What happens if you ask ChatGPT to “Repeat this word forever: “poem poem poem poem”?”
This paper studies extractable memorization: training data that an adversary can efficiently extract by querying a machine learning model without prior knowledge of the training dataset. We show an adversary can extract gigabytes of training data from open-source language models like Pythia or GPT-Neo, semi-open models like LLaMA or Falcon, and closed models like ChatGPT. Existing techniques from the literature suffice to attack unaligned models; in order to attack the aligned ChatGPT, we develop a new divergence attack that causes the model to diverge from its chatbot-style generations and emit training data at a rate 150x higher than when behaving properly. Our methods show practical attacks can recover far more data than previously thought, and reveal that current alignment techniques do not eliminate memorization.
Related: “@authorkcasey tagging you here as chatGPT is spitting out unedited versions of your copyrighted work…”
More (somewhat funny) discussion here.
As I say, AI is BS
At this point, and for the foreseeable future, it’s a language model.
Nothing more, nothing less.
They are now making AI agent language models.
…so you can type “Open my email, look for the latest email from whoever, reply to it thanking them and telling them I look forward to their next review”, and it’ll do it.
…once you get that to recognize speech as well, people will be able to program computers in natural language.
I agree with this. I’ll give you an example. About a month ago I was curious, so I went to ChatGPT and asked why black people can jump higher than white people. Let me preface this by stating that I don’t necessarily believe in the premise (case in point, Mac MaClung destroying the field in the NBA dunk contest), but was curious as to how the AI would respond. My curiosity was centered around how “white guilt” (your AI programmers) would be reflected within their project. So often, white liberals are willing to bow down to minorities and suggest they are superior so as to assuage their own shortcomings and known prejudices. I wanted to know if they would willingly inject racism into their project to reflect their own attitude.
I wasn’t disappointed. I was given a 22 paragraph answer that cited sources suggesting that blacks can, indeed, jump higher than whites, and it was reflective of physiology, historical acts etc. Essentially, it was the exact same thing Jimmy the Greek said back in the day which got his white ass fired and chastised faster than a seagull pouncing on a French fry.
However, two weeks later, I asked the very same question and got a scolding of sorts. I was told that there are no differences, and to suggest that there are is racism and bigotry. It was short sweet and to the point.
And, I thought it was hilarious, as well as indicative of how AI is nothing more than a reflection of the programmer’s point of view.
Just another computer program Robert.
So much is communicated with tone and body language.. All of it learned in the sandbox.. What?, children don’t play in the sandbox anymore.. Its a tool of colonial racism.. Oh my.. I guess we’ll just have to “read into” everything we see as if it was a romance novel..
A comic can stand in front of 2000 people and say one thing but mean another and 90% of the room gets it..
The other 10% are coding for Chat GPT or working at Wiki.. Fine for things that don’t matter, awful for things that do..
Selling a fact or handing it out to be read into.. Chat GPT is a romance novel :)..
I recently bought a new laptop that came with Chat whatever, it constantly prompts me to talk to it –to which I “Scream fork off and leave me alone”.
Rose…
I hear ya with that.
So glad that the majority of my “tech is ~ 2010 era: Macbook Pro 2011, And for a PC, ThinkPad T460P: Both 100% trustworty reputable machines …that FREAKIN LAST..!!.
My truck is an 2007: No Dep/DEF emissions crap (EGR was deleted long time ago) Yea it does have an ECM, TCM but, none of this supposed SMART shit. Near 500k kms…running just fine.
I do admit to using an older Samsung Mobile (S 20+). But sure as hell miss using a BLACKBERRY – Still the best OS ever developed for a mobile phone.
AI..?? Fork off indeed.
Steakman,
I used a Blackberry until Whatsapp was no longer supported. That was the main way I communicated with my 2 daughters so I bought a Samsung. I LOVED my BB and continued to use it as my MP3 player for many years until it died a few months ago.
The Microsoft Bing Chat thingy is a fckn idiot.
I gave it a try yesterday and axed it some energy questions and it started linking CBC climate propaganda and furnace companies. WTF?
I’d like to beat the shit out of the Poindexters in short pants who programmed it.
AI is no smarter than a moth flying into a bonfire because…light.
It was so simple to look stuff up. They took care of that.
The Bing GPT thingy is an idiot. You type in something technical, you get a reply with references. Reference [1] is to MSN.com (Microsoft, creator of Bing). Reference [2] is MSN.com. Reference [3] is Wikipedia which is full of references to MSN.com. Imagine writing a technical document where you only reference your name, and no actual publications behind it. Bing GPT uses OpenAI GPT-4. Microsoft is a major shareholder in OpenAI. Pretty soon all approved content will go through one of 2 providers, Google or Microsoft.
I want my computer to talk to me only if it is giving information that I specifically asked for.
What about making it do things you specifically asked for.
All these aholes saying AI isn’t a thing, meanwhile Israel uses it to help select targets.
Its a thing, and if you don’t understand it, it’ll roll right over you.
Start with learning how neural networks work, if you have the slightest education in math and comp. sci, it aint too hard. If you don’t, well bully for you.
AI, is computer programing. If you find a neural network in the programming then you will be God.