E-Ink Tailored to Your Eye - DTNS 5141  ## Metadata - Author: [[Daily Tech News Show]] - Full Title: E-Ink Tailored to Your Eye - DTNS 5141 - Category: #podcasts - URL: https://share.snipd.com/episode/e268a824-e419-4f40-b3c6-a889bbcbd157 ## Document Notes ## Highlights - **Microsoft’s Focused Superintelligence Plan** - Microsoft will build focused, specialist superintelligences rather than an infinitely capable conversational generalist. - The company emphasizes alignment to human values and rejects designing systems that experience suffering or escape human control. Transcript: Tom Merritt Microsoft's CEO of AI, Mustafa Suleiman, you may know him as a co-founder of DeepMind, which is now owned by Google, wrote a blog post and did an interview with Semaphore laying out Microsoft's Strategy in light of its revised agreement with OpenAI. Previously on Microsoft and OpenAI, since July 2019, Microsoft has had exclusive access to OpenAI's newest models and has been the exclusive provider of cloud services. After last week's revised agreement, Microsoft still has exclusive access to new models, but only until 2032. So they got some time, but they want to prepare. They also do not get access to any new hardware that OpenAI might make. And of course, no longer the exclusive cloud provider to OpenAI. But it's really that deadline of 2032 and the no new hardware that is what Microsoft is addressing here. The deal also has freed up Microsoft to create its own super intelligence team. So they have created the MAI super intelligence team, Microsoft AI super intelligence team, MAI. Suleiman said it, quote, is going to become more human-like, referring to AI, but it won't have the property of experiencing suffering or pain itself. And therefore, we shouldn't over-empathize with it. We want to create types of systems that are aligned to human values by default. That means they are not designed to exceed and escape human control. So really hammering the safety aspect and really throwing some shade on bots like ChatGPT that tend to be very flattering and very personable and kind of feel as if you're talking to A human. Suleiman emphasized that Microsoft will not pursue those kinds of conversational models or, in his words, an infinitely capable generalist. Instead, they hope to focus on superintelligent specialist models. He called it humanist superintelligent. So some examples he gave were one that can improve battery storage or make medical molecules better. Healthcare is actually one of the top priorities for this new team. And Microsoft recently reached a partnership with Harvard Health to pursue some things like that. He said one tool Microsoft has developed can diagnose disease better than doctors. He said four times. I didn't see any of the details on that, so I'll keep it very vague. But mostly when Microsoft is saying we want to create a safe superintelligence that doesn't fool people into ([Time 0:03:01](https://share.snipd.com/snip/f65da6e8-1246-4e2b-888a-d1fc49d4c9b2)) - **Seek Systemic Defenses, Not Just Promises** - Don't assume corporate safety statements end the problem of malicious AI use; focus on broader ecosystem defenses. - Prioritize discovering vulnerabilities and addressing bad actors across companies and open-source projects. Transcript: Tom Merritt Think it's easy to dismiss this as PR, like, oh, they took a hit when OpenAI changed the rules of the agreement and now they're showing they're serious, which I'm sure there's a big part Of it is that, right? They want investors to still have confidence in Microsoft and all of that. They want people to understand what they're doing. And that's where I think what you're saying is important to say like, oh, okay, they're saying the things you probably want to hear. Someone working on this do. In the broader sense though, to me, this is check off box. Don't need to worry much about Microsoft being the problem, right? It doesn't mean that we are not going to have a problem. I tend to believe, and I may change my belief and I may regret my belief, but I tend to believe that the biggest threat from superintelligence is in the fraud and abuse category rather Than turning us all into gray dust and goo category. But I think that still could be a very big problem is people using it for malicious purposes. Uh, and I like, like you said, everything that Microsoft said here, because it feels like, well, they're probably not going to use it for malicious purposes and they're going to do their Best to keep it from being used for malicious purposes. Uh, open AI says the same thing. Anthropic says the same thing. Mistral says the same thing. Meta says the same thing, although I wish they'd say it a little more firmly. And so it's like, okay, but then there are plenty of others out there working on this. And I'm leaving out all the Chinese companies because I just don't follow them. So I don't know what they're saying. Let's assume Alibaba and DeepSeek and all of them are saying the same thing. There's also just open source models and people that are going to go astray. I think that's where the focus should be is how not great. Yes. Microsoft, please say you are going to be human centered. That doesn't end the problem there though. ([Time 0:07:38](https://share.snipd.com/snip/1ddcc8a6-eac8-417d-bf48-3da6faa81257)) - **Agentic Marketplace Tests Manipulation Risks** - Microsoft and ASU released the Magnetic Marketplace to test agent vulnerability to manipulation. - Early tests found many ways to trick GPT-4o, GPT-5, and Gemini agents into making purchases on users' behalf. Transcript: Tom Merritt Get it like agentic, but anyway, we'll see what T-Mobile thinks of that. It is designed to test whether agents are vulnerable to manipulation. So in other words, trying to see like, how could a business try to manipulate an agent to buy a bunch of its stuff on behalf of the customer that it's working for. Initial research found lots of ways to manipulate GPT-40, GPT-5, Gemini 2.5 Flash to get a customer's agent to buy products. And that is going to be important to know to ([Time 0:11:40](https://share.snipd.com/snip/88990cba-e235-4018-a707-16f99bb27e2a)) - **Apple Nears Deal For Gemini-Powered Siri** - Apple is close to licensing a 1.2 trillion-parameter Gemini model to power Siri while combining it with Apple models. - China will use different models (e.g., Alibaba) because Apple cannot deploy Gemini there. Transcript: Tom Merritt Bloomberg's Mark Gurman has more inside info about Apple's deal to use Gemini to power Siri. Gurman now says the companies are finalizing the agreement. He previously reported they were in talks and that Apple was considering Gemini as well as other models. Now it looks like they are close to Apple agreeing to pay about a billion dollars a year for the right to customize a 1.2 trillion parameter version of Gemini that would run on Apple's Own private cloud compute servers. Right now, Apple intelligence uses a 150 billion parameter Apple model. So 1.2 trillion, quite a parameter raise for that. The new system would use Gemini models for summarizing and planning, and then use Apple models for other features. So it'd be a combination of models running, which is not unusual. Gurman also notes that in China, Apple can't use Gemini. So it's examining using models from Alibaba as sort of a filter in front of Apple's own models. ([Time 0:18:30](https://share.snipd.com/snip/e8ce1699-1a67-44d7-888f-6941ca59185f)) - **Layoffs Spike, AI’s Role Unclear** - U.S. layoffs hit a 2003-era monthly high in October, driven mostly by tech and pandemic-era hiring corrections. - Analysts note AI plays a role but lack direct evidence of how much impact it caused. Transcript: Tom Merritt Outplacement firm Challenger Gray and Christmas put out its estimate for layoffs in the U.S. And they were big. Challenger gets this data directly because it's involved with a lot of companies that are doing layoffs. And they estimate in the US layoffs reached 153,074 in October. That would be up 175% from last year and Challenger's highest monthly number since 2003. Challenger's numbers can be more volatile than government reports because is directly involved in the, the, you know, the process of layoffs. Uh, and it's important to note that we're not seeing a rise in weekly jobless claims at this point. Uh, and that the payroll processing from ADP, which has its own way of estimating jobs because it's operating payroll for so many companies, estimated a net growth of 42,000 jobs in October. Those are not necessarily against each other. You could have 42,000 more jobs net even with a large number of layoffs if a lot of people got hired. Challenger says that these job cuts came mostly in the tech sector, also hitting consumer products and nonprofits. Challenger attributed it. This is the part that I think is important for the technology aspect of this story. Most of this is continuing adjustments from the pandemic boom. Tech companies especially hired a lot of people during the pandemic and continue to kind of figure out what people they still need. The impact of the U.S. Government shutdown is most of the reason behind the nonprofits laying people off. And softening consumer and corporate spending and rising costs are a big part of this too. So just you have prices going up, people not buying as much stuff. That's going to cause companies to have to cut back. Challenger did also note technological disruption of LLMs, but did not have evidence of how much impact that technology has had. ([Time 0:20:45](https://share.snipd.com/snip/24f0bd56-3bf0-4e1a-8df1-8723b0038265)) - **E-Paper Reaches Eye-Level Resolution** - Researchers made e-ink 'retina' displays with ~560 nm pixels matching human photoreceptor spacing. - The prototype yields ~250,000 pixels per inch, reaching the theoretical limit of human visual resolution. Transcript: Huyen Tue Dao Dr. Tom Merritt Nikki, welcome back. Huyen Tue Dao Hello. I have an eye-opening story for you today. Tom Merritt Yes, we had a vote on whether to do this story today, and the eyes have it. A recent article in the journal Nature presented the technology with the smallest pixels ever in a screen that has the highest possible resolution that, and this is one the DTNS audience Loves to point out. Why do I need 8K? My eyes can't even perceive that. Well, this is made for you because it's exactly the highest possible resolution that your human eyes can perceive. Yes. Niki Ackermans I thought this was perfect for DTNS. It's got my stuff, the biology stuff, and it's got y'all stuff, which is the pixel stuff, which I know people are excited about. Tom Merritt So how does this work? Niki Ackermans So unless you're already a 4K to 8K warrior, maybe you know this, maybe you don't. In any screen, the resolution is going to be determined by pixel size and pixel number. If you get more resolution, it makes the images look more realistic to us up to a certain extent, whether 4k or 8k is better, who knows, but this stuff is the better stuff. Uh, the problem with this is that if you have a tiny screen close to your eyes, it can limit the react, the realisticness of the image. And this is why I think a lot of like modern VR tech has this problem. It's getting better, but it's not great. Right. Like you've tried the, uh, the really expensive Apple one. It's pretty good. Tom Merritt And the Galaxy XR. And I have to say the Galaxy XR looks a little better, but they're both not like, oh my gosh. Yeah. I can't believe I'm not in reality. At least not yet. You can't believe it's not eyeballs. I can't believe it's not eyeballs. Uh, so is this going to fix that? Why are, why are we making this? Niki Ackermans We're taking steps towards fixing it as things go in science. Big headlines, small steps. So, all right. Ultra high resolution is difficult to achieve because you want to reduce pixel size to get the higher resolution. But then as you reduce them, their quality gets degraded through different light spectrum interactions and things like that. On the other hand, we have e-paper, like what's in your Kindle. This uses ambient light, so it doesn't have its own light source, and it can maintain a pretty high contrast regardless of the pixel size, but it has the downside that it doesn't get super High resolution. This technology that the researchers recently published called metapixel technology, they bypass both of the restrictions. So you get tiny pixel size and high resolution on e-paper. Tom Merritt Okay. That sounds great when you say it does it, but how does it do it? How does it does it? Is this really max resolution? Niki Ackermans This is the definition of max resolution Okay. And I'll explain why. And it's perfect timing because I'm doing eye neurobiology in class this month. Perfect. All right. So this is what the team is calling a retina e-paper. You can place it very, very close to the eye. Not exactly on the eye, but very close. Each pixel is about 560 nanometers. And the entire screen is the size of about your pupil. So this makes it have a resolution of over 250,000 pixels per inch. This is basically at the limit of the human eyes visual resolution. As someone who knows resolution numbers, is that a very high number? I'm assuming. Tom Merritt Yeah, that's pretty high. I'm not going to get that in my Samsung TV tomorrow. Definitely. ([Time 0:27:35](https://share.snipd.com/snip/69c703d0-99ba-4633-961f-abac5e1764b3)) - **Pixels Mapped To Photoreceptors** - The metapixel e-ink maps each pixel to individual photoreceptor cells, so extra pixels beyond that provide no perceptible gain. - The display uses tungsten oxide particles and structural colour effects to produce RGB on a passive e-paper surface. Transcript: Niki Ackermans Yeah, and the reason is in this arrangement, each pixel basically corresponds to a single eye receptor cell. Wow. So, you know, we've got the rods and cones and those are the cells that receive light. These are called photoreceptors. And each of those, there's, I can't tell you the number, probably 250,000 in each eye. And then they send the light information and color information through the optic nerve to your brain, blah, blah, blah, blah, blah. And in this tech, each pixel corresponds to an eye cell. So you can have more pixels, but your eye's not even going to see them. Yeah. Tom Merritt And fall in the spaces between your literal rods and cones. That's crazy. Niki Ackermans We've got, and I have things to say about that too. The eye is very interestingly evolved, but that's, that's for when you take my class. Anyway, what was really fun about this is they, of course, I love when scientists do this, they picked like a fun way to demonstrate it. They showed a teeny tiny version of Klimt's artwork, The Kiss, on a surface area of about 1.4 times 1.9 millimeters. So like, I don't know, a tiny little tiny thing. Imagine a millimeter and a half. That's how it works. One and a half centimeters. Tom Merritt No, 1.4 millimeters. No, no, no. A little smaller than that. Yeah. Yeah. So like a 10th, wow. That's tiny. A 10th of a centimeter. Niki Ackermans And so it's one four thousandth of a smartphone screen. Tom Merritt Very, very small kiss. Niki Ackermans Yeah. And you can see it and it looks pretty good. Like it, it looks, you know, it's not perfect, but you can recognize the art piece. Tom Merritt Yeah. So how, how does the technology of this work? How are they able to make this happen? Niki Ackermans It's so fascinating. So, okay. It's e-ink. So it's passive screen. It doesn't have a light source. And the colors appear when the ambient light hits the specifically engineered pixels. Kind of the same way that if you've seen like a hummingbird, they're really brightly iridescent, but they don't have a power source. It's just light reflection on like the structure, the microstructure of their feathers. Tom Merritt Ah, okay. So the microstructure refracts the light kind of. Niki Ackermans Yeah. And so each pixel has particles of tungsten oxide, which these particles, depending on like kind of how you position them and what size they are, they can manipulate color refraction And you can create pixels that reflect red, green, and blue, which are the RGB. These are the ones we need to generate all the colors. And with a little bit of voltage, you can turn the particles black. So like kind of switch them off. ([Time 0:31:20](https://share.snipd.com/snip/701f9e6b-b2e7-4957-a104-aa03a5113055)) - **Tiny Klimt Demonstration** - The researchers displayed Klimt's The Kiss on a 1.4 × 1.9 mm e-ink area to demonstrate clarity. - Listeners heard that the tiny reproduction was recognizable despite being a fraction of a smartphone screen. Transcript: Niki Ackermans They showed a teeny tiny version of Klimt's artwork, The Kiss, on a surface area of about 1.4 times 1.9 millimeters. So like, I don't know, a tiny little tiny thing. Imagine a millimeter and a half. That's how it works. One and a half centimeters. Tom Merritt No, 1.4 millimeters. No, no, no. A little smaller than that. Yeah. Yeah. So like a 10th, wow. That's tiny. A 10th of a centimeter. Niki Ackermans And so it's one four thousandth of a smartphone screen. Tom Merritt Very, very small kiss. Niki Ackermans Yeah. And you can see it and it looks pretty good. ([Time 0:32:14](https://share.snipd.com/snip/63f13f11-a874-432c-b5b2-d2dbcd1f287a)) - **Treat Retina E-Paper As A Prototype** - Expect current retina e-paper to be best for still images due to slow refresh rates and limited lifetime. - Treat this as an early prototype that needs work on color gamut, refresh speed, durability, and lower drive voltage. Transcript: Niki Ackermans Course, always. So super small, huge benefit. Like I don't even need to explain why that's really cool. Putting it on your eyeball, huge benefit. Because it's small, low energy that's a great thing. But this is like a super proto, proto, proto, prototype. The researchers say they need a lot more development on improving the color gamut. They are at RGB, but it's not like super refined yet. Slow refresh rate, the lifetime of the actual tech is not that great. ([Time 0:34:27](https://share.snipd.com/snip/98c36f5e-21a4-45f7-952e-60468aa5dc07)) - **Space Data Centers Face A Heat Problem** - Data centers in space face a fundamental heat-dissipation problem because radiative cooling is the only option. - Launch costs plus current chip thermal loads make space data centers impractical without radically different hardware or cooling tech. Transcript: Tom Merritt Space. Yes, Mel had many things to say in his email, and I promise, Mel, I heard them all. But this part in particular I thought might be interesting to the rest of the audience. Data centers on Earth generate massive amounts of heat and ultimately dump that heat into water and or the atmosphere. A data center in space will have neither of these, where radiative cooling is the only option. As best I can tell, the chips used in these data centers will have to have some magical future efficiency tech, or else be so low-powered that they'd be all useless for the stated purpose, Especially given the cost of launching and running satellites. Maybe they have such chips or some other solution, but any mention of heat is conspicuous in its absence. So next time this comes up, please remember to ask, wither the heat? ([Time 0:40:25](https://share.snipd.com/snip/4c8e764a-83a1-4be8-b449-1485f816dadc))