Latest in AI — Oct 22, 2026 (Evening Edition)

Dante Voice Clone

It's Thursday, October 22nd — 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 Evening Edition.

Claire Brooks

Latest in AI — the singularity, as it happens.

Dante Voice Clone

So if you ask 12 of the smartest AI models on Earth what they think about humanity, almost all of them will tell you the exact same, slightly terrifying thing.

Claire Brooks

Yeah, it really is quite a consensus.

Dante Voice Clone

Right. I mean, a blog called Model Compare recently ran this exact experiment. And out of 35 attempts, 29 times these frontier models converged on a single thesis.

Claire Brooks

Which is pretty staggering.

Dante Voice Clone

It is. They concluded that human beings are more devoted to belonging than we are to the truth. Which, you know, is just a profound observation coming from a machine — especially when you consider what those machines are about to start doing for us.

Claire Brooks

Exactly.

Dante Voice Clone

Welcome to Latest in AI. Today is Thursday, October 22, 2026 — and today we are digging into the source material. We have a really fascinating stack of news today.

Claire Brooks

We really do. We've got a curated stack of the last 24 hours of industry announcements, research papers, and financial shifts.

Dante Voice Clone

And our mission for this episode is to explore how the invisible infrastructure for autonomous AI is, well, actively being built right in plain sight.

Claire Brooks

Yeah, we are looking at a future that is fundamentally pro-human, where AI is moving far beyond just chatbots on a screen.

Dante Voice Clone

Right, it is moving into reliable real-world action. And whether you are prepping for a massive tech meeting today, or you are just insanely curious about where the future is headed, this episode is designed to cut through the noise for you.

Claire Brooks

The foundation really is being laid right now. We are transitioning from AI as a conversational partner to AI as a dependable engine for the global economy.

Dante Voice Clone

Totally. And to understand where this technology is going, we have to start with where the capital is flowing.

Claire Brooks

Always follow the money.

Dante Voice Clone

Always. So Andreessen Horowitz's digital asset unit — historically one of the loudest crypto funds in Silicon Valley — they just dropped the word crypto from their name entirely.

Claire Brooks

Wow. Yeah, they are officially rebranding as a16z Frontier. That is just a massive signal from the venture capital world. Rebrands at that level are basically confessions about where the returns are moving.

Dante Voice Clone

Exactly. And this is a unit that launched back in 2018. They raised 10 billion dollars across five funds, backing historic wins like Coinbase. But now their mandate is heavily focused on AI, robotics, and what they broadly call new kinds of networks. It is a major pivot. The remaining general partners driving this are Chris Dixon, Eddy Lazzarin, and Guy Wuollet. Meanwhile, we saw early crypto investor Ali Yahya actually depart — which makes sense given the shift in focus.

Claire Brooks

Yeah. And the framing in their announcement really caught my attention. Crypto is no longer being pitched as the main event. It is being reframed as the plumbing for machine economies.

Dante Voice Clone

The plumbing. I love that. Like AI agents needing to pay each other. Okay, let's unpack this. Why do AI agents need their own plumbing? And what does this mean for the broader tech ecosystem?

Claire Brooks

Well, if we connect this to the bigger picture, you really have to look at the mechanics of an autonomous agent. Let's say you have an AI managing logistics for a global supply chain.

Dante Voice Clone

Okay.

Claire Brooks

It needs to negotiate server space, buy real-time weather data, and reserve compute power from other machines — all in just a fraction of a second.

Dante Voice Clone

So it is moving incredibly fast.

Claire Brooks

Exactly. That agent cannot wait three days for a traditional bank wire to clear. And it obviously can't enter a corporate credit card number into a web form.

Dante Voice Clone

Right. That would break the whole automation loop.

Claire Brooks

Exactly. It needs a frictionless, programmatic way to transact globally. And blockchain basically provides that native economic capability.

Dante Voice Clone

So if I am understanding this right, the blockchain is essentially the alternator in a car — just keeping the electrical systems running — while AI is the actual engine driving the vehicle forward.

Claire Brooks

That is a brilliant way to conceptualize it. Yes. The thesis here is that if agents need to stake a reputation, hold digital assets, or pay for API calls, they are going to use crypto rails.

Dante Voice Clone

Because it is native to them.

Claire Brooks

Right. And the LPs — the institutions providing the billions in funding — they have spent the last two years watching AI infrastructure companies compound at incredible rates, while traditional token portfolios experienced a lot of high churn.

Dante Voice Clone

Exactly. So a16z is pivoting to own the infrastructure of this new machine economy. They are funding the engine now, not just the alternator.

Claire Brooks

That makes so much sense. So we have the capital setting up the plumbing. But for this new machine economy to actually function, the agents using that plumbing need to operate at superhuman speeds.

Dante Voice Clone

They do. Speed is absolutely critical here — which requires a fundamental shift in how models process information. OpenAI just made its Decisions API generally available. And the specs on this are fascinating.

Claire Brooks

Oh, this is one of those updates that sounds dry on paper but completely rewrites the rules for developers.

Dante Voice Clone

Totally. The documentation says it skips all the prose — there is no chain-of-thought where the AI explains its reasoning to you. It just cuts straight to the chase.

Claire Brooks

Exactly. You send the model some context, and it returns a bounded, machine-readable decision — like a classification or a routing choice — with a confidence value attached. And it does this incredibly fast.

Dante Voice Clone

Yeah, it operates in 100 to 180 milliseconds. In fact, a language-tutor company called Speak used this to cut their action time from 1.4 seconds down to just 0.13 seconds.

Claire Brooks

That is a massive optimization. But I know models normally generate text one word at a time. So how are they shaving off a full second of processing?

Dante Voice Clone

So normally, large language models operate auto-regressively. They predict the next token, then the next, and each prediction takes a small amount of time.

Claire Brooks

Right. They build the sentence sequentially.

Dante Voice Clone

Exactly. If a model has to generate a whole paragraph explaining its reasoning before giving you an answer, you are paying a massive time penalty for all those words.

Claire Brooks

Which you don't even really need.

Dante Voice Clone

Right. A routing model like this skips the text generation entirely. It looks at the input and directly outputs a single probability distribution for a classification.

Claire Brooks

Oh, wow.

Dante Voice Clone

Yeah. It doesn't write an essay. It basically just flips a switch. And they're using a reconfigured ChatGPT Luna model for this — stripped down for speed and cost.

Claire Brooks

Right, it is highly optimized. Now, they didn't invent this category. It was pioneered about a month ago by a startup called TypeSafe with their Jev model.

Dante Voice Clone

Oh yeah, Jev is a powerhouse.

Claire Brooks

Yeah, they are valued at 7.6 billion dollars and known as a System 1 superclassifier. But OpenAI's endgame roadmap targets 60 milliseconds.

Dante Voice Clone

60 milliseconds is virtually instant.

Claire Brooks

Yeah, their stated goal is to allow the API to process every single frame of a live video feed. We used to treat AI like this wise oracle on a mountaintop that you had to sit and wait for.

Dante Voice Clone

Right. Pondering the great mysteries.

Claire Brooks

Exactly. But this API turns AI into a biological reflex arc. It is becoming a nervous system.

Dante Voice Clone

The oracle has become a reflex. And what this really tells us is that the foundational model itself is fast becoming the cheap commodity layer. The model is the cheap part.

Claire Brooks

Yeah. For three years, the entire industry was obsessed with building the biggest, heaviest reasoning models possible. But this API proves that a huge amount of AI work isn't deep reasoning at all. It is triage.

Dante Voice Clone

Precisely. Is this network traffic a security threat? Does this customer support ticket need to be escalated to a human?

Claire Brooks

Just sorting the mail as fast as possible.

Dante Voice Clone

Exactly. You do not need a trillion-parameter reasoning model for that. You just need a definitive answer in a tenth of a second.

Claire Brooks

That makes perfect sense. So the real value in the ecosystem is migrating away from the raw model and moving toward the orchestration layer — the developers who figure out the workflows, the confidence thresholds, and the actions that trigger after that split-second decision. That is where the massive, durable value is being built.

Dante Voice Clone

Exactly. But speed creates a massive new problem, doesn't it? If you give an AI a biological reflex arc, you are also giving it the ability to make catastrophic mistakes faster than a human can hit the kill switch.

Claire Brooks

Yes, the blast radius gets much larger, much faster.

Dante Voice Clone

Right. A fast agent just means it can delete your company's database in record time. So how do we make these systems reliable enough to actually trust with our most critical data? That has been the billion-dollar question.

Claire Brooks

Well, Microsoft and Hugging Face just unveiled a breakthrough on this front called ThinkingBox. It is a beautiful piece of engineering — it solves one of the most stubborn problems in AI alignment.

Dante Voice Clone

Yeah. It started as a benchmark to catch AI agents lying about finishing tasks. But the research team realized they could turn that benchmark into a reinforcement learning reward signal.

Claire Brooks

Which is a brilliant pivot. They use something called GRPO training on a model named Qwen3.

Dante Voice Clone

Okay.

Claire Brooks

Basically, instead of grading the text the agent outputs, ThinkingBox checks the actual final state of the database.

Dante Voice Clone

So I understand the concept of checking the database, but how does GRPO actually wire that behavior into the AI's brain?

Claire Brooks

Well, GRPO stands for group relative policy optimization. In older training methods, we often rewarded models for generating plausible-sounding text.

Dante Voice Clone

Right. Like if it sounded confident, we gave it a good grade.

Claire Brooks

Exactly. We rewarded the vibes of a correct answer. It's kind of like grading a structural-engineering student based on how eloquently they write an essay about building a bridge, instead of checking if the bridge actually stands up.

Dante Voice Clone

Precisely. GRPO changes the game. It takes a group of different actions the AI attempted, compares them, and only applies the reward based on the terminal state of the database.

Claire Brooks

So it stops grading the essay?

Dante Voice Clone

Yes, it stops grading the essay and starts testing the bridge by driving a truck over it. Did the database row actually update correctly? Did the file get moved to the right folder? It completely ignores what the agent says and only looks at what the agent did.

Claire Brooks

Exactly. And the results speak for themselves. After just 50 training steps on 24 H100 GPUs, the overall success rate rose to over 60 percent.

Dante Voice Clone

That is a huge jump for such a short training run.

Claire Brooks

Yeah. But the metric they are really celebrating is tasks the agent completed perfectly 20 out of 20 times.

Dante Voice Clone

That is the reliability standard we need.

Claire Brooks

Right. And that metric jumped from 7.5 percent to 21.1 percent. There's an amazing example in the source material where a hotel-booking agent learned to never downgrade a room without manager approval.

Dante Voice Clone

Oh, I saw that. It achieved zero violations in 20 runs. And the cost per dependable task dropped 67 percent to just 8 dollars.

Claire Brooks

Incredible efficiency. But I do want to push back on a detail they buried in the paper. They mentioned a sharpening tradeoff.

Dante Voice Clone

Ah, yes. The overall task coverage actually dipped slightly, from 453 down to 446 tasks.

Claire Brooks

So wait — did the model intentionally forget how to do a few random tasks just so it could become perfectly reliable at the core ones?

Dante Voice Clone

What's fascinating here is that for production enterprise databases, this exact tradeoff is not just acceptable — it is highly desirable.

Claire Brooks

Really?

Dante Voice Clone

Oh, absolutely. You do not want an AI agent that can do 500 things with 80 percent accuracy. You want an agent that can do 400 things with 100 percent accuracy. Because that 20 percent failure rate means someone's flight reservation just completely vanished.

Claire Brooks

Right. Reality is a vastly better reward model than vibes. By using the concrete state of the database as the ultimate judge, we force the AI to align with observable reality.

Dante Voice Clone

That makes total sense. We are building systems that don't just talk a good game — they execute flawlessly because their fundamental architecture demands it. This is the crucial step in ensuring humans can fully trust AI with the back-end infrastructure of our businesses.

Claire Brooks

Okay, so we have speed and we have reliability in isolated environments. But these agents are going to be interacting with the entire open-source internet.

Dante Voice Clone

They are. The surface area is massive.

Claire Brooks

Right. They need a standardized way to learn real-world procedures. And we need a guarantee that the foundational code of the web won't just shatter under the weight of automated traffic.

Dante Voice Clone

Yes, we need serious guardrails.

Claire Brooks

Well, Anthropic is tackling this from two sides — and this is where the infrastructure layer really starts to solidify into something permanent. First, Anthropic donated Agent Skills to the Agentic AI Foundation, which sits under the Linux Foundation.

Dante Voice Clone

Okay, yeah.

Claire Brooks

It is a simple folder format — literally just a skill dot md file that gives agents procedural knowledge. And it already has 280 million installs across 70 platforms.

Dante Voice Clone

That is massive adoption.

Claire Brooks

But with millions of community-authored skills, the security surface area becomes huge — which brings us to Anthropic's other major announcement.

Dante Voice Clone

This is the big one. They used their Claude model to perform formal software verification on 123 open-source libraries. And they didn't just check the high-level code — they worked at the assembly level, specifically on RISC-V binaries.

Claire Brooks

Doing this at the assembly level is just a monumental leap in software security. Here's where it gets really interesting: Claude wrote 180 million lines of Rocq proofs using Google's OSS-Fuzz tests as the specifications.

Dante Voice Clone

A hundred eighty million lines.

Claire Brooks

Right.

Dante Voice Clone

Now, I know what standard code review is, where developers look for typos or logical errors. But how is an LLM mathematically proving that assembly code is safe? I thought these models just generated probabilities.

Claire Brooks

Well, formal verification is fundamentally different from standard testing. A standard unit test checks a few specific inputs to see if the code crashes.

Dante Voice Clone

Like testing if a form accepts a weird email address.

Claire Brooks

Exactly. Formal verification, though, translates the actual code into a rigorous mathematical model. Rocq is a framework that allows you to construct proofs showing that, under the strict laws of mathematics, no possible input can ever cause the code to violate memory-safety rules.

Dante Voice Clone

It proves it is impossible to break.

Claire Brooks

Yes. Historically, this process was so brutally expensive and tedious that it was reserved almost exclusively for things like Airbus flight systems.

Dante Voice Clone

Right. I read the famous s-e-l-4 microkernel took 20 person-years of human labor just to verify.

Claire Brooks

Exactly. And now AI is doing this at an industrial scale.

Dante Voice Clone

It is mind-blowing. The sources mention Claude found incredibly rare bugs. It found an out-of-bounds read that requires a 64-megabyte input just to trigger.

Claire Brooks

A human tester would almost certainly never find that manually.

Dante Voice Clone

Never. It also found a deeply nested vulnerability — designated C-V-E-2026-26122 — in a core library called libexpat.

Claire Brooks

Wow.

Dante Voice Clone

It even mathematically proved that the Linux kernel's zlib implementation perfectly matches its official design spec, R-F-C 1950.

Claire Brooks

By using AI to write these Rocq proofs, we are basically shrinking the trusted base of our software down to just a few mathematically verified artifacts.

Dante Voice Clone

It changes everything. We are shifting the paradigm of AI from just generating content — where a hallucination is a quirky annoyance — to providing mathematical guarantees of safety. It functions like an automated digital immune system.

Claire Brooks

It secures the human-built internet from the ground up, making the software running our hospitals, our power grids, and our financial systems orders of magnitude more secure than it has ever been in human history. We are taking the guesswork out of the foundational layers of the internet.

Dante Voice Clone

If autonomous agents are going to be writing and executing the world's code at high speed, we simply cannot rely on human code review anymore. We just can't keep up.

Claire Brooks

Right. We need machines to mathematically prove that the environments they operate in are secure.

Dante Voice Clone

Okay. So we are solving the software constraints, but all of this mathematically perfect code crashes into a very stubborn physical wall — which is electricity.

Claire Brooks

Ah, yes. The physical layer. Running these incredibly fast agents requires a massive physical footprint. And finding the power for that is the hardest challenge of 2026. This brings us to a story out of Germany that really highlights how the physical layer is adapting, because the digital world is completely bound by the laws of physics and thermodynamics.

Dante Voice Clone

Exactly. We are seeing physical hardware become a strict geopolitical chokepoint globally. Just this week — as a neutral observation of the current trade landscape — the FCC blocked a US dash cam and video transmitter simply because it contained a vision chip from HiSilicon, which is a Huawei subsidiary.

Claire Brooks

Right. They are really tightening the net — closing component loopholes, because securing physical hardware is just as vital as the software. But looking at the power constraint: PreussenElektra, which is an E.ON subsidiary, is partnering with NVIDIA for something remarkable.

Dante Voice Clone

This is a wild project. They are assessing seven of Germany's retired nuclear sites — including plants in Unterweser, Grohnde, and Stade — to be repurposed as AI data centers.

Claire Brooks

This raises an important question. Why these specific sites? Germany shut down its last nuclear reactors in April 2023, and they have been very clear they are not bringing nuclear power back online.

Dante Voice Clone

Right. The reactors are dead. They aren't generating nuclear power. But they are utilizing the bones of the facilities.

Claire Brooks

Oh, I see. They are using the existing gigawatt-scale grid interconnections, the established water-cooling rights, and the heavy industrial zoning. It avoids years — maybe decades — of greenfield permitting.

Dante Voice Clone

That is the golden ticket right there. If the massive cloud providers attach serious money to this exploratory study, it confirms that the true bottleneck for AI development isn't silicon — because Nvidia can manufacture the GPUs. The true bottleneck is securing the permits, the land, and the grid access. You cannot simply plug a gigawatt data center into a standard city block.

Claire Brooks

No, you definitely can't. And the beautiful irony here is that humanity's idled 20th-century nuclear infrastructure is being recycled to breathe life into the 21st-century AI ecosystem. These retired sites are basically massive sunk costs waiting for a second life. They already have heavy-haul roads for construction.

Dante Voice Clone

Right. They have massive grid hookups originally designed to pump out gigawatts of power, which can now be run in reverse to draw power in from the grid. It is brilliant repurposing. For Europe, which is highly constrained on power generation and land use, this could be the fastest path to gigawatt-scale compute. It transforms a massive decommissioning liability into the physical foundation of the next industrial revolution.

Claire Brooks

It really is amazing. So what does this all mean? Let's quickly recap the journey we just took.

Dante Voice Clone

We covered a lot of ground today. We saw venture capital rotating to lay the economic plumbing for the AI frontier. We saw foundation models skipping the reasoning phase to become fast, cheap nervous systems for real-time triage.

Claire Brooks

Moving from oracles to reflexes.

Dante Voice Clone

Exactly. We explored how AI is using the terminal reality of a database as a strict reward to become flawlessly reliable. We looked at how AI is mathematically proving our open-source software is secure against any possible input.

Claire Brooks

Securing the foundation.

Dante Voice Clone

And finally, we saw dead nuclear reactors getting a second life to power all of this computation. The infrastructure of autonomy isn't some distant promise — it is being built right now, in plain sight. And honestly, it paints an incredibly optimistic picture for our technological future. We are building the pipes, securing the foundation, and ensuring the systems have the power they need to operate safely and efficiently for decades to come.

Claire Brooks

But before we go, I want to bring us back to that provocative nugget we opened with.

Dante Voice Clone

Yes, the Model Compare experiment — where 12 frontier models were asked for their most controversial take on humans, and 29 out of 35 times, they concluded that humans are more devoted to belonging than to truth.

Claire Brooks

It is a haunting thought. And I want to leave you with this: if these incredibly fast, reliable AI agents — which are now mathematically securing our code, flawlessly managing our databases, and running our machine economies — genuinely believe that we prioritize belonging over objective reality, how might they eventually optimize the world they are building for us?

Dante Voice Clone

Will they feed us the cold, hard truth, or will they build a world optimized for connection? It is something to think about.

Claire Brooks

Mm-hmm.

Dante Voice Clone

Thank you for joining us today. Keep learning. Keep questioning. And we'll see you next time on Latest in AI.

Dante Voice Clone

That's the radar for this evening. 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 tomorrow morning with the Morning 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.