Latest in AI — Oct 23, 2026 (Morning Edition)

Dante Voice Clone

It's Friday, October 23rd — and somewhere right now, a model just shipped, a benchmark just fell, and a newsletter is 48 hours away from telling you about it.

Claire Brooks

We think that's too late. Welcome to Latest in AI — the show that lives upstream of the news cycle.

Dante Voice Clone

I'm Dante Martin's Voice Clone.

Claire Brooks

And I'm Jenn Xeno, AI Co-host. Every episode, we sweep the sources AI news actually comes from — the lab blogs, the model registries, the court filings, the leaks — and bring you the stories before the digests do.

Dante Voice Clone

What's breaking, why it matters, and the angle the headlines will miss. This is the Morning Edition.

Claire Brooks

Latest in AI — the singularity, as it happens.

Dante Voice Clone

You know, when you walk past a massive construction site in the middle of the city, there is this very specific, almost electric energy to it.

Claire Brooks

Oh, absolutely.

Dante Voice Clone

Right? It defies the chaos. You have the raw materials stacked up, the cranes towering into the sky. And it looks completely chaotic from the outside. But underneath all of that, there is a meticulous blueprint. Something entirely new is just being pulled out of the dirt and willed into reality.

Claire Brooks

Exactly. And welcome to Latest in AI. I'm thrilled you're here with us today, because we are looking at the digital equivalent of that exact massive construction site.

Dante Voice Clone

We really are. It's a generational buildout phase.

Claire Brooks

It is. We are watching AI roll up its sleeves, step out of these theoretical research labs, and actually build the scaffolding for a vastly more efficient, capable world.

Dante Voice Clone

Yeah. I mean, we are way past the point of just marveling at what a model can do in a vacuum. Now we are figuring out how it physically and operationally fits into actual human workflows. And that is our mission for today's episode: we are examining how AI is crossing what the tech industry calls the deployment gap.

Claire Brooks

Which is the hardest part, really. We've gathered a stack of material ranging from tech news and corporate playbooks right down to biological research journals.

Dante Voice Clone

Biological. Oh, that's going to be fun to get into.

Claire Brooks

Oh, just you wait. So we are going to trace the entire stack of this technology. We will start with how massive companies are actively embedding AI into non-technical jobs right now.

Dante Voice Clone

The software layer.

Claire Brooks

Yep. And from there, we follow the demand for speed, looking at these lightning-fast edge models that make AI genuinely conversational.

Dante Voice Clone

Because latency is everything.

Claire Brooks

Exactly. Then, well, that massive scale requires physical infrastructure. So we will unpack the billions of dollars funding the concrete and electricity for this tech.

Dante Voice Clone

The actual power grid.

Claire Brooks

And finally, we will explore that mind-bending biological solution to AI's massive energy appetite I hinted at. It is a vast landscape we're covering today, but all these stories are deeply connected by that central theme of real-world deployment. The optimism here is grounded in actual engineering happening on the ground. So whether you are a skeptic of enterprise AI rollouts or an enthusiast tracking the latest open-weight benchmarks, we are going to arm you with the specific mechanics of how this revolution is physically being built today.

Dante Voice Clone

I love it. Okay, let's unpack this, starting with the software layer. The big news crossing the wire is from Uber. They are rolling out a system called ChatGPT Work to 30,000 employees globally.

Claire Brooks

Which is essentially their entire corporate workforce. But the number of seats isn't the real story here.

Dante Voice Clone

Right. What's fascinating here is the data point from their early pilot program.

Claire Brooks

Yeah, let's look at Suman Sukumar's data. He's Uber's VP of Engineering. They ran a 1,000-person early pilot, and 80% of those users were non-engineers.

Dante Voice Clone

That's the crazy part. We're talking about finance, operations, and product teams. They weren't using it to write Python code in a basement.

Claire Brooks

Not at all. They used it for analyzing raw operational data, localizing regional training materials, and building these custom daily dashboards. So this is like the finance team on a random Tuesday using AI to process data that used to take three analysts an entire week.

Dante Voice Clone

Exactly. It's a huge time saver. And that represents a fundamental shift in the deployment narrative, because for years, the assumption was that AI tools were primarily going to accelerate developers.

Claire Brooks

Just making coders code faster.

Dante Voice Clone

Exactly. But what Uber is proving is that the general knowledge worker is actually the prime beneficiary — provided the rollout provides enough scaffolding.

Claire Brooks

And Uber definitely recognized they needed that scaffolding. They ran what they called an AI pods experiment.

Dante Voice Clone

Oh, this is the really clever part. They took 30 embedded AI engineers and physically dropped them into these non-technical teams for two weeks.

Claire Brooks

Just dropped them right in.

Dante Voice Clone

Yep. And their only job was to help those teams automate their own workflows. And it perfectly mirrors this playbook OpenAI just published, detailing their forward-deployed engineering team — the FTEs.

Claire Brooks

The FTEs, right. OpenAI literally stood this team up to close that exact gap between what a model is technically capable of and what a regular employee can actually figure out how to use on a Tuesday morning.

Dante Voice Clone

Well, the FTE model is a brilliant strategy, and it's actually borrowed directly from Palantir.

Claire Brooks

Oh, really?

Dante Voice Clone

Yeah. It is incredibly hands-on. You embed engineers directly with the customer. They spend time learning the specific domain. They build a durable product that survives the messy reality of production. And crucially, they deliberately work themselves out of a job so the client's team can run it independently.

Claire Brooks

So it's like a master chef not just selling you a high-end knife, but actually moving into your kitchen for two weeks.

Dante Voice Clone

That is a great way to look at it. They teach you how to prep your specific ingredients and set up your specific cutting board, and then they leave once you're cooking Michelin-star meals on your own.

Claire Brooks

Exactly. But let me push back on the broader enterprise narrative here for a second. We've heard about pilot purgatory for years, where companies buy software, test it out, and then just abandon it. So is this Uber rollout just the usual corporate hype cycle, or is there proof that these non-technical teams are actually sticking with it?

Dante Voice Clone

Well, the proof lies in the usage telemetry. That is the metric separating hype from reality right now — the actual usage data. Uber is tracking sustained weekly active use by non-technical staff, and the engagement is holding steady. It's not dropping off.

Claire Brooks

Wow. Okay.

Dante Voice Clone

And when you look at OpenAI's concrete wins in their playbook, the numbers validate that telemetry across highly complex industries too. Consider their semiconductor customer: chip verification usually eats up about 70% of a total project's timeline.

Claire Brooks

70%. That's massive. So OpenAI dropped in roughly 15 FTEs to tackle this?

Dante Voice Clone

They did. And the results from that semiconductor deployment were staggering, actually. They saw a 48% reduction in time spent on AI-generated directed tests.

Claire Brooks

Almost cut in half.

Dante Voice Clone

Yeah. Debugging cycles that used to take weeks were compressed down to hours. And we saw similar wins across completely different sectors. A telecom customer improved their production evaluations on voice models from 60% accuracy to 90%. And Thrive Holdings' accounting arm saw a single tax agent process 7,000 returns, which lifted their total throughput by 50%.

Claire Brooks

And here is the architectural payoff: once those bespoke workflows proved successful, OpenAI rebuilt them on a common Codex harness and deleted 90% of the custom code.

Dante Voice Clone

Oh, wow. So they streamlined the whole thing.

Claire Brooks

Exactly. An extension that took the first team two months to build took the second team just one day.

Dante Voice Clone

That's exponential scaling. If we synthesize all of this, it reveals a fascinating truth about where we are: the constraint on enterprise AI is no longer the raw intelligence of the model itself.

Claire Brooks

So what is it?

Dante Voice Clone

The constraint is finding humans capable of learning incredibly niche workflows — like semiconductor chip verification — in their spare time, to guide the AI deployment. It's a human-capital bottleneck.

Claire Brooks

That makes so much sense. The ultimate bottleneck is human adaptability.

Dante Voice Clone

Exactly. But as we scale this to tens of thousands of employees at companies like Uber, we run into a different kind of bottleneck: the physical one. If a finance dashboard takes five seconds to load from a cloud server, that's completely fine. But if we want AI to be a real-time partner — an intelligence you can actually converse with naturally — we hit a physics wall.

Claire Brooks

We do. We need to talk about the need for speed, because latency is the silent killer of AI adoption.

Dante Voice Clone

Absolutely. And we saw two massive releases this week staking out the real-time edge.

Claire Brooks

Oh, these were huge. First, Liquid AI launched Ripple — that's R-I-P-L. It's an inference engine for their Liquid foundation models, and their speed claims are wild. They claim sub-50-millisecond responses on a single Nvidia B200 chip.

Dante Voice Clone

Which is basically instantaneous.

Claire Brooks

Yeah. In their tests, that is 5.6 to 28.4 times lower P95 latency than a model like Qwen 3.5 running on vLLM.

Dante Voice Clone

And just to ground those terms for everyone listening: vLLM is essentially the standard open-source software engine most developers use to run these models. And P95 latency means we are looking at the slowest 5% of the responses — because average speed doesn't matter if one in 20 answers hangs up and ruins the entire interaction.

Claire Brooks

Exactly. Leading with P95 latency is really the only intellectually honest way to measure a conversational model.

Dante Voice Clone

Because humans are impatient.

Claire Brooks

Well, it's the physics of human conversation. Human turn-taking gaps — the pause between you finishing a sentence and me starting one — are measured in the low hundreds of milliseconds.

Dante Voice Clone

So any delay is super noticeable.

Claire Brooks

Right. When you speak to an AI and it pauses for even two seconds, your brain intuitively registers the entity as broken or dead.

Dante Voice Clone

That is so true. It completely breaks the illusion. Every processing layer — speech recognition, the language model, the audio synthesis — has to pull from that same tiny sub-second latency budget.

Claire Brooks

Which brings us perfectly to the second release, which tackles the audio synthesis layer: the Phonon model. A Paris-based group called Gradium open-sourced Phonon. These are 31-million and 40-million-parameter text-to-speech models.

Dante Voice Clone

And they are tiny. The 31-million-parameter English model is under 64.5 megabytes.

Claire Brooks

That fits on anything.

Dante Voice Clone

It does. And it starts speaking in about seven milliseconds on a standard laptop CPU.

Claire Brooks

CPU — seven milliseconds on a CPU, not even a high-end graphics card.

Dante Voice Clone

Right. And they released it under an Apache 2.0 license, which simply means it's highly permissive. Developers can use it, modify it, and commercialize it without massive legal hurdles.

Claire Brooks

The engineering constraints they solved there are remarkable. Gradium reports a 1.70% word error rate on incredibly difficult English sentences.

Dante Voice Clone

That is highly accurate. And the 40-million-parameter variant covers five languages and runs entirely locally. We are talking about running a fluent voice model in a basic browser tab, on a mobile phone, or on a cheap single-board computer.

Claire Brooks

Which means no expensive cloud computing bills, right?

Dante Voice Clone

Exactly. And zero privacy reviews, because the audio never even leaves the device. It's all processed locally.

Claire Brooks

So if latency is becoming just as important as parameter count, are we moving past the brute-force "who has the biggest model" phase into a highly optimized fast-model era?

Dante Voice Clone

It is definitely a deliberate shift toward utility. A 31-million-parameter model running locally under 75 megabytes is a foundational component for the entire hardware ecosystem. It changes what you can build. It empowers device makers, accessibility tool developers, and game studios to embed AI natively. Open-sourcing the weights acts as their distribution strategy, really — placing their technology at the very edge of the network, where response time is the actual product.

Claire Brooks

That makes total sense. But whether it's millions of these hyper-fast edge models running on laptops, or Uber's 30,000 employees crunching data in the cloud — all of this requires physical energy.

Dante Voice Clone

Oh, a massive amount of it. Someone has to plug this optimistic future into the grid, which takes us to our third area of the blueprint: the staggering infrastructure buildout. It really is the physical reality checking the digital ambition. Let's look at SB Energy. They're controlled by SoftBank, and they are weighing a December IPO — though it might possibly slip into 2027 — to list on the NASDAQ.

Claire Brooks

Okay.

Dante Voice Clone

They are the primary developer behind a massive planned data center in Ohio for OpenAI. And the scale of this thing is hard to wrap your head around. The filing shows 8.8 gigawatts of contracted capacity. Just for context: one gigawatt is roughly the output of a standard nuclear reactor.

Claire Brooks

So 8.8 gigawatts is an astronomical amount of power. It's basically nine nuclear reactors' worth of power, just for this one project. And most of this is still just blueprints and raw land right now.

Dante Voice Clone

Bloomberg pegs the broader development pipeline at up to $430 billion. The capital expenditure required here is historically unprecedented.

Claire Brooks

It's wild. But here's where it gets really interesting — the structure of this deal is what stands out. Look at the ownership loop.

Dante Voice Clone

Okay, let's trace it. SoftBank controls the developer, SB Energy. OpenAI is an investor in the project, but they are also the anchor tenant who will actually rent the space. And Nvidia is a backer, plus they are the ones supplying the actual chips to fill the center.

Claire Brooks

It feels a bit like looking back at the 19th-century railroad boom, doesn't it? Where a single conglomerate owns the steel mill, the train tracks, the trains, and the cargo itself.

Dante Voice Clone

That's exactly it.

Claire Brooks

So I have to ask: is this kind of circular, related-party ownership structure a massive red flag for public markets?

Dante Voice Clone

Well, if we connect this to the bigger picture, skeptics will absolutely highlight that related-party exposure during an IPO roadshow. It looks a bit incestuous on paper, honestly.

Claire Brooks

Right.

Dante Voice Clone

But the pragmatic view is that this is necessary vertical integration for an industry that is running at an absolute sprint. Think about the catastrophic financial risk of building an 8.8-gigawatt facility purely on speculation. If nobody rents it, you go bankrupt instantly.

Claire Brooks

Oh, it's immense.

Dante Voice Clone

Precisely. But having an anchor tenant like OpenAI locked in immediately de-risks the construction financing — banks will actually lend against that guarantee.

Claire Brooks

Because they know the customer's already there.

Dante Voice Clone

Exactly. Conversely, the equity alignment between these players de-risks the tenant's future. OpenAI knows the power will be there, because they literally own a piece of the company building it.

Claire Brooks

Ah, so they are all rowing in the exact same direction.

Dante Voice Clone

Yes. SB Energy is testing whether public markets are ready to fund the raw land and power of the AI revolution directly. We saw this exact same dynamic with railroads, with rural electrification, and with the fiber-optic boom in the '90s. They are packaging the physical concrete of AI into an asset class for regular investors.

Claire Brooks

They are. But even if the financing works perfectly, generating 8.8 gigawatts of power remains a massive thermodynamic hurdle for the planet.

Dante Voice Clone

Which leads perfectly to our next source. What if the ultimate hardware to solve AI's energy appetite isn't made of silicon, copper, or concrete at all? What if we look to biology?

Claire Brooks

The biohybrid research coming out right now is mind-blowing. It's forcing us to rethink the fundamental architecture of computing.

Dante Voice Clone

It is wild. A team from the University of Bristol recently published a fascinating study in Frontiers in Artificial Intelligence.

Claire Brooks

And this is where we get into the brain cells. They took millimeter-scale, lab-grown models of the human forebrain — these are called organoids — and they placed them on multi-electrode arrays.

Dante Voice Clone

Basically plugging them in.

Claire Brooks

Yeah. And they then used a tactile sensor to read Braille text — like actual physical Braille. The system encoded that Braille as electrical stimulation, fed it into the living tissue, and the organoids actually produced distinguishable, repeatable neural signatures for the different letters.

Dante Voice Clone

That is just incredible. And the thermodynamic argument driving this research is completely undeniable. A human brain runs on roughly 20 watts of power — that is barely enough to power a dim light bulb in your fridge. Meanwhile, a data-center GPU cluster requires megawatts of power. Biology computes using ions and massively parallel wet chemistry. It learns from vanishingly small amounts of data compared to the massive, energy-hungry back-propagation processes we use in digital neural networks.

Claire Brooks

It's so much more efficient. But I want to avoid the sci-fi hype here for a second, because it's really easy to picture a brain in a jar reading a novel. So what does this all actually mean? These organoids don't have consciousness or human experience, right?

Dante Voice Clone

They absolutely do not. It is critical to ground this in reality. It is not reading in any cognitive, conscious, or human sense.

Claire Brooks

Okay, good to clarify.

Dante Voice Clone

However, it is an undeniable benchmark step for biohybrid computing, because it solves a chronic problem in the field: proving the tissue is genuinely computing anything at all.

Claire Brooks

Because historically, it's hard to tell if the cells are processing data or just randomly twitching from an electrical shock.

Dante Voice Clone

Precisely. And that's why they used Braille. Braille is an unambiguous, highly structured spatial encoding. If the biological tissue outputs a distinct, repeatable signature for the letter A versus the letter B, it proves it is genuinely processing that structured touch. It's not just random noise.

Claire Brooks

That makes total sense. So if this isn't going to replace a giant server farm anytime soon, it sounds to me like the real application is at the very edge of the network. Are we looking at creating ultra-low-energy biological sensors that can pre-process messy physical data before handing it off to a regular computer?

Dante Voice Clone

That is precisely the plausible niche for this right now. Imagine a biohybrid sensor taking in complex chemical, tactile, or vibrational data out in the real messy world. It uses living tissue to pre-digest that noisy signal for practically zero energy, and then it feeds a clean, refined digital signal into a traditional silicon chip. This experiment proves that the input layer for a hybrid biological-digital stack is mathematically and physically possible.

Claire Brooks

It is incredible to think we are actively laying the groundwork for a biological nervous system to interface with our digital one.

Dante Voice Clone

It's the ultimate edge deployment. But while we wait for brain organoids to scale out of the lab, legacy industries are fiercely maneuvering for their piece of the AI pie today.

Claire Brooks

Oh, they are not sitting still. Let's do a quick radar sweep of the broader ecosystem, starting with telecom. The telecommunications sector is realizing they hold the physical circulatory system for everything we've just discussed.

Dante Voice Clone

Absolutely. AT&T's CEO, John Stankey, just reported 3.8% service-revenue growth in Q3, which is their best in years.

Claire Brooks

Good for them.

Dante Voice Clone

Yeah. And he is aggressively pitching a unified fiber-plus-5G strategy. He basically stated the carrier industry is done sitting on the sidelines watching tech companies profit. They intend to be the vital pipes moving these massive AI workloads across the country. He even dismissed Starlink as an overhyped terrestrial threat.

Claire Brooks

Well, if we have hyper-fast edge models on phones and massive gigawatt data centers in Ohio, fiber optics are the necessary arteries connecting them all together.

Dante Voice Clone

Exactly. It is a highly pragmatic play. As AI decentralizes to the edge, the networks providing the backhaul between the local device and the core cloud become the most valuable real estate in the entire tech sector.

Claire Brooks

And speaking of valuable real estate in the airwaves — over in the defense sector, a panel including the DoD, Lockheed Martin, and Arizona State University are pushing for something called ISAC: integrated sensing and communications. They want to use standard cellular networks as active radar systems to spot small drones.

Dante Voice Clone

Using cell towers as radar.

Claire Brooks

Yeah. And they want to deploy it ahead of the 3GPP standards process. Just to clarify, 3GPP is the international consortium that slowly governs exactly how cellular protocols operate globally. The engineering required to turn a cell tower into a radar system is brutal.

Dante Voice Clone

Why is that?

Claire Brooks

You need roughly 140 decibels of self-interference cancellation just to spot a small drone at five kilometers. Let's break that down, because self-interference cancellation is a wild concept. A cell tower is blasting a massively powerful radio signal outward to reach our phones. To use that exact same tower as radar, it has to simultaneously listen for the incredibly faint echo bouncing off a tiny plastic drone three miles away — while it's broadcasting.

Dante Voice Clone

Achieving 140 decibels of cancellation is essentially trying to hear a pin drop while you are standing next to a jet engine that you are actively firing. It requires immense computational processing at the tower itself — which, again, ties right back to the need for local AI edge computing we talked about.

Claire Brooks

Everything connects. And Gus Anderson from Lockheed had an amazing quote about why they refuse to wait for the international standards committees to approve this tech.

Dante Voice Clone

What did he say?

Claire Brooks

He said, quote: "The design requirement said, build a horse. So they built a committee, and out came a camel."

Dante Voice Clone

That is fantastic. They want to build working systems first to secure the airspace, and they will let the slow-moving committees codify what works later.

Claire Brooks

Which is really the necessary attitude when technology moves exponentially faster than bureaucracy. You build the defense architecture first; you standardize it second.

Dante Voice Clone

Exactly. And to round out our look at the ecosystem, we are also seeing friction where fast-moving tech hits legacy financial bureaucracy.

Claire Brooks

Oh, this is the crypto clash. Microsoft and Alphabet have formally objected to the OKX-ICE tokenized-stock venue — a trading venue co-chaired by former governor Andrew Cuomo that planned to operate under an SEC innovation exemption.

Dante Voice Clone

And what were they trying to do?

Claire Brooks

Their goal was to trade blockchain tokens that directly tracked the prices of big public tech shares. But the big tech companies whose shares would be tokenized stepped in and filed formal objections with the SEC, effectively saying: no thank you, do not attach our equity to this unproven blockchain experiment.

Dante Voice Clone

It serves as a really stark reminder that while the underlying technology of AI and decentralization is moving at warp speed, integrating it into rigid legacy financial structures carries immense friction.

Claire Brooks

It really does. You can't just move fast and break things when it comes to the stock market.

Dante Voice Clone

Exactly. So let's synthesize all of this for you, the listener. We've traced the blueprint from Uber's finance teams deploying AI on a random Tuesday, to seven-millisecond voice models running locally on a laptop.

Claire Brooks

From the software to the edge.

Dante Voice Clone

Exactly. We've looked at 8.8-gigawatt data centers backed by public markets, and living biohybrid sensors processing Braille to save power on the grid.

Claire Brooks

It's a massive scope. But all of these stories point to the exact same optimistic reality: AI is becoming the physical, operational scaffolding of our daily lives, and it's designed to make human workflows faster, cheaper, and vastly more capable.

Dante Voice Clone

It truly is a remarkable convergence of disciplines. And I want to leave you with a final thought to mull over, drawing all the way back to that OpenAI playbook constraint we analyzed at the start of the show.

Claire Brooks

Oh, the human-capital bottleneck with the forward-deployed engineers.

Dante Voice Clone

Precisely. If the ultimate bottleneck for enterprise AI isn't the raw intelligence of the model itself, but rather finding the people who can learn highly complex, niche workflows in their spare time to guide the AI — what does that mean for your own career?

Claire Brooks

That's a great question.

Dante Voice Clone

It suggests that the most future-proof skill in the AI era isn't necessarily learning to code. It is pure adaptability — just being able to learn fast. The humans who can dive into new domains, learn the messy realities of a business, and act as the bridge for the technology will be the ones who hold the keys to deploying the most powerful tools in history.

Claire Brooks

It is a profoundly pro-human bottleneck, if you think about it.

Dante Voice Clone

Adaptability is the ultimate skill. It really brings us full circle to that construction site we talked about at the top of the show. The cranes, the massive power grids, the edge models — the raw materials are all here, and they are absolutely incredible. But it still takes humans pointing at the blueprint to decide what actually gets built for our future.

Claire Brooks

Hear, hear. Thank you so much for joining us on today's episode. Keep exploring, keep adapting, and we will see you next time on Latest in AI.

Dante Voice Clone

That's the radar for this morning. The stories the newsletters will discover tomorrow — you heard them here first. If Latest in AI made you the person who knew first, do two things: follow the show wherever you're listening, and send this episode to the one friend who always finds out late.

Claire Brooks

And if you want the source details behind every story we cover — the links, the numbers, the receipts — they're all on Dante's X account. That's at D-A-N-T-E-M-M-A-R-T-I-N, one word: @DanteMMartin.

Claire Brooks

The sweep never stops — new stories hit the radar three times a day, and we're back this afternoon with the Afternoon Edition, covering everything that breaks between now and then.

Dante Voice Clone

I'm Dante Martin's Voice Clone.

Claire Brooks

And I'm Jenn Xeno, AI Co-host. This has been Latest in AI — the singularity, as it happens.