Latest / AI Fire Daily / 🎙️ EP 283: Google’s Wild "Dreambeans" App & Gemma 4 12B Crushes Local Multimodal AI
Transcript
- 0:00it is uh it's pretty strange to think about sometimes
- 0:02yeah it really is your devices are constantly
- 0:05mapping your daily reality but they are also
- 0:08actively trying to reshape it right they record
- 0:11our habits quietly in the background then they
- 0:14subtly try to alter our future behavior and it
- 0:17happens without us even really noticing it welcome
- 0:19to this personalized deep dive we have a uh a
- 0:23truly fascinating journey ahead for you. We really
- 0:26do. We're covering some wild territory today.
- 0:28When you see an app crunching your life's data
- 0:31overnight, it seems, you know, kind of quirky.
- 0:34Yeah, it's a fun little consumer toy. Exactly.
- 0:36But you zoom out and you see the bigger picture.
- 0:39The sheer computing power required is driving
- 0:42a massive infrastructure race. A $700 billion
- 0:45race, to be exact. Right. And that massive race
- 0:48is creating a powerful counter -movement. We're
- 0:51seeing a shift toward... entirely private, open
- 0:54source, local models. Bringing all that power
- 0:57right back to your own laptop. Two sec silence.
- 0:59So we're going to begin at the extreme consumer
- 1:02edge today. This is where massive tech meets
- 1:04your morning routine. Yeah, right at the breakfast
- 1:06table. Google Labs recently dropped a highly
- 1:09experimental new app. It's currently available
- 1:12for both iOS and Android platforms. The project
- 1:15is officially called Dream Beans. Which, I gotta
- 1:18say, sounds like a complete fever dream. It really
- 1:21does sound a bit bizarre. But the technology
- 1:23running behind it is absolutely fascinating.
- 1:25Dream Beans relies heavily on your permission
- 1:27-backed Google data. It curates highly personalized
- 1:31stories every single morning. And we aren't just
- 1:33talking about a couple of alerts. It generates
- 1:36roughly 10 to 14 of these stories. And they aren't
- 1:39just simple text updates either. They're fully
- 1:41animated, narrative -driven stories about your
- 1:44actual life. It is wild. The app crunches your
- 1:48massive data footprint while you sleep. Right.
- 1:50It looks across your entire connected Google
- 1:52ecosystem. It checks your calendar, your maps,
- 1:55your search history. Then it distills everything
- 1:57into a concentrated morning drop. Let's look
- 2:00at a specific example from the source material.
- 2:02Okay, yeah. Say you have... Get a puppy on your
- 2:04calendar. The app actively sees that specific
- 2:07calendar event. It then generates a custom animated
- 2:10guide just for you. Right. It shows exactly what
- 2:13to expect that first week. It basically creates
- 2:15a customized narrative out of your own schedule.
- 2:18Yeah. And it pulls in relevant tips and weather
- 2:20data. It even integrates your location info seamlessly.
- 2:24It's like having a personal Pixar studio on your
- 2:26phone, just animating your daily to -do list
- 2:28every single morning. Yet is a perfect way to
- 2:31visualize the experience. It makes your mundane
- 2:33daily tasks feel much more engaging. Where did
- 2:36that bizarre name actually come from anyway?
- 2:39Product lead Gazda Osner explained the origin
- 2:41quite recently. Okay, what was the reasoning?
- 2:44Well, the dream part of the title is quite literal.
- 2:47The system does all its heavy lifting while you
- 2:49sleep. Ah, so it processes your digital life
- 2:52while you're actually dreaming. Exactly. And
- 2:55what about the beans part of the name? Let me
- 2:57guess. Coffee. You got it. It represents a freshly
- 3:01brewed cup of morning coffee. It's meant to kickstart
- 3:04your early morning with energy. Now, privacy
- 3:06is obviously a massive question right here. You're
- 3:09giving an app access to your entire digital life.
- 3:12That requires an enormous amount of personal
- 3:14trust. It definitely requires a massive leap
- 3:17of faith. But according to Google, the data stays
- 3:21localized to you alone. You maintain total control
- 3:24over your own personal information. Right. You
- 3:27can delete your entire history at any given time.
- 3:29Right now, consumer access is still pretty limited,
- 3:32though. Yeah, it's only available to U .S.-based
- 3:34users at the moment. And specifically, you need
- 3:37a Google AI Ultra subscription to use it. But
- 3:40there is a public waitlist open right now. Yep.
- 3:42Anyone with a personal Google account can actually
- 3:45join it. You can wait in line to see your life
- 3:47cartoonified. I'm curious about the underlying
- 3:49psychology of this tool, though. What's the actual
- 3:51goal of this quirky little app? Google essentially
- 3:54wants to give you a quick burst of inspiration.
- 3:57The idea is to deeply motivate you with these
- 3:59personalized stories. Then you confidently put
- 4:02the phone down to live your life. So a digital
- 4:04push to go live your actual life. Let's move
- 4:08away from personal data curation for a moment.
- 4:10We really need to look at the massive industry
- 4:12consequences here. Yeah, because we rely so incredibly
- 4:15heavily on these intelligence systems now. The
- 4:17shift from personal tools to global infrastructure
- 4:20is staggering. It completely changes how we interact
- 4:23with information daily. Anthropic recently shared
- 4:27a very interesting prompt system regarding this.
- 4:29It's specifically designed to keep humans actively
- 4:33thinking alongside AI. They're growing increasingly
- 4:36worried about our collective cognitive habits.
- 4:39Extremely worried. Letting an AI entirely think
- 4:42for you is quite dangerous. It can quietly weaken
- 4:45your own critical judgment over time. I still
- 4:47wrestle with prompt drift myself, honestly. I
- 4:50catch myself letting AI do my deep thinking.
- 4:53It honestly happens to everyone who uses these
- 4:56tools regularly. You just slowly start outsourcing
- 4:59your own critical thought processes. Right. And
- 5:01Anthropic desperately wants you to remain actively
- 5:03part of it. Yeah, they want you in the cognitive
- 5:05loop. We're also seeing intense new rules emerging
- 5:08globally. Over in the UK, new regulations are
- 5:11pushing back on Google. Content publishers are
- 5:14finally demanding more control over their work.
- 5:16They're tired of their data being endlessly scraped
- 5:18for free. For the first time, they can actually
- 5:21block their content. They can stop it from appearing
- 5:23in AI overviews completely. And they can also
- 5:26block it from AI mode entirely. Right. And block
- 5:29AI -generated search answers as well. It's a
- 5:31massive shift in intellectual property control
- 5:33and digital rights. Speaking of massive shifts,
- 5:36let's talk about Amazon for a second. They are
- 5:38rolling out a very strange new feature right
- 5:41now. Oh, yeah. When you search for items, things
- 5:43might look quite different. You might start seeing
- 5:46AI -generated product images first. These appear
- 5:49before you even see the real physical products.
- 5:52They're showing fake product concepts to help
- 5:55you shop. Yes. Fake product images. to help you
- 5:58find real ones. I have to push back on this Amazon
- 6:01feature, honestly. Why look at fake products
- 6:03to find real ones? It sounds weird, I know. It
- 6:06seems completely counterintuitive to the whole
- 6:08shopping experience. If I want a toaster, I want
- 6:10a real toaster. I don't want an AI hallucination
- 6:12of a toaster. It really does sound deeply confusing
- 6:15on the surface. But they want to help you visualize
- 6:18abstract lifestyle concepts. They think it helps
- 6:21you narrow down a specific aesthetic faster.
- 6:24Well, Google is also aggressively dealing with
- 6:26fake content right now. They're rolling out AI
- 6:29-powered fake call detection software. Which
- 6:32is becoming a highly crucial security feature
- 6:34right now. It can intelligently check if a trusted
- 6:37contact is really calling. It verifies if the
- 6:40familiar voice on the line is authentic. Voice
- 6:42cloning has become a massive global security
- 6:44problem lately. Scammers can perfectly mimic
- 6:47the voices of your loved ones. This defensive
- 6:50feature is now available on Android 14 devices.
- 6:53It runs entirely natively on Android 14 plus
- 6:56devices globally. Doing all of this requires
- 6:59an unbelievable amount of money, though. The
- 7:01financial scale of this specific industry is
- 7:04just staggering. Alphabet recently announced
- 7:07plans to raise $80 billion. They're expanding
- 7:10their AI infrastructure as rapidly as humanly
- 7:13possible. They actually expect to spend up to
- 7:16$190 billion. That's just on AI -related capital
- 7:20expenditures for this year alone. The broader
- 7:22tech sector numbers are even crazier to comprehend.
- 7:25Tech giants could collectively invest an estimated
- 7:28$700 billion. That's nearly a trillion dollars
- 7:32in raw physical infrastructure. It requires massive
- 7:35data centers, cooling systems, and enormous power
- 7:37grids. With all that power, societal impact becomes
- 7:40a huge concern. Why is Anthropic hiring an AI
- 7:44rule and law team for 345K? They urgently need
- 7:48to study how AI affects our society. Specifically,
- 7:51they're researching the impacts on courts and
- 7:54global elections. Building guardrails before
- 7:57the system breaks our civic institutions. Sponsor.
- 8:00We just talked about Alphabet spending $190 billion.
- 8:03What does that kind of astronomical money actually
- 8:06buy? Well, it buys immense computational power
- 8:09and vital research breakthroughs. It completely
- 8:11changes the physical limits of what computers
- 8:13can do. It buys the incredible ability to shrink
- 8:15massive cloud power. You can now fit that power
- 8:18locally on your own laptop. This brings us directly
- 8:21to a major open source release. Google DeepMind
- 8:24just officially dropped the Gemma 412B model.
- 8:27This model handles complex agentic workflows
- 8:30entirely locally. You can run it directly on
- 8:33your own personal machine. It's the exact model
- 8:35many developers have been waiting for. It bridges
- 8:38the huge gap between massive clouds and local
- 8:41hardware. DeepMind completely re -engineered
- 8:44how the model actually processes information.
- 8:47They fundamentally changed how it sees and hues
- 8:50the world. Vision and audio now flow directly
- 8:53into the main backbone. They no longer have to
- 8:55be awkwardly translated into text first. This
- 8:58drastically cuts latency across the entire computing
- 9:00system. It also reduces the heavy memory usage
- 9:03quite significantly. This is their very first
- 9:05mid -sized model to handle audio natively. This
- 9:09multimodal data processing is absolutely fascinating
- 9:11to me. It's like stacking Lego blocks of data.
- 9:14Yeah, that's a great way to describe it. You
- 9:16connect text, vision, and audio directly together
- 9:18without translating them. They simplified the
- 9:21underlying architecture so much to achieve this.
- 9:24raw audio signals project directly into the exact
- 9:26same space they seamlessly occupy the exact same
- 9:29space as text tokens do this allows it to translate
- 9:32and transcribe entirely offline it doesn't need
- 9:35to ping a massive server farm anymore beat. Imagine
- 9:40native raw audio processing running entirely
- 9:43offline on a regular consumer laptop. It's a
- 9:47truly massive leap forward for local computing
- 9:49power. The benchmark performance is also shockingly
- 9:53impressive for its size. It performs very close
- 9:55to Google's much larger, heavier models. Specifically,
- 9:59it rivals the massive 26B mixture of experts
- 10:02model. But it still comfortably runs locally
- 10:05on standard consumer laptops. That kind of efficiency
- 10:08was practically unheard of last year. It also
- 10:10features something called MTP drafters built
- 10:12right in. This keeps the generation of complex
- 10:15text incredibly fast. Let's quickly clarify that
- 10:18concept for a moment. What are MTP drafters exactly?
- 10:21They fundamentally change how the AI writes its
- 10:23responses. Helpers that guess upcoming words
- 10:26faster to speed up the whole system. Exactly.
- 10:28It predicts multiple tokens at once to save crucial
- 10:30time. It doesn't waste time agonizing over a
- 10:33single word choice anymore. The absolute best
- 10:35part is how accessible this model is. It's completely
- 10:38open under a permissive Apache 2 .0 license.
- 10:43the way it's on Hugging Face right now, you can
- 10:45also find them hosted on Kaggle immediately today.
- 10:48You can spin it up instantly in familiar developer
- 10:50tools, tools like Llama, LM Studio, and Llama
- 10:54.cpp. The developer community is already building
- 10:57truly amazing things with it. The open source
- 11:00world moves at an absolutely blistering pace.
- 11:02There is another tool mentioned alongside this
- 11:04release. What exactly is the Gemma Skills Repository
- 11:08mentioned in the source? It's an official developer
- 11:10toolkit launching alongside the main model. It
- 11:13helps you build autonomous, multi -step agents
- 11:15right out of the box. Basically a starter kit
- 11:18for building your own offline AI workers. Beat.
- 11:22We're in a highly weird transitional phase right
- 11:24now. The tech industry is spending nearly a trillion
- 11:27dollars collectively. They're rapidly building
- 11:29out massive, highly centralized cloud infrastructure
- 11:32globally. They desperately want to hoard our
- 11:35cloud data for morning cartoons. Apps like DreamBeans
- 11:38rely entirely on this massive centralized ecosystem.
- 11:42They need your data living on their servers to
- 11:45function properly. But they're simultaneously
- 11:47moving rapidly in the exact opposite direction.
- 11:50They're giving away the keys to the kingdom at
- 11:52the same time. They're democratizing offline
- 11:54open source power globally with Gemma 4. You
- 11:57can now run incredibly smart agents completely
- 12:00on your laptop. The deep tension between centralized
- 12:02cloud and local privacy is fascinating. We're
- 12:06watching. two entirely different philosophies
- 12:08battle for dominance right now. It really makes
- 12:10you wonder about the long -term future of these
- 12:12tools. If models like Gemma 4 get so incredibly
- 12:16good natively, if they can understand our daily
- 12:19reality completely offline, will we eventually
- 12:21abandon cloud -dependent apps like DreamBeans
- 12:24completely? Might we trade them for completely
- 12:27private, entirely local AI companions? It's a
- 12:30profoundly important question about who truly
- 12:32controls our data. And it's a question we'll
- 12:34have to answer very soon. Thank you so much for
- 12:36joining us on this deep dive. We really appreciate
- 12:39you exploring these complex ideas alongside us
- 12:42today.