Google Ad #1 - desktop - 728x90

Google Ad #3 - desktop - 728x90

Google Ad #1 - mobile - 300x200

Google Ad #3 - mobile - 300x200

ATD_Half_Page

Googel Ad #1 - Side Bar - Content pages

ATD Footer Logo

Brian L. Davis

Focuses on the managerial potential of technical employees. Hints on remedying problems stemming from overdeveloped set of analytical skills Ways an expert can become an effective manager Obstacles to effective management of people with an inordinate need to excel.

When Techies Manage

L&D Best Practices: 5 Approaches to Be More Strategic (SU414)

In January, OpenAI’s conversational platform, ChatGPT-3, became the fastest-growing app in internet history. This is just one indication of growing curiosity about GPT (generative pre-trained transformer) and other chatbots in the past few weeks. As one of the first learning professionals to see the potential of conversational AI and one of the few designing the applications for years, I’m excited to see the recent attention paid to this engaging technology.In my book, AI in Talent Development, I outline several ways chatbots can be used as part of a robust learning experience, and ChatGPT-3, the third version in OpenAI’s ChatGPT series, demonstrates most of them. When ATD asked me to write a post on this subject, I struggled to find the value I could add to the exploding conversation. I decided to do what people say I do best: Give you the bottom line and point the way towards practical applications. Let’s start with the basics.How ChatGPT-3 WorksChatGPT operates on an autoregressive model. It’s not as complicated as it may sound. A transformer is a program, model, or algorithm that converts code into something else—in this case, human language. The autoregressive part means the software uses prior events to predict the most likely future events, which is what your brain does all the time. It allows ChatGPT and similar applications to simulate a conversation between two humans. Learn the two keys to making an application like this work:1. Natural Language ProgrammingNatural Language Processing (NLP) allows the bot to recognize input through human speech or writing, eliminating the need for menus or code to engage with it. This same ability enables the bot to reply through text. Before the development of NLP, most chatbots were menu-driven, and the bot responded by selecting the best fit from a database of pre-written options. Conversational AI represents a new way for humans to use computers; it’s a user interface requiring less hardware and prior knowledge than any previous UIs.!CAUTION! Don’t confuse Natural Language Processing with the pseudo-science term using the same acronym. Neural Linguistic Programming is widely disproved and an invalid practice.2. Massive Training DatasetsLike any neural network (including your own), AI models must be effectively trained using a technique called deep learning. Quite frankly, the stunning capabilities of ChatGPT-3 are nothing new. As David Chalmers points out, there is no new technology here. There are many bots with similar abilities, and I described them in detail in 2020. How this bot is trained and how well this approach works is innovative. First, Open AI used the largest training dataset, combining a database containing much of the internet, a huge library of books, and Wikipedia.Then, after refining the bot in-house, Open AI released it free to the world. As we all started to play with it, we were training it further for them. Think of it as a crowd-sourced training effort. The sheer size of the training database resulted in what everyone always expected: Impressive results that can mimic human expression so well it is hard to tell a machine generated it. However, there continues to be an underlying problem: Chatbots are only as good as their training (like us humans). If you consider the source of GPT’s training, you can see why it sometimes makes mistakes.What Can We Do With ChatGPT-3?Because this product is still being explored, there may be new use cases by the time this post publishes. Below are a few applications that can save time and make your work as a learning professional more efficient and productive.Write ContentIf you ask the bot to write a paragraph introducing a subject, it comes back almost instantly with plausible, grammatically correct text. Because its training included unvalidated sources like the internet and Wikipedia, some of the content it produces must be corrected. Still, you can get excellent first drafts and refine your content from there.Summarize DocumentsLooking for a “Reader’s Digest” version of a book or article? Paste the text into a prompt to ChatGPT, and it returns a nice summary that should hit most of the highlights.Write and Troubleshoot Computer CodeBecause ChatGPT was built to analyze and produce output in human language, it is also great at computer languages. Programmers are using it to troubleshoot code. It can even generate new code if you ask it to develop a new capability.Translate DocumentsFar more effective than other translators out there, ChatGPT can take a website or article and translate it into another language.Write and Grade QuizzesYou can feed your content into GPT and ask it to generate a quiz with answers. Or, give it a general subject, like world history, and it will do the same. Feed your student’s quiz responses into GPT, and it will grade them (even essays!).Create GamesI asked GPT to create a game teaching sixth-grade students about the solar system, and the bot came back with a fun adventure through space.Improve Your WritingI didn’t do this for this article, but I could have asked GPT to review my draft and suggest improvements.Conduct ResearchI asked the bot to provide a list of current peer-reviewed studies on chatbot use in education using APA style and received a properly formatted list.Why Does ChatGPT Have Some People Worried?While most of the comments about GPT are positive, you will find a few educators deeply concerned about this new tool’s capabilities.Mimicking Human WritingTeachers are worried about GPT’s ability to write plausible text reading like a human created it and write in a known author’s style. A student could ask the bot to write a paper in the same style as one of their previous works, and it would be hard to tell the GPT paper wasn’t an original student effort. But, again, this mimicking ability is nothing new. AI has generated art in the style of great artists for years. ChatGPT applies the same predictive approach to language.ChatGPT Makes MistakesAnother concern is the bot’s ability to present false information so convincingly that some users forget to validate the response with research. For example, I asked the bot to calculate the mean of five numbers, and the result was interesting. I received a detailed explanation of the statistical definition of “mean” and how to calculate it, which I did NOT request. However, the bot added the five numbers incorrectly.To be fair, ChatGPT wasn’t built to perform math problems. But it didn’t tell me that or report the result might be incorrect. Like any willing assistant, it did its best to provide an answer. If I had been a student relying on GPT’s help, I would have been wrong. The reason for most of these mistakes is the same reason that it is such a stunning improvement in the chatbot experience: It’s primary sources of information are the internet and Wikipedia.I asked Vince Han, one of the pioneers in building chatbots for learning, to sum up the challenges presented by GPT’s new capabilities, and here’s his response:“One of the effects of ChatGPT & increased quality in Artificial Intelligence in general is shifting the skills users need. For example, most users have developed a skill for how to search which is knowing how to craft keywords and how to sift through search results. Now, people will need the skill of discernment, or learning how to decipher truth and views from a chatbot that provides what reads like a definitive answer. So, from a performance support perspective, yes, one can now get an immediate response to a question but how accurate and relevant that response is will be up to them to determine. “As educators, we must now confront a problem that has been boiling for decades—the decline in critical thinking across most domains, demographics, and workgroups. In my opinion, ChatGPT makes this gap more visible and increases the urgency for closing it quickly.Closing ThoughtsBecause Microsoft owns the license for the exclusive use of the model, we’ve all been working for them, revealing exciting use cases that may find their way into AI-enabled Microsoft products. For now, the platform consists of both free and professional versions. It’s an excellent way to begin learning about the promise of AI and our place alongside it. As machine learning expert Santiago Pino says, “AI will not replace you. A person using AI will.”Where Can You Learn More?There are so many articles about this subject right now that I’ve limited myself to a few favorites:Under the Hood of Chat GPT With Myra RoldanHow ChatGPT3 Impacts the Future of L&D in an AI WorldChatGPT Has Educators Scrambling to Keep UpWhat Is ChatGPT and How to Teach With ItWhat Is ChatGPT and Why Does It Matter?What Is ChatGPT? Is It the Beginning of the End?Ger’s Learning Notes #51 - ChatGPTFor more information and resources, Check out ATD's AI Resource page.

ChatGPT-3 Has Everyone Talking

At the beginning of the past two years, I’ve written a post on important points I think you should know. In 2021 I mentioned the importance of context, emotion, and learning to learn. In 2022 I discussed some of the newer cognitive science models of cognition—including distributed, social, and embodied cognition—as well as the role of mental models. This year, I’d like to focus on a trend that we should extinguish: the belief that we are formal, logical, reasoning beings.The cognitive revolution transcended the view of behavioral psychology; we couldn’t talk about what went on in the brain. While many valuable insights were generated by that scientific paradigm, including the role of performance objectives, the move to thinking we could isolate necessary components provided further advances. Results that have stood the test of time—the recognition of sensory, working, and long-term memory as well as core cognitive processes like attention, elaboration, and retrieval—were identified.We tend to adopt existing models as frameworks for thinking about the brain, and the cognitive revolution was no exception. For example, consider the computer as a general-purpose information processing system. It made sense to adopt that model to reflect our abilities to process information and complete a variety of tasks. Thus, the dominant framework for thinking about thinking was to view our brains as formal, logical, reasoning machines.Such models, with some assumptions like mapping the world into formal representations, accomplished impressive tasks. Systems built upon that model could break down problems into components and solve them. They could even learn. Artificial intelligence (AI) advanced as a result. So-called “expert systems” were built upon interviews with the best practitioners. Sadly, these models failed in practical situations, leading to an AI winter. This was a disconcerting time, when behavior naturally observed in humans did not reflect these symbolic systems.The Postcognitive RevolutionOne of the most impressive capabilities of our thinking is the ability to organize the world into flexible frameworks. What David Rumelhart called schemas, Marvin Minsky called frames, and Roger Schank called scripts all reflect our ability to infer knowledge about the world. Thus, we had representations such as wheels that scaled up to vehicles but differently for automobiles versus bicycles. Symbolic logic struggles with this, however. Attempts were made to adapt, such as Lofti Zadeh’s Fuzzy Logic and Douglas Hofstedter’s Slipnets, but suffered in practice.An alternative framework emerged that could be characterized as “subsymbolic,” operating below the symbolic level. John Holland, Keith Holyoak, Richard Nisbett, and Paul Thagard created a genetic algorithm approach in Induction, which borrowed mutation from biology, but wasn’t aligned with neuroscience. Rumelhart, along with Jay McClelland, proposed connectionist models instead in their two-volume set Parallel Distributed Processing with a neuroscientific inspiration. These models were found to exhibit rational behavior similar to humans. They represented the same types of behaviors, including errors, that were empirically obtained from psychology research.These neural net models created not only a psychological revolution, capturing elements of cognition, but also a renewed basis for artificial intelligence, underpinning much of today’s work in machine learning.A Cognitive LegacyOne small legacy of the formal, logical, reasoning model of cognition, however, seems to linger. It appears that many of our practices in learning and development (and business in general), still seem to be based on this idea. This is a potential problem.If we’re formal logical beings, there are several things we should be able to assume. For one, if we give people new information, they’ll appropriately adjust their behavior. Moreover, the person with the greatest access to information should be the one to make a decision, and that decision will be the best one possible. We see this manifest in several ways.For one, our training frequently resembles the presentation of new information with an expectation of behavior change. Too often, we see bullet points of information extracted from (and even presented by) experts. After this, there’s an expectation that we’ll do something differently. Similarly with annual or quarterly reviews. People need feedback at the appropriate time and support for change. We don’t just change when we wish to; it takes time and practice.We also see decisions made by individuals without consultation. Folks will decree a new approach because it’s been determined by the person at the appropriate decision point. Yet the evidence suggests that we get better outcomes when we bring more people into the decision-making process, such as in Keith Sawyer’s Group Genius. When we follow practices that help us suitably explore a problem individually, and then collectively brainstorm solutions, test, and adapt, the outcomes are better.A Problematic LegacySadly, there is also evidence that this approach to work isn’t working. We see it in the outcomes of learning investments, and when we evaluate organizational success as an outcome of our approaches, we see that more informed ones work better.In the 2012 Harvard Business School working paper The Great Training Robbery, Michael Beer, Magnus Finnstrom, and Derek Schrader cite metanalyses that suggest only 10 percent of training investment has an impact. There may be multiple causes for this, but it does suggest that we’re wasting money with our approaches. The Serious eLearning Manifesto suggests that much of traditional learning is based on information presentation, so it’s safe to say that we’re not doing as well as we could.Similarly, efforts like Laura Overton’s Toward Maturity and Lauri Bassi’s Human Capital Capability Scorecard suggest that organizations do better when they’re better aligned with how people think, work, and learn. Organizational performance reliably improves with better employee approaches.Overall, the continued use of ineffective practices undermines our organizational success. We need to effectively achieve our organizational goals, which includes ensuring we can optimally execute what we know we need to do and continually innovate when we have to invent or adapt.Moving ForwardWhat should you do differently?Truly understand how your brain actually operates. Know the basics of the human information processing loop and extensions for learning and performance.Avoid the myths that continue to hinder our industry.Conduct deep needs analyses to find important performance gaps and their root causes. There are many causes of problems besides a lack of knowledge or skills. Stop thinking that if you give people knowledge, behavior will change.Leverage external resources. Decision making is best done collectively and iteratively. To do otherwise is to minimize your potential.Move from legacy practices to evidence-based practices. There are vested interests that continue to exist based on perpetuating approaches that aren’t effective. Many products that are quite lucrative—profiling instruments and development programs —aren’t effective.Start measuring impact. This doesn’t mean what people prefer (there’s no correlation with what people think of a program and its impact). Instead, assess the actual change in the organization. You can’t collect this yourself; you must partner with the business units that require change.Align programs to demonstrable gaps, and then measure until they’ve improved.Ultimately, L&D should be the folks that are responsible for not only optimal execution but also continual innovation. When you’re innovating (problem solving, designing, or researching), you don’t know the answer when you start, so inherently it is learning. It’s informal learning, and no one knows the answer (yet)! Still, it’s an opportunity for L&D to attain a more valuable role in the organization.To make this strategic move and start bringing evidence-based practices from formal to informal learning, take ownership of these practices within your teams. Make it safe to fail; share stories of what you tried; and learn together. Create and understand a learning microculture before taking it to the rest of the organization.Last, identify the organizational legacies that are holding you, and your organizations, back. Then, find alternatives, and implement them. When you’re free of your unexamined past, you can move forward to an evidence-based future that means better outcomes. That’s a trend I’d like to see and encourage you to undertake.

A Trend to Supersede

categories.lvl0

Search results for 'Brian L. Davis'

Categories

Refine By