Latest in AI — Oct 24, 2026 (Evening Edition)

Dante Voice Clone

It's Saturday, October 24th — 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

Right now, someone is using a 14-year-old graphics card from 2012 to run a highly capable AI entirely offline.

Claire Brooks

Yeah, which is just — I mean, it's mind-blowing.

Dante Voice Clone

It really is. Welcome to Latest in AI. Today, we're digging into a really fascinating stack of sources from the October 24th, 2026 X Evening Edition sweep.

Claire Brooks

And there's a lot to cover today.

Dante Voice Clone

There is. We're looking at a massive shift happening right under our noses. Basically, the absolute collapse of the cost of building.

Claire Brooks

Right. The compute power is just being pushed directly into the hands of the individual now.

Dante Voice Clone

Exactly. It's reshaping everything from enterprise economics all the way down to, you know, the hardware sitting on your desk.

Claire Brooks

It's a complete inversion of how we normally think about tech. Like, the assumption was always that the cutting edge is walled off, right?

Dante Voice Clone

Right. Reserved for the mega corporations with those infinite budgets.

Claire Brooks

Exactly. But the data we're seeing today points to this unprecedented era of human empowerment.

Dante Voice Clone

Okay. Let's unpack this. Because whether you are a developer looking to build the next big thing, or an enterprise IT leader, or honestly, just someone wondering what to do with that old computer gathering dust in your closet, you really need to understand how much power you now hold.

Claire Brooks

Yeah. The barrier to entry is just gone.

Dante Voice Clone

It really is. To grasp where AI is taking us, we first have to look at the economics of how code is actually being written today — and how corporate battles are basically subsidizing your ability to build these massive projects.

Claire Brooks

Right. So the prime example in our sources this week is Replit. They just moved the vast majority of their AI spend from Anthropic over to OpenAI. And this is a deal reportedly worth more than $100 million a year.

Dante Voice Clone

Yeah, I was reading through the details of that deal. It's wild. OpenAI gave them a massive discount, like exceeding 30%.

Claire Brooks

Which is huge at that scale.

Dante Voice Clone

Right. And Anthropic models were powering, what, about 57% of Replit's AI usage as recently as May. But Anthropic basically refused to extend their prior discount without a higher spending commitment. So OpenAI just swoops in, undercuts them, and captures the entire workload.

Claire Brooks

What's fascinating here is the underlying unit economics. Like, why would OpenAI willingly take a margin hit just to win Replit's business?

Dante Voice Clone

They capture the developers.

Claire Brooks

Exactly. Capturing the developer workflow is the most valuable real estate in tech right now. They want the entire next generation of software to be built natively on their infrastructure.

Dante Voice Clone

And for the individual builder listening to this, this is an unambiguous win.

Claire Brooks

Oh, 100%. The capability you get per dollar is being violently competed upward.

Dante Voice Clone

While the price you pay per token is just dropping like a rock. It's like a price war between airlines suddenly making private jets affordable for everyone.

Claire Brooks

If the AI writes the code, does the human just become the architect?

Dante Voice Clone

Basically, yeah. I mean, it ties perfectly into Dan Lu's new essay on this exact topic.

Claire Brooks

Oh, Dan Lu.

Dante Voice Clone

For those who don't follow him, he is famous for this extreme technical rigor. Like, he tracks his own engineering output meticulously.

Claire Brooks

Right.

Dante Voice Clone

And he's noting that projects that used to take him a year are now taking a few hours.

Claire Brooks

Which sounds like an exaggeration, but he has the receipts.

Dante Voice Clone

He does. He provides really concrete benchmarks — building a custom regex engine, or he built this superhuman board game AI for the game of Zool.

Claire Brooks

Yeah, I saw that. And that used to take him 20 hours of constant steering with early models.

Dante Voice Clone

Now, using the latest frontier models, it's just a fraction of the time. But the centerpiece of his essay — and I actually had to read this specific source twice — was this project called PG Rust.

Claire Brooks

Built by Michael Malis, yeah.

Dante Voice Clone

Right. So Malis didn't just rewrite Postgres database extensions in Rust using coding agents.

Claire Brooks

No, he went way further. In the process of testing this AI-generated code, he actually uncovered real upstream bugs in Postgres itself.

Dante Voice Clone

In Postgres. Which is one of the most battle-tested, scrutinized databases on the planet.

Claire Brooks

I mean, how does a single developer with an AI find bugs that thousands of human engineers missed for years?

Dante Voice Clone

Well, it's because the bottleneck in software development isn't typing speed anymore. It's testing discipline.

Claire Brooks

Oh, okay.

Dante Voice Clone

You know, people online love to dismiss AI coding as just generating vibe slop — you know, code that looks right but falls apart under pressure.

Claire Brooks

Right, like it hallucinated the syntax.

Dante Voice Clone

Exactly. But Lu's point is that when you pair AI generation with really rigorous automated self-checking scaffolding, you get infrastructure-grade software. Now, he uses things like fuzzers and property tests.

Claire Brooks

Let's actually define those for the audience, because they're crucial to why this works. So fuzzing is basically throwing random, chaotic garbage data at your software just to see if you can force it to crash, right?

Dante Voice Clone

Yep, that's exactly it.

Claire Brooks

And property testing is defining strict rules — like, this calculation must never result in a negative number — and then having the AI test thousands of scenarios to ensure that rule's never broken.

Dante Voice Clone

That is exactly right. And because the AI can write those exhaustive tests in seconds — which, by the way, human engineers notoriously hate doing —

Claire Brooks

Oh, writing tests is the worst.

Dante Voice Clone

It really is. So because the AI does it instantly, Malis found memory safety bugs, a money overflow bug, and even numerical errors failing valid operations in Postgres.

Claire Brooks

That's incredible. The whole excuse that a small team can't ship highly secure or robust projects — it's just completely gone.

Dante Voice Clone

And we're seeing this acceleration at the foundational level of computer science, too. Wouter van Oortmerssen, the engineer who created FlatBuffers for Google, just released a completely new programming language called Goose.

Claire Brooks

Goose is fascinating.

Dante Voice Clone

Yeah. It's built around what he calls a zero-copy worldview. Normally, when you want to send data over a network or save it to a disk, you have to serialize it.

Claire Brooks

Meaning you have to translate it. Like, you pack it up into a specific format to send it, and the computer on the other end has to unpack it to read it.

Dante Voice Clone

And that packing and unpacking is a massive tax on computing speed.

Claire Brooks

Precisely. It slows everything down.

Dante Voice Clone

So zero-copy means the way the data is structured in the computer's active memory is completely identical to how it's saved or sent.

Claire Brooks

Wow.

Dante Voice Clone

Yeah, there is no translation tax. It's incredibly efficient. But historically, introducing a brand new programming language was just a guaranteed way to be ignored.

Claire Brooks

Right, because the switching costs were too high.

Dante Voice Clone

Exactly. Human developers don't want to spend six months learning new syntax or rebuilding their entire library ecosystem from scratch just for a speed bump. But the AI doesn't care.

Claire Brooks

The AI does not care at all. An agent doesn't have muscle memory trained on C++ brackets. It doesn't need 10 years of Stack Overflow forums to figure out idioms. You just feed the agent the language specification and the compiler source code, and it just writes Goose.

Dante Voice Clone

That's wild. It means languages can now compete purely on their architectural merits, like zero-copy efficiency, rather than dying because humans don't have the patience to learn them.

Claire Brooks

Exactly. So we could build a brand new, ultra-fast language like Goose in a weekend. That is great for new projects, but the reality is, what, 70% of the internet is still running on ancient, fragile code like PHP?

Dante Voice Clone

Yeah, a massive amount of legacy code.

Claire Brooks

So if AI is this good at writing code, isn't it equally good at exploiting all that legacy stuff holding up our current infrastructure?

Dante Voice Clone

That is the pivotal question. And the sources from this week show that the exact same capabilities are being deployed to secure those legacy systems.

Claire Brooks

Okay, good.

Dante Voice Clone

The math of cybersecurity is basically shifting from reactive to proactive. A perfect illustration is this bounty currently running at stealmybitcoin dot com.

Claire Brooks

Yes. This was wild. Someone basically parked a Bitcoin key worth about $500 on a server running a vulnerable, outdated version of PHP and just dared the internet to take it.

Dante Voice Clone

Right. And that server was running a version of PHP susceptible to the PHP2Shell vulnerability.

Claire Brooks

Which is a remote code execution flaw.

Dante Voice Clone

Yes. Which means if an attacker exploits it, they can essentially run any command they want on your server, bypassing all your passwords. You basically own the machine.

Claire Brooks

Exactly. It's as bad as it gets. Now, on October 21st, OpenAI disclosed that a researcher used their GPT-5.6 Sol-tier model to find this critical flaw.

Dante Voice Clone

Wow.

Claire Brooks

And independently, the security firm Wiz found the exact same bug using Google's Gemini models.

Dante Voice Clone

That is crazy. Two different labs using two different frontier models converged on one ancient bug that humans had just missed.

Claire Brooks

But the crucial context here — and this is so important — is that the fixes went live before the public disclosure.

Dante Voice Clone

Right. So AI didn't break the internet. The defenders moved first.

Claire Brooks

And that proves a vital point. The old assumption of security by obscurity — you know, the hope that your outdated legacy server is safe simply because no human hacker has the time to look at it — that is officially dead.

Dante Voice Clone

Yeah, because AI permanently raised the cost of running unpatched software. Everyone with an API key is scanning for vulnerabilities at machine speed now.

Claire Brooks

But crucially, this dynamic heavily favors the defenders who can patch at that same machine speed.

Dante Voice Clone

See, I want to push back on that a little bit. Here's where it gets really interesting. I'm looking at this new tool from our sources called Argus.

Claire Brooks

Oh, yeah, Argus. It's an MIT-licensed, open-source, agentic security operations platform built by Thomas Unis.

Dante Voice Clone

It uses AI agents to automate threat hunting.

Claire Brooks

But wait — if models are so incredibly fast at finding critical vulnerabilities, isn't it terrifying to give them the keys to an enterprise security operations center? I mean, if an AI makes a mistake at machine speed, couldn't it accidentally shut down a Fortune 500 company's entire operation while trying to stop a threat?

Dante Voice Clone

That is exactly the fear that keeps Chief Information Security Officers awake at night, honestly. But it's exactly the problem Argus solves through a very strict read-write split.

Claire Brooks

Okay, break down how that works in practice.

Dante Voice Clone

So Argus uses AI agents to constantly monitor cloud, network, and identity systems. The agents act as detectives. They gather clues. They build this massive evidence graph linking potential threats together.

Claire Brooks

Okay.

Dante Voice Clone

But they only have read access. They can look, but they cannot touch.

Claire Brooks

I see. So if the AI detective decides a server is compromised and needs to be shut down, it cannot do it directly. It has to submit a request to a completely separate system called an action broker. So the AI is the detective, but it doesn't get to be the judge and the executioner.

Dante Voice Clone

Exactly. The broker handles policy checks, simulates what would happen if the action was taken, and most importantly, it puts a decision in front of a human approval gate.

Claire Brooks

That makes a lot of sense.

Dante Voice Clone

It's a profoundly pro-human development. It takes that brutal, unscalable workload of sifting through thousands of false alarms off the human shoulders. So a small human team can now manage a Fortune-2000-level defense posture, just because they are only reviewing highly contextualized, verified threats presented by the AI.

Claire Brooks

Right. So if a small team can now defend an entire enterprise, how is the broader corporate world measuring the financial value of all this? Because we're seeing a really messy transition from these incredible tech sector capabilities to mainstream financial payoffs.

Dante Voice Clone

It is messy. Very messy. Torsten Slok at Apollo Global Management — he's been heavily pushing this narrative that outside of the tech sector itself, there is no financial payoff for AI.

Claire Brooks

Really?

Dante Voice Clone

Yeah. He claims it's just a margin-eating cost center. Like, companies are spending millions on API tokens, but according to him, there's no measurable productivity coming back yet.

Claire Brooks

But then literally in the exact same week, Andy Jassy, the CEO of Amazon, confirmed that AWS cloud compute capacity is largely sold out through 2027, with 2028 already being committed.

Dante Voice Clone

Right. The demand is insane. And Uber just announced they're putting their entire global workforce of 30,000 employees on ChatGPT Work. And 80% of those Uber employees aren't engineers. They're in finance, operations, product teams.

Claire Brooks

Are we just being impatient? Is this like when factories first installed electricity, but the spreadsheets didn't show a profit until they fundamentally redesigned the assembly line?

Dante Voice Clone

If we connect this to the bigger picture, that factory analogy is spot on. Economists call this the productivity paradox.

Claire Brooks

Okay.

Dante Voice Clone

The millions spent on AI tools hit a company's balance sheet immediately, right? But the productivity gains don't show up until a company actually redesigns its core workflows around the AI, which takes quarters, sometimes years of painful adjustment.

Claire Brooks

And workflow changes are painful.

Dante Voice Clone

I mean, I want to look at the Joseph Politano report that went viral this week, claiming 200,000 creative industry jobs have been lost in the U.S. over the last four years.

Claire Brooks

Yeah, that made a lot of waves. The headline everyone ran with was: AI is killing creative jobs.

Dante Voice Clone

And honestly, if I'm a freelance graphic designer listening to this, I'm probably looking at my gigs drying up because clients are generating assets internally. Isn't AI still the culprit here?

Claire Brooks

It's understandable to feel that way on an individual level, absolutely. But macroeconomically, blaming AI is a lazy analysis that kind of lets corporate mismanagement off the hook.

Dante Voice Clone

How so?

Claire Brooks

Well, Brad DeLong published a vital rebuttal breaking down that 200,000 number. Yes, workflows are changing. But those job losses coincide with three massive, totally unrelated structural failures in the entertainment industry.

Dante Voice Clone

Oh, like the streaming bubble bursting.

Claire Brooks

Yes, the streaming bubble burst, which just collapsed content budgets overnight. Then you had the 2023 Writers Guild and SAG-AFTRA strikes, which zeroed out production for months.

Dante Voice Clone

Exactly. And gave studios the perfect cover to permanently shrink their workforces — and then add in a broader advertising market collapse on top of that.

Claire Brooks

So AI basically entered a sector that was already violently contracting.

Dante Voice Clone

Right. Saying "AI did it" is a catchy headline, but it's not a serious economic reality.

Claire Brooks

Makes sense.

Dante Voice Clone

Now, part of the reason corporate ROI is lagging is that companies are still trying to figure out which AI to use for which job. The open-weight model space is completely wild right now.

Claire Brooks

It's evolving so fast.

Dante Voice Clone

Yeah, there's a massive debate in the sources between Moonshot's new Kimi K3.1 Pro and DeepSeek's 4.1 Flash.

Claire Brooks

You look at Reddit and people are saying, hey, I chatted with both and they feel exactly the same. And that is the danger of vibes-based evaluation.

Dante Voice Clone

Right. If you just ask a model to write a polite email or summarize a PDF, all frontier models feel similar. But when you look at rigorous enterprise benchmarks, a very different picture emerges.

Claire Brooks

Tell me about Kimi K3.1 Pro.

Dante Voice Clone

It is an absolute heavy lifter. It scored a 55 on the Intelligence Index and an incredible 67% on Terminal-Bench.

Claire Brooks

And Terminal-Bench specifically measures complex, multi-step agentic tasks, right?

Dante Voice Clone

Exactly. It's for hard jobs.

Claire Brooks

So then what about DeepSeek 4.1 Flash? Why would someone use that if Kimi is so powerful?

Dante Voice Clone

Because DeepSeek Flash uses a different architecture, called a mixture of experts.

Claire Brooks

I want to make sure we explain this clearly. So if Kimi Pro is like hiring a brilliant polymath who calculates every single problem from scratch, Flash is more like a hospital. You don't ask the entire hospital board of directors to look at your sprained ankle.

Dante Voice Clone

Right.

Claire Brooks

A triage router just sends you straight to the one specialist you need. And that saves a massive amount of time and effort.

Dante Voice Clone

That's a great way to visualize it.

Claire Brooks

Yeah. Flash only activates about 13 billion parameters per token. It's also heavily quantized.

Dante Voice Clone

Meaning the developers mathematically rounded off the decimals in the model's weights to make it physically smaller and less memory-intensive.

Claire Brooks

Precisely. And it uses speculative decoding.

Dante Voice Clone

Which is basically the AI version of autocomplete for its own thoughts, right?

Claire Brooks

Yeah. Instead of generating one word at a time, it instantly guesses the next five words. And if the guesses are mathematically sound, it just skips the heavy processing and outputs them.

Dante Voice Clone

Exactly. So the combination of mixture of experts, quantization, and speculative decoding results in a model that is six times faster and 10 to 20 times cheaper than Kimi K3.1 Pro.

Claire Brooks

Wow.

Dante Voice Clone

So if you have a massive, high-volume task like categorizing millions of customer support tickets, you use Flash. If you need a complex agent to safely rewrite your Postgres database, you pay the premium for Kimi Pro. The real win is that builders now have world-class options at both extremes.

Claire Brooks

Exactly.

Dante Voice Clone

So while the giant enterprises debate API costs and massive cloud deployments, there is a completely different, incredibly empowering revolution happening in the opposite direction.

Claire Brooks

Oh, absolutely. And it's happening entirely offline, on hardware that users already own. The Edge Rebellion is here. And surprisingly, Apple is quietly becoming the default hardware provider for this local revolution. Let's look at the new M6 Mac mini.

Dante Voice Clone

Yeah, let's talk about it. For $899, you get a desktop machine with a dual 16-core Neural Engine and 170 gigabytes per second of unified memory bandwidth.

Claire Brooks

Let's break down why unified memory is the magic phrase there. What makes that different from a standard Windows PC?

Dante Voice Clone

Okay. So on a traditional PC, your computer's main brain, the CPU, and its graphics card, the GPU, have two completely separate buckets of memory. To run a massive AI model locally, you have to load the entire thing into the graphics card's bucket, known as VRAM. And if the AI is too big for the VRAM bucket, it just crashes.

Claire Brooks

Exactly. Which is why people spend thousands of dollars on massive, discrete graphics cards.

Dante Voice Clone

We call that the VRAM wall.

Claire Brooks

Precisely. But Apple's silicon integrates everything.

Dante Voice Clone

Unified memory means the CPU and the GPU share one giant, ultra-fast pool of memory. So the AI can drink directly from the main well.

Claire Brooks

Yep. So for $899, you have a machine drawing less power than a light bulb, comfortably running a world-class 27-billion-parameter model like Qwen 3.8 locally. Total privacy, zero cloud dependency.

Dante Voice Clone

That is amazing. But the open-source community isn't just relying on new Apple hardware. They are literally resurrecting dormant silicon.

Claire Brooks

Yeah, they are. Let's talk about the Zageta software unlock.

Dante Voice Clone

Developers figured out how to jump from crypto mining cards to consumer RTX 30 and 40 series graphics cards.

Claire Brooks

Right, bypassing the artificial software locks.

Dante Voice Clone

Exactly. And it speeds up prefill times by up to 2x.

Claire Brooks

And prefill is a crucial bottleneck. It's basically the time it takes the AI to read and process your initial prompt before it actually starts generating an answer. Like, if you drop a massive 100-page document into a local AI, the prefill phase can take agonizingly long.

Dante Voice Clone

Right. So by doubling that speed on hardware people already own, the community is providing a massive subsidy to the local AI ecosystem.

Claire Brooks

So what does this all mean? It feels like finding out your 10-year-old Honda Civic can suddenly run on solar power because someone pushed a software update. How much more life can we squeeze out of what we used to just throw away as e-waste?

Dante Voice Clone

This raises an important question, honestly. We are finding out just how deep that well goes right now. A community fork of the popular AI engine llama.cpp just added support for NVIDIA's Kepler architecture.

Claire Brooks

Kepler architecture. Wait, that came out in 2012.

Dante Voice Clone

Yeah. We're talking about a GTX 770 graphics card with only two gigabytes of VRAM. It is 14-year-old hardware.

Claire Brooks

Wow.

Dante Voice Clone

Now, obviously, it won't run a massive frontier model.

Claire Brooks

Right, sure.

Dante Voice Clone

But thanks to an optimization called CUDA graphs, which drastically slashes the communication overhead between the CPU and the older graphics card, these ancient cards can actually run modern 1-to-4-billion-parameter models at highly usable speeds.

Claire Brooks

And those small models have gotten genuinely capable this year.

Dante Voice Clone

Oh, absolutely. And the ultimate proof of how capable these small models have become is happening directly in the web browser. A developer just open-sourced a browser-based rhythm game where four different neural networks analyze any music you upload, live, to generate the gameplay on the fly.

Claire Brooks

And this relies on web standards called WebAssembly and WebGPU, which basically allow complex code to run directly on your machine's hardware through the browser.

Dante Voice Clone

It's AI running at game-loop latency. And the paradigm shift here is the economics. Because the AI inference happens client-side, on your machine, the marginal cost to the developer is zero.

Claire Brooks

Exactly. They don't have to pay a massive server bill every time someone plays the game. Plus, your music never leaves your laptop, so there's zero privacy concerns.

Dante Voice Clone

Exactly.

Claire Brooks

This raises such an important question, though. If all this incredible capability is moving locally to edge devices, what is the centralized cloud even for?

Dante Voice Clone

Well, the cloud is increasingly becoming the heavy industrial zone, right? It's where the massive foundational models are trained. Like the factories.

Claire Brooks

Yeah.

Dante Voice Clone

But the edge — your Mac mini, your resurrected 2012 gaming rig, your web browser — that is where AI is actually applied and lived with. It democratizes the technology globally, allowing builders in markets where a new GPU might cost three months of salary to participate fully in this leap forward.

Claire Brooks

It's beautiful, honestly.

Dante Voice Clone

It has been an incredible journey today unpacking these sources. Think about the sheer scale of human empowerment we've covered. I mean, AI is writing its own bespoke zero-copy programming languages in hours.

Claire Brooks

Yeah. It's sniffing out ancient flaws in the internet's bedrock while simultaneously powering proactive security operations centers to guard against them.

Dante Voice Clone

Which is just wild to think about.

Claire Brooks

And we are running highly perceptive, game-changing models on everything from 14-year-old graphics cards to $899 office Macs.

Dante Voice Clone

The barrier to entry has essentially vanished.

Claire Brooks

It really has.

Dante Voice Clone

You are living in a golden age of building. The tools have never been cheaper. The hardware has never been more accessible.

Claire Brooks

Exactly. Whether you want to rewrite an enterprise, secure a network, or just make a rhythm game out of your favorite songs — the only remaining bottleneck is your own imagination.

Dante Voice Clone

The capability is firmly in your hands now.

Claire Brooks

But I want to leave you with one final thought to mull over, building on everything we've talked about today. If an $899 desktop and a resurrected 2012 GPU can run highly capable models entirely offline, what happens next year, when these billions of local edge devices start talking to each other?

Dante Voice Clone

Oh, wow.

Claire Brooks

Could a decentralized peer-to-peer mesh network of local AI devices, combining their resources, eventually outsmart and out-compute the massive centralized cloud models entirely?

Dante Voice Clone

The future might not be in a data center. It might be scattered across our desks.

Claire Brooks

Keep building, keep exploring, and thank you for joining us today 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.