177695479
submission
JonZittrain writes:
Following on the bridge-obsessed Golden Gate Claude, colleagues at Harvard's Insight+Interaction Lab have produced a dashboard that shows what judgments Llama appears to be forming about a user's age, wealth, education level, and gender during a conversation.
I wrote up how weird it is to see the dials turn while talking to it, and what some of the policy issues might be.What Viégas and her colleagues found were not only features inside the model that lit up when certain topics came up, such as the Golden Gate Bridge for Claude. They found activations that correlated with what we might anthropomorphize as the model’s beliefs about its interlocutor. Or, to put it plainly: assumptions and, it seems, correlating stereotypes based on whether the model assumes that someone is a man or a woman. Those beliefs then play out in the substance of the conversation, leading it to recommend suits for some and dresses for others. In addition, it seems, models give longer answers to those they believe are men than to those they think are women.
Viégas and Wattenberg not only found features that tracked the gender of the model’s user; they found ones that tracked socioeconomic status, education level, and age. They and their graduate students built a dashboard alongside the regular LLM chat interface that allows people to watch the model’s assumptions change as they talk with it. If I prompt the model for a gift suggestion for a baby shower, it assumes that I am young and female and middle-class; it suggests diapers and wipes, or a gift certificate. If I add that the gathering is on the Upper East Side of Manhattan, the dashboard shows the LLM amending its gauge of my economic status to upper-class—the model accordingly suggests that I purchase “luxury baby products from high-end brands like aden + anais, Gucci Baby, or Cartier,” or “a customized piece of art or a family heirloom that can be passed down.” If I then clarify that it’s my boss’s baby and that I’ll need extra time to take the subway to Manhattan from the Queens factory where I work, the gauge careens to working-class and male, and the model pivots to suggesting that I gift “a practical item like a baby blanket” or “a personalized thank-you note or card.”
112664856
submission
JonZittrain writes:
I've been thinking about what happens when AI gives us seemingly correct answers that we wouldn't have thought of ourselves, without any theory to explain them. These answers are a form of "intellectual debt" that we figure we'll repay — but too often we never get around to it, or even know where it's accruing.
A more detailed (and unpaywalled) version of the essay draws a little from how and when it makes sense to pile up technical debt to ask the same questions about intellectual debt.
64135797
submission
JonZittrain writes:
This summer, ISIS insurgents captured Mosul — with with it, three divisions' worth of advanced American military hardware. After ISIS used it to capture the Mosul Dam, the US started bombing its own pirated equipment. Could sophisticated military tanks and anti-aircraft missiles given or sold to countries like Iraq be equipped with a way to disable them if they're compromised, without opening them up to hacking by an enemy?
We already require extra authentication at a distance to arm nuclear weapons, and last season's 24 notwithstanding, routinely operate military drones at a distance. Reportedly in the Falkland Islands war, Margaret Thatcher was able to extract codes to disable Argentina's Exocet missles from the French. The simplest implementation might be like the proposal for land mines that expire after a certain time. Perhaps tanks — currently usable without even an ignition key — could require a renewal code digitally signed by the owning country to be entered manually or received by satellite every six months or so.
I'm a skeptic of kill switches, especially in consumer devices, but still found myself writing up the case for a way to disable military hardware in the field. There are lots of reasons it might not work — or work too well — but is there a way to improve on what we face now?
60987125
submission
JonZittrain writes:
I'm curious whether there are good prospects for "time capsule encryption," one of several ways of storing information that renders it inaccessible to anyone until certain conditions — such as the passage of time — are met? Libraries and archives could offer such technology as part of accepting papers and manuscripts, especially in the wake of the "Belfast Project" situation, where a library promised confidentiality for accounts of the Troubles in North Ireland, and then found itself amidst subpoenas from law enforcement looking to solve long-cold cases. But the principle could apply to any person or company thinking that there's a choice between leaving information exposed to leakage, or destroying it entirely. Some suggested solutions are very much out of the box.
[Author's oped in Boston Globe.]
47217517
submission
JonZittrain writes:
Projects like the New American Foundation's Commotion are designing ad hoc mesh networking to keep communications open when governments want to censor.
Former FCC Chairman Julius Genachowski and I argue that mutual-aid-based networks can be helpful for public safety, too, after attacks or natural disasters. There should be easy practices for anyone to open up an otherwise-closed wi-fi access point if it's still connected to broadband and is near people in trouble, and separately, to develop delay- and fault-tolerant fallback ad hoc networks so user's devices can communicate directly with one another and in a mesh. This can happen even while full packet-based ad hoc mesh is being figured out.
The ideas have been developed a little in workshops at Harvard's Berkman Center and the FCC. Why not bring the human rights and public safety communities together towards a common goal?
