At the start of my career, I would occasionally attend events at my company where people would talk and "network" with each other. Mostly, it would be people from within our company, but occasionally others would join. They could be from our investors, consultancies we worked with, partners or customers... and so on. Any node from our network of associated organisations (bar competitors).
I used to wonder, what was the point? How did speaking to all of these people help anyone? What did we learn? Was it all to suggest some level of prestige for our company? Was it just to make people feel good about working there? Was it just for the extroverts in the building to get their fix? Didn't they get that every day? This was back in the forgotten era, where we all used to have to come into the same building every day.
As a data person, I thought we could learn more from looking at the data - doing interesting analyses. We could dive into why a metric changed, suggest what we could do to improve a metric. I believed this was a better way to learn about our business than talking to people.
However, one thing that I didn't consider at the time, is that I was always talking to other people in the business. In the kitchen, waiting at the coffee machine, in a meeting, at a pub or a bar, at or after an all hands... Then it all stopped.
We watched with concern at work as it spread from country to country, the cases increasing exponentially... One Friday we were told to practice working from home, should we need to do it for a protracted time. The plan was to come back on Monday and then do another "practice" day on Wednesday - I left some things on my desk, thinking I'd get them next week... Even though this era has ended, many of us now work remotely, so we don't habitually occupy the same space as our colleagues. This isn't a post saying I'm suddenly reborn to believe that working in an office is best, and that we should all be in the office 3 days a week, or “it’s not going to work out”.
I, indeed, prefer remote work being the default, and making the choice to see people regularly in a more purposeful format. This is why I run London Analytics Meetup, a big part of why I attend conferences and why I try to meet people in person when they come through town.
Most of the time when I used to spend time with my colleagues in an office-first setting, I would glean little pieces of information that would never be in the data. Buying a drink at the bar with an account executive... who told me the reason we couldn't win more business with that customer is because our competitor, whom they also buy from, was a fellow portco and the CEOs were close. Chatting with our CCO after an all hands, to find out that we lost that big customer because no one bothered to speak to them for two years.
Even though I gained all of this information (not data, real information) through these interactions, I still had my doubts. Is this serendipitous? Can I really rely on this as a way to find things out? How much of what people tell me is true? How much of it is instinct or hearsay? It lacked the scientific rigour I wanted.
What has this all got to do with LLMs? The more I think about LLMs, the more I think about people. LLMs are not sentient, they're not truly able to reason or decide. However, hundreds of millions of us, and a big majority of the knowledge worker population, have interfaced with them. LLMs have broken the Turing test - many people find it uncanny how human they sound. We understand there is value in using LLMs already: we don't know how far the rainbow extends or how big the pot of gold is. We just know it's there and it's big.
The more I think about how LLMs work, the more I think about how human instinct and wisdom works. Why is it that experienced people often (but not always) come to better decisions than those less experienced? The structure of LLMs is inspired by neurones in the human brain. We have seen that the higher the quantity and quality of data an LLM has been trained on, the more powerful the LLM will likely be. Is this how we work, too? Is this why highly knowledgeable and experienced people (with good intentions) can usually come to better decisions than others? Better quality input data (education, reading, work experience) in a higher quantity.
Should we actually trust our instincts more? Provided we have broad understanding of the subject and a good amount of experience? Are these feelings the output of a powerful 11m parameter biological model? Pretty soon, hundreds of millions of us will be using the outputs of LLMs in many contexts. If we can trust these outputs, where the inner working of the model is at least as opaque as the inner working of our own mind, shouldn't we trust ourselves more? I used to think that human instinct, intuition or wisdom was untrustworthy, to be used in the absence of data. In this era of LLMs which are household names - GPT-4, ChatGPT, Bard, Llama2... perhaps Yourself-1 is an even more important model. It's a unique model, truly one of a kind, competitive with GPT-4 on many tests - valuable.
We've started to get LLMs to interact with each other. Their collective ability is not dissimilar to ours. When we come together and interface with each other, we have the collective benefit of all our experience and knowledge. I wasn't harking back to an office-first world before, I was exploring how these interfaces used to work. They were far from perfect - office politics, huge amounts of self-interest and poor alignment plagued them and continue to do so.
This week, we're all gathering together in my home city - more people I know, from further afar, are coming than I can remember. This new remote-first era, with purposeful gathering, is better than the past. We're gathering with intent - to learn from each other, to catch up with our friends from afar, to understand possibilities, to drive our thinking as an industry forwards... Surely this is better than happenstance interactions at the coffee machine? Better than the interaction polluted by office politics and self-interest?
Sure, we’re going to be talking a lot about AI and LLMs, but we are talking to each other.
Come say hi if you see me on the floor!
Thanks to Ravit Jain and Sanjeev Mohan for getting it started early! I really enjoyed the session yesterday and the talks were great. Catch the slides or recording if you can.