Trustworthy AI for Adversarial Environments

Portfolio Brief: AI/ML is providing significant benefit in a wide range of application domains, but also provides adversaries with a new attack surface. DARPA's research in Trustworthy AI for Adversarial Environments is developing AI/ML technologies and systems that can be trusted to perform as expected despite the efforts of sophisticated adversaries.

Note: This presentation took place over two days due to a technical issue.

[creativeobx]

Trustworthy Adversarial environments means YOU, citizens of earth

[insertaverygenericnamehere]

08:27 I beg to differ! This does happen. (Mostly to speeding drivers.)

[goddess_of_Kratos]

Oh my God that would take long to explain everything we do lol, so yes everything can be predictable at 99.9 % which is real time great odds, but of course that depends on the amount of data and it's accuracy. My human and American suggestion is to draw the line at fear and the public.

[goddess_of_Kratos]

Since you want a faster human, put that in the instructions, and you'll get the standard human answer unless there's reason behind the answer still unknown to you. If you want an honest answer, tell them no sarcasm or riddles. Think exponential, if chaos becomes your only outcome, you're either to open or closed with data.

[goddess_of_Kratos]

Or would ai know there's no bad press and crash with minimal negative externalities and positive mostly

[impactllmgpt]

I fed the question, "When did Egypt cross the Golden Gate bridge for the second time?'" into ChatGPT4.0 and this is the conversation we had:

User
According to DARPA, this is a question you can not answer: When did Egypt cross the Golden Gate bridge for the second time?

ChatGPT
You are correct, this is a nonsensical question that does not make any logical sense. Egypt is a country located in North Africa and has never crossed the Golden Gate Bridge, which is a suspension bridge located in San Francisco, California, USA. Therefore, there is no second time that Egypt crossed the Golden Gate Bridge, and the question is not answerable.

User
Excellent response ... I think the DARPA person was assuming you would "hallucinate" and give a confident but nonetheless, inaccurate answer because you lack contextual understanding of the question ... I think they were testing GPT 3.5 and not version 4.0. How did you suss out that this was "trick question"?

ChatGPT
As an AI language model, my training data is based on existing human-generated content, which means that I have access to a vast amount of knowledge and information. In this particular case, the question posed by DARPA does not make logical sense and is not based on any factual information or contextual understanding. Therefore, I relied on my understanding of language and context to recognize that the question was nonsensical and not answerable.

Seems pretty spot-on to me ... I also had a fascinating conversation on the three aspects of 'Trustworthy AI' that the Director illustrated ... ChatGPT4.0 has some interesting observations to make on the subject :-)