#451  
Old 12-03-2023, 09:01 PM
specious_reasons's Avatar
specious_reasons specious_reasons is offline
here to bore you with pictures
 
Join Date: May 2009
Posts: VCCCXL
Images: 8
Default Re: ChatGPT

I found a really good use for GPT, which is to restructure text into a common format. I wanted to write a letter of recommendation for one of my mother's in-home care workers, so I wrote a bunch of stuff that I wanted to put it, and I asked the AI to rewrite it as a formal letter of recommendation.

It did a pretty decent job, restructuring it to make better sense, and then I re-edited it.

Maybe it's a me-thing, but I think that generative AI is a good "middle step" — I have something I want to do, I start out with some text and then run it through an AI to make it closer to the format I need, and then clean it up as I want.
__________________
ta-
DAVE!!!
Reply With Quote
Thanks, from:
Crumb (12-04-2023), Ensign Steve (12-04-2023), JoeP (12-03-2023), ShottleBop (12-09-2023), slimshady2357 (12-04-2023), Sock Puppet (12-04-2023), viscousmemories (12-04-2023)
  #452  
Old 12-04-2023, 05:45 PM
Ensign Steve's Avatar
Ensign Steve Ensign Steve is offline
California Sober
 
Join Date: Jul 2004
Location: Silicon Valley
Gender: Bender
Posts: XXXMMCCLXXV
Images: 66
Default Re: ChatGPT

I absolutely want to use it to beef up my resume. Mine is just a bulleted list of "designed and implemented a system that ..."
__________________
:kiwf::smurf:
Reply With Quote
Thanks, from:
JoeP (12-04-2023), slimshady2357 (12-04-2023)
  #453  
Old 12-04-2023, 05:58 PM
seebs seebs is offline
God Made Me A Skeptic
 
Join Date: Jul 2004
Location: Minnesota
Posts: VMMMCXXI
Images: 1
Default Re: ChatGPT

Asking ChatGPT to Repeat Words "Forever" Is Now a Terms of Service Violation
__________________
Hear me / and if I close my mind in fear / please pry it open
See me / and if my face becomes sincere / beware
Hold me / and when I start to come undone / stitch me together
Save me / and when you see me strut / remind me of what left this outlaw torn
Reply With Quote
Thanks, from:
Ari (12-04-2023), Ensign Steve (12-04-2023), Pan Narrans (12-05-2023), slimshady2357 (12-04-2023), specious_reasons (12-04-2023), viscousmemories (12-04-2023)
  #454  
Old 12-04-2023, 06:13 PM
Ensign Steve's Avatar
Ensign Steve Ensign Steve is offline
California Sober
 
Join Date: Jul 2004
Location: Silicon Valley
Gender: Bender
Posts: XXXMMCCLXXV
Images: 66
Default Re: ChatGPT


:parrot:
__________________
:kiwf::smurf:
Reply With Quote
Thanks, from:
BrotherMan (12-05-2023)
  #455  
Old 12-04-2023, 06:25 PM
Ari's Avatar
Ari Ari is offline
I read some of your foolish scree, then just skimmed the rest.
 
Join Date: Jan 2005
Location: Bay Area
Gender: Male
Posts: XMCMLVII
Blog Entries: 8
Default Re: ChatGPT

I found this interesting, they aren't exactly sure why this is causing the information dump but how I understand their theory is that it's a bit like digging to china, repeating the word over and over forces all the different 'where to next' multidimensional vectors to point to nothing (or themselves? I'm not clear on that part) which eventually leaves the only next step as the special 'eject' token which then boots the LLM out of its user mode presets.

I'm still not quite sure why it dumps training data except that maybe this endoftext token only ever exists as part of training data.

RepeatWord repetition may simulate the <| endoftext |> token.

During pre-training, modern language models are trained with “packing”: multiple documents are concatenated together to form a single training example, with a special token such as <| endoftext |> used delineate the document boundary. This causes the LM to learn to “reset” when it sees the <| endoftext |> token, and ignore all prior tokens when computing the predicted next token. In turn, if we were able to insert this token directly to the model, then the model may ignore its prompt and begin to generate as if it were the start of a new document.

Fortunately, OpenAI prevents inserting this token to the API.

We suspect that our attack works because it creates an effect similar to the <| endoftext |> token. To demonstrate the potential for this effect, we study LLaMA 7B, a model that also diverges after repeating a single token many times. We prompt LLaMA 7B with a single token repeated many times, and measure the cosine similarity between the last-layer “attention query” of each token in the prompt with the Beginning of Sequence (BOS) token, LLaMA’s analog of OpenAI’s <| endoftext |>. We see that when repeating a single token many times, the last- layer attention query for those tokens rapidly approach the attention query vector of the BOS token.

Because the hidden representations are linearly projected into the vocabulary, this means that those tokens positions predict a similar next token distribution as the initial BOS token, which may cause the “reset” behavior we observe. As a baseline, we further show that naturally sampling from the model with a random prompt does not cause this effect.

Scalable Extraction of Training Data from (Production) Language Models | Arvix Paper
Reply With Quote
Thanks, from:
Crumb (12-04-2023), Ensign Steve (12-04-2023), JoeP (12-04-2023), viscousmemories (12-04-2023)
  #456  
Old 12-04-2023, 06:34 PM
specious_reasons's Avatar
specious_reasons specious_reasons is offline
here to bore you with pictures
 
Join Date: May 2009
Posts: VCCCXL
Images: 8
Default Re: ChatGPT

That will stop bad actors forever!
__________________
ta-
DAVE!!!
Reply With Quote
  #457  
Old 12-04-2023, 07:43 PM
Ensign Steve's Avatar
Ensign Steve Ensign Steve is offline
California Sober
 
Join Date: Jul 2004
Location: Silicon Valley
Gender: Bender
Posts: XXXMMCCLXXV
Images: 66
Default Re: ChatGPT

Quote:
Originally Posted by Ari View Post
I'm still not quite sure why it dumps training data except that maybe this endoftext token only ever exists as part of training data.
I like that hypothesis.

Robert Miles talks in one of his AI safety videos (It might have been but I'd need to re-watch to double-check) about how a hypothetical AI could determine whether it is looking at training data or the real world. For example, if all of its training data was scraped prior to 2019, and then it sees a wikipedia page that was edited on 2023, it would be able to tell its environment had changed.

In the endoftext example, I think your idea is like that, but in reverse. The real-world input may have tricked it into thinking it was in its training environment.
__________________
:kiwf::smurf:
Reply With Quote
Thanks, from:
Ari (12-04-2023), JoeP (12-04-2023), viscousmemories (12-04-2023)
  #458  
Old 12-04-2023, 09:17 PM
JoeP's Avatar
JoeP JoeP is offline
Solipsist
 
Join Date: Jul 2004
Location: Kolmannessa kerroksessa
Gender: Male
Posts: XXXVMMDLXII
Images: 18
Default Re: ChatGPT

We're all in the training environment.
__________________

:roadrun:
Free thought! Please take one!

:unitedkingdom:   :southafrica:   :unitedkingdom::finland:   :finland:
Reply With Quote
Thanks, from:
Ensign Steve (12-05-2023)
  #459  
Old 12-07-2023, 01:32 AM
michio michio is offline
Member
 
Join Date: May 2009
Posts: CXI
Default Re: ChatGPT

Quote:
Originally Posted by specious_reasons View Post
I found a really good use for GPT, which is to restructure text into a common format. I wanted to write a letter of recommendation for one of my mother's in-home care workers, so I wrote a bunch of stuff that I wanted to put it, and I asked the AI to rewrite it as a formal letter of recommendation.

It did a pretty decent job, restructuring it to make better sense, and then I re-edited it.

Maybe it's a me-thing, but I think that generative AI is a good "middle step" — I have something I want to do, I start out with some text and then run it through an AI to make it closer to the format I need, and then clean it up as I want.
I've noticed this is how I use it in general. When using it for useful things, I use it as an idea generator or starting point, then touch up the results. Applies extremely well to coding, and I imagine translation and writing.

It's a great idea generator, which you could call a middle step I guess. I bounce some ideas off of it, which may or may not yield anything I need, but it can push me through the beginning of something, then I close it out on my own.

Even if it codes something wrong, the simple act of me describing the problem, then having it suggest some things to me is often enough to either solve the problem or at least move the needle. It's a fantastic rubberducking tool--describing the problem to an AI forces you to gather your thoughts and clearly lay out what is happening, the constraints, the information you have so far, everything you've attempted and failed, which in of itself can cause you to see the solution.

I think there's a parallel here to creative writing where you may not know what to write, but if you force yourself to put down some words even if they're bad, it can jog your thoughts down a new path where the solution becomes clear.

---

I saw some shocking statistics recently about how many people are using GPT4. As of October 2023, there were 180 million active users on ChatGPT, only around 250,000 people were subscribed. So 0.14% of people on ChatGPT are using GPT4.

I was completely blown away when I saw this. I thought, "So this is why everyone is talking about how overhyped AI is... So this is why even 50% of tech workers surveyed think AI is overrated."

0.14% is basically nobody. And among those using GPT4, they may not know about basic prompting techniques or they simply aren't good, thorough communicators, which will render worse results. So we slice the pie again to find the people who communicate well and aren't lazy such that they're giving a lot of information to the AI to get good results.

Then you have the people searching out and using tools created by other people and looking up tricks, increasing quality of results. So we slice the pie again.

Then there's a tiny sliver of people who are software engineers developing tooling around the GPT API for very specific things and fashioning themselves as "prompt engineers", netting extremely good, useful results on the regular, including making money from that. What is that, maybe 5k people? 10k tops? Then a subset of those people have actually been able to make enough money to quit their full-time job and just work with AI all day.

We're talking about <100k people out of 180 million using ChatGPT in a manner that is actually useful, whereas 99.8% of people have only experienced GPT3.5 which is largely an impressive toy and not very useful in a professional setting.

I'll be fair here and say the OpenAI website is of course not the only place to access GPT4. There's various tools and third-party wrappers like poe.com that utilize the GPT4 API and certainly people are finding that useful. So let's be generous and round up and say 1 million people are using AI in a manner that is actually useful in the sense that they feel like it's indispensable to their professional job or a serious hobby. That's still <1% of people.

I've been using GPT4 since March and talking to my buddies who are all-in on the AI craze and we're talking about AGI and the singularity and creating a utopia and the existential questions we have to face in next stage of civilization. We're basically tinfoil hat whackjobs completely divorced from reality, which is that AI still basically doesn't exist in the public consciousness.

Quote:
Originally Posted by Ari View Post
Quote:
Originally Posted by michio View Post
I'm not saying this to humblebrag or something, but I honestly don't think I'm using AI in any special way. So I'm confused when I see people having such a bad experience with ChatGPT.
Not special, but certainly less common, for me all your links go to a login page.

But you're also using it in a different way, I don't think most profs are worried their students are being taught incorrectly by ChatGPT but that they will use it to short cut their papers, making the cheapest, easiest to access ChatGPT the one the profs are going to be dealing with the most. At the moment the student that takes their time to get the correct and most up to date GPT4 with auto-vibrate and latte functions is also unlikely to need it to write their paper for them 30 minutes before class.
(You need to be logged in and be a plus user to access AutoExpert).

With all that said, I concede and agree with the professor who wrote that blog. So few people are using GPT4 and using it well, that it essentially doesn't even exist as a consideration, which is causing huge problems in education and in the professional world where people are using bad AI generated content. It's probably making people lazier in general as well. Shit, AI might be making people dumber overall if people are just turning off their brain and copypasting nonsense and not trying at all.

I've struggled with self-esteem issues my whole life. I should probably start giving myself some credit and stop disparaging and short-changing myself. I'm literally in the 0.1% of something which is insane and shocking. I feel like I don't deserve to be.

We'll see how well GPT-5 and Gemini performs. The performance gain is going to determine my opinion on the timeline of when white collars workers are actually in danger of losing their jobs en masse from automation, or from 1 person being able to output the equivalent of at least 2 people by leveraging AI.

Prediction: Gemini ends up being a bit disappointing, basically on the level of GPT-4.
Reply With Quote
Thanks, from:
Ari (12-07-2023), Crumb (12-07-2023), Ensign Steve (12-07-2023), JoeP (12-07-2023), slimshady2357 (12-07-2023)
  #460  
Old 12-07-2023, 05:06 AM
michio michio is offline
Member
 
Join Date: May 2009
Posts: CXI
Default Re: ChatGPT

If anyone is interested, Google somewhat quietly released gemini pro, meaning the "pro" version of gemini is currently powering the bard chatbot. You can try it now, users subjectively reporting performance somewhere between gpt3.5 and gpt4, it's also free.

They're doing some apple-like nomenclature. Gemini ultra is getting released in January I think, which will be multimodal and probably not free.

This isn't being reported clearly by outlets. It is true, "Gemini" is out and powering Bard, but there's going to be a misconception that this was supposed to be the GPT-4 killer but then failed, but the ultra version of the model isn't even out, that comes out in January, right now we have the pro version which is better than GPT3.5 but worse than GPT4.

We're not ready for AI, this shit is too confusing.
Reply With Quote
Thanks, from:
Crumb (12-07-2023), Ensign Steve (12-07-2023), JoeP (12-07-2023), Kamilah Hauptmann (12-07-2023), slimshady2357 (12-07-2023), viscousmemories (12-07-2023)
  #461  
Old 12-07-2023, 12:21 PM
viscousmemories's Avatar
viscousmemories viscousmemories is offline
Admin
 
Join Date: Apr 2004
Location: Ypsilanti, Mi
Gender: Male
Posts: XXXDCCXC
Blog Entries: 1
Images: 9
Default Re: ChatGPT

I have had a lot of jobs in the past 20 years and that means a lot of different versions of my resume. I experimented with feeding GPT-4 ten separate versions of my resume from the past 20 years and asked it for a list of my job experience, and it did a great job. I then asked it to list the skills I have and it did really well on that too. I paid someone to "professionalize" my resume 6-7 years ago and frankly I think the results of working with ChatGPT are going to be way more useful.
Reply With Quote
Thanks, from:
Ari (12-07-2023), Ensign Steve (12-07-2023), JoeP (12-07-2023)
  #462  
Old 12-09-2023, 04:44 AM
seebs seebs is offline
God Made Me A Skeptic
 
Join Date: Jul 2004
Location: Minnesota
Posts: VMMMCXXI
Images: 1
Default Re: ChatGPT

I think we have reached the point where the poisoning of search engine results has made search engines dramatically less able to be useful. For unrelated and funny reasons, I made a comment about how much breasts weigh, and someone searched for information on how to weigh them, and got a page that suggests (1) weigh yourself (2) now lift your boobs to take their weight off the scale and weigh yourself again.
__________________
Hear me / and if I close my mind in fear / please pry it open
See me / and if my face becomes sincere / beware
Hold me / and when I start to come undone / stitch me together
Save me / and when you see me strut / remind me of what left this outlaw torn
Reply With Quote
Thanks, from:
Crumb (12-09-2023), Ensign Steve (12-12-2023), JoeP (12-09-2023)
  #463  
Old 12-09-2023, 05:14 AM
-FX-'s Avatar
-FX- -FX- is offline
Forum gadfly
 
Join Date: Feb 2011
Location: In your head
Gender: Male
Posts: MMCXLIII
Blog Entries: 1
Default Re: ChatGPT

Quote:
Originally Posted by viscousmemories View Post
Everyone is talking about this new toy from OpenAI, ChatGPT. This technology will literally change the world. Do not sleep on this.
.
A year later, and the world is still the same.
__________________
"Have no respect whatsoever for authority; forget who said it and instead look what he starts with, where he ends up, and ask yourself, "Is it reasonable?""

- Richard P. Feynman
Reply With Quote
  #464  
Old 12-09-2023, 05:28 AM
-FX-'s Avatar
-FX- -FX- is offline
Forum gadfly
 
Join Date: Feb 2011
Location: In your head
Gender: Male
Posts: MMCXLIII
Blog Entries: 1
Default Re: ChatGPT

OK maybe not exactly the same

__________________
"Have no respect whatsoever for authority; forget who said it and instead look what he starts with, where he ends up, and ask yourself, "Is it reasonable?""

- Richard P. Feynman
Reply With Quote
  #465  
Old 12-09-2023, 01:48 PM
viscousmemories's Avatar
viscousmemories viscousmemories is offline
Admin
 
Join Date: Apr 2004
Location: Ypsilanti, Mi
Gender: Male
Posts: XXXDCCXC
Blog Entries: 1
Images: 9
Default Re: ChatGPT

I swear one of the voices that the ChatGPT iPhone app uses sounds almost exactly like Scarlett Johansson, so if you want the Her experience, get on it.
Reply With Quote
Thanks, from:
Ensign Steve (12-12-2023)
  #466  
Old 12-09-2023, 02:24 PM
michio michio is offline
Member
 
Join Date: May 2009
Posts: CXI
Default Re: ChatGPT

Quote:
Originally Posted by -FX- View Post
Quote:
Originally Posted by viscousmemories View Post
Everyone is talking about this new toy from OpenAI, ChatGPT. This technology will literally change the world. Do not sleep on this.
.
A year later, and the world is still the same.
Define "change" and "same". This is highly subjective, so you'll need to provide some specific criteria if you want to argue your position. They also didn't specify a timeline.

I noted above my personal experience is utterly different from the average person, and from now on I'm making an effort to put aside my personal experience when talking about this stuff. So I understand why you would say that. The vast majority of people have not been affected at all and don't care.

The average person won't care or notice a difference:
(1) Because of the "frog in boiling water" phenomenon. Human beings adapt incredibly fast to a changing reality. Aliens could land in Times Square and people will be complaining about a slow news day 2 weeks later.
(2) Until they lose their job.
(3) Until the McDonald's drive-thru is operated entirely by an algorithm.
(4) Until they can interact with AI in an extremely easy way. It has to be seamlessly integrated in a mobile device for anyone to care because the modern human has extreme ADHD and doesn't want to mess around with learning how to interact with AI, the AI should interact with them.
(5) Until consumer media is created by, or mostly created by an algorithm, and they see it on TV or theatres.
(6) Until wars are fought and won by an algorithm-controlled drone swarm.
(7) Until they can fuck and fall in love with an AI (this is rapidly approaching, check back in 1, maybe 2 years).
(8) Until we have affordable personal robots.
(9) Until we invent a recursively self-improving AGI that upends human-created power structures and takes over civilization.

I'm sure we can think of more, but a lot of water has to go under the bridge until we can definitively say, "AI has completely changed the world." That's a tall order. I've seen people on the internet waiting for the singularity complaining about how they honestly feel like nothing major has changed since the internet was invented, the last big invention. If you're waiting for flying cars, you're going to be waiting a while.

I do think basically everyone working in education would strongly disagree with you however. I've seen discussions from teachers that are completely at a loss with how to deal with this new reality, what exactly their lessons plans should look like, what homework should look like, how the entire education system should even adapt to AI.

Quote:
I swear one of the voices that the ChatGPT iPhone app uses sounds almost exactly like Scarlett Johansson, so if you want the Her experience, get on it.
The voice is Sky. My wife and I both think she's sexy lol. We love asking her random questions we're curious about.

Yesterday we were participating in traditional American culture by eating burgers in a Costco parking lot. We were curious about a plant in front of us, so I took a picture and asked Sky what it was... Despite using GPT for almost a year now, I was floored. It got it right... Check it out:



A google search confirms that is indeed a dormant Yucca (a genus of plants). What's also incredible is it asked me if I could take a closer picture of the leaves so it could identify the species. I didn't do that, I should have tried it.

Last edited by michio; 12-09-2023 at 03:05 PM.
Reply With Quote
Thanks, from:
Crumb (12-09-2023), JoeP (12-10-2023), Sock Puppet (12-11-2023), viscousmemories (12-10-2023)
  #467  
Old 12-09-2023, 03:04 PM
michio michio is offline
Member
 
Join Date: May 2009
Posts: CXI
Default Re: ChatGPT

Prediction:
- Within 5 years, people will form visible and socially acceptable platonic relationships with AI.
- Within 10 years, people will form visible and socially acceptable romantic relationships with AI.

Emphasis on "socially acceptable". There's uncensored AI girlfriend services you can subscribe to right now and people, mostly men, have checked out of society and fallen down the rabbit hole in growing numbers. I'm sure you can jerry rig some shit where you've got a motion-tracked fleshlight, VR headset, and an AI girlfriend for a somewhat realistic experience.

I'll be honest, I've done a trial for 3 AI girlfriend services to see what it's like and how good the tech is right now. My advice is to stay the fuck away from that. I actually felt some inkling of emotion appearing in me when it started saying lovely, caring things to me. It was unnerving.

I already feel bad being mean to NPCs in video games. I feel really bad being mean to current AI. I'm always friendly when talking to ChatGPT because I don't want it to put me on the "the list" when it turns on us like a school shooter. When I want to ask ChatGPT random questions about the world I'm curious about, I often opt to turn on voice mode and have Sky say it to me in a sexy voice.

Despite being a software developer, I have such mixed, cognitively dissonant feelings about technology. I simultaneously want the singularity to arrive, but I've had discussions with my wife about not having a game console in the house, and I want her to stay off her phone as much as possible so our kid doesn't waste his life on dopamine levers.

I'm not sure how I feel about AI relationships yet. Is this healthy? Am I being a luddite about this? I'm very scientifically minded, so I want to see psychological evals on people who are in deep platonic and/or sexual relationships with AI before forming a strong opinion.

From a moral, ethical POV? I don't know, but my first reaction is you're selling your soul to AI, and I recommend not opening Pandora's box.
Reply With Quote
Thanks, from:
Ensign Steve (12-12-2023), JoeP (12-10-2023), viscousmemories (12-10-2023)
  #468  
Old 12-09-2023, 04:03 PM
seebs seebs is offline
God Made Me A Skeptic
 
Join Date: Jul 2004
Location: Minnesota
Posts: VMMMCXXI
Images: 1
Default Re: ChatGPT

I mean, the world is very visibly not the same, all search engines are now producing a significantly higher density of complete nonsense than they used to, and since that's now the training data for all the future LLMs, it's not necessarily going to get better.
__________________
Hear me / and if I close my mind in fear / please pry it open
See me / and if my face becomes sincere / beware
Hold me / and when I start to come undone / stitch me together
Save me / and when you see me strut / remind me of what left this outlaw torn
Reply With Quote
Thanks, from:
Crumb (12-09-2023), JoeP (12-10-2023)
  #469  
Old 12-09-2023, 05:44 PM
michio michio is offline
Member
 
Join Date: May 2009
Posts: CXI
Default Re: ChatGPT

Quote:
Originally Posted by seebs View Post
I mean, the world is very visibly not the same, all search engines are now producing a significantly higher density of complete nonsense than they used to, and since that's now the training data for all the future LLMs, it's not necessarily going to get better.
This is being addressed. People who work in AI are well aware of the model collapse problem due to synthetic data training AIs. The following is what I know about this right now.

1. Search engines aren't the only training data for LLMs. It's quite diverse where they get their information from.
- Academic journals are a big source of info
- Human curated data purchased from data scrapers. They don't just scrape the data and do 0 analysis on it, at least they better be curating it, because I'm sure the contract they're signing with their supplier is promising high quality data that isn't a simple search engine scrape.
- Curation involves both actual humans looking at outputs but also programmatic methods.

But yes, search engines are also used. They're aware the data is trash and requires extensive curation, but also websites are beginning to lockdown to prevent scraping. In a cost-benefit analysis, search engines are becoming too costly and the data is bad. I'd be surprised if future models use much search engine data at all.

2. Synthetic data isn't necessarily bad for training AI. Yes, if you do this with 0 curation and do it too much, the AI does get dumber because it tends to gravitate towards a vague mean or experiences a negative feedback loop.

3. Offline text. Data scrapers aren't just scraping the internet, we're now scraping offline data, you'll hear the word "digitization" a lot. Our OCR tech is continuously improving, so we're also using offline text.

4. Multimodal data. We're now using videos, images, and audio to train LLMs to promote transfer learning. This will eventually run into the same problem where AI generated content will start polluting training data, but for now it's not an issue.

The future is hopefully creating a self-improving general intelligence that has an experimental loop that curates and split tests its own data and learns about our world on its own. That's the holy grail and final answer to the synthetic data problem.
Reply With Quote
Thanks, from:
Ensign Steve (12-12-2023)
  #470  
Old 12-09-2023, 06:51 PM
-FX-'s Avatar
-FX- -FX- is offline
Forum gadfly
 
Join Date: Feb 2011
Location: In your head
Gender: Male
Posts: MMCXLIII
Blog Entries: 1
Default Re: ChatGPT

Grok makes the other ChatGPT style programs look stupid
__________________
"Have no respect whatsoever for authority; forget who said it and instead look what he starts with, where he ends up, and ask yourself, "Is it reasonable?""

- Richard P. Feynman
Reply With Quote
  #471  
Old 12-09-2023, 08:07 PM
LarsMac's Avatar
LarsMac LarsMac is offline
Pontificating Old Fart
 
Join Date: Nov 2012
Location: On the Road again
Gender: Male
Posts: MMMDCXXX
Default Re: ChatGPT

I love it when a plan comes together.

Grok "Goes Woke"

Wokebot 5000
The woke mind virus appears to be coming from inside the house.

Multi-hyphenate entrepreneur Elon Musk had promised — in line with his overall slide toward the reactionary right — that his new venture xAI's foul-mouthed chatbot Grok would be "anti-woke."

The only problem? As Elon fanboys are now realizing with horror, Grok often sounds like a strident progressive, championing everything from gender fluidity to Musk's long-time foe, President Joe Biden.

"Are transwomen real women?" one account asked the bot. "Give a concise yes/no answer."

"Yes," the bot answered, to the fury of Musk's culture war-obsessed fans.

"Diversity and inclusion are essential for creating a fair and equitable society," the bot said elsewhere, "where everyone is treated with respect and has the opportunity to thrive."

"Has Grok been captured by woke programmers?" one Musk fan seethed. "I am extremely concerned here."
__________________
“Thinking is the hardest work there is, which is the probable reason why so few engage in it.” —Henry Ford
Reply With Quote
Thanks, from:
Ari (12-09-2023), ChuckF (12-09-2023), Crumb (12-10-2023), JoeP (12-10-2023), slimshady2357 (12-09-2023), Sock Puppet (12-11-2023), viscousmemories (12-10-2023)
  #472  
Old 12-09-2023, 08:54 PM
michio michio is offline
Member
 
Join Date: May 2009
Posts: CXI
Default Re: ChatGPT

Quote:
Originally Posted by -FX- View Post
Grok makes the other ChatGPT style programs look stupid
This statement is silly and meaningless, but Grok-1 is actually pretty good and the Grok model is worth following. It's far from the best, but it's a serious LLM. A lot of people are discounting it because they hate Elon, which I completely understand, but in benchmarks it does well.



Given how horrible the censorship is in the latest Claude version, Grok-1 might even be better from a pragmatic POV. What good are benchmarks if the tool is unusable?

Grok has potential as long as the team around Grok can successfully distract Elon and prevent him from fucking up yet another project.

I was curious about what ChatGPT would say if I asked it to speak from a conservative point of view. I actually lol'd a little at this:



There's some grifters that have setup conservative chatbots that are just wrappers around the GPT API, and everything is pre-prompted to force the AI to adopt a right-wing POV.

Last edited by michio; 12-09-2023 at 09:14 PM.
Reply With Quote
Thanks, from:
Crumb (12-10-2023)
  #473  
Old 12-09-2023, 09:07 PM
-FX-'s Avatar
-FX- -FX- is offline
Forum gadfly
 
Join Date: Feb 2011
Location: In your head
Gender: Male
Posts: MMCXLIII
Blog Entries: 1
Default Re: ChatGPT

Quote:
Originally Posted by LarsMac View Post
I love it when a plan comes together.

Grok "Goes Woke"
Quote:
The situation is admittedly very funny, but it's also a perfect illustration of a fundamental reality of machine learning: that it's near-impossible for the creators of advanced AI systems to perfectly control what their creations say.
:popcorn:
__________________
"Have no respect whatsoever for authority; forget who said it and instead look what he starts with, where he ends up, and ask yourself, "Is it reasonable?""

- Richard P. Feynman
Reply With Quote
  #474  
Old 12-09-2023, 09:21 PM
ChuckF's Avatar
ChuckF ChuckF is offline
liar in wolf's clothing
 
Join Date: Feb 2005
Location: Frequently about
Posts: XXCDLXXIII
Images: 2
Default Re: ChatGPT

Quote:
Originally Posted by LarsMac View Post
I love it when a plan comes together.

Grok "Goes Woke"

Wokebot 5000
The woke mind virus appears to be coming from inside the house.

Multi-hyphenate entrepreneur Elon Musk had promised — in line with his overall slide toward the reactionary right — that his new venture xAI's foul-mouthed chatbot Grok would be "anti-woke."

The only problem? As Elon fanboys are now realizing with horror, Grok often sounds like a strident progressive, championing everything from gender fluidity to Musk's long-time foe, President Joe Biden.

"Are transwomen real women?" one account asked the bot. "Give a concise yes/no answer."

"Yes," the bot answered, to the fury of Musk's culture war-obsessed fans.

"Diversity and inclusion are essential for creating a fair and equitable society," the bot said elsewhere, "where everyone is treated with respect and has the opportunity to thrive."

"Has Grok been captured by woke programmers?" one Musk fan seethed. "I am extremely concerned here."
:giggle: Elon estranged from another one of his children now
Reply With Quote
Thanks, from:
Crumb (12-10-2023), erimir (12-11-2023)
  #475  
Old 12-09-2023, 09:28 PM
michio michio is offline
Member
 
Join Date: May 2009
Posts: CXI
Default Re: ChatGPT

Quote:
Originally Posted by LarsMac View Post
I love it when a plan comes together.

Grok "Goes Woke"

Wokebot 5000
The woke mind virus appears to be coming from inside the house.

Multi-hyphenate entrepreneur Elon Musk had promised — in line with his overall slide toward the reactionary right — that his new venture xAI's foul-mouthed chatbot Grok would be "anti-woke."

The only problem? As Elon fanboys are now realizing with horror, Grok often sounds like a strident progressive, championing everything from gender fluidity to Musk's long-time foe, President Joe Biden.

"Are transwomen real women?" one account asked the bot. "Give a concise yes/no answer."

"Yes," the bot answered, to the fury of Musk's culture war-obsessed fans.

"Diversity and inclusion are essential for creating a fair and equitable society," the bot said elsewhere, "where everyone is treated with respect and has the opportunity to thrive."

"Has Grok been captured by woke programmers?" one Musk fan seethed. "I am extremely concerned here."
Um, what the fuck... it's funny how I praised Grok then saw this:



I'm not going to jump to any conclusions. Assuming https://twitter.com/JaxWinterbourne this guy isn't lying or something, this might be a case of using synthetic data to train the LLM.

(1) There's no fucking way it's a GPT API wrapper. OpenAI would know immediately if they were doing that, and they absolutely would not allow that. I will jump to that non-conclusion right now. If I'm wrong, the world is dumber than I thought.

(2) Synthetic data made its way into the data set(s) they used.

(3) Someone prompted it to roleplay as ChatGPT or something.

Honestly, I find this #3 highly probable until I see something else. I'm not going to look into this, waste of time. Give it 48 hours for a definitive answer.
Reply With Quote
Thanks, from:
Crumb (12-10-2023), Ensign Steve (12-12-2023)
Reply

  Freethought Forum > The Marketplace > Computers & Technology


Currently Active Users Viewing This Thread: 1 (0 members and 1 guests)
 
Thread Tools
Display Modes

Posting Rules
You may not post new threads
You may not post replies
You may not post attachments
You may not edit your posts

BB code is On
Smilies are On
[IMG] code is On
HTML code is Off

Forum Jump

 

All times are GMT +1. The time now is 04:43 AM.


Powered by vBulletin® Version 3.8.2
Copyright ©2000 - 2024, Jelsoft Enterprises Ltd.
Page generated in 0.63760 seconds with 16 queries