As you wander around the exhibition at Learning Technologies 2024 exhibition or attend the conference, remember that all is not as it seems. Forget the demand for ‘more training’ on AI, more ‘skills training’ from L&D on AI. That’s not how this is playing out. I spoke alongside the main keynote Daniel Susskind at the CIPD Conference and he said then what he said now – train people on how to use AI. To be fair he’s an economist and doesn’t know much about workplace learning. But this is the misfire I’m talking about. It’s the hammer and nail problem. If you carry a (course mindset) around like a hammer, everything will look like a nail (course)
As L&D try to get to grips with AI, they misfire for one simple reason. AI, like water, is a rising tide that never ebbs. It seeps and soaks into the workplace, and workplace learning, like an invisible force. It has bypassed L&D, and it is everywhere.
The best technology is invisible
Outside of work, it is what search is, mediates social media, catches harmful content, is the interfaces to your streaming services, your passport at border gates, scanners in supermarkets, sensors in cars, ANPR on our roads. You are rarely out of sight or the mind of intelligent AI but you rarely see it as AI. The world is now awash with invisible AI.
The same is now true in learning. At work if you use Microsoft, Google or most other services, AI has saturated into almost everything. Not just Copilot but predictive text, spellchecking, keeping spam out of your email. Behind the scenes things are becoming more automated through AI – they just happen. Learning is now happening all the time with technology that is almost invisible.
The mistake, of course, is to see AI as a course content tool. I don’t mind this as it will and is happening, disrupting the whole e-learning content market. But that it is not what we used to call the ‘killer app’. There’s lots of this at Learning Technologies, as it is how L&D thinks and the vendors sell to what people think they want, not necessarily to waht people are doing.
The real killer app is something we’ve known works for many years, with a long history back to the 70s and 80s but substantiated by great work done by Marsick, Watkins, Gery, Rossett, Cross, Wallce and others for many years. It hasn’t happened on scale, as it has been something that has been very hard to implement – until now.
It is PERFORMANCE SUPPORT.
Almost all current use of AI in learning is actually ‘performance support’. The hundreds of millions using ChatGPT and similar services use it for task and performance support. No one is typing – ‘Give me a short course on X’ into ChatGPT or Claude. Everyone is saying ‘Can you tell me, help me, show me… what, when where or how to do things, get things done…’
We have it open as a tab as it is do damn useful. I use it as much as I use search. Once you learn how to move from single queries to real dialogue you start to use it in your workflow, the whole workflow learning thing suddenly makes sense, it becomes a reality. As you start to use it for more sophisticated tasks, data analysis, coding, critiquing work, creating text and other media, you feel the power of it as a performance assistant.
Perfect performance interface
You know what you want, you know what context you are in at that moment and, as Papert said, the perfect interface is not some clumsy menu system in a VLE, LMS or forward and back buttons on an e-learning course. Turns out to be a box into which you ask it something. Papert described this is a ‘low floor, high ceiling, wide walls’ interface. Low floor – amazingly simple to use. High ceiling – gives you back far more that you expected. Wide walls – seems to know everything. The most successful interfaces are becoming simple and as frictionless as possible.
Personal agency and engagement
What makes it work, unlike an LMS or e-learning course, is that it gives you personal agency. You feel in control and the reward is feeling that you’re learning and getting things done. It so often exceeds your expectations. The excitement of using these tools is what made GenAI the fasted adopted technology ever. You don’t have to worry about ‘engagement’. The whole world has ‘engaged’ with this technology, with billions of uses per month – at home, in schools, Universities and, of course – the workplace.
Digital agents
GenAI has already expanded into performance support through agents. I have my Digital-Don. A good test for any expert at the conference is to ask if they have a persona GPT or agent. If they don’t ask why not. It takes about 30 minutes to build one with zero coding skills.
I really like those who are doing things with digital coaches and assistants and there are a few of these around. They get it. Agents are the next evolution in using AI for learning. We had hints with early LLM use that specific types of prompts improved performance, things like, telling it that the results really mattered or expressing emotions. This direction of travel is now being built in GenAI.
Why is this big news? It gives teaching and learning potency to an LLM. Imagine the agent as a human, that does not just prompt and wait on an answer, but engages in dialogue about what is wanted, initiates an actual learning experience or actual task as a workflow.GPT3.5 with an agentic workflow already outperforms GPT4 and prmise3s to amplify the effectiveness of LLM services,
Agentic workflows reflect on the output you give it, improves it and tells you what you’ve need to know or do. An agent can look at output 1, critique it, find errors and improve the next output. You can have specific types of agents, such as a critic agent, teaching agent, domain specific task agent.
Agentic tools can be used such as a lookup tool, such a solve a complex maths problem or do some data analysis, things the LLM cannot do. Examples of tools would be Wolfram Alpha for mathematical analysis, searching the web, searching Wikipedia, email and calendar productivity tools, image generation, image captioning or object detection. We have seen all of these incorporated into AI services. They are there now and resuable.
Planning is another capability. Agentic approach allow the tech to plan and explain things step by step to improve performance. Agents can break things down into steps, as well as iterate within those steps. Send to a job support agent or research agent and this will be executed.
Multi-agent collaboration is hugely promising. You prompt an LLM, saying you are a CEO, supervisor, state a job role, then generate complex outputs. The difference is that different agents can debate each other to improve performance. This is a bit like creating a team suited to the task, with different characteristics and capabilities.
Agentic reasoning will matter this year. Agentic workflows (all of the above) where agents do things with the output from other agents.
Conclusion
If I were an investor or someone buying at LT – I’d start with ‘performance support’. Learning pool’s last acquisition was a Performance Support company, others are looking for such companies. The revolution should be around performance in workplace learning. This is real paradigm shift but you won’t see it much on the exhibition floor or conference talks. I say it again - If you want to move a graveyard you will get no help from the incumbents.
No comments:
Post a Comment