Large Language Models Are The Next iPhone Moment
We’re in a new iPhone moment.
The world of technology has always moved quickly, but every so often, there’s something so new and innovative that it totally changes the way we work, and sometimes, even the way we live. The internet was one of those things. So was the iPhone. Now, I’m betting ChatGPT is on that list as well.
By now, you’ve probably heard of (and used) ChatGPT, a large language model (LLM) from OpenAI, which allows anyone to leverage the power of near-human level language and reasoning skills in an automated tool. There’s been so much hype around this new technology that you could easily be forgiven for being skeptical and writing it off as just another fad. I wasn’t paying it much attention myself not too long ago. But ChatGPT, and large language models in general, aren’t a fad, and the hype is well justified. ChatGPT and other large language models are likely to entirely transform the way we work because they’re able to do something that machines have never been able to do before: reason.
For those in the world of machine learning and A.I., you’ll likely scoff at that last statement. We know that ChatGPT isn’t really “reasoning” like humans do. In fact, all they’re really doing is predicting which words are most likely to come next in a series of other words. But this basic premise disguises the world of possibilities that these A.I. systems open up. The outputs that they generate, while not truly “thoughtful” in the traditional sense of the word, are logical, sensible, and meaningful.
Up until now, machines have been terrible at making intelligent decisions. In general, we’ve had to program in the specific set of rules that we want them to follow, which means that we, as humans, have been limited to using them for well-defined processes and tasks. Despite the enormous productivity boosts we’ve had using that old paradigm, it was still highly constrained, because when we as humans need to create the rules for what needs to happen in any given scenario, the systems we create can only generalize up to a certain point before they become limited in their capabilities.
But now, with machines that can act reasonably in a wide variety of scenarios, we are presented with a much wider, “blue sky” world where many more things become possible. We can begin to automate things that we may have never been able to reasonably automate in the past. We can now rely on machines to make reasonable “decisions” in more complex and nuanced situations. This allows us to scale up tasks that originally required human-level intellect and creativity, which will undoubtedly lead to large productivity gains. We’re likely to see a new world where many processes are automated, sophisticated and complicated systems are built with LLMs as key components, and business productivity will grow.
But there will be challenges
Despite the new capabilities that we’ll get, there will certainly be many speed bumps and challenges that we’ll need to overcome, both as individual employees, businesses, and societies as a whole. More and more white collar workers are worried about the security of their jobs. Existing businesses might see a dramatic rise in low-cost competition from those that leverage these tools. And even businesses that want to use ChatGPT will find themselves up against engineering challenges, cost issues, and data privacy and security concerns.
Despite how easy it is to use ChatGPT today, there are still problems with integrating it with business data. ChatGPT has a limited “prompt size”, which means you can’t just feed it a huge library of knowledge that it has never seen before and ask it to answer all of your questions. You need to develop tactics and systems that feed it the right information at the right time, or you need to develop a way for fine-tuning its underlying models using your own data.
And even if you can reliably connect your own data to ChatGPT and feed it the right snippets of information at the right time, you need to ensure data privacy, security, and permissions are respected. It’s too easy for a business to connect an internal employee database to ChatGPT, only to forget to limit which parts of that system can be used when querying the LLM, accidentally opening up sensitive employee data for all employees to see.
Leveraging large language models properly will take time, investment, and expertise. Companies that do it well will see a boost to their gains. Companies that don’t use it at all will risk becoming irrelevant.
Yes, I believe ChatGPT (and other LLMs) will be that impactful.
The engineering requirements are huge
It’s clear (to me at least) that there will be significant benefits that come from incorporating LLMs into your business processes, but teams that find themselves wanting to do anything more than simply going to the ChatGPT website to ask questions will find themselves building a world of infrastructure to handle their use cases. Our team at Locusive has already written tens of thousands of lines of code for creating batch processes on top of ChatGPT, creating chatbots, creating contextual information-retrieval systems that minimize hallucinations, and that enable new chat interfaces on top of existing workflows, and we’ve done it while focusing on maintaining existing data privacy, security, and permissioning systems that businesses have already invested in.
Creating applications for enterprises requires a lot of infrastructure. We’ve been working on that, and will continue to do so in the future. In fact, I’m so excited by the possibilities of LLMs and so convinced that businesses will need tools to manage their own deployment of LLM-supported features that I’ve pivoted my entire career to building the middleware and application layer for new LLM-enabled applications.
So far it has been a blast, and I look forward to helping more businesses talk to their data.
