The Freezer light nobody noticed for four years

I asked an LLM a simple question about defrosting my freezer. It answered, and then pointed out a small yellow light that had been on, unnoticed, for 4 years.

Defrosting with minimal mopping

I simply wanted to find out if there were any features on my (inherited) freezer that could help me defrost it.

Yes, I know I can just turn it off.

I have heard the tales of appliances out there that offer various levels of help with defrosting. Ideally, I would like to not have to remove frozen produce from the freezer to avoid the risk of them being defrosted completely (and then refreezing). Even if there is a process that takes weeks but is somewhat unattended and safe, I still consider it superior. If I can avoid the puddle of water forming, that would be welcome too. I know how to use a mop, I just choose to do as little mopping as possible.

One question, a couple of photos

Normally, I'd read the manual, but these days I ask an LLM.

On this occasion I, half-assedly, uploaded a snap of my freezer's front (including the control panel). Sure enough, within seconds, the LLM responded.

It discovered that this manufacturer released multiple models with an indistinguishable front panel, so, as instructed, I needed to fetch the model from the nameplate on the inside wall, right beside the bottom drawer. It knew exactly where the nameplate was. Once the model was pinned, the verdict arrived within seconds, and it was disappointing, as this appliance does not have any magic helping with defrosting. I confirmed that later in the manual.

The part I didn't ask about

I was hoping for a different outcome, but I expected this to be the most likely one, given the age of the appliance. The most surprising part for me was not the ability of the LLM to recognise details in an image, its ability to search for the manual or its ability to reason with me.

It was when it identified that one of the control lights was turned on, bringing to my attention a fact that I may have missed. The LLM had already matched the supplied model to the manual in its context window, so it knew that the light indicates a fast freeze being turned on. This is a mode of operation that does not cycle the compressor but keeps it running continuously.

The advice was to have it turned on for hours (up to a day) only to help with quickly freezing the bulk of fresh produce.

Ours was on for 4 years!

The moment I turned the fast-freeze mode off, the room became pleasantly silent.
We must have stopped picking up on that constant compressor buzzing within the first weeks.

Could I have spotted this control light earlier, identified the model, found the manual and checked what the light means?
Yes.

Did my brain happen to register a yellow-coloured light as something urgent enough to move my focus to it?
No.

I never asked about that light. Spotting it without being asked is what I value immensely when I code with these models.

Artificial due diligence

I have noticed that when I work on code with good models, they often find scope for improvement with little or no extra prompting. I think it is mainly anchored in the coding guidelines acknowledging the need for widely defined engineering qualities. It helps when the requirements are clear and existing code isn't a mess either.

In my experience, it often happens unprompted (no pun intended), when decent models capable of long-horizon reasoning are given breathing space (such as "ultra" flows with abundance of agents).

It also helps to clarify in your guidelines that running real experiments is encouraged if it aids the discovery of a correct approach.

I've stopped treating the extras as noise. Sometimes they're the yellow light.

Wait a minute

You might ask: how does the code get to a point where it even needs improvement, especially if the same LLM is generating it? There are limits to how much information can directly influence the output.

The context window is of limited size, and once breached, truncation occurs. Compact your sessions when the moment is right.

Code familiarity is achieved with basic, low-level tools which can bring back erroneous results. I like to maintain documentation that LLMs can use as a guide to the software.

The documentation may be incomplete, outdated or simply wrong. Use LLMs to keep docs complete, up to date and correct. I am now starting to lean towards an approach that reduces the verbosity of the docs too.

Finally, our prompts can also seem crystal clear to us because the full story exists in our heads. They are, however, interpreted outside the context of that story. Too often, a prompt reuses a previous session with similar-sounding but unrelated threads of activity, polluting the reasoning.

All the more reason to welcome a model that looks past the question it was asked.