Hacker News Reader

Comment by almost | Hacker News Reader

LLMs are really good when your primary goal is just to fob people off. So customer support when you don't care about your customers etc. I can imagine it's very appealing when you just want to come up with semi-plausable reasons to deny insurance claims. Who cares if it's wrong? They might give up anyway and you've saved paying out!

Not saying they're not good for other things, but the "waste people's time and make them go away" use is just so perfect for the capabilities of even cheap models.

View on HN

Believe it or not, that's not even the topic of the article.

The article is entirely about the AI doing the first pass of the claim intake and causing a lot more work for everyone involved instead. Mis-routing, mis-categorization, hallucinating details, etc.

Anybody that has worked on a phone exchange knows they already do this sort of thing. Some customers just never get through to a person. This is by design.

Can confirm. Until I worked on an honest phone system tuned to actually get people connected with call centre reps, I never quite had a feel for how malicious many contact funnels have become. A good one runs like a Swiss watch, and is trivially tunable to sub-5 to 10 minutes wait times. Anything longer is a conscious choice on the company's execs part. It doesn't "just happen", and to the degree it does, it shouldn't be happening on the regular. Not if you're watching your metrics and tuning appropriately.

On the flip-side, LLM's can be pretty good at coming up with almost water-tight stories for the insurance claim!

I read and article on here some months ago titled something like "Leftist usecases for LLMs" and one of them was negotiating bureaucracies, and that reminds me, I need to call back JCP&L about my Board of Public Utilities complaint that an LLM helped me out together.

You see, I live with in a rural area and especially since LED light bulbs have become the thing, it's been clear that we are delivered far less than 120v service a lot of the time. We live about 200 yards from the transformer that we share with several other houses, too many I believe. About two years ago I set up Home Assistant and bought one of those power monitoring rigs that you put on each circuit in your breaker box, mostly I just wanted data about where my power bill money was going. I work in FinOps in case anyone was curious.

Anyhoo, couple years later I'm analyzing several months worth of power bills vs the circuit meter readings to quantify where I can make some upgrades and save some money. After a day or so of this I mention to Claude about the voltage issues and sure enough, it puts together a whole packet for me documenting brownouts down to 95v and routine brownouts below 110v. It helps me fill out the BPU complaint form.

2 days later I get a call from JCP&L (unfortunately to my wife's phone because it's her name as the primary on the account). I want a new, larger transformer for myself and my neighbors, so I need to call them back today.

If you were monitoring the power coming into your house, why did you need claude to "put together a whole packet documenting brownouts"? Shouldn't you just have the raw data, maybe highlighted with the offending parts? Or was the prose of your complaint the issue?

Frequently you have to mention specific laws and describe how your situation matches the law or your complaints get ignored. At least this is how it works when you call your insurance company and make similar complaints.

There is still a lot of work to take the raw data and put it into a useful format. LLMs are really good at that (when they don't hallucinate data).

This. Putting together a complaint in the language of a electricity service professional(?) with proper citations etc.

In short, a rock solid data analysis artifact takes a lot longer than just running the queries.

How big is the transformer and how many residences are sharing it? It’s probably stencil painted on the enclosure.

I’ve seen half a dozen houses on a single 7.2kV -> 240V 25kva transformer which can lead to problems given the available current on the secondary side is only 104A. As the transformer overheats you get voltage drop, as you’ve noticed ;)

Your supply voltage should be +/- 5% of nominal, 115-125V line to neutral and 230-250V line to line, appliances are generally OK with +/- 10% from nominal voltage.

I don't know "how big", but I do know that it's a 7.2kV line, and that there are 6 houses on it, and that mine is pretty far away while the other houses are quite close to the transformer.

Like I said, it routinely drops to or below 100v. It's most noticeable after an outage is restored when everyone's hot water heaters and fridges kick back on at once.

Reply on news.ycombinator.com