The technology itself isn’t new. Deepfake software to “nudify” a photo has been around for at least five years, and is increasingly becoming a major cyber-bullying problem in high schools. Photoshop has been used by cottage perverts to make non-consensual porn for decades. And of course you can find examples of edgy online predators “playing” with women’s avatars and representations all the way back in the LambdaMoo days. The only thing new about the Grok image editor is how deeply integrated it is with the X’s social media features: it’s almost tailor-made to sexually harass women.

All of which is to say that this “trend” should surprise no one — not even a company that has made it clear that trust and safety aren’t a priority. Given Musk’s response (“So what if Grok can put people in bikinis?”) and general posture on harassment I expect the only thing the company didn’t expect was just how massive the blowback and outrage would be. The company’s current “solution” is to limit the sexual harassment features within X conversation streams to premium subscribers, though it’s reportedly still available for free in the Grok app and on their website. Given this kind of response, I think the deepfake porn site formerly known as Twitter moniker is going to stick.

Before I get into what they did, here’s a quick refresher on how we see color. White light is made up of a broad spectrum of wavelengths, but our eyes only have three types of color detectors (called cones), each sensitive to its own overlapping range of wavelengths. What we see as color is the relative response of just those three type of cones at a certain point, so light around 575 nm triggers the L cones the most, M cones a little less and hardly any S, which we perceive as yellow. But you can get the exact same response from a mixture of light around 549 nm and 612 nm (green and red), which is how RGB monitors get away with displaying practically any color with just three color subpixels. It’s also why we can perceive magenta, a color that isn’t in the rainbow at all but results from a combination of blue and red light (that is, a high response from S and L cones but low response from M cones). Notice that it’s possible to trigger just the S cones with short-wavelength light and just the L cones with long-wavelength light, but there’s normally no way to trigger just the M cones without also triggering L or S cones as well because of the overlap.

The Berkeley and UW team has developed a prototype system, called Oz, that scans a subject’s retina to identify the exact location of S, M and L cones and then uses a laser to excite just the ones to produce a desired color at any given location. In theory such a system could render every possible color, including ones that are impossible to see in nature because they’re outside the range of S/M/L responses you can get with natural light. In practice they estimate up to two thirds of the light “leaks” over to neighboring cones, but that’s still enough to produce a large range of colors including ones outside the natural gamut. One such color, the one produced by stimulating only M cones and no others, they’ve named olo (from 010 — get it?).

In theory only five subjects in the world have seen olo, which they describe as a “blue-green of unprecedented saturation, when viewed relative to a neutral gray background.” But that’s not very satisfying. If you’ve just discovered a new color naturally the first thing anyone is going to ask is “what does it look like?” — it’s much nicer if you can answer “here, take a look” instead of “come back to my lab and I’ll shoot you in the eye with a laser.”

Luckily, I think I’ve found a way to see olo without any of the complex set-up. The Oz team creates olo by selectively stimulating M cones in a region of the retina, but we should be able to get the same effect by first staring at a magenta color field for 20-30 seconds (which suppresses responses from the S and L cones) and then quickly shifting over to a pure green. The difference is analogous to additive vs subtractive color: Oz works by stimulating only the M cones, while color adaptation involves the suppression of the stimulus color (in this case magenta, the complement of green).

Stare at the center dot for 30 seconds. Then without moving your eyes, move your mouse pointer into the square (or tap on mobile). You should briefly see a super-saturated blue-green image.

Cute effect, but is it olo — are you and I actually seeing the same color the Oz team sees with their device? Short answer is… maybe? For an effect that’s so readily testable there’s still surprisingly little consensus about exactly what causes negative after-images, or even exactly what colors one is expected to see. Explanations range from simple cone adaptation to the currently dominant theory that after-images are caused by some kind of opponent process between different color responses higher up the processing chain (probably in the retinal ganglion cells). If something like the cone adaptation model is correct then I’d definitely expect the two methods to produce the same color, modulo how much leakage is in Oz vs. how much the S and L cone responses are suppressed in the after-image. But even if the processing is higher up the chain it wouldn’t surprise me if the effects are essentially the same, because regardless of the underlying mechanism it’s clear that when an after-image is mixed with a real image (e.g. when viewed against a colored background) the result is as if the original stimulus was partially subtracted from the background. That’s why the after-image from magenta appears green on a white background, but greenish-yellow on a yellow background and brownish-red on a red background.

One way to test whether the two colors really are the same would be to do the same tests the Oz team did but with the after-image + green, using a tunable laser plus white light to match the perceived color. Alternatively one could turn the experiment on its head and use Oz itself to match the perceived color directly, and see how close it gets to their olo.

I plan on reaching out to the Oz team to see what they think, and I’ll update if they write back.

Does it make toast?

One researcher who had just joined our wearables group said he wanted to design a “general purpose wearable.” I asked him “does it make toast?” He looked at me a little perplexed and said no, of course it doesn’t. “Then it isn’t really general purpose,” I responded. “Now, what’s the set of activities that you really intend for it to be used for, so we can design something that covers that set?” Whenever I’m designing a new device or application I try to come up with one to four usage scenarios. These scenarios form the corners of the design space. They also let me ask “if I built this, would anyone care” before I put lots of effort into designing.

(On a related note: whenever a project says “we’re trying to get our device into the hands of developers so they can figure out how to use it,” I worry. It’s fine if you already have some clearly compelling applications and you just want to find more, but too often it’s a tacit admission that you don’t really know why anyone would need your technology and you’re just hoping someone else can figure it out before your funding runs out.) 

Approximate the future.

One of the things the MIT Media Lab was good at was looking at technological trends and asking “OK, so where does this wind up in 10 years?” The wearables group was a good example: we knew that computers were getting smaller and more powerful, that microphones could soon be the size of a tic-tac, displays the size of your fingernail, etc., and we wanted to explore what we could do in that world. So we mocked it up as best we could, attaching displays the size of harmonicas to glasses frames or hats and putting computers in large purse-bags with camcorder batteries. It looked strange and was unwieldy, but we knew if we could look past the limitations that we knew would be solved by Moore’s Law then we could understand the challenges and opportunities that would remain.

Communications is always the killer app.

I remember one tutorial that Thad and I taught on wearable computing had a slide showing a cell phone, a GPS, a walkman, a camera/camcorder and a PDA (which back then meant notepad, calendar, to-do list and maybe a calculator), and we explained that soon all these devices would be converging into one. Thad and I disagreed on which of these applications would be the most important though — which application will convince people to buy the device, and which ones can you skimp on when you need to make trade-offs? Thad thought it would be home electronics like the camera and the music player. I remembered back to reading about Minitel and similar systems that always seemed to be subverted to be a communications device and declared that (human-to-human) communication would always be the killer app. Not all the interface decisions have gone my way (I was particularly surprised when the iPhone got rid of the physical keyboard, trading off more screen for harder input) but I’d still contend that if people had to choose between a phone/IM/SMS/Facebook device or all other smartphone apps they’d choose the former (especially if you throw the Web into that list).

What you should know once you have your degree.

Not really related to research but it’s something I tell people thinking of going into grad school:

When I got my Bachelor’s degree, I knew everything.
Then I got my Master’s degree, and realized that I actually knew nothing.
Then I got my PhD, and realized that no one else knew anything either.

• 1970s: Spoof email
• 2000s: Virus that sends email to all your contacts, virtual kidnapping scam
• 2010s: Fake caller-ID and robocalling, twitter bots, fake facebook profiles, cat phishing
• 2019: Deep fake videos, bot-assisted cat phishing, IRS-impersonation scam
• 2023: Impersonate voice and phone number in realtime phone call <– You are here
• 202?: Impersonate loved one in realtime video
• 202?: Automated voice-chat scam robocalls

She has more interesting experiments with DALL-E, AR and video processing over at her Instagram. (h/t to Boing Boing)

[Migrated from Google+]

[Migrated from Google+]

[Migrated from Google+]

Introduction to Artificial Intelligence – Fall 2011

[Migrated from Google+]

Scripting Gmail to snooze emails (so they come back later) is cool! (Actually, scripting Gmail is cool).
http://gmailblog.blogspot.com/2011/07/gmail-snooze-with-apps-script.html

But I bet Boomerang (http://www.boomeranggmail.com/) doesn’t like it so much. On the other hand, how smart was it to create an entire business that could be obviated by a single 48-line script?

Gmail Snooze with Apps Script – Official Gmail Blog

[Migrated from Google+]