Thunderbird: Free Your Inbox
Open discussion:
How are you feeling about AI right now, specifically in terms of generative Large Language Model AI tools (such as #ChatGPT) being used as a work assistant? Could it be beneficial if done properly? Or is it just a privacy nightmare waiting to happen?
It's a divisive topic, and there are bound to be many heated but valid opinions.
Let's talk about it.
(❗No, we don't have any current plans for #AI in #Thunderbird.)
Jason Evangelho
@markmcelroy @BenjaminNelan @thunderbird
Holy cow. I asked ChatGPT to write a 5-paragraph biography of me, and I'd estimate the same percentage of flat-out wrong details.
I wasn't born in 1976, never went to Youngstown State, never wrote for Ars Technica, and never joined Red Hat as a Senior Linux Evangelist.
Although that sounds like a kickass job ?
Mark McElroy
@BenjaminNelan @thunderbird That transparency dimension is a frequently overlooked criteria. Thanks for calling it out.
A ChatGPT-generated biography of me contains about 40% completely made up and inaccurate material. It'd be cool to know where that came from and/or why it was included, especially since the correct facts persist from attempt to attempt, while the falsehoods always vary.
Benjamin
@thunderbird It's scary to think of the unintentional chaos it'll cause when used be people who put too much confidence in it - not realising the limitations it has, the biases it has and what it's actually outputting. *(For example, someone getting AI to code a website that handles personal information. Many examples of insecure AI-generated code floating around.)*
Benjamin
@thunderbird I personally wouldn't mind something like ChatGPT in Thunderbird or even in #ObsidianMD - being able to ask a more natural question of my own data, such as, "What is this email thread about?" or, "Can you give me a summary of last month?" I feel would be incredibly useful.
However, my main contention would be that a solution like this would have to be:
- On-device, no cloud processing
- Transparent, explains where it got that information from
Jason Evangelho
@BenjaminNelan @markmcelroy @thunderbird Well, the facts are much closer to being facts! The birth year is right, the home state is right.
And I'm impressed that it detected my shift to Linux and the resulting content (wrong year, but close).
Also, I LOVE this fictitious rock group I'm in!
Benjamin
@killyourfm @markmcelroy @thunderbird Here it is with the GPT4 model, how'd it do?
Jason Evangelho
@BenjaminNelan @markmcelroy @thunderbird (But I won't admit it until you call me out)
Benjamin
@killyourfm @markmcelroy @thunderbird Wooooooow. ?
Can't wait to employ this in future footnotes: "The information in this sentence was completely fabricated (maybe)."
Jason Evangelho
@BenjaminNelan @markmcelroy @thunderbird I'm really curious too. I am 99.9% positive there are no other Jason Evangelhos on earth.
Benjamin
@killyourfm @markmcelroy @thunderbird I can't wait for *Silent Uprising's *next album. ? Where is it even getting this from I wonder? I thought maybe there's another Jason Evangelho it's getting mixed up with but google doesn't seem to turn up anything.
On the plus side, Bing AI plays it a bit closer to the chest:
Jason Evangelho
@Magess brb have to try this...
Magess :heart_ace:
It's pretty great at producing form letters because those all DO have a set structure and are likely to contain the same words in very similar orders.
You can generate a "statement from the CEO" that's totally usable cause all statements from CEOs are the same vacuous buzz words and contain no information.
Thunderbird: Free Your Inbox
@PictoPirate @jladata It's not just wrong, but CONFIDENTLY wrong.
But you're absolutely right: until YOU can be confident that it's always going to be 100% factual, it's potentially a waste of time.
Magess :heart_ace:
Ultimately that's because it's not a knowledgebase. It's not referencing facts, it's referencing sentence structures. The way they've built ChatGPT is very much to convince you that it's 1) a mind and 2) a knowledgebase. Cause that people will pay for.
You ask it questions (requests for information).
It pauses and produces results that look like a person typing (mimic something with a mind).
But it's just filling in statistically likely words next to each other.
Shaun Dyer
@jladata @thunderbird Do you check the results yourself? I’ve seen loads of examples of it being just plain wrong. If you have to check if it’s right that won’t save time. If you don’t check there’s a risk it’s wrong…
Jindra Lacko
@thunderbird I have used it in "dumb intern" mode, and was pleasantly surprised.
No Skynet, and at present incapable of independent work. Tons of big question marks - starting with privacy (sending my prompt to some API) to legality of training materials (it has scraped Wikipedia and Stack Overflow, both under CC-BY-SA licenses).
But having said that that I have found it very helpful in my line of work, quite worth the $20 subscription price. Mostly as a faster alternative to Google search.
Jason Evangelho
@Magess You were right! I created a couple fake letters about various decisions, and instructed it to use "conversational but confident" language.
The results were honestly pretty solid, and convincing.
GNU social JP is a social network, courtesy of GNU social JP管理人. It runs on GNU social, version 2.0.2-dev, available under the GNU Affero General Public License.
All GNU social JP content and data are available under the Creative Commons Attribution 3.0 license.