
Thoughts and Observation on the ๐๐ฉ๐๐ง๐๐ ๐๐ซ๐๐ฉ๐๐ซ๐๐๐ง๐๐ฌ๐ฌ ๐ ๐ซ๐๐ฆ๐๐ฐ๐จ๐ซ๐ค
Nov 24, 2024
2 min read
0
32
0
Author - Anshu Gupta
I just read up the ๐๐ฉ๐๐ง๐๐ ๐๐ซ๐๐ฉ๐๐ซ๐๐๐ง๐๐ฌ๐ฌ ๐ ๐ซ๐๐ฆ๐๐ฐ๐จ๐ซ๐ค (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 ๐ฎ๐ฉ๐๐๐ญ๐๐ ๐๐ฏ๐๐ซ๐ฒ ๐ช๐ฎ๐๐ซ๐ญ๐๐ซ/6 ๐ฆ๐จ๐ง๐ญ๐ก๐ฌ 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 โ๐๐ฅ๐ฅ๐ฎ๐ฌ๐ญ๐ซ๐๐ญ๐ข๐ฏ๐โ which means that they may not be in place. More clarity is needed on this.
3. The risk assessment and reporting framework is very ๐ฉ๐ซ๐จ๐๐๐๐ฎ๐ซ๐๐ฅ ๐ข๐ง ๐ง๐๐ญ๐ฎ๐ซ๐ 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 ๐๐ฌ๐ฉ๐ข๐ซ๐๐ญ๐ข๐จ๐ง๐๐ฅ ๐ฏ๐๐ซ๐ฌ๐ฎ๐ฌ ๐ซ๐๐๐ฅ in the Preparedness Framework esp. given the Beta designation of the framework.
5. Some of the positions like the OpenAI ๐๐๐๐๐ญ๐ฒ ๐๐๐ฏ๐ข๐ฌ๐จ๐ซ๐ฒ ๐๐ซ๐จ๐ฎ๐ฉ (๐๐๐) chair mentioned in the Preparedness framework, should be named positions for ownership, accountability and visibility perspective. It seems that there is new committee called ๐๐๐๐๐ญ๐ฒ ๐๐ง๐ ๐๐๐๐ฎ๐ซ๐ข๐ญ๐ฒ ๐๐จ๐ฆ๐ฆ๐ข๐ญ๐ญ๐๐ 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. ๐๐๐๐๐๐ซ๐ฌ๐ก๐ข๐ฉ ๐๐๐ง ๐จ๐ฏ๐๐ซ๐ซ๐ฎ๐ฅ๐ ๐๐ฉ๐๐ง๐๐ ๐๐๐๐๐ญ๐ฒ ๐๐๐ฏ๐ข๐ฌ๐จ๐ซ๐ฒ ๐๐ซ๐จ๐ฎ๐ฉ (๐๐๐) - 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, โLeadership can also make decisions without the Safety Advisory Group (SAG) SAGโs participation, i.e., the SAG does not have the ability to โfilibusterโโ
7. ๐๐๐ซ๐ฅ๐ฒ ๐๐๐๐๐ฌ๐ฌ ๐ญ๐จ ๐๐จ๐ฏ๐๐ซ๐ง๐ฆ๐๐ง๐ญ ๐๐จ๐ซ ๐๐จ๐๐๐ฅ ๐๐๐ฅ๐๐๐ฌ๐๐ฌ - The framework does not specify which Government. In my opinion the language should be โearly access to model releases to relevant legal and governmental bodiesโ in the jurisdictions we operate, as authorized by the BoD.
8. ๐๐ก๐ข๐ฅ๐ ๐๐๐๐๐ญ๐ฒ ๐๐ง๐ ๐๐จ๐๐ข๐๐ญ๐๐ฅ ๐๐๐ฅ๐ฅ ๐๐๐ข๐ง๐ (e.g. someone ideating on self harm) and related mitigations are not considered in the the fundamental โTracked Risksโ Categories in this framework. Historically, both these areas come into focus only when something bad happens.