I started at Filtered in July 2016 because I wanted to work at the business farthest along in artificial intelligence in our industry. I also liked the Shoreditch location and the serious lack of a dress code, and those things seemed related to the cool tech.
Yet, as sophisticated as they were, I was struck by the simplicity of our machine learning algorithms then. We were awarded the first U.S. patent for an algorithmically personalised syllabus. It was powerful but, relying on unsupervised clustering, it required a fair tranche of user feedback data (which we gained by selling an Excel course to 1 million people) to produce meaningful results (which it did!).
Thus, I became fond of saying that if you wanted to do something with AI you had better start, as Filtered does in each case with our new products, using what is known as symbolic, classical or ‘good old-fashioned AI’ (GOFAI). This is a method of programming predetermined mathematical logic (or a sequence of IF THIS THEN THAT combinations) into machines as old as the Antikythera mechanism. Laboriously hand-tune your robot to follow its determined path and you’ll end up with a functional automation and, thus, you’ll garner enough feedback data to build a machine learning algorithm on top.
For a while I ceased to believe that AI would ever achieve something that looked like human reasoning, even as we deployed more and more sophisticated natural language processing algorithms. The earliest GPT models seemed weak and vapid.
My personal ‘aha’ moment with ChatGPT was when, with aching legs, I asked it to help me make a massage roller from the stuff I had at home. It suggested that I freeze a bottle of water and use that, noting that the cooling effect of the ice would also be useful.
I tracked down sources for this wisdom to Reddit (Reddit is one of the major data sources for GenAI models). But it had also understood the additional benefits of using ice. It didn’t just search: it had connected dots to solve a simple problem, recontextualising the experience of others.
In other words, it was reflecting extant knowledge back to me in a helpful new way.
I think the technology most alike to generative AI is the mirror. Not the dull, ‘in-a-glass-darkly’ reflectors of ancient times, but the crystal-clear lead-backed mirrors of the Renaissance. These perfect mirrors unlocked incredible new perspectives and problem-solving powers, leading to new movements in art and new scientific discoveries, such as photography and optics.
GenAI is such a reflector. But rather than light, GenAI reflects knowledge. Like mirrors did for light, GenAI turns knowledge into movable material that can be refracted, redirected and spun out in new configurations. GenAI produces new forms of reach for our reasoning capability, in sharp contrast to most human knowledge’s tendency to get bogged down in bodies and mushed up with weird emotions and feelings.
Early studies show that GenAI is moving the productivity needle in certain areas for knowledge workers. But I think that its main significance is to change our sense of what it is possible to do with knowledge. In talent development, the knowledge and reasoning agent that is generative AI promises to accomplish three main things.
(I don’t expect AI to do any of them as well as a human. But that isn’t the point. Think about the time before gramophones and radio, when the only way to listen to a Beethoven symphony was to physically be in the concert hall. The two-dimensional version broadcast to millions on the radio was not, admittedly, the same as the three-dimensional sound object experienced in the concert hall. It was a reflection, not the real thing. But nevertheless, an experience became available reliably and at scale)
All of these things do not happen reliably and at scale in organisations. We have proven that GenAI plus the Filtered search algorithm can already do them all:
1. Coach through a specific problem to find content and get advice
This is the use of GenAI at work as it is being used in personal life by millions already. The difference is that GenAI hooked into your existing learning content can point your people to specific, verified resources as sources, rather than Reddit (which I think we agree is a good thing). We have accomplished this by linking GenAI to the Filtered search algorithm which is now available as an API.
2. Manage skills in an
external system
This one is so mundane and trivial that I love it. One of the most annoying things in HR systems life is getting people to update skill profiles and goals. Even when the UX is lovely, it’s still a time consuming process to rate your skills in one assessment tool or another, transfer results and so on. We found something amazing: hook up GenAI to an LXP API and it can use and update those skill priorities for you. It figures out how to use the API to make those changes, without hard programming.
3. Skills based career goals and how to achieve them
This last is the most exciting. We found that, with access to an LXP API for data on the current role, it was incredibly easy for generative AI to figure out a likely career goal with almost no prompting (actually: no prompting). Generative AI doesn’t solve the engagement issue: for lack of time and enthusiasm, people will refuse to give a bunch of information to a bot just like they don’t update their skill profiles. But an AI interface that could infer likely plans from the existing data, build a career goal, and curate a pathway against it would make something available to the masses that has been the preserve of a select few up until now.
Effectively, this is GenAI + APIs as a way to integrate skills and learning data across different systems in the talent technology ecosystem. Skills data can be retrieved, updated and synchronised everywhere in the course of a conversation with an AI that is capable of reaching into each system. Accomplishing this would mean we had solved one of the thorniest challenges faced by HR: how to connect our different skill systems. As one of our pilot customers says:
“The potential the AI Agent has to unlock talent mobility, career planning, reskilling and upskilling across the business at scale is incredible. Every one of our thousands of employees will have their own ultra-personalised career coach, available 24/7, helping them understand and access all the opportunities available to them.” -
Learner Experience Manager, global beverages business
Filtered has a bot that can do all three things already. But we have also encountered the main obstacles to deploying it:
- Corporate worry about copyright and GenAI: the nervousness is that the AI will produce material that pirates protected intellectual property (recently a big issue for Figma). (That shouldn’t be a big problem for the three use cases above though. Pointing GenAI only at the metadata from a verified content database goes some significant way to mitigating copyright theft risk, too).
- The potential cost of doing this at scale with the most powerful models, which is still a big issue for GPT-4 equivalent AI. Managing the currently high costs per API request will require some careful licensing and optimisation
- Managing the hand-off between AI and human: Recently a customer told me about this issue with an OpenAI-based HR chatbot: what if someone mentions a bereavement? At that point you need to bring in a real human advisor, fast. It’s likely corporate AI systems will require a human in the loop.
It will take us all a while to work through these issues through pilots, tests and mitigations. But I am confident that we will overcome them because the pay-off, to finally solve some of the most intractable system usability issues with coaching and skill development through technology, is enormous. Just like many other big problems: the adoption of cloud technologies is a good analogy. Even ten years ago it was seen as far too risky or expensive to use the internet to host HR services.
But if we want something to be a reality in three to five years, we have to start planning for it now, investing in it now. It won’t be free. It won’t be easy. It may not pay off immediately. But is it worth starting now? Absolutely.
If you want to start now, get in touch.
