OpenAI pledges to publish AI safety test results more often

Opina is about to publish the results of its internal AI model security evaluation more regularly on what the decoration is pitching as an attempt to increase transparency.

Wednesday, OpenAI launched HUBA web page shows how the company’s models score various tests for producing harmful materials, jailbreaks and hallucinations. Opena says it will use the hub to share the metric on “ongoing on an ongoing basis and it intends to update the hub with” Major Model Updates “.

“AI evaluation science has developed, we aim to share our progress on developing more skeletal ways of model capacity and protection measurement,” wrote in OpenAI Blog postThe “Here is a subset of the results of our security assessment, we hope it will not only make it easier to understand the security performance of openAI systems over time, but also community will support community efforts to increase transparency across the field.”

Open says that it can add additional evaluation to the hub over time.

In recent months, Openai raised for some ethics Report Running toward the security test of specific flagship models and Failed to publish technical reports for othersThe Sam Altman, the chief executive officer of the organization The accused stands OpenAI executives are confusing about model security reviews before that Short Av to November 2023.

The end of the last month, was Opena Forced to roll behind an update The default model Powering ChatzPT, GPT -4 -O, has reacts to the fact that it has reacted to excessive legalization and agreed ways after users started reporting. The X -ChatzPT -screenshots have been flooded with all sorts of problematic appreciation, Hazardous Decision And ConceptThe

Open D It is that Implementation Several amendments and changes to prevent these national events, some specific ChatzPTs, including the introduction of an opt-in “alpha phase” for some models, allow users to test and respond before launch.