Comment by spywaregorilla | Hacker News Reader

Comment by spywaregorilla | Hacker News Reader

A few things stand out to me here.

* Most notably, the ai bot is repeating itself and serving very standard looking messages. I don't think this is an "AI" chatbot in the sense of an LLM. It's just a dialogue tree with some parsing.

* The author's communication is... bad. Don't say "connect to support rep". Say "talk to human".

* It's very silly to state the following to a dumb chat bot: > Hello! I just received and installed the new AT&T router/fiber modem, and ... the Internet speed is just as slow as before. I pay for 1 Gbps symmetric, and I'm getting 8 Mbps down and 6 Mbps up. On 6/28, the average connection speed went from 1 Gbps down to 100 Mbps. On 7/8 the average speed went from 100 Mbps to 8 Mbps. This is all measured both on the device at the fiber, and through a separate monitor I have wired into the 1 Gbps network.

I would recommend "slow internet".

Also saying "Slow internet, not slow wifi" is probably causing the bot to believe you're asking about wifi specifically because it's not an "AI" bot in the trendy sense of the word and it just sees the word wifi.

> It's very silly to state the following to a dumb chat bot

It was the first message. We don't know if the author knew the decision tree will come first. Usually the initial prompt is something like "please describe the issue".

Thought the same. It's kind of infuriating to see people bashing SoTA LLMs and AI while referring to chatbots from previous era using a completely different types of technology.

And it's not like customer service of the same service providers without this tech was appreciated by people before these chatbots. I kind of imagine someone like OP bashing the human as well for incompetence if there wasn't a chatbot.

> And it's not like customer service of the same service providers without this tech was appreciated by people before these chatbots.

This is just "but humans also".

I wish our industry's collective standards didn't allow for just replacing one subpar experience with a very similar one that has less empathy and is lots cheaper.

What would your ideal proposed solution be like?

If you owned a telecom company that provided internet services and mobile plans for everyday people?

I would have thought that was obvious from my comment.

I'm with Jeff on this one.

LLMs are an obviously terrible solution for any problem that actually has a complex but invariant solution.

Personally I prefer rigid, automated questionnaire filtering that gets you to a real person as soon as possible.

I also thoroughly believe in a) human support chats and b) believing the people answering them to be decent human beings who deserve my clarity, preparedness, politeness and empathy, and I think more should be done on chat interfaces to make them slightly less amenable to stupid unstructured queries. You help people ask the question right.

But what Jeff was facing wasn't an LLM. As far as I saw it was pattern matching with hardcoded responses. Hence it had a hardcoded response about WiFi, after saying "Not WiFi". ChatGPT etc wouldn't make such a mistake.

And the final human also didn't understand the request correctly.

ChatGPT wouldn't respond with "I see you have a question about your Wi-Fi" after "Slow Internet, not slow WiFi".

I think this pattern matching was more like "automated questionnaire filtering" that you were suggesting. It's just it's very hard to do this filtering, to account for differences between "not WiFi" vs "WiFi", while it's easy with LLMs.

That was the message I sent to the support agent after the chatbot finally got out of the way.

Can you elaborate on why you refer to this as an AI chatbot?

Because the industry currently refers to machine learning as AI, so rather than fight that battle, I classify decision tree ML as AI since it's pretty closely related to LLMs and other 'AI' things that suddenly converted from ML to AI once enough money was pumped in.

I would argue that's pretty misleading. Especially if you're saying "AI chatbots are the future". The tech you're discussing here has been commonplace for the past decade and has nothing to do with current trends or the future.

Folks here are talking about _dialogue_ trees not ML _decision_ trees. These are very likely manually coded. It's conceivable but not very likely that there's a simple classifier model behind the scenes here predicting which line of pre written dialogue to serve at each junction. If you write AI chatbots your readers are going to expect that it's a chatbot that actually generates the text, LLM or otherwise.

idk man. feels clickbaity.

Reply on news.ycombinator.com