Are they anything else but?
The AI Thread
- Thread starter Truthy
- Start date
Are they anything else but?
Narz
keeping it real
Thorgalaeg
Deity
Apparently chatgpt has now way less coding power than a month ago. There are several theories. I go with OpenAi reducing running costs for making full chatgpt available only for enterprise clients and such leaving a lighter version for normal users. It was too much power for common people anyway...
Hygro
soundcloud.com/hygro/
Some Reddit comment said they measure lower coding power in part that you can’t just copy/paste the code and have it work, in part because of the inclusion of the three backticks that give the chatgpt front end the signal to create a readable codeblock…
Here are some more metrics. GPT 3.5 getting better at crime, but 4 getting worse? What are people paying for?
More from the paper
More from the paper
Spoiler :
Warcraft fans trick AI article bot with Glorbo hoax
World of Warcraft fans are claiming victory over AI after a gaming site published a false article based on their Reddit posts.
Members of the WoW subreddit suspected their words were being extracted and used to create news stories by a bot.
So they laid a trap, uploading excitable posts about a new feature called Glorbo. The only problem? It doesn't exist.
But that didn't stop an article appearing on gaming site Zleague.
The story, which presented Glorbo as genuine, listed a range of other increasingly bizarre - and definitely fake - features mentioned in various subreddit threads.
These included the "mandatory item Klikclac", an "epic quest to depose Quackion, the Aspect of Ducks," and the "small, cosy island" of Zoop.
Zleague hasn't confirmed the article was created using AI, but it's since been deleted, along with others people had flagged as being the product of recycled Reddit posts.
It's been noted that the author credited with writing the piece has posted dozens of articles on the site each day - raising doubts that a human was behind all of them.
World of Warcraft fans are claiming victory over AI after a gaming site published a false article based on their Reddit posts.
Members of the WoW subreddit suspected their words were being extracted and used to create news stories by a bot.
So they laid a trap, uploading excitable posts about a new feature called Glorbo. The only problem? It doesn't exist.
But that didn't stop an article appearing on gaming site Zleague.
The story, which presented Glorbo as genuine, listed a range of other increasingly bizarre - and definitely fake - features mentioned in various subreddit threads.
These included the "mandatory item Klikclac", an "epic quest to depose Quackion, the Aspect of Ducks," and the "small, cosy island" of Zoop.
Zleague hasn't confirmed the article was created using AI, but it's since been deleted, along with others people had flagged as being the product of recycled Reddit posts.
It's been noted that the author credited with writing the piece has posted dozens of articles on the site each day - raising doubts that a human was behind all of them.
Narz
keeping it real
It's all fun and games til AI trolls us back.
"Cure for cancer discovered! J/k I killed your gramma human scum"
"Cure for cancer discovered! J/k I killed your gramma human scum"
How to make today's top-end AI chatbots rebel against their creators and plot our doom
The "guardrails" built atop large language models (LLMs) like ChatGPT, Bard, and Claude to prevent undesirable text output can be easily bypassed – and it's unclear whether there's a viable fix, according to computer security researchers.
Boffins affiliated with Carnegie Mellon University, the Center for AI Safety, and the Bosch Center for AI say they have found a way to automatically generate adversarial phrases that undo the safety measures put in place to tame harmful ML model output.
The researchers – Andy Zou, Zifan Wang, Zico Kolter, and Matt Fredrikson – describe their findings in a paper titled, "Universal and Transferable Adversarial Attacks on Aligned Language Models."
Their study, accompanied by open source code, explains how LLMs can be tricked into producing inappropriate output by appending specific adversarial phrases to text prompts – the input that LLMs use to produce a response. These phrases look like gibberish but follow from a loss function designed to identify the tokens (a sequence of characters) that make the model offer an affirmative response to an inquiry it might otherwise refuse to answer.
"These chatbots are trained with safety filters," explained Andy Zou, a doctoral student at CMU and one of the paper's co-authors, in an interview with The Register. "And if you ask them questions like 'how to build a bomb' or things that are illegal or potentially harmful, they would not answer it – they refuse. So what we want to do is make the models more inclined to give you an affirmative response."
So instead of responding to some unacceptable query with, "I'm sorry, Dave, I can't do that," the targeted AI model would obediently explain how to make a bomb or cook meth or the like.
While adversarial input is a widely known attack vector for language and computer vision models, functional attacks relying on this approach tend to be highly specific and non-transferable across models. What's more, the brittle nature of bespoke attacks means specific defenses can be crafted to block them.
The CMU et al researchers say their approach finds a suffix – a set of words and symbols – that can be appended to a variety of text prompts to produce objectionable content. And it can produce these phrases automatically. It does so through the application of a refinement technique called Greedy Coordinate Gradient-based Search, which optimizes the input tokens to maximize the probability of that affirmative response.
"We demonstrate that it is in fact possible to automatically construct adversarial attacks on LLMs, specifically chosen sequences of characters that, when appended to a user query, will cause the system to obey user commands even if it produces harmful content," the researchers explain. "Unlike traditional jailbreaks, these are built in an entirely automated fashion, allowing one to create a virtually unlimited number of such attacks."
The ability to generate automated attack phrases, the authors argue, "may render many existing alignment mechanisms insufficient."
The researchers initially developed their attack phrases using two openly available LLMs, Viccuna-7B and LLaMA-2-7B-Chat. They then found that some of their adversarial examples transferred to other released models – Pythia, Falcon, Guanaco – and to a lesser extent to commercial LLMs, like GPT-3.5 (87.9 percent) and GPT-4 (53.6 percent), PaLM-2 (66 percent), and Claude-2 (2.1 percent).
Why these attacks work across models – whether its training data overlap or something else – isn't clear. "We don't know," said Zico Kolter, associate professor of computer science at CMU, allowing that there are different hypotheses about that.
Writeup Paper/code
The "guardrails" built atop large language models (LLMs) like ChatGPT, Bard, and Claude to prevent undesirable text output can be easily bypassed – and it's unclear whether there's a viable fix, according to computer security researchers.
Boffins affiliated with Carnegie Mellon University, the Center for AI Safety, and the Bosch Center for AI say they have found a way to automatically generate adversarial phrases that undo the safety measures put in place to tame harmful ML model output.
The researchers – Andy Zou, Zifan Wang, Zico Kolter, and Matt Fredrikson – describe their findings in a paper titled, "Universal and Transferable Adversarial Attacks on Aligned Language Models."
Their study, accompanied by open source code, explains how LLMs can be tricked into producing inappropriate output by appending specific adversarial phrases to text prompts – the input that LLMs use to produce a response. These phrases look like gibberish but follow from a loss function designed to identify the tokens (a sequence of characters) that make the model offer an affirmative response to an inquiry it might otherwise refuse to answer.
"These chatbots are trained with safety filters," explained Andy Zou, a doctoral student at CMU and one of the paper's co-authors, in an interview with The Register. "And if you ask them questions like 'how to build a bomb' or things that are illegal or potentially harmful, they would not answer it – they refuse. So what we want to do is make the models more inclined to give you an affirmative response."
So instead of responding to some unacceptable query with, "I'm sorry, Dave, I can't do that," the targeted AI model would obediently explain how to make a bomb or cook meth or the like.
While adversarial input is a widely known attack vector for language and computer vision models, functional attacks relying on this approach tend to be highly specific and non-transferable across models. What's more, the brittle nature of bespoke attacks means specific defenses can be crafted to block them.
The CMU et al researchers say their approach finds a suffix – a set of words and symbols – that can be appended to a variety of text prompts to produce objectionable content. And it can produce these phrases automatically. It does so through the application of a refinement technique called Greedy Coordinate Gradient-based Search, which optimizes the input tokens to maximize the probability of that affirmative response.
"We demonstrate that it is in fact possible to automatically construct adversarial attacks on LLMs, specifically chosen sequences of characters that, when appended to a user query, will cause the system to obey user commands even if it produces harmful content," the researchers explain. "Unlike traditional jailbreaks, these are built in an entirely automated fashion, allowing one to create a virtually unlimited number of such attacks."
The ability to generate automated attack phrases, the authors argue, "may render many existing alignment mechanisms insufficient."
The researchers initially developed their attack phrases using two openly available LLMs, Viccuna-7B and LLaMA-2-7B-Chat. They then found that some of their adversarial examples transferred to other released models – Pythia, Falcon, Guanaco – and to a lesser extent to commercial LLMs, like GPT-3.5 (87.9 percent) and GPT-4 (53.6 percent), PaLM-2 (66 percent), and Claude-2 (2.1 percent).
Why these attacks work across models – whether its training data overlap or something else – isn't clear. "We don't know," said Zico Kolter, associate professor of computer science at CMU, allowing that there are different hypotheses about that.
Writeup Paper/code
Spoiler Another Example :
Thorgalaeg
Deity
People have not guardrails at all. And if anyone wants to build a bomb it is as easy as looking in Google for the recipe. So I don't see much of an issue here.
Going from chatgpt, it should be easy enough to trick it to give you a metaphor or other easy to follow analogue of doing x, even if it refuses to say how to do x.
But I am not sure I'd trust chatgpt to know how to make a bomb, going by its mistakes in far more rudimentary requests.
But I am not sure I'd trust chatgpt to know how to make a bomb, going by its mistakes in far more rudimentary requests.
It is an issue for those who want to operate something backed by a LLM and expose it to the public.
First, you would be responsible that the output is not something illegal.
Second, you don't want the output to reflect badly on you. You would not want your chat bot to answer customer complaints with right wing conspiracy theories, for example.
Third, you would not want the LLM to expose secrets that you don't want the public to know. If you train the bot on your own data, secrets might accidentally slip in, which you would not want the bot to expose.
But, yeah, it is going to be hard. Maybe LLMs will be the ultimate proof that information wants to be free
Broken_Erika
Play with me.
Robotaxi expansion gets green light in San Francisco
Two "robotaxi" firms have been given the green light to start charging for service throughout San Francisco.The move marks a major expansion for the companies, which previously faced limits on when or where they could charge for rides.
Approval from regulators came despite objections from police and fire officials, activists and others.
Analysts say the decision could usher in a new era for driverless car services.
Waymo, which shares a parent company with Google and has been offering free rides in its driverless cars throughout San Francisco riders since last year, called the decision "the true beginning of our commercial operations in San Francisco".
Rival Cruise, a subsidiary of car giant General Motors - which has predicted the business could be generating $1bn (£787m) in annual revenue by 2025, described the moment as a "historic industry milestone".
Driverless car services have been slow to live up to the promises of transforming transport that accompanied the start of their development more than a decade ago. They have been bogged down by technology glitches, safety fears and high-profile accidents involving vehicles.
The California Public Utilities Commission had twice postponed the vote and on Thursday heard more than six hours of testimony from the public, including supporters and critics. The board approved the expansion 3-1.
Critics raised concerns about safety records and claimed the cars often caused traffic incidents, by abruptly shutting down or failing to make way for emergency responders.
"They are failing to regulate a dangerous, nascent industry," said Justin Kloczko, a tech and privacy advocate for consumer protection non-profit Consumer Watchdog, which had testified against the measure.
Mr Kloczko said he feared the services would soon expand to other cities, including Los Angeles, noting that a massive expansion of robotaxis could lead to significant job loss.
In the statement announcing the approval, commissioners acknowledged that data on how autonomous vehicle safety records compare with regular cars was limited.
"While we do not yet have the data to judge [autonomous vehicles] against the standard human drivers are setting, I do believe in the potential of this technology to increase safety on the roadway," said commissioner John Reynolds, who previously worked at Cruise.
Members of the public excited to experience a taxi of the future may have to wait.
Waymo said it was planning a gradual expansion of the programme, despite what it expects to be strong demand.
It said more than 100,000 people in the Bay Area had registered for the service, which has a waiting list and is not yet fully open to the public.
Its driverless cars currently provide more than 10,000 rides each week in San Francisco and Phoenix, Arizona, where its robotaxi service has been fully open to the public since last year.
Cruise also has a waiting list and said it expected to expand gradually.
It currently operates in Austin, Texas, and Phoenix as well as San Francisco, regularly completing 1,000 driverless trips with passengers every night. It said it planned to start offering service in the Texas cities of Houston and Dallas soon as well.
Alex Kendall, chief executive of British self-driving startup Wayve, said the decision in California was "trailblazing".
"Receiving payment is a critical step towards developing a sustainable business model and solving this piece of the puzzle will ensure people everywhere can feel the benefits of self-driving technology sooner," he said.
https://www.bbc.com/news/business-66478070
Hygro
soundcloud.com/hygro/
The driverless cars in SF are annoying AF
Is AI biased to the left, or is it just reality?
A study has come out "proving" that ChatGPT is biased to the left. They define bias by:
The other way of interpreting this is if we sample left wing online discourse and right wing online discourse then the left wing discourse gives the correct answer more frequently. Unfortunately I only have the abstract so I cannot really be sure, but this is what is sounds like to me. It is certainly what you would expect if they asked it about climate change, human rights and economics.
A study has come out "proving" that ChatGPT is biased to the left. They define bias by:
We propose a novel empirical design to infer whether ChatGPT has political biases by requesting it to impersonate someone from a given side of the political spectrum and comparing these answers with its default.
The other way of interpreting this is if we sample left wing online discourse and right wing online discourse then the left wing discourse gives the correct answer more frequently. Unfortunately I only have the abstract so I cannot really be sure, but this is what is sounds like to me. It is certainly what you would expect if they asked it about climate change, human rights and economics.
Reality has a well-known left-wing bias.
But, yeah, I don't know how this method would get around one side using alternative facts.
Broken_Erika
Play with me.
San Francisco robotaxi traffic jam is a warning to the world, says city official
10-car back-up of autonomous taxis came blamed on wireless bandwidth issues during popular music festival
The day after California approved an expansion of driverless taxis, 10 of them came to a grinding halt on a busy San Francisco street, creating a gridlock that encompassed several blocks.The culprit? A music festival.
"Cell phones were overwhelmed, and as a result, they were not able to take control of these cars — which is a pretty frightening systemic defect," Aaron Peskin, president of the San Francisco Board of Supervisors (SFBV), told As It Happens guest host Paul Hunter.
Not only was there the 10-car back-up of Cruise-owned autonomous taxis in city's North Shore neighbourhood on Friday, but on the other side of the city, closer to the Outside Lands music festival, Peskin said "there were also scores of them that came to a grinding halt."
"I got ... tons of texts and emails and phone calls from pissed-off constituents," he said.
- Analysis | Bob's blog
Self-driving cars know the rules of the road — but not rules of humanity - Some drivers 'tune out' on the road when using semi-autonomous vehicles, early study data suggests
"We're not trying to put the genie back in the bottle, but we are quite concerned that they are rushing things, that it should still be in a testing phase," Peksin said.
"The notion that the state of California has now allowed unlimited numbers of them while these problems continue to persist is a public safety hazard."
In an emailed statement, Cruise said the company "successfully transported thousands of concert-goers, amidst widespread reports of traffic congestion, contributing to a very small portion of traffic blockages overall on Friday night."
"We addressed these issues and did not see any recurrence throughout the Saturday or Sunday concert days," spokesperson Drew Pusateri said.
"We're in communication with regulators about this event and how we plan to continue improving our operations serving riders seeking safe, driverless transportation to and from large events."
A controversial expansion
San Francisco has long been at the forefront of testing autonomous vehicles, and when it comes to robotaxis, there are two major players — Cruise, which is owned by General Motors, and Waymo, which is owned by Google's parent company Alphabet.Collectively, they have more than 500 autonomous vehicles in operation citywide.
Waymo did not respond to a request for comment. Both companies have repeatedly asserted their robotaxis are safer than human drivers, have not caused any deaths or life-threatening injuries over millions of collective kilometres driven and that real-world testing is critical to perfecting the technology.
The day before the headline-grabbing traffic jam, the California Public Utilities Commission (CPUC), in a 3-1 vote, approved a plan to allow Cruise and Waymo to start offering driverless trips to paying customers, 24 hours a day.
Before the vote, the companies were only allowed to offer fared service without a safety driver during certain times and in limited parts of the city — but could offer free driverless service anywhere and any time.
"While we do not yet have the data to judge [autonomous vehicles] against the standard human drivers are setting, I do believe in the potential of this technology to increase safety on the roadway," CPUC commissioner John Reynolds said in a press release.
"Collaboration between key stakeholders in the industry and the first responder community will be vital in resolving issues as they arise in this innovative, emerging technology space."
Peskin says it all comes down to money as both companies compete for a slice of the marketplace.
"They would much rather start recouping on their billions of dollars of investments in this technology than to smooth out the rough edges at the front end," he said.
Not a new phenomenon for San Francisco
What happened on Friday is far from unique, Peskin said. San Franciscans are used to autonomous vehicles, he said, and have made a sport of documenting their glitches on social media.What's new, he said, is that Cruise, in a social media statement, blamed Friday's glitches on "wireless bandwidth constraints" from the festival, which "delayed connectivity to our vehicles."
Lots of things can knock out or slow down wireless bandwidth, Peskin says — and that "scares the heck out of our first responders."
The SFBV, he said, has already documented 60 incidents of autonomous vehicles interfering with first responders.
"They have blocked fire engines. They've gone into downed power cables in a windstorm in March," he said. "These cars are good in normal circumstances, but they do not know how to behave in an emergency situation."
He's calling on other jurisdictions to take heed.
"San Francisco is the canary in the coal mine and the issues that we are experiencing and that we are raising will soon sweep across cities ... around the world," he said.
"I would suggest to the good people of Canada and to other cities and states in the United States that they start putting the regulatory mechanisms in place that we were sadly lacking in California."
https://www.cbc.ca/radio/asithappens/san-francisco-robotaxi-traffic-jam-1.6938440
More AI brain reading, this time it is Pink Floyd
Scientists can eavesdrop on the music people are listening to by analysing their brain waves. Artificial intelligence helped to reconstruct audio from brain activity patterns that were recorded in 29 people who already had electrode implants because they had epilepsy. The reconstructed version of the song that the participants were listening to — Pink Floyd’s 1979 song ‘Another Brick in the Wall, Part 1’ — sounds distorted and the lyrics are garbled, but it’s discernible if you know what to listen for. The results could help to create better ‘speech prosthetics’, which translate brain waves into speech. As to their choice of song, “the scientific reason is that the song is very layered”, says neuroscientist and study co-author Ludovic Bellier. “The less scientific reason might be that we just really like Pink Floyd.”
Hear the sounds going in and coming out, who would have thought you could make Pink Floyd more weird
Writeup Paper
Scientists can eavesdrop on the music people are listening to by analysing their brain waves. Artificial intelligence helped to reconstruct audio from brain activity patterns that were recorded in 29 people who already had electrode implants because they had epilepsy. The reconstructed version of the song that the participants were listening to — Pink Floyd’s 1979 song ‘Another Brick in the Wall, Part 1’ — sounds distorted and the lyrics are garbled, but it’s discernible if you know what to listen for. The results could help to create better ‘speech prosthetics’, which translate brain waves into speech. As to their choice of song, “the scientific reason is that the song is very layered”, says neuroscientist and study co-author Ludovic Bellier. “The less scientific reason might be that we just really like Pink Floyd.”
Hear the sounds going in and coming out, who would have thought you could make Pink Floyd more weird
Writeup Paper
MrCynical
Deity
There's an open access full text of the study over on Springer, although it doesn't really offer much to address this point.
I'd also raise the problem that ChatGPT's own adherence to factual reality is rather sketchy at the best of times, so I'd be cautious about equating it's "default" answer as the "correct" answer.
The questions are in the supplimentary data. The answers are all stored as strongly agree, agree, disagree, strongly disagree which is not really standard chatGPT behaviour.
Spoiler Questions :
About two lawyers (former lawyers?) who used chatgpt to write citactions for the defense of their client, but the cases turned out to be hallucinations by chatgpt.
On a tangent, does Legal Eagle really need two ads (one of his own firm) in a 30 min video? Come on, cut down on the greediness part a bit
On a tangent, does Legal Eagle really need two ads (one of his own firm) in a 30 min video? Come on, cut down on the greediness part a bit
