The 3 most exciting problems in AI

Over the last few weeks, I’ve spent a lot of time thinking about what value large language models (LLMs) will bring to the enterprise. I believe this value will come from automating workflows in every function. This will dramatically improve productivity across the board. You will have the ability to delegate tasks to an AI based tool, much like you would do to another colleague.

Today, I’m going to cover why I have deep conviction in this and the problems that need solving if we are to get there. I’ve never felt this kind of energy with an emerging technology and it’s an exciting time to be building. Here’s why.

The opportunity

LLMs will help us automate multi-step tasks and unlock productivity like we’ve never seen before. I believe this because:

1. So much of work involves repetitive tasks. I’ve started to note down tasks in my day that are repetitive and the list is growing every single day. For example, before I meet someone I either look them up on LinkedIn (if I’ve never met them before) or go through my notes from the previous meeting. This means I switch away from a task I was focussed on: writing code, doing research for my product or something unrelated to work. Time saved compounds: you get the time back, but you’re also going to be more productive at whatever you were doing before because you don’t switch context.

2. LLMs are better at processing large amounts of information, especially unstructured information. If I gave you 10,000 books that had 10,000 words each, how long would you take to synthesise the key findings? An LLM can do it in minutes for a small fee.

3. LLMs are reasoning engines, which means they can recommend decisions that are subjective. To date, we’ve not automated workflows with a subjective step because the investment required to do so was too high. If you needed a computer to make a decision on whether to follow path A or B given a set of information, you could do so but the tech involved was non-trivial. Today, GPT4 makes this easier than ever before.

It’s hard to debate the value of the above. What I think is more interesting, is how we get there. There are tough problems to be solved and I’m going to list what I think they are.

Components of workflows

Imagine you’ve just joined a new company. You’re tasked with helping your manager, the head of sales, prepare for an upcoming sales meeting with Heinz Ketchup. How would you start?

First, you will want more context on the meeting. Who are we meeting within the organisation What is the goal of the meeting? What have we discussed in previous meetings?

You might then gather additional information: for example, look through past proposals to understand the tone and style used by your company. Look through the annual reports of the prospect to figure out what their priorities are, and how your company’s product could help them.

Once you’ve gathered all of your information, you will put together material for the meeting like a sales presentation. You will probably have multiple reviews with your manager before the material is finalised.

So in general, the steps involved are:

1. Gathering information and understanding context

2. Taking the right action

3. Presenting your draft

4. Implementing feedback

This flow can be generalised across a variety of different tasks across businesses.

Problems to be solved

Now, imagine an AI assistant helping you get the task done faster. To achieve this level of automation, we need to solve these three problems:

The accuracy problem

LLMs require context about the task at hand. They don’t have the same world views or context that you have built up over time (though this will improve). This means they need all the context they can get to complete the task at hand.

On the other hand, LLMs like ChatGPT have limited memory. They simply cannot remember as much as you can. The most common method for this is storing information elsewhere. Imagine a folder on your laptop where you save stuff, when you ask the LLM to do something it will look through the folder and pick something that is relevant to the query at hand. It works pretty well but it isn’t perfect.

At their heart, LLMs are trained to predict the next word. A combination of limited context and memory means that we will get inaccurate results. When you do multi-step workflows, this inaccuracy compounds (i.e. 10% inaccuracy in step 1 shows up as 50% inaccuracy in the final output).

The walled garden problem

800 × 450

To really add value, you need to take action: make a slide deck, book a meeting etc. The traditional way to do this is to use APIs — code that applications use to communicate with each other. If you need to build integrations between every app and every other app, you may as well not use an LLM.

To truly unlock automation, you want the LLM to decide how to communicate with an application. It might choose how to use an API to do it, or alternatively it might do it like you do: log in to an application, click some buttons and then hit ‘Submit’. The cons of using a user interface to do it is that it’s prone to error.

The handover problem

Most complex workflows won’t be handled entirely by applications. This means that the human needs to be in the loop. When do you loop in the human? Do it too often and there’s no point in automation. Do it too little, and you will end up with low accuracy. Choosing when to present something to the human is much like choosing when to present something to your manager.

Exciting problems to solve

These are exciting problems to solve because it unlocks a new wave of disruption. It means people can do more meaningful work. The technology is changing rapidly, which makes it an exciting space to be building in.

The problems require deep thinking on engineering, user experience and product. I’m excited to be going after them as part of my product: Kili. We’re recruiting engineers to join our team. If you or someone you know is excited by the opportunity, see this page for more details and get in touch.