Hello Reader!

For a little while, Trevor Miles and I turned our tiny corner of the internet back into a place of productive, civil debate.

Remember the early days of social media? No algorithm decided what you saw. You saw what the people you followed posted. And the content had more depth.

Trevor and I brought that back, quite organically.

From an intriguing metaphor to a full debate

I saw a post from Trevor that put AI into Daniel Kahneman’s famous framework of System 1 and System 2. The idea: your brain judges situations in two ways. System 1 is fast, so you escape the saber-toothed tiger. System 2 is slow and analytical. It takes longer, but it is less prone to prejudice and bias, so it can second-guess System 1. (If this intrigues you, read Kahneman’s Thinking, Fast and Slow. It is the original source.)

I shared the post on LinkedIn and then added my view on the role of optimization a few days later. A core feature of optimization models: they never violate a constraint you gave them. That makes them hallucination free. An LLM sometimes ignores an instruction, no matter how often and how prominently you state it. Likely you've all witnessed this already.

Both posts were well received. Then Trevor published a longer article on his blog and shared it on LinkedIn. It gave me a good look at his approach to planning under uncertainty. He contrasted it with optimization, but he presented optimization inaccurately. So I wrote a response.

What writing the response taught me

My article had two missions:

• Correct the claim that optimization is overconfident. Trevor conceded the point after reading my response.
• Describe the modern approach to decision-making under uncertainty. This matters more to me than the overconfidence question.

While writing, something interesting happened. I had read parts of Trevor’s article with System 1. Writing a precise response forced me to read it with System 2. And I found surprises in the details.

Trevor and I come from completely different angles. Yet we both want realistic plans despite uncertainty. And the plans our tools produce look very similar. Only the inner workings differ.

My preferred approach is to tune an optimization formulation and use it as a decision policy in Warren Powell’s sense of Sequential Decision Analytics. This makes Trevor's suggestion possible, but not mandatory: generate lots of training data offline with the tuned formulation and train a neural network on it. Evaluating a neural network cuts online computation time vs solving an optimization problem, he suggests. In many cases, I agree. When pushing the boundary of what's computationally possible, this is a viable approach.

But in my experience, a decent optimization formulation already solves fast enough for real-world decisions. When it is fast enough, I would not add a neural network. It might hallucinate, no matter how carefully you curate the training data.

What did we miss?

The debate has several chapters by now, all linked for you above. Read along, see where we missed interesting tangents, then hit reply: Which question should we tackle next?

Until the next iteration! Tim Varelmann ​Bluebird Optimization​ Complicated Decisions - Simply Automated! Follow me on LinkedIn​

​