On 11 September, somebody put a 744-billion-parameter model online
On 11 September, somebody put a 744-billion-parameter model online. No launch event, no blog post, no keynote. The paper turned up four days later.
It came from Shanghai AI Laboratory. The model is Atria Dawn Preview, released under an MIT licence, so you can download it, modify it and use it commercially. Weights went up on HuggingFace on the 11th, an FP8 checkpoint on the 12th, and only on the 15th did the paper appear on arXiv.
143 authors. The title is not shy: "Atria Dawn: The Dawn of Agentic Superintelligence".
What kept me reading was not the title. It was how they measured themselves.
One section of the paper takes the development of the model itself as the object of study: 56 participants, 769 task records, and the agent logs alongside them. The participants were then asked to assess the work they had finished. About a third of completed tasks, they said, would have been infeasible without AI.
Not "how much faster". "Could not have been done at all".
The part worth copying is not the model. It is that method.
When I teach this, the same question comes up every time: how do you know it is actually worth anything? The easy answer is "it is faster". But faster loses the argument against anyone who did not want to change in the first place, because they can always say you were slow to begin with. "This could not have been done without it" is a different kind of line. It is one you can say out loud, and it is the thing you should be able to answer before you spend the budget.
About a third of completed tasks, they said, would have been infeasible without AI.
The caveats, plainly:
・That one-third figure is the team who built the model rating their own work. It is not third-party research. They marked their own paper.
・They report the highest score on 5 of 16 benchmarks, but the abstract names neither which five nor by how much.
・No independent verification yet. By one count, third-party coverage reaches 14 of the 435 benchmarks relevant here.
・744B free to download and 744B affordable to run are not the same sentence. This does not fit on one machine.
・The title says superintelligence. The evidence underneath it is the work log of 56 people. Those are not the same order of claim.
The method is right and the marking is not independent. Both can be true at once.
So the question I am left with: counting backwards over the past month, how many finished pieces of your own work could genuinely not have been done without AI?
More in AI news
All →
In my line of work there is one rule nobody argues with
Microsoft has written its AI a rulebook, and the rule that matters most is this
Some of the work people hand to AI outside their own job turns into a habit. Some doesn't
On Saturday, the head of an AI company proposed something very few companies volunteer