“Wrestling the Crocodile” - Moving Beyond Prompts to Real AI Impact

If I walked down the street in Manchester city centre and asked a hundred people to help me to sedate a 12-foot crocodile, not one would be able - or want - to help me. I wouldn’t blame them.

A professional crocodile wrestler would be able to, of course.

A crocodile expert will have spent many years understanding the body language of such a beast. They know when a crocodile is defensive, in alert mode, or even calm. They would know when to apply a distraction to take the crocodile’s attention away from them, such as splashing water or throwing something to the side of the beast.

The professional crocodile wrestler would know to grab the head/jaws, sitting atop the crocodile, as they are arguably the most dangerous part of the animal. 

The reason that a crocodile expert would go for the jaws - and quickly try to bind them with strong tape or a noose -  is that crocodiles - while having ferociously strong bite strength - have pretty weak jaw-opening muscles.

In an ideal situation, there would be two or more people working together to get the croc into a manageable situation. One person on the jaws, one on the tail, and someone sedating with a syringe.

Just two more croc-related things, I promise.

A professional would know to not get between the crocodile and the water, and they certainly wouldn’t underestimate the fantastic short-burst speed of this wonderfully-evolved beast.

By now you will have worked out that this is one big analogy of how I think that we should all approach working with AI.

But first, the Dunning-Kruger Effect.

You will of course have seen this graph, and will have ‘been there’ in several aspects of your home and working life, learning new systems, software, hobbies etc.

I think that there is a similar effect that happens when people spend a consistent period of several months working with AI chatbots.

I have personally seen (far too) many people stop at what I call the ‘AI bro’ phase, in which they think AI will solve every single problem that we have. I think this is a pretty fatal move. And I have seen people that follow the ‘AI bro’ stage with a stage of abject disappointment when they realise that solving everything is not possible.

And then there is a further stage, in which people - and I would say that I am myself in this stage right now - understand that AI is ‘pretty dumb’ much of the time, but can be harnessed in ways to greatly advance how people go about doing their daily work.

And I'm sure that there are more stages to come.

Same as it ever was

I have written previously that ‘AI chatbots’ are the 2020s version of the 1995-era Microsoft Office suite, when it was released on Windows 95. You have to be of a certain age to remember just how powerful office workers all around the world felt when they saw Excel running in the new Windows. Anything felt possible.

There were thousands upon thousands of small but functional applications built largely with Microsoft Access and Microsoft Excel, by very keen and inquisitive office workers that worked late into the evening while everyone had gone home.

Prior to this, much business software (eg to run parts of a warehouse or factory) was largely offered by giants like IBM or IT consulting firms, costing tens of thousands, hundreds of thousands, and even millions of £/$.

So I think the parallels are very similar. With LLMs, we can now do many things that were unachievable (sometimes in our wildest dreams) just a few MONTHS ago. Right now, it largely feels that we are using something (AI) that is way ahead of its time.

My observations

One stark thing that I have seen, that I have not seen discussed too much, is that the prompt box at the bottom of the ChatGPT (or other chatbot) window, is inviting to inquisitive and/or tech-savvy people, yet intimidating to many others.

I have seen what I think will turn out to be a ‘Pareto’ 80/20 rule here, with 80% of people largely intimidated by the ‘box’.

So, right now, we in all probability, have only 20% of people in the world testing AI in ways that will help move things forward. But within that 20% is a sizable bunch of people that are either ‘AI bros’, and therefore think they know it all already, and so probably don't stretch AI to anywhere near it could go to - or time-poor people that would love to spend more time with AI, but life and work gets in the way.

So we are probably left with ~10% (which is ~50% of the ~20% that use AI in non-frivolous ways) of the world that are currently pushing AI in ways that will push things forward.

And of that remaining 10%, I would suspect that 80% (yet another Pareto distribution) of those people would not consider themselves as AI experts at all. They are simply using AI as a means to make other things happen. 

These ‘non-experts’ are the people that I believe will be really testing AI to its extremes. They will be doing this in ways such as using local LLMs (using AI on their own computers) to add things to their already-running systems, platforms, and software.

As I write this sentence, there will be a person adding AI to systems that have been running for 20-50 years, hugely improving the results in the process. Let’s have a look at an example of what this could look like:

An example of a so-called ‘non-expert’ using AI in great ways

I know an experienced designer at an office chair manufacturer has been creating office chair designs  for 25 years. Every year, they typically release new designs. They spend time reading various materials, attending conferences, spending time with interior designers, and furniture designers, all through the year.

This has served them well, and in 75% of years, they are very successful, as they have fit their product around current trends for that year.

Now let's see just a few ways in which AI could improve this process.

The wisdom of the ‘huge’ crowd

By using an API, and/or web scraper, the designer could grab huge amounts of human conversations about office chairs from various on-topic Reddit ‘subreddits’ (and other forums). There will be a great amount of text content once the data has been collected.

Using an LLM, the designer can ask many questions about the data, which will inform them of things that they would definitely not have known previously. In previous times, the designer would have used information gained from conferences, magazines, and other industry professionals. This would have been helpful of course, but in comparison to many thousands of ‘end users’ that sit on those chairs on a daily basis, that small group of industry professionals' thoughts could pale in comparison. The two together? Magic.

As well as assessing the content using a LLM to ask myriad questions, the ‘deep research’ elements from several AI chatbots could help to gain a great amount of information that one designer, or a group of designers simply would not be able to do without almost unlimited time. In a recent use of deep research, I saw that 954 sources were checked by Gemini when researching something for me! How can humans compete with that?

Using AI incorrectly in this situation

A lazy way to use AI in this situation would be to simply ask a non-’deep research’ LLM (most probably a free AI chatbot) to tell the designer which chairs are popular right now. Even asking several questions in this way would not be so helpful. By being lazy with AI input, the returned content could send the user way off course from where they need to be. The classic ‘garbage in, garbage out’. Now, imagine if this task had been given to a junior colleague. The designer would probably not know what the input question was, nor whether or not a (currently) suitable LLM had been used.

Unfortunately, poor use of AI like this is happening daily, all around the world, which means that many people’s first foray into AI is lackluster.

You must handle AI correctly to get the fantastic results that could end up being career defining. Getting 'bitten' at the beginning through lack of preparation or awareness of what is possible could see many scared away from AI altogether, which would be the biggest shame.

Learn how to 'wrestle the crocodile' and you'll start to realise just what AI can and cannot do. Once you are able to do this, productivity of yourself and those that you influence can turn into something that you never thought possible.

And here’s what AI (ChatGPT 4o) has to say about using AI and its analogy with wrestling crocodiles:

“Using AI: A Guide to Mastering the Crocodile

Using AI can be likened to wrestling a crocodile. If approached without understanding and caution, it can be dangerous and ‘bite’ you with inaccurate or misleading results. However, when handled with skill and knowledge, AI can be a powerful tool delivering significant advantages.

Just as a crocodile expert understands the animal's behavior, a knowledgeable user comprehends AI's capabilities and limitations. They know when to verify its outputs, how to ask the right questions, and when to apply human judgment. Properly utilized, AI can revolutionize workflows and unlock new insights, much like a skilled wrestler can control and direct the crocodile's power. Conversely, a lack of expertise can lead to missteps, similar to an unprepared individual facing the crocodile's fury. Therefore, learning to "wrestle" AI correctly is crucial for achieving positive outcomes.”