This instrument is FRONTIER AI models for intelligence

Executives: Artificiality Companies can LOVE TO TAKE USE that Ague It’s almost here, but the latest models still need extra extra teaching to be able to be as smart.

Professional AI, a company that played a key role in helping Frontier AI companies developed a platform that can automatically check the model through thousands of criteria and tasks that need to contribute to their skills. The scale will, of course, will provide the required data.

The scale rose until important importance to human labor, training and testing AI models. Large language models (LLMS) are trained on books, network and other sources of the text Oodles. Turning these models into useful, consistent and well-prepared conversations requires additional “further training” in the form of people who give feedback on the outcome of the model.

Scale Accessories Workers who are expert in the expert models of problems and restrictions. The new tool, called a scale assessment, automates some of these works using algorithms of owner car training.

“There are all these pedigree ways in large laboratories to follow the exemplary weaknesses,” says Daniel Berri, the head of the scale assessment. The new tool is “a means (model manufacturers) to cross and cut them through the results and pieces to understand how the model doesn’t do well.”

Berri says that several basic model companies in AI are already using the tool. He says most of them use it to improve the basic capacity of their best models. The reasoning of AI includes a model that tries to sever the problem in the components so that it will be more effective. The approach largely relies on post-election training users to determine whether the model is right to solve the problem.

In one case, Berrios says that the scale scale found that the skills of the model caused the skills were fed in non-English prompts. “Meanwhile (model) opportunities were quite good and were doing well on the benchmarks, they were prone to decompression when the instructions were not English,” he said. The evolution of the scale highlighted the question and gave the company to collect additional learning data to solve it.

Jonathan Frankli is the main literate of databases, a company that builds large AI models, says it is fundamentally useful for one foundation model. “Everyone who pushes the ball forward helps us to build aii better,” says Franklen.

In recent months, the scale has contributed to the development of several new benchmarks designed to become smart AI models, and study how to make mistakes. They include EnigmaevalTo be in style CrowdTo be in style Maskand The last test of humanityA number

The scale says it becomes more difficult to measure improvements in AI models, but they better exist in existing tests. The company says her new tool offers a more comprehensive picture, combining many different benchmarks and can be used to customize customs tests in different languages. Own AI can have a given problem and make more examples, allowing more comprehensive test model skills.

The new company’s new tool can also notify efforts to standardize AI models for incorrectly. Some researchers say the lack of standardization means that Some model jailbreaks do not fallA number

In February, the National Institute of Standards and Technology, the National Institute of Technology, announced that it would help develop methods to test models so that they are safe and reliable.

What bugs did you notice in the results of Generive AI tools? What do you think are the biggest blind spots of models? Let us know too. By mail Hello@wired.com or interpreting below.