Thoughts and Observation on the 𝐎𝐩𝐞𝐧𝐀𝐈 𝐏𝐫𝐞𝐩𝐚𝐫𝐞𝐝𝐧𝐞𝐬𝐬 𝐅𝐫𝐚𝐦𝐞𝐰𝐨𝐫𝐤

Author - Anshu Gupta

I just read up the 𝐎𝐩𝐞𝐧𝐀𝐈 𝐏𝐫𝐞𝐩𝐚𝐫𝐞𝐝𝐧𝐞𝐬𝐬 𝐅𝐫𝐚𝐦𝐞𝐰𝐨𝐫𝐤 (Beta) which was released last year on Dec 18 2023. To my knowledge it has not been updated since then. As they prepare for AGI, I think it is time that OpenAI updates it and releases the final version.

Here are my thoughts & observations

1. The Preparedness Framework should be 𝐮𝐩𝐝𝐚𝐭𝐞𝐝 𝐞𝐯𝐞𝐫𝐲 𝐪𝐮𝐚𝐫𝐭𝐞𝐫/6 𝐦𝐨𝐧𝐭𝐡𝐬 due to the rapidly evolving nature of AI Threats and Safety issues. This was released in Dec 2018 and still marked as Beta.

2. The controls on Page 16-19 have been watermarked as “𝐈𝐥𝐥𝐮𝐬𝐭𝐫𝐚𝐭𝐢𝐯𝐞” which means that they may not be in place. More clarity is needed on this.

3. The risk assessment and reporting framework is very 𝐩𝐫𝐨𝐜𝐞𝐝𝐮𝐫𝐚𝐥 𝐢𝐧 𝐧𝐚𝐭𝐮𝐫𝐞 and will eat up hours in paperwork which no reads or uses. They need to develop a tool for the purposes and open source the code which might be both cost efficient and easier to use.

4. Given some of the transitions in the OpenAI Safety team in May/June time frame, I am unsure of how much is 𝐚𝐬𝐩𝐢𝐫𝐚𝐭𝐢𝐨𝐧𝐚𝐥 𝐯𝐞𝐫𝐬𝐮𝐬 𝐫𝐞𝐚𝐥 in the Preparedness Framework esp. given the Beta designation of the framework.

5. Some of the positions like the OpenAI 𝐒𝐚𝐟𝐞𝐭𝐲 𝐀𝐝𝐯𝐢𝐬𝐨𝐫𝐲 𝐆𝐫𝐨𝐮𝐩 (𝐒𝐀𝐆) chair mentioned in the Preparedness framework, should be named positions for ownership, accountability and visibility perspective. It seems that there is new committee called 𝐒𝐚𝐟𝐞𝐭𝐲 𝐚𝐧𝐝 𝐒𝐞𝐜𝐮𝐫𝐢𝐭𝐲 𝐂𝐨𝐦𝐦𝐢𝐭𝐭𝐞𝐞 led by directors Bret Taylor (Chair), Adam D'Angelo, Nicole Seligman, and Sam Altman (CEO). There should be clarity if SAG and SSC are one and same. I also think that the SAG Chair needs to be a tactical role Vs a strategic role.

6. 𝐋𝐞𝐚𝐝𝐞𝐫𝐬𝐡𝐢𝐩 𝐜𝐚𝐧 𝐨𝐯𝐞𝐫𝐫𝐮𝐥𝐞 𝐎𝐩𝐞𝐧𝐀𝐈 𝐒𝐚𝐟𝐞𝐭𝐲 𝐀𝐝𝐯𝐢𝐬𝐨𝐫𝐲 𝐆𝐫𝐨𝐮𝐩 (𝐒𝐀𝐆) - Not sure if this is the right way, especially to ensure SAG can not be unduly influenced to make way for commercial interests over ensuring safety, security and privacy. A per the framework, “Leadership can also make decisions without the Safety Advisory Group (SAG) SAG’s participation, i.e., the SAG does not have the ability to “filibuster””

7. 𝐄𝐚𝐫𝐥𝐲 𝐀𝐜𝐜𝐞𝐬𝐬 𝐭𝐨 𝐆𝐨𝐯𝐞𝐫𝐧𝐦𝐞𝐧𝐭 𝐟𝐨𝐫 𝐌𝐨𝐝𝐞𝐥 𝐑𝐞𝐥𝐞𝐚𝐬𝐞𝐬 - The framework does not specify which Government. In my opinion the language should be “early access to model releases to relevant legal and governmental bodies” in the jurisdictions we operate, as authorized by the BoD.

8. 𝐂𝐡𝐢𝐥𝐝 𝐒𝐚𝐟𝐞𝐭𝐲 𝐚𝐧𝐝 𝐒𝐨𝐜𝐢𝐞𝐭𝐚𝐥 𝐖𝐞𝐥𝐥 𝐁𝐞𝐢𝐧𝐠 (e.g. someone ideating on self harm) and related mitigations are not considered in the the fundamental “Tracked Risks” Categories in this framework. Historically, both these areas come into focus only when something bad happens.