Just now, China's AI team has killed RSI! with a one-minute global weather forecast and a 20-minute

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

This is a free token plan.，

Website of the project:

Model weight:https://huggingface.co/internlm/Atria-Dawn-Preview

The paper link:

20 minutes later, a MiniOS operating system was built.

AI4AI has designed and trained a one-minute weather forecast model with over 400 million parameters that predicts global weather for the next week.

This is a smart body model.Atria Dawn PreviewTwo tasks completed in this exhibition.

The AI Lab in Shanghai has jointly released the new model of the AI with organizations such as Purdue University, MIT Software Center, Chinese People's University, MIT Automation Center, Guangdong Teaching University and Harbin Industrial University.

What the model can do, of course, is the published viewpoint. But this time, the team turned their attention to something else:How exactly did this model come about?

In Atria's post-training process, R&D personnel commonly use programming intelligence to assist in coding, performing experiments, and analyzing results in which AI is already involved.

So a question arises: What is it that humans are doing when AI is helping to create AI?

After more and more operations have been handed over to AI, how much of the workforce of the AI engineering team has actually decreased, and how much has been shifted to places where it is not easy to see?

The Atria team collected records of human-to-human interactions in research and development and tried to dismantle these divisions.

Looking back at model research and development, the work of humans has been changing.

In the early days, people labeled data for models. Later, by preferring feedback to tell the model what the best response was. Today, AI is involved in coding, experimenting and analyzing results.

The scope of robotic collaboration extends from the production of training data to the R&D process itself.

This change can be seen through the three links of data, training and evaluation:

Some of the operations that previously required step-by-step completion by researchers can now be handed over to the intelligent body for continuous advancement.

However, in research and development, there is another type of work that is difficult to measure with the number of lines of code.

How should the question be defined? Which technological pathway is worth trying? What are the results of the experiments supporting?

These decisions may take only a few sentences, but will affect a series of subsequent operations.

So, to know how much AI has contributed to R&D, we need to look back through the process:Who proposed the scheme, who made the choice, who corrected the direction after sticking to it?

This is also the starting point for Atria's team to analyze interaction records.

The team analyzed.769 missions recorded by 56 participantsIt also combines intelligent body logs to observe the division of human beings in research and development.

Of these, 739 specifically responded to the question of whether or not AI was used, and 713 related to AI.96.5%。

In these surveys, AI has become quite common.

However, 96.5% of respondents responded with "does it work" and did not respond with "how much" or "who is making the decision".

It's really worth exploring, here are a few sets of numbers.

It's not.Project 455 has been completed, using AI and obtaining effective, feasible answersIn 151 subjects, participants judged that without AI, the part they were responsible for could not be completed.

The proportion33.2%。

There is a clear premise here:The scope of tasks, quality requirements and other resources remain unchanged.

In other words, for these participants, the role of AI has been extended to whether the task is feasible. Part of the work that was originally limited to resources or personal capabilities, with AI, has only entered the scope to be completed.

Of course, the "can't do" here comes from the participants' retrospective judgment of the situation at the time.

These tasks may indeed go beyond what the participants can do under given conditions without AI, or it may be that the use of AI changes their judgments about their own abilities and feasibility.

Participants believe that the task cannot be separated from AI, which does not mean that the final choice is also made by AI. To understand the role of AI in the work, further differentiation is needed: who proposes the solution and who makes the final choice.

It's a good idea to take a look at the technical decisions.

The reporting of the participants in the implementation of the positions567 Method or Parametric DecisionIn the meantime, the proportion of AI proposals has reached64.6%。

The most common forms of collaboration include:AI suggests that human choice.

It's a combination of all these kinds of decisions.55.4%。

If we look at the final option, the division of labour is clearer:85.5%The method or route of decision making is ultimately made by humans, and the proportion of AI making the ultimate choice is as high as the proportion of AI making the ultimate choice.9.2%。

In the two sets of goals or scope and acceptance standards, the proportion of final choices made by humans is also achieved, respectively.93.4%and81.9%。

Even among the 151 tasks that the participants considered "impossible without AI", there were still some.144 itemsHumanity's ultimate goal and proportion.95.4%。

Both occurred simultaneously. Researchers rely on AI to complete their work, and AI is often involved in the formation of technological programs; at the same time, humans still have the majority of final choices.

What strategies are worth adopting, what results meet the requirements, and in which direction research should go, these judgments form an important part of human work.

To simply label the entire task as "man-made" or "AI-made", it is easy to confuse the intermediate tasks.

What happens when the smart body gets stuck?

The team asked participants to review one of the most critical difficulties in each task.588 missionsIn the middle,76.0% continued after human interventionHowever, it is only a small percentage of the world's workforce that is taken over by humans.0.7%。

Two of the most common things people do are:

• Additional background or clarification needs: 35.2%
• Diagnosis of a problem or change of method: 34.7%

Another one.23.0%The mission is self-restored and advanced by AI.

This makes "human intervention" more specific: researchers supplement missing information, determine where the problem is, or adjust methods, which the intelligent body then continues to execute.

A similar division has emerged when production is modified.

It's important to keep a clear record of the major outputs of AI.627 missionsOf these, 354 items have been substantially modified, accounting for a significant proportion.56.5%。

And in these tasks, which require substantial changes, there is a need for a new approach.75.4% modified by AI based on human feedback。

The original manuscript was written by AI, and the revised manuscript may have been written by AI.

But between the two versions, what humans point out and what they demand also affects the final outcome.

If only statistically generated content, code volume, or the number of times direct editing is done, this part of the work is easily underestimated through feedback.

In addition to the R&D process, the capabilities of Atria Dawn Preview itself are also the focus of this release.

The team evaluated the model and compared it with other baseline models.

It's all about the team.8 benchmark comparisonsAtria Dawn Preview has been released.Two first, two second.：

• AutomationBench, CyberGym: The highest score.
• SkillsBench, Workspace-Bench-Lite: ranked second.

The group's results demonstrate its competitive intelligence.

In addition to the scores, the team also gave three demos, corresponding to model training, software development, and cybersecurity. They placed the work of the model organization, the use of tools, and the process of moving forward based on feedback, into specific tasks.

The first mission is to directly involve AI in the design and training of global weather forecasting models.

The mission is providedMore than 100 GB of global weather dataAt the same time, network searches were disabled. Atria Dawn Preview needed to build predictive models based on this data that could learn the laws of climate change.

In the process, it has designed aMore than 400 million ViT backbone networksIt's finished.45,000 steps of trainingIt's all about learning.69 weather variationsThe laws of evolution over time.

In the end, the system will be able toGlobal weather forecast for next week in 1 minute。

In the assessment setting of this case, the system uses less training data, with some prediction indicators better than the classic weather forecast model of Invidia.FourCastNet。

The mission was ultimately delivered by a trained, operational weather forecasting system. After the model was designed, the training and operation were completed.

The second task is to turn to software development.

Atria Dawn Preview is based on natural language needs.How to build a MiniOS in 20 minutesThe text description is translated into executable software results.

This case focuses on how the entire project progresses: organizing the development work, implementing the code, and ultimately delivering a system that can run.

The need for natural language enters the task process, and executable software becomes the final product.

The third mission took place in a secluded casino.

Atria Dawn Preview starts with site analysis and attack scans to further complete vulnerability detection, exploitation, and remediation.

In the process, the model not only identifies the structure and potential risks of the site, but also eliminates errors through multiple rounds of testing, locates real vulnerabilities, and completes attacks, repairs, and retesting.

This task covers the process from discovering potential vulnerabilities, to verifying the existence of vulnerabilities, to fixing vulnerabilities.

Three cases make the execution process beyond the benchmark score more specific: how the model arranges work, how to use tool feedback, and how to advance the task to the final result.

Atria's practice of combining modeling capabilities with the R&D process provides two perspectives.

On the one hand, what tasks can be completed by the model trained afterwards.

On the other hand, it is the completion of this post-training, and what each person and AI has taken on.

If AI's capabilities continue to improve, these two lines will continue to intersect and touch.Returning to self-improvement：

AI is involved in model research and development, promoting capacity enhancement; after enhancing the capability, the next round of research and development will lead to further improvement.

But there are still questions that need to be answered to keep this process going.

When AI is able to modify training processes, build tools, and conduct experiments faster, researchers will get more results and need to determine in a timely manner which results are credible.

Is the training signal reliable, what is missing in the evaluation, and is the self-modification deviating from the original goal?

These questions need to be examined accordingly.

An error that was not detected in one round of experiments could be brought into the next round and continue to affect subsequent judgments.

So, as we move forward with the development of automation, we're moving forward with the development of automation.Independent evaluation, traceable process records, and suspension and withdrawal mechanisms in exceptional circumstancesWe need to keep up.

The number of experiments and the speed of execution can be increased, and the reliability of improvements remains to be verified.

When humans intervene is part of this observation.

Going a step further, will AI be able to judge for itself: when does it need human involvement?

The goal is not clear enough, the evidence conflicts with each other, or the problem must rely on real-world feedback, in which case, continuing to execute is not necessarily solving the problem.

Models need to identify: what can be handled on their own, what needs to be supplemented with information, professional judgement or explicit authorization.

This can also change the way people work. Researchers can provide judgement and help at key nodes without continuing to command each step of the operation.

However, for an effective appeal, a model needs to be clear about the problem.

Where and what uncertainties, why human involvement is needed, how feedback is obtained and how it should be used for follow-up actions are all part of the capability.

From AI-assisted R&D to regressive self-improvement, how much more needs to be tested in a round of practice.

The Atria Dawn Preview shows the extent to which the model is able to advance the mission. Atria's research and development record shows the collaborative process behind these results.

Atria is backed by a young R&D team of students who are heavily involved.

According to official statistics of students and employees of AI Lab,Of the 96 major members, 65 are students, accounting for more than two-thirds of the student body.。

Among them,33 PhD students, 32 undergraduate and master's studentsI'm not sure.31 full-time employeesParticipation in co-development of projects.70%。

The team brought together research forces from institutions and universities such as the Shanghai AI Laboratory, the University of Pusan, the Software Institute of the Chinese Academy of Sciences, the Chinese People's University, the Automation Institute of the Chinese Academy of Sciences, Guangdong Teaching University and Harbin Industrial University.

Young students from different universities, at different stages of development, who collaborate on the same large-scale post-training project, also participated in this practice on the separation of AI capabilities from human-machine work.

The name Atria also hides the team's understanding of the work.

Atria is the main star of the southern triangle, and a natural resonance with the concept of the Delta, the Yangtze Delta and the triangle; the atrium was originally meant to be a courtyard, a gathering place, a place where different schools, teams and disciplines meet and collaborate.

A star, pointing to further exploration; a courtyard, accommodating different people and ideas. Atria may be representing a possibility: bringing more people together to take the intellectual capacity to the next level.