ChatGPT just got destroyed at chess by a 46-year-old Atari 2600 console.

And as someone who owned that exact console as a kid, I find this absolutely hilarious.

My Atari 2600 was a hand-me-down from my cousin. And even as a kid in the early 1990s, I could see it wasn’t great.

By today’s standards, though, it seems far, far worse.

The Atari 2600 has about 1/250,000 the processing power of an iPhone 15 Pro. By comparison, ChatGPT runs on data centres worth hundreds of millions of dollars.

Yet, when a software developer recently pitted them against each other at Video Chess (beginner level), ChatGPT got “absolutely wrecked.”​

ChatGPT “confused rooks for bishops, missed pawn forks, and repeatedly lost track of where pieces were on the board.”

And here’s the best part: ChatGPT was the one who suggested the match in the first place. It confidently claimed it would “easily beat” the ancient Atari.

Now, I’m not trading in my ChatGPT subscription for my old Atari anytime soon. Yet, this embarrassing loss highlights three critical lessons for data scientists:

1. Match your solution to the problem, not the problem to your solution

The Atari won because its solution was purpose-built for chess - a search problem requiring evaluation of decision trees and outcomes.

ChatGPT is a token prediction model. Using it to solve a search problem is like using a screwdriver to hammer in a nail. It might work eventually, but it’s the wrong tool.

How often do you see data scientists force deep learning solutions onto problems that a simple regression would solve better?

Or build elaborate pipelines when a spreadsheet would suffice?

Start with the problem. Then pick your tool.

2. Complex doesn’t mean better

All those data centres. All that processing power. All that sophistication.

None of it mattered against a simple, well-designed solution.

Your stakeholders don’t care how sophisticated your model is. They care whether it works and whether they can trust it.

Sometimes the “boring” solution is the right one.

3. Know when uncertainty matters (and when it doesn’t)

LLMs are built to handle uncertainty - they’re trained on patterns and probabilities across massive datasets.

But chess is a game of pure skill with deterministic rules. There’s no randomness to model. Introducing probabilistic thinking to a deterministic problem just creates unnecessary noise.

The same applies to your work: not every problem needs a probabilistic model.

If you’re dealing with fixed business rules or deterministic processes, model them as such.

Adding uncertainty where none exists doesn’t make you sophisticated. It makes you wrong.

---

In summary

Before you reach for the latest, most powerful tool in your arsenal, ask yourself:

• What type of problem am I actually solving?
• What’s the simplest solution that would work?
• Am I introducing complexity to serve the solution, or just my ego?

Sometimes the best answer is the one that’s been sitting around, gathering dust since 1979.

Talk again soon,

Dr Genevieve Hayes