufficiently general, sufficiently capable and sufficiently autonomous that we stop thinking of it as a tool we operate and start treating it like an intelligent colleague to whom we delegate responsibility.
If that’s the definition, Astra and Fable are getting awfully close.
Maybe they have already crossed it.
History has a nasty habit of beginning before historians agree that anything happened.
And someday people may look back at 2026 and ask:
“When did AGI arrive?
The answer may turn out to be: It arrived when we stopped telling the computer how to do the job—and started simply telling it what job needed to be done.
The rest of us were just too busy arguing about the definition to notice.
Who is ahead?
This is not just Astra or Fable—it is that OpenAI, Anthropic, xAI, Meta, Google, and the Chinese model ecosystem are all converging on the same kind of autonomous agent.
Here is the comparison I would use right now, September 6, 2026:
Company
Current system/model
What makes it interesting
AGI-like characteristic
OpenAI
GPT-6 Astra
Best end-to-end computer worker
Takes a goal, operates software, researches, codes, verifies, adapts
Powerful coding, long-horizon work, computer control; open ecosystem component
Can be embedded into custom autonomous systems
1. OpenAI Astra — probably the best general digital worker
Astra’s distinction is not simply reasoning. It is computer use.
OpenAI explicitly designed it for complex workflows involving browsers, code, research and professional software. OpenAI reports Astra at 41.4% on AutomationBench versus 31.4% for Fable 5.1, and it substantially beats GPT-5.6 Sol on that benchmark. (OpenAI)
In other words:
Give Astra a computer and a destination. It increasingly figures out the road.
That’s perhaps the strongest practical AGI argument of the group.
2. Anthropic Fable 5.1 — perhaps the best independent thinker/engineer
Fable is fascinating for a different reason.
Anthropic says it is specifically designed for jobs lasting hours and spanning multiple applications. It plans the job, uses whatever tools it needs, recovers after failures and can operate unattended. It can also conduct multi-day autonomous coding sessions and build its own tests to verify its work. (Anthropic)
That sounds less like software and more like:
“Here is the problem. Come back when you’ve figured it out.”
Fable also currently scores higher than Astra on at least one broad intelligence composite cited by OpenAI—65.7 versus Astra’s 61.2—so “best” depends heavily on what you’re measuring. (OpenAI)
3. xAI Grok — potentially the most interesting multi-agent system
This deserves much more attention than I gave it before.
xAI has Grok 4.6, explicitly optimized for long-running agents. (SpaceXAI)
But xAI also has something arguably more important:
Grok Multi-Agent.
Instead of one AI doing everything sequentially, Grok can orchestrate several agents simultaneously:
Agent A: research this.
Agent B: analyze the numbers.
Agent C: investigate competing explanations.
Agent D: verify everybody else’s conclusions.
Then combine the results. xAI officially describes the system as multiple specialized agents collaborating in real time. (X.ai Documentation)
And then there’s Grok Bot.
xAI describes Bots as always-on AI teammates with their own computer that can work across applications and keep operating 24/7, returning when human approval is necessary. (SpaceXAI)
That is a huge conceptual step.
It isn’t:
AI as chatbot.
It’s:
AI as workforce.
4. Meta — the personal superintelligence approach
Your transcript mentions Meta and says:
Watermelon is coming soon.
That appears to refer to Meta’s reported internal Watermelon project. But I would not present Watermelon as a released product yet. Public reporting describes it as an internal/unreleased model rather than something we can meaningfully benchmark today. (Datallm Lab)
What Meta actually has publicly now is Muse Spark 1.3.
Meta describes it as trained for long-horizon agentic workflows, capable of maintaining context and prior results and working through conflicting or messy information. (AI Meta)
Meta’s strategic difference is important.
OpenAI seems to be saying:
“Give AI your work.”
Anthropic:
“Give AI your difficult intellectual problem.”
xAI:
“Give a team of AIs the job.”
Meta increasingly seems to be saying:
“Give AI your life context.”
Meta explicitly calls its destination personal superintelligence. (AI Meta)
That may become incredibly powerful because Facebook, Instagram, WhatsApp, Messenger and Meta’s devices potentially give it something other models struggle with:
personal context.
5. Google Gemini 3.8 — the potentially enormous scale play
The transcript doesn’t emphasize Google nearly as much, but we shouldn’t leave Google out of the real comparison.
Google released Gemini 3.8 Flash on September 2, describing it specifically as a model for agentic workflows, software engineering and complex multi-step reasoning. (blog.google)
That matters tremendously for agentic systems because agents can consume vast numbers of tokens while repeatedly planning, checking and retrying.
Google doesn’t necessarily have to have the absolute smartest single agent.
If Google can make 100 competent agents economically practical, that’s another road toward something resembling AGI.
6. GLM-5.3 — perhaps the wildcard
This part of your source needs one correction.
The transcript says GLM-5.3 is available as open weights.
The situation appears more nuanced now.
Z.ai’s materials describe GLM-5.3 as very strong in coding and long-horizon agentic work, while GLM-5.3-Flash is the version clearly associated with the newer open-source multimodal release. There has actually been community criticism over GLM-5.3 itself not using the same FLOSS licensing.
But the larger point in the transcript is absolutely correct:
Open models change the AGI equation.
Because then somebody can take the model and build:
their own memory
their own tools
their own databases
their own computer control
their own agents
their own permission system
their own persistent objectives
without depending completely upon OpenAI, Anthropic, Meta or xAI.
That becomes particularly interesting for the kind of local-AI systems you’ve been experimenting with.
What I now think the real AGI story is
Forget model rankings for a moment.
Look at what is happening independently: Astra → one extremely capable autonomous worker.
Fable → one extremely capable autonomous intellectual worker.
Grok → teams of autonomous workers.
Meta → an autonomous worker that increasingly knows you.
Gemini → inexpensive agents deployable at enormous scale.
GLM/open models → autonomous intelligence anyone can build into their own infrastructure.
And Anthropic is explicitly discussing teams of agents working autonomously for days rather than humans supervising every individual action.
The strongest argument for “Is AGI Here Now?” may actually not be:Astra is AGI.
It may be: “AGI may already be here because six competing laboratories have independently arrived at machines that can understand goals, make plans, use tools, operate computers, recover from mistakes, remember context and increasingly work without continuous human supervision.”
That is a much stronger thesis.
Because then AGI isn’t dependent upon whether one company happened to make one magical model.
It’s a technological transition happening across the entire industry simultaneously.
