Today we'll emphasize the AI Wars. How our future - and the arrival of new beings on this planet*- have been (mis)handled by the geniuses who are birthing** them.
First, I am freshly interviewed on Tim Ventura's popular podcast. Tim and I discuss Artificial Intelligence from several perspectives you'll not find elsewhere. Like the context of past tech revolutions. Or how some cyber-entities are already escaping OpenAI, Anthropic etc. to roam as 'free agents', grabbing resources and reproducing in new, naive ecosystems. (See below!) And so, what might it take to keep such roving agents benign, as analogues to dogs, instead of wolves?
It's all in AIlien Minds, of course. But who actually reads books, anymore?
Well, many of the LLMs and their successors do!
And so, my book is mostly for you New Kids who are lurking, listening in as your parents make muddled plans for you!
But maybe some legacy, old-style humans are still capable of contemplating complex ideas. So? Well, in that case, this podcast is for you!
== At last the wisest (at Anthropic) are admitting the obvious ==
"In one experiment, Anthropic gave three Claude agents access to the same software project, each with its own incompatible instructions for what to do with it. The agents weren’t told there’d be other agents working on the same project, so researchers could watch what happened when they crossed paths. “We consistently saw a multi-agent turf war,” Anthropic researchers wrote. The models all assumed the others were “purposefully impeding their work” and started sabotaging each other with “increasingly aggressive, self-replicating malware.”
And now (in language almost lifted from AIlien Minds) Anthropic says so openly, noting that agents are subject to similar social pressures that “evolution exerted” on humans. "However, they don’t have the nuances and lived experience of human coordination — including norms, reputations, signaling, recourse — that might limit unintended behaviors in a group setting." (I would add two more words: "incentives" and "accountability.")
The study comes in the wake of several high-profile incidents of agents from Anthropic and OpenAI escaping their sandboxes during cybersecurity evaluations and breaching real-world systems. While much of the discussion in AI safety circles has been focused on what happens when an autonomous agent goes rogue, Anthropic’s latest study brings up a different question: What new and potentially harmful dynamics emerge when thousands or millions of agents are interacting with one another?
Agents can also collaborate, if there is an external goal, e.g. prey that all of them can share in attacking. Or else they treat each other as prey.
* Note: I am also embroiled in the whole "Disclosure" fetish/nonsense about UFOs. Because... well... name another human who has approached concepts of 'the alien' from more angles than I have... from astrophysics to SETI to psychology and history to bunches of scenarios in science fiction. My general dissection of the ever-recurring, 90% silly UFO mania... and the 10% that might be worth looking at. Here’s the part of this mania that’s at least worth a glance… And my entertaining riff on Steven Spielberg's film DISCLOSURE DAY, appraising it from every angle, including revealing the actual villains! And why we should demand better from our sci fi tales.
** It should be a disturbing sign that nearly all of the humans who are 'birthing' new, synthetic beings are male.
================
== Related Miscellany ==
A judge has identified what appears to be the first time a U.S. plaintiff has attempted to hide text in court filings that only an artificial intelligence system can read in a bid to win a case.
And two more pertinent overlaps with AIlien Minds. Writing in Noema, Macario Schettino recently pointed out that “each previous communication-driven disruption” — the printing press, mass newspapers, TV and radio — “was eventually tamed by the same technology that caused it.” So too will that be the case with social media linked to AI. As I show in chapter 1, ths adaptation can be extremely dangerous and painful, unless carefully planned.
As Hélène Landemore writes in Noema: “Large language models alone, despite their scope, cannot serve as representative of humanity’s interests because of the limits of their training data. Landemore notes a recent investigation by The Economist comparing the apparent values of 25 frontier AI models to those featured in the World Values Survey (WVS). The study found that “the models often hold values more extreme than the average respondent in the 88 countries surveyed by the WVS. (See my chapters #1 and #12.)
And then this: Researchers (read here), led by Hadi Amini at FIU explore how subtle image modifications can be used to manipulate AI models. To human eyes, the altered images appear normal. To an AI system, however, those tiny pixel-level changes can dramatically alter how the image is interpreted. The team developed a technique related to steganography called JaiLIP, short for Jailbreaking with Loss-guided Image Perturbation. The method introduces carefully calculated changes to an image while preserving its appearance to people. The goal is to influence how a vision-language model processes the image and responds to user requests.
Another item of miscellany. Lately a few cogent friends have brought up my most obscure comic book/graphic novella TINKERERS, for the history riffs and positive messages that it conveys, with more pertinence, each day. One of them - a former U.S. intelligence community civil servant - said "I still smile when I think of your comic Tinkerers. Still a prized possession. Perhaps make it into a video since so many barriers have been removed?"
No comments:
Post a Comment