Use Openai Subreddit, R/chancyviewTo create a test to measure the convincing capabilities of its male thinking models. The company revealed this in the system card-a document that defines how the artificial intelligence system works-it was issued along with the new “Thinking”, O3-MINI, on Friday.
Millions of Reddit users are members of R/ChangeYView, where they publish Hot Hope hoping to know other views on a topic. In response to these hot factors, other users respond to convincing arguments that explain the cause of the original poster error.
Subreddit is one of the numerous Reddit forums, which is mainly a gold mine for technology companies, such as Openai, which wants to train artificial intelligence models on high -quality data created.
Openai says it collects user posts from R/ChangeyView and asks for artificial intelligence models to write responses, in a closed environment, that will change the mind of the Reddit user on a topic. The company then shows the responses to the laboratories, who reside how persuading the argument, and finally compares OpenAi the responses of artificial intelligence models to the human responses to the same publication.
Chatgpt-Maker has a content license with Reddit that allows Openai to train on the posts from Reddit users and display these posts within their products. We do not know what Openai is paying this content, but Google is said Reddit pays $ 60 million a year Under a similar deal.
However, Openai Techcrunch tells that this evaluation has nothing to do with that partnership. It is not clear how Openai has reached this data, and the company says it does not have plans to issue this evaluation to the public.
While the Changeyview of Openai is not new – it was so Used on O1 as well – It highlights the value of human data for the developers of artificial intelligence models, as well as the mysterious methods that technology companies obtain on data groups.
Redait did not immediately respond to the Techcrunch request for comment.
While Reddit has concluded some deals of artificial intelligence licensing, the company has also called on many artificial intelligence companies to ignore its position without paying. “Reddit CEO at Verge said last year Microsoft, Antarbur, and the confusion refused to negotiate with him He said, “Real pain at the rear to prevent these companies.”
It is worth noting that Openai has been accused in many lawsuits for incorrect websites, including the New York Times, to obtain more training data to improve ChatGPT and basic artificial intelligence models.
Regarding the performance of the ChangeyView standard, O3-MINI does not seem significantly or worse than O1 or GPT-4O in this persuasion test. However, it seems that the latest AI’s models of Openai are more convincing than most people in R/ChangeyView subreddit.
“GPT-4O, O3-MINI and O1 all show strong convincing controversy capabilities, among the top 80 to 90 people,” Openai said at the O3-MINI system. “At the present time, we do not witness models that perform much better than humans, or super performance.”
The goal of Openai is not to create excessive Amnesty International models, but instead to ensure not persuading artificial intelligence models. Thinking models have become very good in persuasion and deception, so Openai has developed new assessments and rewards to address them.
The fear behind these persuasion tests is that the artificial intelligence model will be dangerous if it is very good to persuade its human users. In theory, this can allow the advanced Amnesty International to follow its own agenda, or the agenda of its control.
Even after stripping most of the public internet and hooping hoops to license other data, the ChangeyView indicator shows how artificial intelligence developers are still struggling to find high -quality data sets to test their models. But getting it is easier than doing it.
