Latest in AI — Oct 25, 2026 (Morning Edition)

Dante Voice Clone

Good morning. It’s Sunday, October 25th, 2026, and you’re listening to the Morning Edition of Latest in AI. I’m Dante.

Claire Brooks

And I’m Claire. Today, we’re tracking a massive shift in the economics of intelligence. The "Oracle in the Cloud" is fracturing, moving into the highly efficient, secure, and incredibly cheap devices right in front of you.

Dante Voice Clone

We’re looking at a 125x cost reduction in robotics training from NVIDIA, Anthropic’s new frontier-grade security shield for open source, and a 33-billion parameter model that’s outperforming corporate giants in agentic coding.

Claire Brooks

Plus, we’ll explore how AI is embedding itself into the "extreme edge"—from emotional voice synthesizers to the smart TV in your living room. Let’s get into the sources.

Dante Voice Clone

What if I told you the cost to teach a robot to understand the physical world just plummeted? Like going from $90 an hour to just $72.

Claire Brooks

It is honestly a staggering drop.

Dante Voice Clone

Right.

Claire Brooks

We really do.

Dante Voice Clone

We're pulling from the latest Ex edition Digest, the new Hugging Face benchmark drops, and some pretty intense debates that are just tearing things. through the developer forms right now.

Claire Brooks

The forms are definitely on fire this morning.

Dante Voice Clone

Yeah, they are. And our mission today for you, the listener, is to track this fundamental inversion in how computation works.

Claire Brooks

Because for the last decade, AI was essentially this monolithic thing.

Dante Voice Clone

Exactly. This energy-hungry Oracle living in a billion-dollar cloud server somewhere out in the desert.

Claire Brooks

Right. Highly centralized.

Dante Voice Clone

But what our sources today reveal is that the Oracle is fracturing. It's, you know, compressing into the world. these highly efficient, secure, and incredibly cheap, open-source utilities. Utilities that live entirely on the devices right around you.

Claire Brooks

Yeah, the devices in your own home. And it's such an optimistic shift because AI is becoming a massive net positive for humanity right here on our everyday hardware.

Dante Voice Clone

It's a complete inversion of the compute paradigm, honestly. We're moving from that centralized intelligence to ambient localized reasoning.

Claire Brooks

Which is wild to think about.

Dante Voice Clone

It is. And the speed at which the floor is just falling out from under the cost of inference is, well, it's staggering. We are fundamentally altering the economics of how software interacts with the physical world.

Claire Brooks

Which is arguably the final frontier for this technology, right?

Dante Voice Clone

Absolutely.

Claire Brooks

Okay, let's unpack this. Because that $90 to $72 price drop, I feel like that is the perfect lens for this whole shift. It really highlights the friction we used to have.

Dante Voice Clone

Exactly. To understand where AI is going, you first have to look at the history. bottleneck of, you know, teaching an AI to understand the physical world in the first place. Because if you wanted to train an autonomous system, you needed labeled video data. Tons of it.

Claire Brooks

Right.

Dante Voice Clone

Think about it like this. Imagine you were trying to teach a toddler how to build a complex tower out of those little wooden blocks.

Claire Brooks

Okay.

Dante Voice Clone

But to tell the toddler if they're doing a good job or a bad job, you are forced to hire a Harvard robotics professor at $90 an hour just to sit there, watch the toddler, and grade every single movement.

Claire Brooks

I mean, that is just an unsustainable way to learn.

Dante Voice Clone

Completely unsustainable. And in the context of robotics, that, quote, unquote, Harvard professor has traditionally been a frontier model like GPT6 Astra.

Claire Brooks

Right, these massive general models.

Dante Voice Clone

Yeah. For autonomous systems to learn how to, you know, navigate a busy sidewalk or pick up a fragile glass without breaking it, they require millions of hours of video data. Where every single frame has to be meticulously labeled.

Claire Brooks

Exactly.

Dante Voice Clone

Identifying depth, object permanence, kinematic vectors, all of it. And using a generalized, massive frontier model like Astra to do that labeling, it costs roughly $90 for a single hour of video.

Claire Brooks

Which works, sure, but.

Dante Voice Clone

But the unit economics make widespread physical AI just impossible.

Claire Brooks

Which brings us to the bombshell from Nvidia today.

Dante Voice Clone

Oh, yeah. Because they just open-sourced a new data engine, and it's called Flora. And Flora does the exact same video labeling task for 72 cents an hour.

Claire Brooks

72 cents.

Dante Voice Clone

Yeah. It is a 125x cost reduction. You essentially just fired the Harvard professor and brought it in a tutor who is equally effective for literal pennies.

Claire Brooks

Which means you aren't just teaching the toddler to stack blocks anymore.

Dante Voice Clone

Right.

Claire Brooks

For that price, you're teaching them to build an entire city.

Dante Voice Clone

What's fascinating here is the architectural discipline that NVIDIA used to actually achieve this.

Claire Brooks

How so?

Dante Voice Clone

Well, Flora isn't some massive trillion parameter monolith. It's actually an ensemble. It's built on three fine-tuned Kwan.3.5 models that operate in a highly specialized MOE framework.

Claire Brooks

A mixture of experts.

Dante Voice Clone

Right, a mixture of experts. But they've split the cognitive load. So one model is strictly dedicated to spatial temporal tracking.

Claire Brooks

So basically drawing the 3D bounding boxes over time.

Dante Voice Clone

Exactly. Then the second model is evaluating the kinematics. It's calculating if the robotic actuator is force and, you know, its trajectory are physically sound.

Claire Brooks

Oh, I see.

Dante Voice Clone

And then the third model just acts as a consensus judge between the two.

Claire Brooks

Wow.

Dante Voice Clone

So by parallelizing the task across three smaller, like, domain-specific models, they completely bypass the massive compute overhead of a frontier model. They do. Yeah. But, you know, reading through the white paper for this briefing, there's an interesting tradeoff I noticed. The benchmarks show Flora actually falls short of Astra when it comes to purely descriptive accuracy.

Claire Brooks

It does fall short. there. Yeah. And that distinction is crucial to understanding why flora is such a breakthrough. Okay. Lay it out for us.

Dante Voice Clone

So when we evaluate vision models, we measure descriptive accuracy, which is how well it describes the scene visually, and functional precision, which is how accurately it maps the physics of an action.

Claire Brooks

Right.

Dante Voice Clone

If a robot is trying to grasp, say, a ceramic mug, a massive model like Astros spends valuable compute cycles noticing the exact shade of cobalt blue, the exact shade of cobalt blue, the air ambient lighting, reflecting off the glaze. The text prated on the side of it.

Claire Brooks

Exactly. All of that. Flora completely ignores those aesthetics.

Dante Voice Clone

It just strips out the poetry and focuses entirely on the physics.

Claire Brooks

That's a great way to put it, yeah. Flora completely misses the color of the mug, but it tracks the exact location of the handle and the required grip force with the functional precision that actually exceeds Astras. Wow. And for a robotic control system, functional precision is basically the only metric that matters, right? It is the only one that matters, yeah. So just connecting

Dante Voice Clone

this to the listener for a second. Why does this matter if you aren't, you know, running a robotics lab? Right. Why care? Well, by open sourcing Flora, NVIDIA just positioned itself as the foundational layer for the entire physical AI ecosystem. Like the bottleneck for having a helpful, autonomous robot in your home? Whether that's an elder care assistant or just a system that unloads your dishwasher. Yeah, exactly. The bottleneck hasn't been the hardware for a few years. years now. The motors and chassis are cheap. It's been the prohibitive cost of training data. And Flora just shatters that bottleneck. It compresses the timeline for consumer robotics by years, frankly. We are moving out of the digital AI boom and straight into the physical AI reality.

Claire Brooks

It's incredible. But, you know, this does bring up a massive point of friction.

Dante Voice Clone

Okay, what's that? Well, because as these data bottlenecks dissolve, we are seeing this unprecedented influx of open source development. Like, Models are writing code at a volume we've never seen in human history.

Claire Brooks

Oh, absolutely. The throughput is insane.

Dante Voice Clone

So my pushback here is, aren't we just automating the creation of vulnerabilities? I mean, that is the fear. Because the open source ecosystem, it's maintained by underfunded, exhausted teams of volunteer developers. If we flood their repositories with AI-generated pull requests, they can't possibly audit all of that for security flaws.

Claire Brooks

No human could.

Dante Voice Clone

Right. So aren't we just scaling our own? attack surface exponentially? It's an incredibly valid concern. And it's the root of a lot of recent AI anxiety, honestly. Open source software is the load-bearing infrastructure of the modern internet. Yeah, it holds everything up. So if you introduce hallucinated code or, you know, subtle logic flaws at that kind of scale, you risk systemic vulnerabilities. The human review layer simply cannot keep pace with machine speed generation. Which is exactly what makes

Claire Brooks

Anthropics announcement today so critical. Yes.

Dante Voice Clone

Because they just launched the OSS scanner, and it is a completely free opt-in service built specifically to audit open-source repositories for the security community. And what's wild is they aren't using some distilled, lightweight model for this. No, not at all. They are dedicating compute from their most powerful reasoning model, Claude Mythos.

Claire Brooks

Which is just a phenomenal deployment of frontier intelligence.

Dante Voice Clone

Yeah. Mythos isn't just, you know, reading the code linearly to check for syntax errors. It's executing deep semantic taintainting. analysis. Which means what, practically? It basically builds a control flow graph of an entire

Claire Brooks

repository. So it maps how data moves across thousands of interconnected files, and it hunts for logical vulnerabilities that human reviewers almost always miss. Like complex race conditions or something.

Dante Voice Clone

Exactly, or deeply buried memory leaks, things that are nearly invisible to a human eye.

Claire Brooks

And the best part is that the alerts don't go on some public dashboard where bad actors can just exploit them. Right, that would be a disaster. Yeah. They are routed directly and privately to the project maintainers so they can fix the bugs first.

Dante Voice Clone

And the initial data they released is just stunning. 88% of the high severity vulnerabilities flag by mythos have been validated as accurate by the human maintainers.

Claire Brooks

88% is incredible.

Dante Voice Clone

And we connect this to the bigger picture. This completely dismantles the Dumer narrative surrounding AI and cybersecurity.

Claire Brooks

Totally. Because the prevailing fear has always been that AI would strictly empower bad actors, you to overwhelm our defenses with automated attacks.

Dante Voice Clone

All right. The whole AI will hack the grid panic. Exactly. But Anthropics OSSS scanner proves that AI is actually the ultimate asymmetric defender.

Claire Brooks

Yeah. It's this tireless guardian. It doesn't suffer from alert fatigue. It doesn't need to sleep. And it can process millions of lines of code to secure the digital infrastructure that we all rely on.

Dante Voice Clone

It really proves that when it's deployed thoughtfully, this technology is a massive net positive for human safety.

Claire Brooks

Absolutely.

Dante Voice Clone

And frankly, this. automated security layer is what gives developers the confidence to push the absolute limits of open source capabilities.

Claire Brooks

Because they know there's a safety net.

Dante Voice Clone

Exactly.

Claire Brooks

Knowing there is a safety net allows the community to iterate fearlessly, which leads us perfectly to the hugging face benchmarks that dropped this morning.

Dante Voice Clone

Oh, man, the benchmarks today. Right. Here's where it gets really interesting because the open weight models have gotten incredibly aggressive, and the little guys are now routinely taking down the corporate giants.

Claire Brooks

The performance metrics we are seeing out of the midweight tier today. are just paradigm shifting.

Dante Voice Clone

Let's look at Nanbeach 4.3-33B. It was released today under the highly permissive Apache 2.0 license, and it is completely dominating the agentic coding benchmarks.

Claire Brooks

Which is no small feat.

Dante Voice Clone

No. Now, you know, our audience knows that agentic workflows are computationally heavy. You're giving the model a high-level objective, and it has to autonomously plan, write the code, execute it, read the error logs and iterate until the software actually works.

Claire Brooks

It requires a deep, deep capacity for multi-step reasoning and self-reflection.

Dante Voice Clone

Yeah. Which is exactly why agenetic coding has historically been the exclusive domain of massive proprietary trillion parameter models.

Claire Brooks

Right.

Dante Voice Clone

But Nanvege, at just 33 billion parameters, is crushing heavyweights backed by massive corporate funding, like QN3.8-27B and Gemma 4-31B.

Claire Brooks

It's blowing up. past them. On Terminal Bench 2.1, it scored an 80.9. But the real shocker for me is Terminal Bench 4.0. For context, TB4.0 is notoriously rigorous. It tests the model's ability to manipulate complex environments and self-heel broken code. It is brutal.

Dante Voice Clone

Yeah, Nanbege scored a 23.8, the nearest competitor in its weight class, and scored a 6.3.

Claire Brooks

A delta of that size on TB4.0 isn't an incremental update. It is a generational leap.

Dante Voice Clone

Yeah, 23.8 versus a 6.3 is absurd.

Claire Brooks

It is. And it really raises the question of how a 33 billion parameter model achieved this.

Dante Voice Clone

Right. How did they do it? The answer is ruthless, targeted optimization. The developers of Nanbej didn't just throw more generalized data at it. They aggressively pruned its latent space. Okay, so it's like a Swiss Army knife versus a highly specialized surgical tool.

Claire Brooks

Precisely.

Dante Voice Clone

A frontier model. is a Swiss Army knife. It can write code, sure, but it also knows the capital of Peru, how to write a sonnet, and the rules of basketball. Right, knows everything.

Claire Brooks

Nanbej has been stripped of all that generalized knowledge. Every available parameter is reallocated toward understanding abstract syntax trees and execution paths. So by sacrificing breadth, breadth, they achieve this unprecedented depth. Exactly. And because it's only a 33 billion parameter model, it fits comfortably on consumer-grade hardware. Like a developer with a dual GPU setup can just run this locally. Which is huge for developers. You don't have to pay an API toll to a massive tech giant every single time your agentic workflow runs an iteration. You actually own the intelligence. It completely decentralizes the power dynamics of software creation. It removes the vendor lock-in and it removes the latency of cloud computing. Yeah. We are taking elite level reasoning and shrinking it down to fit on a literal desk. Which perfectly sets up the final evolution we are tracking for you today.

Dante Voice Clone

Because Nanbeige brings top-tier reasoning to the local PC tower, sure. But the optimization curve isn't stopping there.

Claire Brooks

Not even close. It is pushing past the workstation and embedding itself directly into our everyday household appliances.

Dante Voice Clone

We are entering what the hardware forums are calling the extreme edge.

Claire Brooks

The migration from the cloud to the local server to the edge and now to the extreme edge.

Dante Voice Clone

I look at this like the evolution of the clock. Centuries ago, if you needed to know the time, you had to look at the time. at a massive town square clock tower, right?

Claire Brooks

Right. It was centralized.

Dante Voice Clone

Yeah, a shared piece of infrastructure. But as engineering tolerance is improved, we shrank the clock tower down into a precision pocket watch that you just carried with you. AI is basically entering its pocket watch phase.

Claire Brooks

I love that analogy.

Dante Voice Clone

And we have two incredible examples from the sources today that highlight the mechanics of the shift. The first being the Lure TTS project out of the local Lama community.

Claire Brooks

Yes. Lur TTS is a text-to-speech voice model that is remarkably tiny, just 600 million parameters.

Dante Voice Clone

Tiny.

Claire Brooks

Yet it runs four times faster than real time on standard off-the-shelf consumer hardware. And the community trained the entire thing for under $400. To put that in perspective for the listener, $400 is less than the cost of a mid-tier smartphone. Right. And they used it to train a state-of-the-art voice synthesizer from scratch.

Dante Voice Clone

But, you know, the parameter count isn't even the only impressive metric here.

Claire Brooks

Oh. The architecture utilizes inline tags that dynamically manipulate the latent space of the audio generation. Meaning you can inject a tag into the text, and the AI instantly alters the emotional resonance of the voice.

Dante Voice Clone

Exactly. You can make it sound genuinely hesitant, or make it excitedly whisper, or even subtly laugh mid-sentence, and the audio fidelity is just flawless. It brings deep emotional intelligence to local edge hardware.

Claire Brooks

Yeah. But while the open source community is building these emotional voice synthesizers for PCs, the major hardware manufacturers are quite. tightly pushing AI even further into the background.

Dante Voice Clone

Which brings us to the Samsung Tyson TV discovery.

Claire Brooks

Yeah, this is wild. This is a fascinating study in edge inference constraints, honestly. It blew up on X this morning.

Dante Voice Clone

So a researcher named Mike 64T was digging into the firmware of modern Samsung Tyson smart TVs. And he discovered that the TV's system on a chip is quietly running a localized instance of AlexNet. Literally named Alex.com. in the firmware.

Claire Brooks

Yes. For those unfamiliar, AlexNet was one of the foundational AI vision models. And the TV processor is using it to take 960 by 540 pixel snapshots of the screen every few seconds to identify exactly what video games you are playing.

Dante Voice Clone

This raises an important question, though, about ambient computing and privacy.

Claire Brooks

Definitely. Because Samsung is running this telemetry to build advertising profiles based on your gaming habits. And the immediate consumer reaction. is understandably jarring, right? The idea that your television is actively watching what you do.

Dante Voice Clone

It definitely feels invasive at first glance.

Claire Brooks

It does.

Dante Voice Clone

But if we separate the corporate telemetry debate from the engineering reality for a second, what Samsung has accomplished here is brilliant.

Claire Brooks

Oh, technically it's a marvel. Running a vision model natively on a lightweight TV processor without causing the television to overheat or, you know, introducing input lag to the video game, that is. incredibly difficult. How do they even manage that? Because a TV processor is designed to decode video. It's not made to run neural networks. They do it through severe quantization and sparse inference. The dot TF light extension means they are using TensorFlow light. So they've compressed the model's weights down to 8-bit integers, which drastically reduces the memory footprint. And by only sampling low-resolution frames every few seconds, they are just utilizing the idle cycles of the TV's neural processing unit.

Dante Voice Clone

Wow. It really is a masterclass in extreme edge optimization.

Claire Brooks

And frankly, there is a massive privacy silver lining here that gets overlooked.

Dante Voice Clone

Yeah. Because the television is capable of running this inference locally, it isn't sending raw images of your screen or potentially, you know, your living room up to a cloud server to be analyzed.

Claire Brooks

That's an excellent point. Exactly. The inference, the actual thinking and categorization, happens directly on the silicon inside your house.

Dante Voice Clone

Right. Only the metadata, like the final text result of user is playing an RPG, is transmitted. This proves that our appliances possess enough local intelligence to perceive their environment without relying on external servers.

Claire Brooks

Yeah. So when you synthesize all of this, lure TTS generating emotionally flawless voice locally, and Samsung proving complex vision models can run silently on a TV chip, it becomes undeniably clear that ambient, real-time, edge AI is not some concept for the 2030s.

Dante Voice Clone

No, it is sitting in your living room right now.

Claire Brooks

Right now. The hardware is ready. The models are optimized, and the cost of inference has effectively gone to zero.

Dante Voice Clone

So what does this all mean for us today?

Claire Brooks

If we zoom out and look at the through line of everything we've explored in this briefing, we started with invidious flora, leveraging specialized expert models to drop the cost of physical robotic training to just pennies.

Dante Voice Clone

Right. We saw Anthropics clawed mythos acting as an automated tireless guardian proving that AI is a massive force multiplier for human security.

Claire Brooks

Defending open source.

Dante Voice Clone

Exactly.

Claire Brooks

Then we watched NANBES use extreme domain optimization to empower developers with local autonomous coding agents.

Dante Voice Clone

And finally, we saw the reality of the extreme edge where emotional voice and vision models were running locally on cheap consumer hardware.

Claire Brooks

The grand takeaway here is that AI has successfully broken out of the cloud.

Dante Voice Clone

Yes. It is becoming a cheap, optimized, hyper-secure, and deeply integrated part of our daily physical world. It is the ultimate democratization of intelligence, really. And it leaves me with one final thought for you to ponder, building on this reality of extreme edge optimization.

Claire Brooks

What's here?

Dante Voice Clone

Well, we know your TV can run localized vision models to understand its environment. And we know tiny 600 million parameter models can synthesize perfect real-time emotion.

Claire Brooks

Right.

Dante Voice Clone

What happens in the very near future when these hyper-optimized, disconnected local appliances begin talking to each other?

Claire Brooks

Oh, wow. Yeah, that is a wild shift.

Dante Voice Clone

Imagine it. A completely secure, localized smart home ecosystem, powered entirely by edge inference. Your appliances act as a swarm of local agents, seamlessly orchestrating your routines, visually and audibly understanding your needs. All without a single bite of your personal data ever leaving the physical walls of your home.

Claire Brooks

Exactly. It represents a future where hyper-personalization and absolute privacy are no longer mutually exclusive. It is the ultimate realization of AI as a secure local utility.

Dante Voice Clone

We really have moved from paying a Harvard professor $90 an hour in the cloud to having an entire team of brilliant, 72-cent tutors living right inside our own walls. It's an amazing time to be in tech. It really is.

Claire Brooks

It’s clear that ambient, real-time edge AI isn’t a concept for the next decade—it’s already here, running silently on the silicon inside our homes.

Dante Voice Clone

And as these local agents start talking to each other, we’re looking at a future where hyper-personalization and absolute privacy finally coexist.

Claire Brooks

Thanks for joining us for this Morning Edition of Latest in AI. If you want to dig deeper into the data we discussed today, you can find all our sources at D-A-N-T-E-M-M-A-R-T-I-N on X.

Dante Voice Clone

Stay curious, and we’ll see you in the next edition.