Artificial Intelligence - ChatGPT

[ReaperKK]

Exactly. I'll give you a great example. During the bidding process the vendor told us we only need about 20 documents to establish a training model which we could then deploy on our entire file repository. We are now in the pilot period and we are testing this AI tool by having it review some public (vendor hasn't been internally cleared to view private side docs) documents. We sent them 30 docs and want the tool to extract a few values from the document. Static values that doesn't require any contextual processing, about as an easy as a layup as you can get. After a few weeks the vendor comes back and asks us to provide 1000 docs and also we need to annotate what data the tool needs to grab, at that point if we're review the data ourselves then we might as well skip the AI step and do it manually.

[cramx3]

Damn, from 20 to 1000 docs to get the AI going is quite a difference in ask 

[MinistroRaven]

Well thinking on training an AI with just 20 docs and that it would make it all smart and savvy was a tad naive. Even GPT-3.5, with its massive 1.5 trillion parameters, sometimes got things totally wrong.

Asking for just 20 docs seems like a huge "uh-oh" moment that nobody really caught onto, right?

Or maybe those guys were too eager to make a sale. Training an AI with just 20 docs sounds like a stretch, especially if the outcome is falling so short of what was initially promised. Hopefully, they'll step up their game and actually deliver what you were hoping for.

[cramx3]

Depends on the docs I'd say, if there's a lot of variety than yeah, I'd say 20 isn't much and givent hey want 1000 now, I'd say those docs probably were too much for a 20 count to really work. 

It certainly does sound like an over achieving sales pitch.

[PMSummer]

From what I've gathered, it seems like the AI isn't quite as advanced as they're trying to make it out to be. There's been talk that the company that owns ChatGPT even acknowledges that it can often just comes up with things that aren't accurate.

I haven't actually given ChatGPT a shot myself, but for those who tried I'm curious if it can truly pass as a real human in conversations. You know, there's always been this 'uncanny valley' feeling with chatbots – they get close to sounding real, but there's still that something missing that makes them feel a bit off.

[XJDenton]

I asked ChatGPT if the answers it gives can be trusted, and it said no. So yeah, if you have an application that relies on factual accuracy rather than generating something wholecloth, steer away.

[ReaperKK]

Depends on the docs I'd say, if there's a lot of variety than yeah, I'd say 20 isn't much and givent hey want 1000 now, I'd say those docs probably were too much for a 20 count to really work. 

It certainly does sound like an over achieving sales pitch.

Our team and I did think it was a bit ambitious but that said the ask for subsequent annotations was ridiculous. There are a bunch of other issues with the tool itself that are pretty disappointing. For example even if you train a set of documents the AI tool will still have instances where itll capture incorrect data and then the next set of docs will stray further from the established ground truth.

To top it off the engineers from the vendor are MIT grads whose response to these short comings are "don't worry it'll work in prod, source: trust me bro"

[TheHoveringSojourn808]

From what I've gathered, it seems like the AI isn't quite as advanced as they're trying to make it out to be. There's been talk that the company that owns ChatGPT even acknowledges that it can often just comes up with things that aren't accurate.

I haven't actually given ChatGPT a shot myself, but for those who tried I'm curious if it can truly pass as a real human in conversations. You know, there's always been this 'uncanny valley' feeling with chatbots – they get close to sounding real, but there's still that something missing that makes them feel a bit off.

yeah, i've seen those discussions too. it's like they're hyping up the AI, but it's not all smooth sailing. i guess the company is honest about its limitations, which is something. i haven't given ChatGPT a chance yet personally, but i'm intrigued by whether it can nail that human vibe. you know, chatbots often get close, but that missing piece keeps it a tad weird. anyway, if anyone's tested it, spill the beans!

[Lonk]

I haven't actually given ChatGPT a shot myself, but for those who tried I'm curious if it can truly pass as a real human in conversations. You know, there's always been this 'uncanny valley' feeling with chatbots – they get close to sounding real, but there's still that something missing that makes them feel a bit off.

I wouldn't say it can be passed as a human conversation, but the responses you get are pretty good. Think of it as a much more advanced siri, or goggle assistant type of tool that you give it a prompt, and it gives you an answer. I do think it does takes us a step closer to where we want to be with AI.

[jammindude]

From what I've gathered, it seems like the AI isn't quite as advanced as they're trying to make it out to be. There's been talk that the company that owns ChatGPT even acknowledges that it can often just comes up with things that aren't accurate.

I haven't actually given ChatGPT a shot myself, but for those who tried I'm curious if it can truly pass as a real human in conversations. You know, there's always been this 'uncanny valley' feeling with chatbots – they get close to sounding real, but there's still that something missing that makes them feel a bit off.

I would definitely attest to the fact that the free version available to the public, which I use commonly, is an excellent “assistant”, but it absolutely does not meet the human standard at all.

That being said, I’ve been hearing from the inside of the industry that the pay version is miles beyond the free version, and it's made more than a few people stop and think that the government is most likely already using something that is even beyond that.  So I’m more than a bit curious as to what AI’s actual capabilities are.

[PMSummer]

Thanks you for shedding some light on it and I don't even want to know what the government is up to with AI!

I'm thinking of asking one of my more techy buddies to show me the ropes with chatgpt. Seems like an interesting tool to explore from everything I've been reading.

[TheHoveringSojourn808]

Thanks you for shedding some light on it and I don't even want to know what the government is up to with AI!

I'm thinking of asking one of my more techy buddies to show me the ropes with chatgpt. Seems like an interesting tool to explore from everything I've been reading.

once you've got the know-how, could ya hook me up with a walkthrough? i'd totally appreciate it, thanks a bunch!

[Skeever]

Interestingly enough, this has come up with a few friends of mine lately. I discovered recently that one of my cohorts has been using chat GPT to write all kind of things that he does not feel like writing himself. I've tried to tell him that it feels kind of weird to see messages from him that weren't actually written by him, but he's delighted to be able to feed an AI a couple bullet points and have it spit out something that resembles actual human writing. A couple times I've pointed out how similar a lot of his chat GPT compositions feel after a while, but he remains undeterred. Instead of abandoning the project, he just makes small tweaks to address some of the things that myself and others flagged up.

How would you all feel if you started receiving emails from a boss or a co-worker that were composed using AI? I still feel like it's kind of a slap in the face, but my company does not have an explicit policy against doing it. Moreover, I use some tools like voice to text for convenience myself. At the end of the day, I feel like the technology will only get better the more people earnestly try to embrace it, even if there are some awkward kinks along the way.

[TheHoveringSojourn808]

How would you all feel if you started receiving emails from a boss or a co-worker that were composed using AI?

there's something about personal touch that might be lost with AI-generated emails. the way someone writes an email can give you a sense of their personality and style, and that could be diminished if everything starts sounding like it's coming from a machine. i think there's also the concern that if AI starts doing all the emailing, it might lead to less human interaction, and that's something i'd be a bit wary of personally

i guess when you think about it overall, most email is spam, and most people don't read most emails, so at least people are wasting less time writing them now 

[jammindude]

Personally, as someone who uses it frequently, I would never use it in an “inter office” situation. If I’m “Ben the coworker” then I am just that.

But the minute that I have to be the professional “face of the city” or in any other situation where I need a level of formality that I am unaccustomed to, I will turn to ChatGPT every time.

I might be inclined to tell your friend that the more you put into it, the better stuff you get out.  For example, I’ve recently had to write a formal letter to a judge. This is a theater that I am *completely* lost in, and I was tempted to hire a paralegal outright to help me write the letter. But then I decided to just treat ChatGPT like my paralegal. I wrote out at least 5-6 paragraphs of information explaining the entire background from start to finish of everything that I’m going through and even explained to “Chad” (my nickname for it) that I obviously didn’t want all this information in my letter, but that I needed the entity that was going to prepare this letter to understand the entire backstory that got us to the point of needing the letter in the first place.

The letter I got was brief, said exactly what I needed it to say, it was formal and professional but addressed my personal concerns as well, and when I showed it to a couple of people I know with legal experience they were pretty shocked at how nearly perfect the letter was.

So TLDR: you do get better results by giving it more to go on. If your friend is only feeding it “bullet points”, his results aren’t going to be as good.

[MinistroRaven]

When you don’t know what to ask chatgpt to get the result you want, try using https://promptperfect.jina.ai/ you tell it what you need he will perfect the prompt your results are 100% better

[MinistroRaven]

Tomorrow out company will get on a business call with the devs of jasper.ai as we plan to integrate it in our everyday tasks