A two-hour walking tour with ChatGPT: 'I'd not suggest that to my worst enemy' | Hacker News Reader

A two-hour walking tour with ChatGPT: 'I'd not suggest that to my worst enemy'

theguardian.com

11 pointsEvgeniyZh14 comments

This author clearly knows nothing about the way an LLM works and did not bother to do any actual research either.

Articles like this just make me go shrug, ok, you and those that believe you can just go on thinking that it's nothing but hype. More benefit for those that do understand what they are good for and what not.

Anything article that uses GPT 3.5 should be automatically disregarded. They couldn’t pay $20 for the much more superior GPT 4?

"I want ChatGPT to know it’s not OK to casually seed lies to maintain an aura of omniscience. But telling off an app for fibbing feels unhinged and demeaningly human. So all I manage to do is contribute free labour to ChatGPT by refining its data on Sydney’s public transport options. Given my long tenure as an unpaid content creator for Meta, that’s probably the least of my big tech complicity."

Internet journalist trashing generative AI with less research done than GPT3.5.

I built a GPT-4 powered walking tour app during a hackathon. It uses the Wikipedia and Google Maps APIs to find interesting facts about things you're walking by and generates scripts to read aloud. It's fun and I learned a lot about places I go to all the time.

Unfortunately all the APIs are pretty costly so I can't post it for the group. Now that OpenAI's TTS and GPT-4 Turbo is out, I could probably revisit it. Still, not sure if people would be willing to pay enough for something like this.

Older models had little understanding of geography. Do you know if that has been addressed?

You don’t need AI for that.

https://apps.apple.com/app/hewell-personal-tour-guide/id1535...

You appear to be missing the point. The point is to play with GPT in an environment where we have good baselines.

No, the point is that this person built something useful and expensive to run to the degree they didn’t publish it. Which for anyone but the author makes it as meaningful (meaningless) as it not existing. Yet without AI you can have the same thing, that everyone can use, at zero recurring cost.

That's your point, not theirs.

You can start your own thread about how we could do all this stuff without LLMs in your opinion.

> That's your point, not theirs.

That’s how conversations work. Someone makes a point then you reply with another point which responds to the original.

> You can start your own thread about how we could do all this stuff without LLMs in your opinion.

Or I could reply in context, in the place that makes sense. This has a real, existing, relevant counterpoint example which the original commenter may not have realised was an option.

While ChatGPT has obvious limits and flaws, the conclusion that it is a robotic tour without any unexpected detours seems unfair.

It is also something I barely ever came across in traditional itineraries by humans on the various websites I used for our most recent London trip.

No, it was basically (just like here) popularity rankings tuned for good SEO rankings. Even personal blogs barely ever surprised me where exciting detours were spots such as the tourist infested “unknown” Neal’s Yard. I got so many Needles, so many Sky Gardens. Discovering the ticket-free option The Garden at 120 took much more prodding.

The flaws and lies from GPT 3.5 are common, but our unfair bias against AI equally so. We often seem to place these services on a higher pedestal than ourselves.

Oh and please don’t use budget models for articles in high-profile newspapers and draw conclusions on AI from those. Also, remember they follow your commands; if you want lesser known and unexpected spots, be sure to say so!

Can confirm, would only suggest such things to my worst enemy.

In trying to push chat GPT to its limits, I aimed to build a website from scratch, with no prior experience.

I very very quickly gave up on trusting its output and instead had 2 weeks of learning everything via YouTube and docs. ChatGPT got relegated to being a better Wikipedia; If I wanted to understand why things were they were, ChatGPT was faster.

I also have copilot - in the cases where I watched an LLM help, it’s invariably an experienced individual selecting the useful parts from the absolutely incorrect.

You need the domain knowledge to make the right calls, the model is more than happy to let you build an ocean liner in a desert.

"While it's had its moments of inaccuracy, this icy retraction of a straight-up lie that's been presented with the same veneer of competency as the correct parts of my itinerary is a little chilling. It's just a train station, no one's lost an eye, but embedding lies within truths whiffs of gaslighting."

Reply on news.ycombinator.com