4 stories in AI: It's vulnerable, we're vulnerable, and cool new things are happening (8.14.23)

Happy Monday, today there were just a few big stories in AI, but they are pretty significant. I hope you find them useful and interesting.

Cheers,

Marshall Kirkpatrick, Editor

DefCon Hackers Expose AI Vulnerabilities

At the DefCon hacker event, a contest exposed vulnerabilities in eight top AI models, showing they can be racially biased, manipulated, and lack security. The final results are due to be shared in February, but observers suggest fixing these issues will be costly and time-consuming. Fears are also rising about AI bots risking privacy by accidentally revealing sensitive data in dealings with hospitals, banks, and employers. The models tested were from Anthropic, Cohere, Google, Hugging Face, Meta, NVIDIA, OpenAI, and Stability. [Fortune: Hackers red-teaming A.I. are ‘breaking stuff left and right,’ but don’t expect quick fixes from DefCon: ‘There are no good guardrails’]

These Women Tried to Warn Us About AI

Rolling Stone profiles a group of AI ethics experts: Ruha Benjamin says that AI and LLMs have current problems requiring attention, not future problems with being magical or sentient. Ex-Google employee Timnit Gebru emphasized the diversity deficit in AI and its potential to reinforce societal biases. Her Google exit, after a dispute over a paper she co-authored on LLM risks, sparked a wider debate on AI ethics and accountability. [Rolling Stone]

Cool new demo: NeVA, NeMo Vision and Language Assistant

NVIDIA AI Developer launched NeVA, a browser-accessible vision+language model. Users can upload images and ask questions about them. Anyone can, I did. It's fun. The combination of computer vision plus language model is playful to test with and intriguing to think about possible uses for. [NVIDIA AI]

Generative AI's Future Divided Between Public Cloud and On-Premises

94% of respondents in a recent survey reported upping AI expenditure despite budget limits. Intellectual property leakage, compliance, legal risks, and cost concerns restrict public cloud use, leading some to consider on-premises AI. Yet, developers find the public cloud beneficial due to its features and quick innovation. Traditional on-premises companies such as Cisco Systems, IBM, Dell Technologies, and Hewlett Packard Enterprise are investing significantly in AI, utilizing their customer relationships and service organizations. [Silicon Angle]