Cassie Goff
The City Journals
Brian Acord, founder of Acord.AI, showed members of West Valley’s local business community how to “Unlock the Power of AI” during a curated workshop by teaching attendees how to make work processes more effective and efficient with new technology.
Acord works as an AI specialist, educator at Salt Lake Community College, an entrepreneur for the Goldman Sachs 10,000 Small Business program, and a startup consultant through Acord.AI. He is a proponent of working with generative AI to help businesses grow.
“AI is going to destroy the world and it’s going to save the world,” Acord said. “We just don’t know which one it’s going to be first.”
He described how AI has been integrated into many types of web tools that companies are already using. Now, it’s going to be about developing specialty processes for generative AI to get more and more accurate.
“The solutions to AI problems will be found using AI,” he said.
Acord overviewed a handful of tasks AI can be programmed to complete, instead of spending human time and resources.
AI can be asked to respond to emails (while keeping in mind which ones are most important); record meetings (and trained to recognize specific voices); create images; write code; re-work information for presentations to boards/councils; summarize a five-star review to immediately share on LinkedIn; create “zaps” for applications that weren’t originally designed to work together; and record, summarize and organize information from a “braindump” or “brainstorm” session.
AI can be asked to play devil's advocate, explore new perspectives, and address blind spots while analyzing data. Acord suggested asking AI the following questions: What am I missing? What could I be wrong about? What is this process missing?
He also recommended programming AI for the different “hats” or “identities” users might wear. He has programmed separate AI pages, by uploading his previously recorded conversations for the multiple “hats” he is constantly wearing.
“Look at the different roles AI can play,” Acord said. “It opens up how much you can really use these roles for.”
When utilizing AI to help speed up processes, Acord emphasized the importance of thinking through the deliverables users want, in what format, and the data security. Prompts should be simple, straightforward and based on the anticipated result.
“The more data you can give it, the more accurate the response is,” he said.
He provided a basic prompt structure to input into text-generative AI programs to yield relevant and impactful results: “Acting as [ROLE] perform [TASK] to create a [DELIVERABLE] for [AUDIENCE] in [FORMAT].”
Acord pulled back the curtain to quickly overview how AI thinks. He likened the processes by which AI learns to those of how the human brain learns.
“We have a lot in common with AI,” Acord said.
The human brain constantly takes in massive amounts of data and builds neural networks based on that data — the more frequently a specific learned association occurs, the deeper the associated neural pathway becomes. Humans are constantly re-evaluating and questioning their own neural networks and looking for inaccuracies to determine if the brain needs to deepen a neural pathway or re-route.
“We are interpreting patterns and we use those patterns to tell stories to help explain something,” Acord said. “Based on those stories, we have a world view.”
Like the human brain, AI identifies nonlinear data, interprets complex unstructured data, identifies patterns, and evaluates meaning. It learns continuously, but can find it difficult to go back into a previously learned pathway and troubleshoot.
“ChatGPT goes through all of this just like you do,” he said.
Acord cited an AI process that medical professionals tried to train and develop to streamline the early detection of cancer screening. Their idea was to train AI to identify an image of a skin lesion as either malignant or benign. A total of 25,000 images were fed into generative AI. When cross-checking the results of the AI model, researchers realized that every image including a ruler was being flagged as malignant. They had to go back and ask AI to remove images that included rulers and re-evaluate.
Eventually, this AI model began identifying correctly with a mid-90 percent rate for images with primarily white skin. Images showing skin from people of color reported only 70 percent accuracy.
“It does have bias built into it — as we all do,” Acord said.
Another example Acord used was asking for a book recommendation. If he asked a stranger for a book recommendation, they’d probably recommend their favorite book or ask for more data and alter the response. AI works similarly.
Acord suggested that the crowd explore a handful of AI systems beyond ChatGPT, including Perplexity, Claude 2.0 and Zapier.