Pebble Watch Is Back, Baby! - DTNS 4944  ## Metadata - Author: [[Daily Tech News Show]] - Full Title: Pebble Watch Is Back, Baby! - DTNS 4944 - Category: #podcasts - URL: https://share.snipd.com/episode/7e65133e-457a-4464-ab60-043b518117e9 ## Document Notes ## Highlights - **Pebble's Comeback** - Pebble founder Eric Migikovsky announced the smartwatch's return. - A small team is working to revive the original hardware and software experience. Transcript: Tom Merritt Need to know with the big story. All right, if you thought the Pebble watch was a throwback to a simpler time in the world of wearables, well, you would be right. But people are very nostalgic, I'm sure you know this, Tom, for the trend-setting smartwatch brand that launched way back at 2012. Those people might actually be very happy to hear that the Pebble watch is coming back, baby. Eric Mikikovsky is Pebble's founder, and he announced on the RePebble launch page that the smartwatch is returning thanks to the efforts of a small team who's working closely with Eric to rekindle the glory days of the original hardware. Some of the features of this new smartwatch effort and what they're chasing anyways include an always-on e-paper screen, Pebble's notably long battery life, those wonderful hardware Buttons that you don't need to look at the screen in order to know what you're doing, and a robust and comprehensive user experience that includes many of the functions that we've come To expect from smartwatches since Pebble's launch. Because you've got to remember, Pebble was kind of there in the beginning, and so set the trend for a lot of other wearables that we use today. Migakowski says the new watch is in development and will have similar specs and features as the Pebble watch, along with quote, some fun new stuff as well. That's not all. Migakowski also announced that Pebble OS, which is the operating system that drove the unique smartwatch line, is being open sourced by Google. And just to kind of explain this a little bit, Fitbit bought Pebble's IP back in 2016 when Pebble kind of started to go away, followed by Google's acquisition of Fitbit in 2021. So essentially, Google has all of that Pebble IP. Megakovsky says he was able to convince, quote, friends at Google to make the open source transition of Pebble OS a reality. And this also means that anyone who wants to create their own Pebble-like watch running the open source OS is free to do so. So that's kind of neat. Or any type of hardware, really, that's running the OS. This punctuates another key component of the news that Migikovsky pointed out, that the new OS is, quote, hackable, meaning custom watch faces can be expected. Migikovsky says that an open source Pebble compatible app called Cobble is already running on Android, coming soon to iOS. Existing Pebble apps and watch faces will also be compatible with the new offering. There is no timeline, though, set for the new hardware currently. ([Time 0:02:40](https://share.snipd.com/snip/ab56815a-118a-4620-ba37-8c69b98b870d)) - **Niche Appeal** - The new Pebble watch likely won't outcompete Apple Watch or Wear OS. - It serves a niche market of Pebble enthusiasts like Brian Brushwood. Transcript: Tom Merritt Tom, what do you think? Did you have the Pebble watch way back in the day? I never had one. Jason Howell No, I never had one. My friend Brian Brushwood is still an enthusiast. He held onto his Pebble until the very end. So there's so many interesting things about this. It is interesting to see Mikakovsky go off Beeper to this because it's kind of a fight with Apple in both cases, right? Beeper clearly was a fight with apple in trying to include iMessage compatibility i wonder if he learned a lot there to help with compatibility here however he told Ars Technica that Compatibility would be even worse now because apple has like clamped down even more on that sort of thing. But he clearly is up for the fight on that end. And more importantly than that, I think, is just the leaning into a niche market where you can tell that he's not worried about this becoming the biggest smartwatch in the world. He wants to serve the people like Brian Brushwood who are really excited about Pebble and would love to have it back. Yeah. Tom Merritt I mean, the Pebble, what is the right word? The Pebble excitement never really died down from the people who were real true fans. I mean, over the years, even ever since it went away, we would have stories come up on All About Android and later Android Faithful where the Pebble story continued. And really, when you think about where wearables and smartwatches and all that stuff is right now, you really do have to credit Pebble for so much, which is kind of a testament to how unique They were back in 2012. This was prior to the Apple Watch. It was doing a lot of things that people wanted out of a wrist computer before there was a true market for it. And that's what was really kind of remarkable about this up and coming hardware platform that didn't start on Kickstarter. I mean, it was just one of those examples of like, holy cow, you captured lightning in a bottle at a moment in time where this whole platform, this whole landscape of wearables was really Just forming. So we have a lot to thank Pebble for when it comes to the wearables on a wrist. Yeah. Jason Howell The OLED wearables we have now are always on. They have longer battery life. Some of the things that Pebble was so far in front of because of the e-ink screen, uh, aren't necessarily as big of a problem, but still you can't get days of life out of a current smartwatch. You, you, you're always on is got some compromises here and there, especially when your battery starts to run low, it's going to be like, yeah, we can't be always on anymore. I think he's right that, yes, again, I'll repeat it. You're never going to overtake the Apple Watch or Wear OS or anything like that, but you are going to find a group of people who really want this. And I like that that's what matters here. Not every enterprise has to become the winner in its space. You can build really good businesses serving people who want a particular thing. And that's pebble to a T. So I don't know, 15 years later, 13, 13, I'm bad at math, 13 years later, still, that's a long time. Still um i hope i hope it takes off i and i actually texted uh brian once as soon as i saw this story i'm like you're gonna get this right uh so i'll let you know if he if he does of course he is Tom Merritt If he was a huge fan i could i could actually already hear brian's like voice going are you kidding of course i am i ordered three yesterday and they're not even being built yet. Jason Howell I'm going to build my own. So we want to hear from the Pebble fans, especially if you're like, I might not do it this time. Or if you're somebody who never was into Pebble, but you're like, you know what? Now in 2025, I want this exactly thing. Feedback at DailyTechNewsShow.com. Let us know what you think. ([Time 0:05:20](https://share.snipd.com/snip/be2159cd-c7ae-43a9-a8b1-7e55d55cb470)) - **DeepSeek's Janus Pro** - DeepSeek's new Janus Pro models offer efficient multimodal imaging. - The largest, Janus Pro 7BM, reportedly outperforms Stable Diffusion and DALI 3 in benchmarks. Transcript: Tom Merritt There is more we need to know today, so why don't we get right to the briefs, Tom? Deep Seek. Jason Howell Have you ever heard of it? Have you ever heard of anything but it for the past two days? But there's something new with Deep Seek already. A new model family was released on Hugging Face. It's called Janus Pro. It's a set of multimodal models for imaging, ranging in size from one to seven billion parameters. The largest model is called Janus Pro 7BM, which apparently can beat stable diffusion and DALI 3 in Genival and DPG benchmarks. According to DeepSeq, anyway, I haven't seen any independent verification of that yet, but I'm sure that's coming soon. The models can analyze and create new images, though analysis is limited to resolutions of 384 by 384, at least at the current moment. That's tiny. Tom Merritt That's tiny. OpenAI announced ChatGPTGov, a self-hosted version of its platform designed for government agencies to confidently handle, quote, non-public sensitive information. The tool will be available through an agency's own Microsoft Azure commercial cloud or Azure government community cloud. That makes it more, even more allied off than ChatGPT Enterprise anyways. The self-hosted solution is compliant with security frameworks, including IL5, CJIS, ITAR, and FedRAMP. Hi, but only for public information at this point anyways. Testing of the product is expected within the month with wider availability in the near future. Jason Howell Yeah, this is why you cozy up to a new president so that you can sell them a new product. Yeah, but ChatGPT Enterprise is very secure. It gives a lot of control to the enterprise. This gives all the control to the enterprise. This has got to be solid. And if you're concerned about this sort of thing right now, only public information supposed to go in here. Now, granted, your weakest point is always humans. But yeah, this is a very big win for open AI. And it could be beneficial in a lot of agencies in speeding up things and making things more productive. Although it'll be difficult to speed up everything when you're like, oh, wait, I can't put that in there. Or I think more often, I don't know if I should put this in here. So I'm not gonna, because I don't want to get in trouble. Tom Merritt Yeah, you got to have certainty around that. And, you know, OpenAI is in deep trouble because of DeepSeek. So, you know, they've got to they've got to carve out new new pathways. ([Time 0:13:08](https://share.snipd.com/snip/2001e687-a932-4ea9-a50b-7137763e440f)) - **GPU Demand** - DeepSeek's efficiency doesn't diminish the need for powerful GPUs like Nvidia's H100. - Increased efficiency enables more complex AI tasks, driving demand even further. Transcript: Tom Merritt Bet you can't guess what it is. It starts with D, ends with K, and has Eepsie in the middle. With the pop with the popularity of the super efficient deep seek skyrocketing over the weekend i think if you go to tech meme it's like it's not even 100 deep seek it's like 220 um a lot Of folks bigger yes exactly a lot of folks thought this might spell doom for the need for NVIDIA chips. We actually found someone who can explain why that might not be so. Here's you, Tom, talking with Andrew Main. Jason Howell I'm talking with Andrew Main, co-founder of Interdimensional. And were you the original prompt engineer at OpenAI? Is that fair? Andrew Main We never had an official title prompt engineer at OpenAI, but I was probably the first person hired because of that. Jason Howell Okay. Andrew Main Yeah. And not like it was necessarily super competitive. I'm like, oh, I'll work on this. Like, okay, cool. Sure, sure. Jason Howell So I found out about DeepSeek a couple of weeks ago when you mentioned it on the attention mechanism, it's not new. It's not overnight. But tell us what it is and why you think it suddenly got all this attention. Andrew Main So you have DeepSeek, which is the Chinese lab, the company that built the DeepSeek models, which you've heard about. And they've got a couple models. So over a month ago, they released V3, which was their new model, which competed with the GPT-4. And it was worth the attention because they came up with some very, very good optimization strategies. I mean, very clever. And anything I say after this, I don't want to sound like I'm downplaying what they've done, because they did some amazing optimization things, were able to figure out how to train it On, you know, using restricted et cetera. I also pointed out that if you asked the model what model you are a couple times, it would say GPT-4. If you asked it for its API, it would spit out the GPT-4 API. And there is a habit among some companies to bootstrap their training data by basically taking a ton of outputs from other models, which arguably takes compute and should be accounted For. But it strongly looked like, hey, they did some very, very clever strategies to make this. And now they've created the V3 model. Then they just released the R1 model last week, which is their version of 01. And it's a reasoning model. It takes time to think about things. And by all accounts, R1 is a great model. I've seen people having wonderful results, whatever. And it's a fraction of the cost to use than OpenAI's 01. Although we don't know the real cost of OpenAI's 01 because OpenAI determines that. I would say by comparison, though, Google has their own flash thinking model, which is a very, very, very inexpensive thinking model that they're using offering for like, I think free On their API right now. So we know there are some really efficient thinking models out there. Jason Howell And this is what's called chain of thought, right? Where they sort of iterate on the answer before they give a final answer. Andrew Main You give it a question, it breaks it down. It says, okay, I've got to think about this. And it might be if you give it a problem, like, you know, write a paragraph about memory without using the letter E, it'll write something, go, I can try it again. And R1 will fail at that in my experience, but it will try to solve it. And you can see that process that works. Jason Howell Okay. So chain of thought isn't new, but they are one of the – I guess fewer people doing it than LLMs. Yeah. Andrew Main It's a chain of thoughts. The O1 paper was published a few months ago, and so we have other people who are doing chain of thought-like things. The big thing was the efficiency is that they're doing chain of thought much, much, much more cheaply than O1 is. But I want to give context in that part of the reactions, Ben, is this idea that they trained this model for only like $5 million. But even if you read their paper, they talk about one component of it took that. There was a lot of other work into it. And also, if you're training on top of even Hugging Phase, which is trying to recreate this right now, is going to use the V3 model to output a bunch of data to then train on top of, which Is kind of going to question like, well, where does that data come from originally? Jason Howell So there's verifiable efficiencies that you see that you can say like, yeah, they've actually done some really good stuff that make this work with less compute. On the end, we can see there's their claim that they were able to train it more efficiently that we can't really verify. We just have to take their word for it that like they weren't using H100s or other types of chip. Andrew Main Yeah. I don't, I don't know. There's been some, yeah, some people questioning that they're using secret data. And that's, I mean, it's a Chinese company possible, but I, I'm going to accept them that they're being sincere because on the training part, hugging face is trying to recreate this Right now. And we'll find out pretty quickly. Jason Howell OK, so that would find out real quickly. What would we have to see for DeepSeek to live up to the hype that is specifically like you won't need as many GPUs? Let's not get to the investor part of it yet, but just, yeah, in the future, nobody will need as many GPUs or as powerful GPUs to do the equivalent amount of training here. Andrew Main For training, I don't know that they're saying that, you know, because, you know, they use something like 800,000 high quality examples to train the reasoning model, but we, and that Came from their other model, but we don't know where, like, give you an example. OpenAid has GPT-4 and they have GPT-4 Mini. GPT-4 Mini is what you basically call, what you consider, a distilled model. You take GPT-4 and then you have it output millions of higher quality stuff. Instead of just random internet text, it's questions and answers. How do I calculate this? What's the answer to that? So instead of it, if you want to learn a language, there's two ways to do it. I could give you 100 books in Japanese and 100 books in English that are the same book, and you could take time to figure out how to translate. Or I give you one book that's a translation book, and you'll learn much faster. Think of a distilled model as basically having the bigger model create that translation for it so it doesn't have to learn. And so you still need that bigger model and that data originally to come from. Got it. Jason Howell So is everyone going crazy by saying, oh, NVIDIA is not going to be as valuable now because of this? Andrew Main It's literally like somebody came up with a more efficient light bulb in 1899. And we're like, we don't need electricity providers anymore. I advise VC firms and people like this. And I would tell you that in my opinion, the people making a lot of these investment decisions know a lot less than I do. And I have friends that work at labs who are shaking their heads going, don't people understand that this is just going to increase even more usage? And full disclosure, I am an N shareholder and even more so today, not financial advice. But I don't think Wall Street's never been good at predicting where ([Time 0:21:47](https://share.snipd.com/snip/71deaceb-8d24-4e5f-b725-5a1889cd2d37)) - **Humanoid Robot Training** - Humanoid robots are easier to program for human tasks using videos of human motion. - America's Test Kitchen demonstrated this with a barista arm trained to create latte art. Transcript: Tom Merritt We end every episode of DTNS with some shared wisdom. And today, Andrew, not Andrew Bain, is helping us understand why we want robots that look like humans. Jason Howell Yeah. He writes, humanoid robots may seem creepy, but as America's Test Kitchen's CES video demonstrates, it's faster to program them with videos of other humans doing those jobs. They visited with a barista arm that was trained in a real barista, and it did a bang-up job at Latte Art, as well as the battery oven that you could talk about on the show, which we won't Be talking about the battery oven. But yeah, good stuff from Frozen Cusser, aka Andrew. Thank you for pointing that out. You can train the algorithm, if it's humanoid, to do human things based on human motions. Tom Merritt I guess that makes sense. That makes a lot of sense. Thank you, Cusser of the Frozen Variety. Thanks to Andrew Main and, yes, of course, Frozen Cusser for contributing to today's show. Thank you for being along for Daily Tech News Show. The show is made possible by our patrons at patreon.com/DTNS. DTNS has a live version called DTNS Live on YouTube and Twitch. You can find details on that and more on dailytechnewsshow.com. We will talk to you tomorrow. ([Time 0:33:20](https://share.snipd.com/snip/f89bab45-ee5c-4792-88fb-a25f701e613c))