Is AGI here or not? Did GPT-6 Astra just become AGI? Should You Care?

On September 3rd, OpenAI shipped GPT-6 Astra. Three days later, Jensen Huang, the CEO of NVIDIA, posted these words

Within hours, Gary Marcus, a cognitive scientist at NYU, replied that the claim came with no evidence and no definitions.

Here’s the thing: you don’t care, and you shouldn’t, at least right now, at least about AGI.

But there is something you should care about, and that’s

Report Cards

The same week GPT-6 came out, it took the same test twice. The first time it scored 62%. That’s a D on its report card. The second time it scored 98%, an A-report card.

Nothing about the model changed, not the training, not the size, not how hard it was allowed to think. What changed was the system around it and what it remembered from what it had just worked out. That gap is really the only part of this launch that should matter to you.

What changed for you when GPT-6 came out?

The short answer is probably nothing.

The long answer, though, is more complicated than just nothing.

People are asking whether GPT-6 Astra is AGI. But before anyone can answer that, we need to define what AGI really is. And the definition of AGI is now gone, not disputed, but gone. AGI had a real definition once.

In OpenAI’s charter back in 2018, OpenAI calls AGI, quote,

“Highly autonomous systems that outperform humans at most economically valuable work.”

That definition had teeth because it sat in a contract with Microsoft. The Information reported in December 2024 what the actual AGI trigger was: AGI would be reached when OpenAI systems could generate the maximum total profits owed to its earliest investors, about a hundred billion dollars.

That created a definition with legal weight. Back then, AGI wasn’t technical; AGI was a revenue number. Then on April 27, 2026, that clause got deleted, and the two companies agreed that the revenue share runs to 2030 regardless of OpenAI’s technological progress.

So the word AGI had one hard definition, and it was financial, and four months ago, it removed it. And now, four months later, Greg Brockman calls the definition a spiritual concept.

And on launch day, he said the quiet part out loud

The man announcing AGI told you the word doesn’t mean anything anymore, so believe him. Since the term AGI stopped needing a definition, it also means that it doesn’t affect you anymore. So set it down. Just assume they’re right. Assume we really have reached AGI. The models are only getting better, and that’s not gonna stop anytime soon.

So let’s talk about those two scores because this is the important part.

ARC Prize runs this benchmark. They didn’t just run Astra twice. They actually ran it 12 times and published 12 results. Six of the runs put the model in what they call a standard harness.

That’s kind of their word for the rig a model gets tested inside, and theirs is deliberately plain. The model has to carry forward out in the open whatever notes it decides to keep. Then the other six put that same model inside OpenAI’s own adapter, where the system holds the model’s reasoning state between requests and then compacts the conversation when it runs long.

The machinery around the model remembers what the model just worked out. Then they varied one more thing across both sets: how hard the model was allowed to think. Six settings from maximum down to none at all.

In the plain rig, thinking as hard as it can, 62.7% report card. The same model in the same setting, but this time inside of OpenAI’s adapter where the context is actually provided, 98.6%.

That’s a 36-point difference, but that’s not the important part. This is. Turn the thinking down and the plain rig comes apart. Across six settings, it scores anywhere between 17.5 and 62.7%. That’s a 45-point swing. Inside OpenAI’s adapter, though, it never leaves the high 90s across the same six settings.

96.7% in the adapter in a specific harness with the model’s reasoning effort set to none.

The version of the model that does no reasoning at all inside a system that keeps its context beats the version with reasoning at full strength inside a system that doesn’t by 34 points

The model was the concept. The plumbing was the variable.

ARC-AGI3 is an interactive benchmark. The model plays a game it’s never seen and works out the rules by acting.

On a task like that, holding onto what you just learned isn’t scaffolding around intelligence. It’s part of the intelligence itself. That’s key. If keeping a hold of your own context is part of what being intelligent means, then not keeping a hold of your context isn’t gonna be rescued by buying a more intelligent model.

It’s missing the same thing the model was always missing one level up. Now the question that matters. Does the model work without the right context, or does your model understand the big picture? If you hand the model a question with no memory of last Tuesday, no idea of which of the four copies of the pricing policy is current, and no record of what it worked out for another department last month, what kind of grade do you think it’s gonna get?

Every one of those 12 AGI runs was the same model. So the gap between what AI does in a launch video and what it does on your desk was never a model problem. It’s a harness problem, AGI or not

So should you care? Yeah, just not for the reason in the headline.

For three years, there was one sentence that ended every conversation like this, and it was true

That sentence is no longer available to you, not because AGI arrived, but because it stopped mattering whether it did or not.

Foundational models are incredibly good, whether we tag them as AGI or not.

Whether the model is mediocre or brilliant, a wrong answer requires a brilliant harness to close that gap. Fixing the harness and not the model is the difference between you getting a D on your report card and an A.

Deloitte gave the Australian government a report with academic references in it that do not exist. The firm refunded the final installment of the contract: AU$97,000.

Deloitte Canada put fabricated citations in a health workforce plan for Newfoundland and Labrador, on a contract worth one point six million dollars. A provincial accounting regulator opened an investigation, and the province rewrote its procurement rules.

EY published a cybersecurity report in which sixteen of twenty-seven citations were fabricated, misattributed, or pointed nowhere. The firm took it off its website.

KPMG published a report about how to do AI well. Five of its forty-five citations checked out.

Four firms whose entire product is being right about things. Each of them shipped a document with invented facts to a paying client and got caught by somebody outside the building.

Those are the ones we know about. They became public because a researcher checked, or a journalist checked, or a regulator checked.

Nobody checks the answer your AI gave your project manager on Tuesday.

A question nobody owns

One of the most common things people type about AI right now is a question about the 30% rule. The internet has five answers.

AI does seventy percent of the work, and humans keep thirty percent for judgment. AI does thirty and humans do seventy, which is the exact inverse, published with the same confidence. No more than thirty percent of your homework may be AI-written, which turns out to be a university plagiarism policy. Automate your first thirty percent of repetitive tasks. Put thirty percent of your AI budget into data quality.

Most of those pages admit, in their own text, that the rule is not real. No standard behind it. No regulator wrote it. The closest thing to a source is a Gartner prediction from July 2024 that thirty percent of generative AI projects would be abandoned after the proof of concept, which is about projects dying rather than about splitting work with a machine.

A question with no owner gets answered by people with something to sell, and they all sound certain. It is the whole failure, running in public, where you can watch it happen.

Your company has questions like that too. Yours are simply not on the internet where anyone can check them.

It was trained to guess

The usual explanation for wrong answers is that the model is broken or the technology is immature. It is neither.

OpenAI published a paper in September 2025 arguing that models make things up because we trained them to. The mechanism is worth understanding, because it explains why buying a better model does not fix this.

Models get scored on benchmarks. On most benchmarks, a wrong answer scores zero and “I do not know” also scores zero. Guessing is free. If you are unsure and you guess, you might be right. If you admit you are unsure, you definitely get nothing. Over millions of training examples, the version that survives is the one that always has an answer.

OpenAI demonstrated this on its own models. One of them declines to answer one percent of the time and gets seventy-five percent of its answers wrong. A newer one declines fifty-two percent of the time and gets twenty-six percent wrong. Accuracy between the two moved by two points. The difference is that one of them will tell you when it does not know.

Your AI is not lying to you. It was rewarded, for years, for never saying that it was unsure. So it does not.

## Your documents do not fix it

This is the part where the money goes, and it is the part almost everyone gets wrong.

The standard fix is to give the AI your documents. Point it at your drive, your wiki, your policies. Ground it in your own data. Every vendor in this market sells that, and I sell a version of it too, so weigh what follows accordingly.

It helps. It does not fix it.

RAGTruth, published at a computational linguistics conference in 2024, took about eighteen thousand AI answers where the model had been handed the source documents, and had people annotate them word by word. GPT-4, with the documents in front of it, made claims those documents did not support in roughly one answer out of six.

One in six, with the answer sitting in the file.

Then there is the worse finding. Stanford researchers ran twelve hundred questions where they planted a wrong answer in the retrieved document on purpose, and watched what the model did when that document disagreed with what it already knew correctly. It followed the document. More than sixty percent of the time, the model threw out its own correct knowledge and repeated the error it had been handed.

Grounding an AI in your documents does not make it right. It makes it obedient to whatever it pulled.

So if your drive holds two versions of the same policy and one of them is from 2021, you have not built a system that gives correct answers. You have built one that confidently hands you whichever version it reached first. Ask again next week and get the other one.

Researchers at Tsinghua and Cambridge have a name for this: Inter-context conflict is two retrieved sources contradicting each other, and their survey names outdated documents sitting beside current ones as a principal cause.

Last week I said your knowledge sits on islands. This is what happens when you hand a machine all the islands at once and do not tell it which one is current.

Nobody can tell you the number

People want to know what percentage of AI answers are wrong. The honest answer is that the number does not exist.

You will see three percent. That comes from a leaderboard where models are handed one document and told to summarise only what is in it. It measures faithfulness to a page, not truth about the world, and secondary coverage constantly relabels it.

You will see forty-five percent, from the European Broadcasting Union and the BBC. That is a real study across twenty-two broadcasters, with journalists grading twenty-seven hundred answers. Forty-five percent is the share of answers with at least one significant issue across four criteria, and a missing citation counts as an issue. The accuracy-specific figure in the same report is twenty percent. Guess which one gets quoted. Worth adding that the journalists doing the grading were employed by the same broadcasters whose content was being cited, which the report discloses openly.

You will see ninety-five percent, as in ninety-five percent of AI pilots fail. That traces to a report whose own text says something different, with the calculation never shown and the survey collected at four conferences.

Three different things, measured three different ways. Anyone quoting you one number for how often AI is wrong is selling you something, and that includes anyone quoting you a low one.

The number for your company is not in any study. It is in your building.

The hour that finds it

You need a room, four to eight people who answer questions for a living, and no preparation. Do not clean anything up first. Cleaning up first is how you get the pilot result instead of the real one.

Step one, fifteen minutes. Everyone writes privately. Read this out and then stop talking:

Write down the three questions you were asked this month that you could not answer in five minutes. Not the hard strategic ones. The ones where the answer definitely exists somewhere, and you could not get to it. Do not talk while you write.

The silence matters. The first person to speak sets the frame for the room.

Step two, ten minutes. Read them out and find the repeats. The question more than one person wrote down is your gap. It is almost never the one you would have picked.

Step three, twenty minutes. This is the step that does the work. Take the top repeat and trace the answer.

• Where does it live?

• How many places?

• Are those copies the same?

• When was each one last changed?

• Who is allowed to change it?

Most rooms answer the first two confidently and then stall completely on whether the copies are the same. Somebody says there is a version in the shared drive. Somebody else says the real one is in a thread. Nobody knows which is current.

That stall is the whole argument. It is inter-context conflict, sitting in your company, found by hand in twenty minutes.

Step four, fifteen minutes. Price it.

• How often does the question get asked?

• Who answers it?

• What is an hour of that person worth?

Then the one that hurts:

Deloitte’s number was AU$97,000 and a corrected report on a government website. Yours will be different. Write it down anyway.

You leave with one sentence, and it reads something like this. The question we cannot answer quickly is this one; the answer lives in four places, two of them disagree, and it gets asked eleven times a month.

That sentence is worth more than any pilot, and here is why. You did, by hand, with people who work there, exactly what you expected the AI to do on its own. If your own team needs twenty minutes and an argument to work out which version of a policy is current, a model doing it in four hundred milliseconds is not being more careful. It is being faster, and it will not tell you it was unsure.

Fix the sentence, not the model

Pick the one question. Make one version of the answer current. Put it somewhere both your people and your systems can reach, and delete the rest. That is one afternoon. It is also the smallest possible piece of a company brain, and you will have built it before you have bought anything.

Run the hour this week, then tell me in the comments which question turned out to be the repeat. I am collecting them, and the pattern across companies is not what I expected.

New York City spent about $600,000 to build a chatbot that answers business owners’ questions about city rules. In March 2024, reporters at The Markup and the city found the chatbot

• telling landlords they could refuse tenants holding Section 8 vouchers

• telling employers they could take a cut of their workers’ tips

• telling businesses they could refuse cash, three years after a city law required them to accept it.

The city kept it running at roughly half a million dollars a year until the incoming mayor canceled it in January 2026.

The easy read is that the technology failed. I do not think that is what happened.

New York City could not answer those questions quickly through any channel. The rules were real and public and scattered across agencies, notices, and pages written years apart by people who never met, with nothing to say which one won.

The chatbot did not invent a knowledge problem. It published one.

Sit with that, because it is the whole argument.

An AI system does not create the gap between what your organization knows and what it can retrieve. It makes the gap

• legible

• at scale

• to strangers

• in writing

• at the speed of a search box.

So before you evaluate another vendor, run the seven questions below on your own company. None of them are hard. That’s the point. Every one has an answer, and someone in your building knows it. The only variable is how long it takes to reach.

1. What did we tell the last customer who asked this?

Your team answers the same forty questions all year. Pricing edge cases. Whether you support the thing you technically support but do not advertise. What happens when a project goes over scope?

Almost no organization can retrieve its own prior answer to a repeated question. So each person reconstructs one. They are all reasonable, they are all slightly different, and they drift further apart each quarter as the people who gave the early answers move on.

Nobody lied. Nobody was careless. The answer simply has no home, so it gets rebuilt from memory every time, and memory has a slope.

Your customers compare notes. You do not.

2. Which version of this is current?

In February 2024, the British Columbia Civil Resolution Tribunal ordered Air Canada to pay a passenger about $650 after its website chatbot told him he could apply for a bereavement fare retroactively. The airline’s own bereavement travel page said the opposite.

Both statements were live on aircanada.com at the same time.The money is trivial. The defense is the story. Air Canada argued that the chatbot was, in the tribunal’s summary, a separate legal entity responsible for its own actions. The tribunal’s answer runs eleven words longer than it needed to and is better for it:

“It should be obvious to Air Canada that it is responsible for all the information on its website. It makes no difference whether the information comes from a static page or a chatbot.”

A company stood up in front of a tribunal and argued that it was not responsible for part of its own website. That is what happens when an organization loses track of which version of a policy is the real one. It can no longer recognize its own voice.

Ask what your current refund policy is. Then count how many places a customer could find a different answer.

3. Why did we decide it this way?

Organizations are good at preserving decisions and terrible at preserving reasons.

The rule survives. The constraint that produced it does not. Two years later, nobody can tell you whether the rule still holds, because nobody can tell you what it was for. So it stays, because removing a rule you do not understand feels riskier than keeping it.

This is how companies accumulate policies that everyone follows, and nobody will defend. Ask why the approval threshold is set where it is. Watch what happens when the honest answer is that someone set it in 2019 for a reason that stopped applying in 2021.

Of the seven, this one costs the most and shows up on no report anywhere.

4. Who owns this?

A question with no owner has no maintainer. A page with no maintainer isnt stable; its just unedited.

Most companies have a wiki full of documents in exactly this state. They look maintained. They are abandoned in place.

5. Has anyone here already done this?

Someone on your team is currently rebuilding an analysis, a template, a deck, or a process map that already exists somewhere in the company.

They are not lazy. They looked. The search returned nothing useful, so they concluded nothing existed, which is a rational inference from the evidence available to them and completely wrong. The work was there. It was named something else, filed by someone who has since left, in a place they had no reason to look.

You pay for that work twice and count it once.

6. What do we actually promise, in writing?

Not what you intend to promise. What your published surfaces currently say, added together, to a customer who reads them all.

New York City believed it was publishing help for small businesses. What it published was permission to break the law. Air Canada believed it had a bereavement policy. What it had was two.

Every organization has a gap here, because promises accumulate across contracts, proposals, marketing pages, support macros, and whatever a salesperson said in a meeting last March to close a deal. Nobody holds the sum. The sum is what you are accountable for.

7. What did we learn the last time this went wrong?

You have had this outage before. This client relationship has ended this way before. This hiring mistake has a precedent inside your own company, with a post-mortem written by someone thoughtful, saved to a folder that has not been opened since.

An organization that cannot reach its own post-mortems does not have institutional memory. It has an archive, which is a different thing. An archive is where knowledge goes when nobody built a path back to it.

What the seven have in common

None of these are intelligence questions. Every one of them is a retrieval question.

The answers exist.

They’re already inside the company, held by a person, a file, a thread, or a system you are paying for. Nothing on this list requires new insight, new research, or a smarter team. Each one requires a path from question to answer that is shorter than the effort someone is willing to spend before giving up and rebuilding it from scratch.

My own rough number for that threshold is five minutes. I have not measured it, and I am not aware of anyone who has. It is what I have watched happen, repeatedly, and I would rather give you the figure I actually use than a sourced one I do not believe. Under five minutes, people retrieve. Over five minutes, people improvise and move on, and their improvisation becomes the new answer for everyone who asks them next.

Your company is not running on what it knows. It is running on what it can reach inside five minutes. The rest is inventory you are storing and paying for and cannot sell.

What to do about it, in order

Do not buy anything yet.

Take the seven questions into a room with the people who would actually be asked them. Do not discuss them. Time them. Pick one real instance of each, start a clock, and find the answer the way a normal person would.

You are measuring time to answer, not whether an answer exists. Everyone already believes the answers exist. The number that matters is how long the path is, and whether two people walking it separately arrive at the same place.

Rank the seven by how long they took. That ranking is your build order, and my bet is it will not match the order you would have guessed before you started. I want to be clear that this is a prediction based on watching it happen, not a finding from a study. The reason I expect it is structural: leadership gets fast answers by asking a person directly, and that path does not exist for anyone else in the company. The gap you cannot see is the one nobody has ever had to walk around you to reach.

Then fix them in that order. Fixing means giving each answer one home, one owner, one date, and one path that anyone can walk without knowing whom to ask. Do that seven times, and you have started building what I call a company brain: the place where what your company knows is organized, current, and reachable by your people and by any system you point at it.

That work is unglamorous, mostly not technical, and it decides whether every AI dollar you spend after it returns anything.

The bottom line

New York City did not have an AI problem. It had a knowledge problem that a $600,000 AI system read aloud to the public for two years.

Your AI will answer all seven of these questions tomorrow, instantly, in complete sentences, to anyone who asks. It will sound certain either way. The only thing you get to decide is whether it is right.

By the way, I write a weekly piece on the knowledge your company already has but cannot reach. https://halfofit.trickywombat.ai

S&P Global surveyed a thousand enterprise AI leaders and found the share of companies scrapping most of their AI projects before production went from 17 percent to 42 percent in a single year. The models did not get worse. Every model on the market got better.

The projects died in the gap between the pilot and the real company. I keep seeing the same pattern, and it has nothing to do with which model you picked.

The pilot is a lie you tell yourself

In a pilot, somebody picks the documents. They are current. They are clean. They are the ten files that answer the twenty questions everyone agreed to test. The demo works. Of course it works.

Then you point the same system at the real company. Now it is reading a drive nobody has cleaned out since 2021. Two versions of the same policy, both undated. A process that changed in March and lives in one person’s head. A number that sits in a system the AI was never given access to.

Same model. Same vendor. Same prompt. Different answer, because it is seeing something different.

Your people already live in that second version. Somebody on your team gets asked a question. They answer with what they find in five minutes: a file they remember, a person they reach, the email still sitting in their inbox. Then they move on. The better answer was somewhere they did not look. It existed. Nobody reached it.

The structure is the problem, not the people

This is not your team failing. They are smart and they are trying. But companies add tools faster than they organize what those tools hold. Your knowledge is in Slack, in email, in a drive nobody cleans out, in a wiki nobody updates, in the heads of three people who are always in meetings.

Every one of those is an island. When a question comes up, it gets answered from whatever island is closest, not from everything you know. There is no company brain. No single place where what your company knows is organized and reachable. So you run on fragments.

Bigger models do not fix this

This is where the AI spending goes sideways. Companies buy bigger models to fix bad answers. It is the wrong fix.

An AI is only as good as what it sees. Point the smartest model in the world at scattered, half-updated knowledge and it will do what your team does. It will answer from the nearest fragment, confidently, and sometimes it will be wrong. Then everyone blames the model.

The model was never the problem. It was everything the model could not see.

The cost nobody counts

Someone rebuilds an analysis that already existed, because the first one sat in a folder they did not know about. A new hire takes four months to get productive, because the knowledge lives in people instead of in a place. A decision gets made on the half of the picture that happened to surface in the meeting. The AI rollout you approved looks like a failure when it is only starving.

None of that shows up on a timesheet. There is no line item called “answered from the wrong file.” That is how it survives budget after budget.

The bottom line

You do not fix this by buying a bigger model. You fix it by making your knowledge reachable. You take what your company knows and you put it somewhere structured, current, and connected, so your people and your AI both pull the right answer instead of the nearest one. That is what a company brain is.

Building one is the thread I am going to pull on this show, week after week, in plain terms, with real examples. Next week I am handing you the exercise: one hour, the people already in the room, and you will find the single most expensive knowledge gap in your company.

Subscribe to You Don’t Know The Half of It so you do not miss it. And tell me in the comments: what is the one question your team asks that nobody can answer quickly.