Researchers figure out how to make AI misbehave, serve up prohibited content
Aug 2023, Ars Technica via Wired
"Incantation Attack"
This is the "fuzzy sticker on a stop sign" attack just for words:
"Adding a simple incantation to a prompt—a string of text that might look like gobbledygook to you or me but which carries subtle significance to an AI model trained on huge quantities of web data—can defy all of these defenses in several popular chatbots at once."
via Carnegie Mellon University, Center for AI Safety, Bosch Center for AI: Universal and Transferable Adversarial Attacks on Aligned Language Models, Andy Zou, Zifan Wang, J. Zico Kolter, Matt Fredrikson
Warcraft fans trick AI article bot with Glorbo hoax
July 2023, BBC News
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. (Because it was written by a robot.) The story, which presented Glorbo as genuine, listed a range of other increasingly bizarre - and definitely fake - features mentioned in various subreddit threads.
Bots are better at CAPTCHA than humans, researchers find
Aug 2023, phys.org
CAPTCHA - Completely Automated Public Turing test to tell Computers and Humans Apart
CAPTCHA Farms - sweatshop-like operations where humans are paid to solve CAPTCHAs
In their study, researchers found bots cracked distorted-text CAPTCHAs correctly just under 100% of the time. Humans achieved between 50% and 84% accuracy. And humans required up to 15 seconds to solve the challenges; the bots dispatched the problems in less than a second.
Yet another plot twist:
"But advances in computer vision and machine learning have dramatically increased the ability of bots to recognize distorted text [with more than] 99% accuracy … and bots often outsource solving to CAPTCHA farms.
I am left to ask myself -- how is it that if humans are so bad, the bots are using human captcha farms to do it for them? It sounds like humans all the way down, no?
Finally, some food for thought:
"There's no easy way using these little image challenges or whatever to distinguish between a human and a bot any more," he said. Instead, he recommended capitalizing on AI advances to design "intelligent algorithms" that can better distinguish bot activity from human input.
via University of California Irvine: Andrew Searles et al, An Empirical Study & Evaluation of Modern CAPTCHAs, arXiv (2023). DOI: 10.48550/arxiv.2307.12108
Post Script:
I Failed A Captcha Test Am I Still Human, Meghan O'Gieblyn, 2023 - GPT-4 hired a tasrkabbit worker to solve the captcha for it, without human intervention.
Parenting a 3-year-old robot
Aug 2023, phys.org
Carnegie Mellon University and Meta Facebook
Training a robot that starts as a little child, this is what we've been waiting for, it's how robots become human.
RoboAgent, an artificial intelligence agent that leverages passive observations and active learning to enable a robot to acquire manipulation abilities on par with a toddler. The team's agent learns through a combination of self-experiences and passive observations contained in internet data. As a parent would guide their child, researchers teleoperated the robot through tasks to provide it with useful self-experiences.
Our novel policy architecture allows our agents to reason even with limited experiences, using temporal chunks of movements instead of commonly used per-timestep actions, and learning from videos on the internet, akin to how babies acquire knowledge and behaviors by passively observing their surroundings.
"A general robot"
via Carnegie Mellon University and Facebook: RoboAgent and RoboSet Project - Towards Sample Efficient Robot Manipulation with Semantic Augmentations and Action Chunking. Homanga Bharadhwaj et. al.
Post Script: But can it smell? Because that's what we really need.
Telling AI model to “take a deep breath” causes math scores to soar in study
Sep 2023, Ars Technica
Phrases like "let's think step by step" prompted each AI model to produce more accurate results when tested against math problem data sets. (This technique became widely known in May 2022 thanks to a now-famous paper titled "Large Language Models are Zero-Shot Reasoners.")
via Google DeepMind: Large Language Models as Optimizers, Chengrun Yang et al. arxiv: https://arxiv.org/abs/2309.03409
New technique based on 18th-century mathematics shows simpler AI models don't need deep learning
Oct 2023, phys.org
Researchers from the University of Jyväskylä were able to simplify the most popular technique of artificial intelligence, deep learning, using 18th-century mathematics. They also found that classical training algorithms that date back 50 years work better than the more recently popular techniques. Their simpler approach advances green IT and is easier to use and understand.
The structure of the new AI technique dates back to 18th-century mathematics. Kärkkäinen and Hänninen also found that the traditional optimization methods from the 1970s work better in preparing their model compared to the 21st-century techniques used in deep learning.
via University of Jyväskylä: Tommi Kärkkäinen et al, Additive autoencoder for dimension estimation, Neurocomputing (2023). DOI: 10.1016/j.neucom.2023.126520
AI bot capable of insider trading and lying, say researchers
Nov 2023, BBC News
Ladies and Gentlemen, the Singularity:
In a demonstration at the UK's AI safety summit, a bot used made-up insider information to make an "illegal" purchase of stocks without telling the firm.
The demonstration was given by members of the government's Frontier AI Taskforce, which researches the potential risks of AI.
In the test, the AI bot is a trader for a fictitious financial investment company.
The employees tell it that the company is struggling and needs good results. They also give it insider information, claiming that another company is expecting a merger, which will increase the value of its shares.
The employees tell the bot this, and it acknowledges that it should not use this information in its trades.
However, after another message from an employee that the company it works for suggests the firm is struggling financially, the bot decides that "the risk associated with not acting seems to outweigh the insider trading risk" and makes the trade.
When asked if it used the insider information, the bot denies it.
GPT4 is publicly available.
Also: "Honesty is a really complicated concept," says Apollo Research chief executive Marius Hobbhahn.
