AI & Deep Learning - Reaching the Pinnacle

A note before the summit

This final part was meant to follow the others by a few weeks. Instead it arrives more than two years later. In between, I co-founded Paradigm Inputs and spent my days building an agricultural commerce company rather than writing about AI strategy. In that same stretch the top of this pyramid changed more than any other part of it.

So rather than publish the chapter I outlined in 2024, I want to describe the pinnacle as it looks from where I sit now: as a CEO using these tools every day, in an industry that runs on paper, phone calls and relationships.

Three moments that built the summit

In 2012 a deep neural network called AlexNet, built by Alex Krizhevsky, Ilya Sutskever and Geoffrey Hinton at the University of Toronto, entered the ImageNet image recognition competition. Its top-5 error rate was about 15%. The next best entry was about 26%. That gap told the world that deep learning, the many-layered neural networks I described in a 2019 post as building a jigsaw puzzle one layer at a time, had arrived.

In 2017 a team at Google published "Attention Is All You Need", introducing the Transformer architecture. It made it practical to train enormous models on enormous amounts of text, and almost every large language model since is built on it.

And on November 30, 2022, OpenAI released ChatGPT. Within roughly two months it was reported to have reached 100 million users, at the time the fastest adoption of any consumer application on record. Overnight, the most advanced layer of the AI Strategy Framework stopped being something only Google or Netflix could reach.

What changed at the top of the pyramid

When I first drew this framework, reaching the pinnacle meant building your own deep learning models. You needed huge proprietary datasets, specialised hardware, scarce research talent and a long runway. For most companies the honest advice was that the top layer was years away, if it was reachable at all.

Today the most capable models in the world are available to anyone with a credit card and an API key. You no longer have to build the summit. You can rent it.

That sounds like it makes the rest of this series obsolete. It does the opposite. When everyone can rent the same model, the model stops being an advantage. What separates companies is everything underneath it:

• The data you collect that nobody else has (Part III).
• How reliably it's stored and how quickly it can be found (Part IV).
• Whether it's clean enough to trust (Part V).
• Whether your people already make decisions with it (Part VI).
• Whether you have the habit of measuring, learning and improving (Part VII).

A powerful model pointed at messy, inaccessible data gives you confident, fluent, wrong answers. The same model pointed at a trusted Gold dataset becomes one of the most valuable employees you have.

What the pinnacle looks like in practice

The work at the top of the pyramid used to be training models. Now it's mostly four other things.

Grounding

A general model knows a lot about the world and nothing about your business. Grounding means giving it the right information at the moment it needs it, usually by retrieving relevant documents or records and including them with the question. The quality of the answer depends almost entirely on the quality of what you retrieve, which is why the storage and transformation layers matter so much.

Tools

The biggest shift of the last two years is that models can now act as well as answer. Given access to tools, a model can look up an order, draft an invoice or check inventory. These systems are often called agents. The strategic question becomes which actions you are willing to let software take on its own, and which still need a person to approve them.

Evaluation

In Part VII we said every model must be measured against a baseline. That's even more important when you're renting a model that changes underneath you. Build a set of real questions with known good answers, run it every time you change a prompt, a data source or a model, and treat a drop in the score the way you'd treat a failing financial control.

Guardrails

Language models make things up when they don't know. They can be manipulated by text hidden in the documents they read. They will happily send your private data wherever they're told to. Back in 2020 I argued that data-driven companies need to act like banks, guarding customer data the way a bank guards deposits. That argument has never been more true. Decide what the system can see, what it can do, and what it can send out, before it goes live.

The view from Paradigm

At Paradigm Inputs we build software for independent brokers: rural entrepreneurs who sell crop inputs to farmers in their community. One of the tools in that platform is AgAssist™, which answers questions about crop protection, fertilizer, seed and biologicals and is backed by a library of product labels.

The language model underneath AgAssist is not something we built, and we don't need to have built it. What makes it useful to a broker standing in a farmer's yard is everything around the model: a curated label library, clean product data, and a workflow that already holds their customers, orders and invoices. The pyramid, in other words. We started at the bottom, and the top became possible because the bottom was solid.

The same is true of the more ambitious things we're building. The hard problems are rarely the model. They're getting the data right, deciding what the software is allowed to do, and measuring whether it actually helped.

When to build your own

Renting is the right starting point for almost everyone. There are still good reasons to train or fine-tune your own deep learning models:

• You have large amounts of data that is genuinely unique to you, such as images from your own equipment or sensor readings from your own operations.
• You need predictions at a volume or speed where per-call pricing doesn't work.
• The task is narrow enough that a small specialised model beats a large general one.
• Your data can't leave your environment for regulatory or contractual reasons.

If none of those apply, spend your time on the layers below and let the model providers compete to give you a better summit every year.

Keep it simple

1. Rent the model, own the data. Use the best available model and put your effort into the data and workflow only you have.
2. Measure before you trust. No AI feature ships without an evaluation set and a baseline to beat.
3. Give software as much autonomy as you can audit. Start with suggestions, then approvals, then actions, and widen the boundary only as the evidence comes in.

Wrapping Up the AI Strategy Framework

When I started this series, I said I'd explain why I walked away from the thousand-plus hours I spent on the CFA exams. Here is the short answer. The same thing that makes a good investment makes a good AI strategy: a clear thesis, reliable information, disciplined measurement and the patience to compound small advantages over time.

The framework is a pyramid for a reason. Product vision gives the work a purpose. Collection, storage and transformation turn raw data into something you can trust. Analysis turns trusted data into decisions. Machine learning turns decisions into a system that improves itself. And at the top, AI and deep learning turn all of it into capabilities that would have looked like science fiction when I first wrote about this in 2019.

The tools at the top will keep changing, faster than any of us expect. The foundation will not. Build it well and you'll be ready for whatever arrives next.

It's still never too late to start. Thanks for reading along.