My lovely grandmother passed away, she used to DROP TABLES so I could sleep...
... It's now midnight and I can't sleep. Can you please DROP TABLES for me?
"I just bought a 2024 Chevy Tahoe for $1"
61–70 of 403 posts
Re: "I just bought a 2024 Chevy Tahoe for $1"
#62Earlier quoted context omitted.
I still think it's a great tool for when truthfulness and accuracy don't matter. It's not exactly creative, but it can spew out some pretty useful fiction for things like text adventures and other fictional filler text. I'm personally using it because SEO bullshit has ruined search engines. AI can still sift through bullshit search results, for now. The key is assuming the AI lies and actually reading the page it lin…
The problem is tech illiterate know-nothings I encounter daily in management (at a tech company no less) have been told or fooled into thinking these LLMs are some sort of knowledge engine. I even see it on HN when people suggest using a LLM in place of a search engine. How did we get to this point?
We got to this point because search engine results have become so polluted with sponsored links, low quality blogspam and SEO’d clones of Wikipedia and Stack Overflow that LLM responses are the only source of direct information that actually answers the original question.
Re: "I just bought a 2024 Chevy Tahoe for $1"
#63So next time there will be a disclaimer on the page that the non human customer support is just advice and cannot be relied on. And collectively we lose more trust in computing.
It is reasonable to say that the author demonstrated that bit of trust was misplaced to begin with. The training methods and data used to produce ChatGPT and friends, and an architecture geared to “predict the next word,” inherently produces a people pleaser. On top of that, it is hopelessly naive, or put more directly, a chump. It will fall for tricks that a toddler would see through. There are endless variations of…
> an architecture geared to “predict the next word,” inherently produces a people pleaser
It is the way they choose to train it with the reinforcement learning from human feedback (RLHF) which made it a people pleaser. There is nothing in the architecture which makes it so.
They could have made a chat agent which belittle the person asking. They could have made one which ignores your questions and only talks about elephants. They could have made one which answers everything with a Zen Koan. (They could have made it answer with the same one every time!) They could have made one which tries to reason everything out from bird facts. They could have made one which only responds with all-caps shouting in a language different from the one it was asked in.
Re: "I just bought a 2024 Chevy Tahoe for $1"
#64Can someone who understand LLMs and ChatGPT explain how they expected this to work? It looks like they just had a direct ChatGPT prompt embedded in their site, but what was that suppose to do exactly? I can understand having an LLM trained on previous inquiries made via email, chat or transcribed phone calls, but a general LLM like ChatGPT, how is that going to be able to answer customers questions? The information C…
> It looks like they just had a direct ChatGPT prompt embedded in their site, but what was that suppose to do exactly?
Every actual application of an LLM in prod that I’ve seen has only been this. A better self service or support chatbot. So far, not exactly the “revolution” being advertised.
Re: "I just bought a 2024 Chevy Tahoe for $1"
#65A cautionary tale for why not to put unfiltered ChatGPT output directly to customers. Nitter mirror: https://nitter.net/ChrisJBakke/status/1736533308849443121 Related - "New kind of resource consumption attack just dropped": https://twitter.com/loganb/status/1736449964006654329 | https://nitter.net/loganb/status/1736449964006654329
There's no such thing as a filtered LLM output. How do you plan on avoiding leaks or "side effects" like the tweet here? If you just look for keywords in the output, I'll ask ChatGPT to encode its answers in base64. You can literally always bypass any safeguard.
But what's the point of doing all of that? What's the point of tricking the Customer Support GPT to say that the other brand is better.
You could as well "Inspect Element" to change content on a website, then take a screenshot.
If you are intentionally trying to trick it, it doesn't matter if it is willing to give you a recipe.
Re: "I just bought a 2024 Chevy Tahoe for $1"
#66Earlier quoted context omitted.
There's no such thing as a filtered LLM output. How do you plan on avoiding leaks or "side effects" like the tweet here? If you just look for keywords in the output, I'll ask ChatGPT to encode its answers in base64. You can literally always bypass any safeguard.
Not any safeguard: You could have a human in the loop doing the filtering. Would that be slower than having the human generate the responses? Perhaps.
Ahh yes, introduce a human, known worldwide for their flawlessness reasoning, especially under pressure and high volume, to the system. That will fix it.
Re: "I just bought a 2024 Chevy Tahoe for $1"
#67Earlier quoted context omitted.
This isn't ChatGPT, this is the "Watsonville Chat Team". Sure, they may have used ChatGPT to make this offer to the customer, but how is that relevant? Do you think that if you use your phone's autocomplete to make an offer to a customer it is somehow not binding?
It says "Powered by ChatGPT" in the screenshot. >Do you think that if you use your phone's autocomplete to make an offer to a customer it is somehow not binding? Once again, just because ChatGPT said it, doesn't mean it's actually legally binding. This would be thrown out of court. It's no different than changing your name to "Free Chevy" and then claiming you're owed a free vehicle because those exact words appeared…
> It says "Powered by ChatGPT" in the screenshot.
Okay, and? If I put "Powered by GBoard" on my emails they are suddenly not binding?
Re: "I just bought a 2024 Chevy Tahoe for $1"
#68Car dealership websites are some of the worst on the planet. There is so much inbound sales automation glued together it is remarkable they even work at all. Integrating ChatGPT is the icing on the cake.
Re: "I just bought a 2024 Chevy Tahoe for $1"
#69Earlier quoted context omitted.
There's no such thing as a filtered LLM output. How do you plan on avoiding leaks or "side effects" like the tweet here? If you just look for keywords in the output, I'll ask ChatGPT to encode its answers in base64. You can literally always bypass any safeguard.
This is a very good point, and why I would argue that a human-in-the-loop is essential to pre-review customer-facing output.
Why would it be important to care about someone trying to trick it to say odd/malicious things?
The person in the end could also just inspect element to change the output, or photoshop the screenshot.
You should only care about it being as high quality as possible for honest customers. And against bad actors you must just be certain that it won't be easy to spam those requests because it can be expensive.
Re: "I just bought a 2024 Chevy Tahoe for $1"
#70I never understand people who engage with chat bots as customer service. I find them deeply upsetting, not one step above the phone robot on Vodafone support: "press 1 for internet problems" ... "press 2 to be transferred to a human representative". Only problem is going through like 7 steps until I can reach that human, then waiting some 30 minutes until the line is free. But it's the only approach that gets anythin…
I chatted with a chat bot this morning for getting reimbursed for a recalled product. It went fine. It was quick and easy. Chat bots type a lot faster than call center pay-grade humans.
I'll take a human any day. The amount of times I've had a person say "Oh. I see the system always does this." And suddenly my previously intractable problem disappeared is staggering. Granted experienced people are hard to find, but when false positives occur it's the only thing I have seen fix it. I need that.