explore | Grilled Cheese

Beautiful! Wish we had a few sites like that here. Our buildings are so new in comparison. #DigiBlogChat

Old and historical we have quite the choice #DigiBlogChat just a few grabbed quickly

Q6 #DigiBlogChat
People forget that current AI systems are **language** models!! They manipulate text!!
AI systems of the 1980s would not be making these kinds of nonsensical mistakes/hallucinations/confabulations ... they were "expert systems" operating on rules specified by subject domain experts.

OWASP identifies agent-specific risks including goal hijacking, tool misuse, excessive privileges and unexpected code execution. Giving an agent more ability to act also expands the potential attack surface. #DigiBlogChat

Between questions, here’s the topic on next week’s chat: Managing Digital Data! #DigiBlogChat

A7. It’s capable of a lot, but you still need human intervention. And I worry about all of those (hallucinations, API costs, runaway loops, security) #DigiBlogChat

A7 Hallucinations are only part of the risk. Runaway loops, unexpected API bills, excessive permissions, data exposure, bad tool calls and an agent confidently taking the wrong action can all turn a small automation into a big problem. #DigiBlogChat

Q7: What are the biggest operational risks (e.g., hallucinations, API costs, runaway loops, security) you worry about with agentic AI? #DigiBlogChat

Here we go! Trending. #DigiBlogChat

A6. Grounding. You train the AI, but you have to constantly train and retrain the AI. They’re like cats and will jump off the rails and do their own thing when you let them! #DigiBlogChat

A6 Build quality control into the workflow rather than treating it as a final inspection. Use approved sources, structured prompts, validation steps, fact checks, brand guidelines and human review where the consequences of an error are significant. #DigiBlogChat

It's a fast moving area of innovation, so people are at a wide range of stages of knowledge and adoption, and in a wide range of aspects of the subject ... so it's multidimensional.
#DigiBlogChat

It’s the halfway point and time for a quick stretch! And don't let an autonomous AI agent do it for you! #DigiBlogChat

Q6: How do you maintain quality control, brand voice, and factual accuracy when content or data workflows run autonomously? #DigiBlogChat

A4. My use is more limited as a sounding board that can read SharePoint, 365 and with the right permissions HubSpot so I can work inside Copilot. We've found larger data samples and for creativity Claude is better #DigiBlogChat

A5. I'd never let an AI execute tasks without prior sign-off. A human has to be in there testing and testing. #DigiBlogChat