AI Safety / Compliance

[robmoffat]

Feature Request

Was discussing with @pcheong-lbg last week - she's interested in exploring how her OSPO can help developers in her firm make the most of AI. Specifically around language models: training data, off-the-shelf models, algorithms etc.

• What are the licensing implications of using these?
• What are the Supply-chain security implications? (it is perhaps easier to smuggle unwanted behaviour in an LLM than a regular algorithm).
• What do safe engineering practices look like for people deploying LLMs?

@finos/open-source-readiness-contributors welcome to pitch in their ideas on this.

[psmulovics]

Tagging @tenatzen and @padraignix

[jstclair2019]

Great discussion/idea - I think the OSPO can be an enabler for what the organization's AI policy is. There is plenty in the open source AI space to leverage. If the OSPO is there to answer "what can we do in Open Source?" then AI is a great fit, whether adversarial tools, open LLMs, training, etc

[tenatzen]

We're going to be covering some of this on the Zenith SIG call tomorrow. We have a POC project going through proposal on a fully open source, secure and stable datalake for training sustainably (predominantly focused on Regulatory data first), but showing best practice for industrial collaboration on data sharing purely for training.

Zoom details: Dec 14th, 10am EST / 3pm GMT
https://zoom.us/j/98538837792?pwd=ZHg3QVVnaCtOT0tEZSs5c0hrZTg1dz09
Meeting ID: 985 3883 7792
Passcode: 152455

[robmoffat]

@tenatzen yes I'm really hoping Zenith can do some of the heavy lifting on this and provide us with some of the thinking we're going to need around this. I have a clash at that time but I'll happily digest any minutes from the meeting!

cheers,
Rob

[psmulovics]

Meeting notes would come later - for now, here is the deck: https://github.com/finos/zenith/blob/main/Governance/SIG%20meetings/2023-12-14/Zenith%20-%20Dec%2014%202023.pdf

[robmoffat]

Closing, since we now have AI Readiness.