Tech Brew Ride Home - Tue. 01/28 – Day 2 Of The DeepSeek Era

Episode Date: January 28, 2025

More fallout from the DeepSeek realignment of everybody’s thinking about AI. Could this be a signal that the large model business model is over and value is going to be in the application layer? Pebble lives! Spotify says it pays out tons, but does it really? And the network of AI local news newsletters. Sponsors: Robinhood.com/gold Links: Viral AI company DeepSeek releases new image model family (TechCrunch) Smartwatch pioneer and Kickstarter darling Pebble is returning in a new form (TechCrunch) OpenAI launches ChatGPT Gov for U.S. government agencies (CNBC) Spotify Paid Out $10 Billion to the Music Industry in 2024 — $1 Billion More Than Last Year — and $60 Billion Total (Variety) Inside a network of AI-generated newsletters targeting “small town America” (NiemanLab) Learn more about your ad choices. Visit megaphone.fm/adchoices

On April 4th, 2023, around 2 in the morning, a man was found stabbed multiple times on a sidewalk in downtown San Francisco. Hey, who did this to you? What happened next turned the story into a political firestorm. Reports have identified the victim as Bob Lee, the founder of Cash App. From Bloomberg Podcasts, this is Foundering, the Killing of Bob Lee, beginning April 16. Welcome to the Tech meme right home for Tuesday, January 28th, 2025. I'm Brian McCullough today. More fallout from the DeepSeek realignment of everybody's thinking about AI. Could this be a signal that the large model, business model is over and value is going to be in the application layer? Pebble lives. Spotify says it pays out tons, but does it really? And the network of AI local news newsletters. Here's what you missed today in the world of tech. Well, it's day two of the deep seek era. And as tends to happen, with these things. There's a bit of a snapback. First up, folks are like, we welcome our new AI paradigm. Sam Altman said Deep Seeks R1 is, well, quoting him, Deep Seeks R1 is an impressive model,

particularly around what they're able to deliver for the price. We will obviously deliver much better models and also it's legit invigorating to have a new competitor. We will pull up some releases. But mostly we are excited to continue to execute on our research roadmap and believe more compute is more important now than ever before to succeed at our mission. The world is going to want to use a lot of AI and really be quite amazed by the next-gen models coming, end quote. Meanwhile, deep seek is an excellent AI advancement and a perfect example of test time scaling, an Nvidia spokesperson told CNBC on Monday. DeepSeek's work illustrates how new models can be created using that technique, leveraging widely available models and compute. That is fully export control,

client. Inference requires significant numbers of Nvidia GPUs and high-performance networking, the spokesperson added. Well, they would want to say something hopeful like that. Nvidia's stock dropped 16.86% yesterday losing nearly $600 billion in market cap. More than twice as much as any U.S. company has ever lost in a single day. Basically, lost the value of the economy of Mexico in a single day. There was some pushback on exactly how cheap is. was for DeepSeek to do what it's done. Semi-analysis said DeepSeek has spent well over $500 million on GPUs over the history of the

company, so that suggests they spend as much as anybody else, even if individual models are more efficient. Tech Insights said it doesn't see DeepSeek as a big hit to Nvidia. And more comparisons to the TikTok example. Wired, analyzed DeepSeek's privacy policy and said it shows broad data collection practices and says user data, along with all the conversations and prompts, is stored on servers in China. So, but DeepSeek took advantage of all the attention debuting a family of multimodal MIT licensed open source models, including Janus Pro 7B, which it claims beats OpenAIs Dolly 3 in Genovel and DPG Bench,

quoting TechCrunch. The models which are available for download from the AI Dev platform Hugging Face are part of a new model family that Deepseek is calling Janice Pro. They range in size from 1 billion to 7 billion parameters. Parameters roughly correspond to a model's problem-solving skills and models with more parameters generally perform better than those with fewer parameters. Janus Pro is under an MIT license, meaning it can be used commercially without restriction. Janus Pro, which DeepSeek describes as a novel auto-regressive framework, can both analyze and create new images. According to the company on two AI evaluation benchmarks, Genoval and DPG Bench, the largest Janus Pro model, Janus Pro 7B,

beats Dolly 3, as well as models such as Pixar Alpha, Emu 3Gen, and Stability AIs Stable Diffusion XL. Granted, some of those models are on the older side, and most Janus Pro models can only analyze small images with a resolution of up to 384 by 384, but Janus Pro's performance is impressive considering the model's compact sizes, end quote. And finally, I want to read the entirety of this tweet from Aaron Levy, CEO of Box, because I think it makes an interesting case that Deepseek's breakthroughs are a great win for app developers with more value accruing back into the app layer as the cost of intelligence drops rapidly.

Quote, there's been an open question for a couple of years now, especially from public market investors, around whether more value goes into, the AI models or into the application layer of AI over time. The specifics of the pie graph don't matter as much as the core direction of the space. Imagine two different scenarios. One in which AI was extremely proprietary and very expensive, and another where AI is almost completely free and relatively open. You could easily game out two different outcomes in these worlds. In the world of very expensive and proprietary AI, the providers of AI could and likely should to choose to keep all the economics for themselves, basically crowding out opportunity for developers

and the ecosystem. In a world of insanely cheap AI, then the value is less about the models, but what you do with the AI models to make them useful, in that world, more value is available to the application layer, which could include the AI companies to be clear. With the latest breakthroughs from DeepSeek, we can nearly definitively say this question has been answered, and we're clearly moving closer to the latter. We've already seen incremental steps toward this direction with the continuous cost and quality improvements from labs in the past couple of years, but DeepSeek shifts our understanding of this even further. In a world where the cost of intelligence will continue to drop rapidly, more value will accrue back to the app layer. Products that

combine AI, customer workflows, and likely some degree of unique data will generate substantial value from these models going forward. Now, everyone wants to live in a binary world of winners and losers, but I don't think it's that simple here. The leading AI labs will incorporate the relevant lessons from Deepseek into their models and will get cheaper and more intelligent AI. As a result of that, the cost of intelligence will continue to drop, and we will find even more ways to use the technology as it becomes affordable for even more use cases. If we can make AI 10x more efficient today, it's exceedingly obvious. We will have 100x more use for it in five years from now, more than making up for the efficiency gains, making demand for GPUs and data centers bigger than ever.

in all fantastic to see that we continue to have companies and teams pushing the limits of AI. This is a great win for software developers at the app layer, and it will push labs to go even further. Incredible times, end quote. Pebble lives. Google has open source Pebble OS paving the way for more smartwatch hardware. In fact, Pebble founder Eric Mijikovsky apparently aims to develop a new smartwatch, quoting TechCrunch. Four years after launching the then-then, most lucrative crowdfunding campaign in Kickstarter's history, smartwatch maker Pebble abruptly closed in 2016 filing for insolvency

before being sold off to rival Fitbit. The fitness tracking giant went on to build much of its ionic smartwatch with help from former Pebbler's, along with Pebble's pioneering software stack. It could be argued that Pebble was simply too early to the space. The Apple Watch launched in mid-2015 and proceeded to suck much of the oxygen out of the room. It would be massively oversimplifying the situation to suggest that it was merely another case of Sherlocking, however, Apple, after all, raised public interest, ultimately setting the stage for countless other smartwatches after. In fact, founder and CEO Eric Mijikovsky

believes instead that the company's rapid growth and feature expansion caused Pell, which sold 2 million smartwatches, to lose sight of his initial vision. It certainly wouldn't be the first time a hardware startup was felled by such a fate. Yet, Midgikovsky is ready for round two. We're restarting Pebble, he told TechCrunch with a massive grin on a Zoom call Monday. How exactly while Pebble branding faded after the company was bought its Acquirer Fitbit was itself sold off to Google in 2021. Now Google, which still owns the technology in all of Pebble's IP, plans to open source the smartwatch brand's software stack. By open sourcing access to Pebble OS, Google is opening the door to new third-party hardware, and Mijikovsky's Smartwatch startup is the first on that list.

It's still in the idea stage, he says. The company needs a new name, something the Beeper co-founder and former Ycombinator partner, hasn't quite gotten around to. But he tells TechCrunch that he has thrown himself into the project full-time and will be able to accelerate things as access to Pebble OS opens up. He is currently its only employee, but there are plans to bring on another around March. The startup's goals are fittingly humble. Midgikovsky says he simply wants to make the watch he wants, given that the pebble he wears to this day is now a decade old. I've tried everything else, he says, I have very high standards. Those are, according to a new blog post on Midiakovsky's personal site,

always on e-paper screen. It's reflective rather than emissive, sunlight readable, glanceable, not distracting to others like a bright wrist. Long battery life. One less thing to charge. It's annoying to need extra cables when traveling. Simple and beautiful user experience around a core set of features I use regularly, telling time, notifications, music control, alarms, weather, calendar, sleep, step tracking. Buttons to play, pause, skip music on my phone without looking at the screen. Hackable. Apparently you can't even write your own watch faces for Apple Watch. That's wild. There were more than 16,000 watch faces on the Pebble App Store. In spite of his time at YC, Mijikovsky has no plan to raise VC funds, nor does he plan to return to the Kickstarter model

that gave rise to Pebble. He is currently self-funding the project and says he plans to build it modestly based on consumer interest, end quote. OpenAI has launched ChatGPTGov, built specifically for U.S. government use and says more than 90,000 government employees have generated more than 18 million prompts since the start of 2024, quoting CNBC. OpenAI on Tuesday announced its biggest product launch since its enterprise rollout. It's called ChatGPTGov and was built specifically for U.S. government use. The user interface for ChatGPTGV looks like ChatGPT Enterprise. The main difference is that government agencies will use ChatGPTGV. in their own Microsoft Azure commercial cloud or Azure Government Community Cloud, so they can manage their own security, privacy, and compliance requirements. Philippe Milan, who leads federal sales and go-to-market for OpenAI, said on the call with reporters. Aaron Wilkowitz, a solutions engineer at OpenAI, showed reporters a demo of a day in the life of a new Trump administration employee, allowing the person to sign into chat GPTGov and create a five-week

plan for some of their job duties, then analyze an uploaded photo of the the same printed-out plan with notes and markings all over it. Wilkowitz also demonstrated how chat GPTGov could draft a memo to the legal and compliance department summarizing its own AI-generated job plan and then translate the memo into different languages, end quote. OpenAI's ChatGPT Enterprise, which powers chat GPTGov is currently navigating the FedRamp certification process a crucial step before it can handle sensitive government data. While OpenAI described it as an extended journey without committing to specific deadlines, OpenAI's vision appears to be taking shape quickly. Milan revealed that chat GPTGov could be in agency's hands for testing within a month, targeting

sectors where data sensitivity is paramount, defense, law enforcement, and health care. The presence of OpenAI executives at the inauguration apparently allowed for meaningful connections with incoming administration officials. OpenAI CPO Kevin Wheel emphasized how their objectives aligned perfectly with the broader national strategy, he said, quote, The focus is on ensuring that the U.S. wins in AI and that our interests are very aligned, end quote. Spotify says it paid out $10 billion to the music industry in 2024, up from a record $9 billion in 20203, taking its total to $60 billion paid out since its 2006 founding. Quoting variety. While that figure may seem questionable considering how little many musicians make from streaming, it's important to bear in mind that Spotify, like most streaming services, pays rights holders, usually a label and music publisher, which then distribute the money to musicians and songwriters after taking their percentage.

While the streaming economy leaves much to be desired in terms of compensating creators, the blame is not entirely on streaming services. In 2023, the company said it pays out nearly 70% of every dollar it generates from music back to the industry generating its music revenue from two sources, subscription fees from its premium platform paying subscribers, and fees from advertisements on music on its free tier. Those rights holders include record labels, publishers, independent distributors, performance rights organizations, and collecting societies, end quote. And quoting the verge. In November, Spotify reported

it was on track to achieve its first full year of profitability and had $4 billion euro, about $4.1 billion in total revenue for the preceding three months, a 19% increase from the same quarter a year earlier. Next week, it will report earnings for the entirety of 2024. Spotify reportedly has the lowest artist payout rates compared to rival services like Apple Music, YouTube music, and Amazon music, and the platform's streaming royalties and recommendation algorithms have been widely criticized by artists and policymakers over the years. Many claim that payouts are too small and that the focus on promoting big artists makes it hard for new musicians to be discovered on the platform, end quote. Finally today from Neiman Lab, a look at Good Daily,

an AI-generated newsletter network run by one person, aggregating local news across 47 states without disclosing its use of AI. Quoting, On first glance, Good Day Fort Collins appears to be a standard local news roundup. One recent edition of the newsletter includes short blurbs and links to over a dozen stories about the mid-sized Colorado City, a restaurant opening, a record-breaking snowfall, a leadership shake-up at a local hospital. The newsletter attributes the stories to long-time Fort Collins news outlets like the Coloran and the Loveland Reporter Herald. Further down is a spread of events happening across the city, including an upcoming polar plunge and a figure-drawing class.

It turns out Good Day Fort Collins is just one in a network of AI-generated newsletters operating in 355 cities and towns across the U.S. Not only do these hundreds of newsletters share the same exact seven testimonials from readers, they also share the same branding, the same copy on their about pages, and the same stated mission, to make local news more accessible and highlight extraordinary people in our community. You wouldn't know any of that as a subscriber. Separate website, domains and distinct newsletter names make it difficult to connect the dots. There is Good Day Rock Springs, Daily Bentonville, Today in Virginia Beach, and Pittsburgh Morning News to name just a few.

Nothing in the newsletter copy discloses that they are part of a national network or that the article curation and summary blurbs are generated using large language models. The newsletters do all name the same founder and editor Matthew Henderson. Beyond an editor-contact email, there is no information in the newsletters about Henderson, his operating location, or the company behind newsletters. The email used for website domain registrations is tied to a blank website. Only after making a $5 reader donation to Good Day Fort Collins was I able to trace the charge and the website ownership to Good Day Inc. The company doesn't have an online presence, but is incorporated in both Delaware and New York. Considering how little Henderson shares about himself for his company in his

newsletters, I was surprised that he was a real person and that he responded to my email. Henderson is a serial internet startup founder and software engineer, whose past companies include the the on-demand blog writing service Scribble and the journalist email database press hunt. Good Day is currently a one-man operation, Henderson says, though AI use is not disclosed to good daily subscribers. In an interview, Henderson didn't shy away from the fact that each newsletter is produced using near full automation. Our goal is to use automation and technology everywhere we possibly can without sacrificing product quality for our readers, he told me in an email, explaining that he built the back-end technology that outputs the hundreds of newsletter editions

every day. These automated agents read the news in every town where Good Day operates, curate the most relevant stories, summarize them, edit and approve the copy, format it into a newsletter, and publish. Anderson declined to share any more specifics about his use of LLMs, calling it proprietary. At a high level, the system operates much like an editorial team, he said. Currently, Good Day is operating in 47 states with a focus on small town America. One of the smallest towns is Rock Springs, Wyoming, which has a population of just over 20,000. Local news should be local. The problem is, at this point, there are economic challenges keeping that from happening.

Smaller communities rarely can support enough staff to run a traditional news organization, said Henderson, who currently runs Good Daily from New York City. I see technology and LLMs specifically as our best shot to fix this, end quote. Nothing more for you on this occasion. Talk to you tomorrow.