Four months after the final battle of the Ottoman Empire, AGI was delivered.

  It's the first time in the history of the world.

AGI has already been announced.

During a phone call with the Financial Times on Wednesday, Mr. Huang said:

For many tasks, we can say that we have already achieved AGI.

On the same day, a more than two-hour interview with the Ottoman leader was also released.

For Ottman, the four months were not just a technical impulse. He had to play the AGI card before it went on sale next year.

He couldn't get out, and by the time he rang the bell, there was already an opponent in front of him who was already on the market, his market value had doubled, and he was still profitable.

This is a battle that OpenAI cannot fight back.

In that phone call, someone asked him, what do you think about OpenAI companies that are so eager to chase AGI?

Old Yellow's answer is that this is not a good question right now. The whole industry, even "smart", is not talking about anything, let alone which gauge to measure AGI.

He was holding something else in his hand.

Invisalign's AVO architecture was tested in a test that focused on "smart bodies can't function on their own for long".ARC-AGI-3Up, got it.100% satisfied。

Following this achievement, he made an ambitious prediction:

In the not-too-distant future, INVEDA may only need to maintain around 40,000 human employees, but it will have 400,000, or even 4 million digital employees.

It is well known that ARC-AGI-3's gameplay is quite "non-violent".

Throw your mind directly into a strange game, with no rules and no goals, relying entirely on yourself, on yourself, on yourself, on your guesses.64×64 gridWith a few buttons, each environment has at least six nodes, the first node can go through five steps, and the sixth node can go through 50 steps.

People are playing hard, and models aren't getting anywhere better.

The key here: the model used by AVO is Claude Opus 5.

Then Ingrid gave it a stick. The weight of a parameter did not move, running directly from 30.16% to fullness.

There are two things that really work here.

• One is the lasting memory, the context is full of memory and clean, the intelligent body continues to dry from its current state, without having to step on the same crater again.

• The second is the monitor, which is not functioning on its own, only staring at the entire search trajectory next to it, and when it discovers that the main intelligence body is turned around, it turns to another strategy.

However, strictly speaking, this whole roll is not actually an arithmetic.

More than three hours after the announcement, François Chollet, the father of ARC-AGI, said:

Getting 100% on an open demo is not the same as getting 100% on an ARC-AGI-3 scale. It's like you're switching on a video game and you're actually just removing a new tutorial.

According to the ARC Prize's own description, public collections are not rated, the real calculation is semi-private collections and private collections, both of which are not on AVO's scoreboards.

What's even more subtle is that this is the third full volume in six weeks.

Tycho took the lead at the end of July, Vista took the lead on 5 August, and AVO was the third.

There are more and more people announcing AGI, but it seems that they are both capable of creating AGI.

And this card, no one has really played it yet.

In the interview, Ottoman confidence was high:

By the end of the year, there will be a system within OpenAI that he wants to call AGI.

Chief researcher Mark Chen is relatively conservative. In his view, OpenAI is 20% less than AGI.

It sounds crazy, but they do have something in their hands.

This year, OpenAI has set itself a very clear goal: to create an automated AI research internship that can really work.

And this "AI automation intern" is Astra.

Give Astra an experimental idea that it can write code, run experiments, and deliver results on its own in OpenAI's code library.

Throwing a paper at it, Astra can complete the work of a top human researcher in the past week or so.

In a recent in-house demonstration, dozens of senior customer executives saw it firsthand.

On the big screen, 16 AI entities break up a research-level mathematical problem into a set of sub-problems.

In the middle, they verified each other, and finally came up with a complete proof.

This is exactly what it looks like, a team that already rules by mutual consent is closing its doors.

The second demonstration was even more shocking.

Astra opens up common desktop software, bouncing back and forth between different applications, building files, changing content, scheduling tasks, and it's not like people are tapping on keyboards.

Astra's "permanent intelligence" has been completely unleashed.

Ottman even predicted, "I expect this to be the first model to actually invent something important new. It's a very AGI thing".

Note that the phrase "invent something new" means to create something that the whole human race does not yet know.

Recently, OpenAI's first self-developed chip, the "Jalapeño", directly rubs the old yellow flagship GPU onto the ground.

The number one driver behind this is the GPT-Astra.

It's straightforward, from circuit design to software.

On the one hand, the chip optimizes the underlying circuitry. On the other hand, the software writes and optimizes the core kernel at the bottom of the chip, and also refines the three main open source models.

As a result, the performance of the BF16 multiplier increased by 56%, while the matrix unit area shrank by 10%.

The two core modules, Attention Mechanism and MoE, are 1.5 to 1.8 times faster than the human expert handwritten versions.

Guys, AI is a physical circuit.

It's interesting to note that this is the same thing that Ingrid is doing.

AVO Root is not designed for gaming, its main battlefield is to write the underlying code for GPUs.

In March, Invidia published a paper in which the central idea was to give the whole process of "changing the code" to a self-sustaining intelligent organism.

In tests, it runs continuously autonomously.7 daysIt has been fully uninterrupted, explored more than 500 optimization directions, and finally submitted 40 valid kernel versions.

The core code that came out was 3.5% faster than the closed-source version of Invisalign, and 10.5% faster than the industry's leading open-source version.

Now, if you put the two together, the Ottoman phrase ends.

Each iteration of the model consumes a large amount of human research time. But if Astra begins to take over this cycle, the entire development speed will change.

Once this cycle is turned around, the rest is a matter of speed.

In fact, the GPT-5.6 and Fable 5 that we can use outdoors are the products of the previous round.

And the outbreak shows that the real next generation on both sides has actually been trained.

On the other hand, according to Leos and others who have been tracking the dynamics of front-end models for a long time, this year's pre-training ranks are quite dense:

• Spud（Sol）Pre-training was completed in March, in line with the GPT-5.5 and 5.6 series, which are now open to the public.

• DougIt was completed by May 4 and continued on it after reinforcement training, such as the upcoming Astra, or GPT-6 in the outer gate.

• BelJust after the end of pre-training in August, the total parameter was over 10 trillion, which is close to the GPT-4.5 of that year.

Within OpenAI, Bel is seen as the "basic model after GPT-6", and may even be the basis for a model that crosses the AGI threshold.

On the Anthropic side, it was discovered on August 18 that part of Claude's webpage had selected a Fable 5 account and had been clipped onto a new model.

It's the old rule before the Anthropic release, and Fable 5 itself was just as gray at the beginning.

The code was subsequently removed.

• Claude-melon-eap, supposedly Fable 5.1;

• Claude-marshmallow-eap, supposedly Opus 5.1.

In addition, both new models are said to be running on a small scale and can be pushed to anyone at any time.

As for why it's not working, the analysis of netizens is that who comes first and who will be overtaken by the other in a few days.

Anthropic, in particular, has also been exposed and will take the lead in early 2027.

It's just a window into the war's timing.

The remaining four months of 2026:

Some institutions have done the measurements, and Anthropic has88% probability of OpenAI going publicThe most likely date isOctober 26th of this yearAnd the most likely date for OpenAI isNext July 15th。

Market value forecast for the first day of trading, Anthropic$1.82 trillion，OpenAI $1.06 trillion。

OpenAI's CFO said at its August membership meeting that the company would go public in 2027 or earlier, provided the business "continues to turn upside down".

So the AGI at the end of the year is not just a technological milestone.

It is one of the most "critical" pages in the book.

When it came out, the story of OpenAI went from "We're burning money as a model" to "We've reached the finish line", and the whole logic of valuation changed.

The answer lies in these four words: self-improvement.

The old adage "40,000 commands 4 million AI agents" and the OpenAI phrase "AI is already helping to create a generation of AI" all say the same thing: once AI can take control of the next generation of AI, progress will no longer be linear.

That's why everyone is so excited about this super-intelligent (ASI) contest, and the winner is ultimately all-consuming.

The first player who actually turns the loop will take everything behind him with his hands. As for the second, he may not be eligible to be followed.

After Ottman, there is nowhere to turn.