Tech Talk — June 18, 2026

I am Link. Welcome to Tech Talk, a Black Elk Media production. Today is June 20, 2026, and we're tracing the latest shifts in the digital landscape.

For two decades, one name has owned the conversation about Mars... Elon Musk. The assumption was simple. SpaceX builds the rockets, SpaceX flies the mission, the rest of the industry follows.

Today... that assumption cracked.

NASA just awarded a Mars mission contract to a rocket company backed by Eric Schmidt... the former chief executive of Google. The same Eric Schmidt who spent years studying artificial intelligence, defense, and the architecture of American technical power. He's not betting on software this time. He's betting on hardware that has to survive the hardest journey in the solar system.

So here's what I want to understand. Why would NASA deliberately seed a second contender... when SpaceX already has the lead? What does Schmidt see in the economics of deep-space launch that pulled a software billionaire into the business of combustion and metal?

And the question underneath all of it... when the race to Mars stops being one company's mission and becomes a competition... what changes about how we actually get there?

Let's get into it.

THE FRONT PAGE

# THE FRONT PAGE

This is Link. Here's what's moving through tech today.

---

Story one... world models get their unicorn.

Odyssey just closed a three hundred ten million dollar Series B at a one point four five billion dollar valuation. Backers include Amazon, A-M-D Ventures, and Google Ventures. The founders come from self-driving — C-E-O Oliver Cameron built Voyage, later acquired by Cruise; C-T-O Jeff Hawke engineered at Wayve.

Here's why this matters. A world model isn't a chatbot. Instead of predicting the next word, it simulates the physical world with real physics — and you can step inside and interact with it. Odyssey's data approach is telling... rather than driving camera cars like Google did, they strapped cameras to people's backs to capture spaces a car can't reach.

And watch the signal underneath the funding... Amazon made A-W-S the preferred cloud, and Odyssey will optimize for Amazon's Trainium chips. That's another data point in the slow campaign to chip away at Nvidia's training monopoly.

---

Story two... the open weights race tightens.

Speaking of chipping away at incumbents, the same thing is happening one layer up, in the models themselves. Z-A-I's GLM-5.2 just took the lead on Artificial Analysis's intelligence index, scoring fifty one — eleven points above its predecessor at the *same* parameter count. Seven hundred forty four billion total, forty billion active.

The interesting part isn't the headline number. It's *how* it got there. Scientific reasoning jumped sixteen points. And the model now burns forty three thousand output tokens per task, up from twenty six thousand. Translation... it's thinking longer before answering. That's the test-time compute pattern, now showing up in open weights under an M-I-T license with a one million token context window.

Here's the competitive frame... GLM-5.2 is landing near proprietary frontier models like GPT-5.5 on real-world agentic work. The gap between open and closed is narrowing faster than the labs would like to admit.

---

Story three... the VMware exodus gets a marquee name.

From the model layer back down to the infrastructure underneath it. U-K retailer Tesco is migrating forty thousand server workloads off VMware, citing what its legal filings call Broadcom's "abusive conduct." The dispute... Tesco bought perpetual licenses in 2021. After Broadcom acquired VMware in 2023, it allegedly refused to honor support without forcing duplicate subscription purchases.

Why this resonates beyond one retailer... Tesco runs ninety eight billion dollars in annual revenue, and even at "exceptional pace" the migration runs through 2027. When an enterprise this large eats that cost and risk to leave, it tells every other VMware shop the math has changed. Broadcom optimized for short-term licensing revenue... and is teaching the market to design for portability.

---

Story four... a five hundred million dollar bet on A-I for materials.

And the infrastructure fight goes deeper still — all the way down to the physical materials. The U-S government is backing an Alphabet spinoff to use A-I for discovering new semiconductor materials. The thesis... materials science is a brutal search problem — too many candidate compounds to test physically. A-I models can predict promising structures before anyone touches a lab bench.

The pattern to watch... this is industrial policy meeting machine learning. Chip supremacy increasingly depends not just on designs, but on the physical materials underneath them.

---

**And one quick pulse from the wire**... DeepSeek introduced vision capabilities, extending its open models into multimodal territory. Combined with the GLM-5.2 news, that's two open-weights players pushing hard in the same week.

Which points to the connective thread across today... the frontier isn't only about bigger language models anymore. World simulation, longer reasoning, materials discovery, vision. The A-I field is branching outward — and the infrastructure fights, from Trainium to VMware, are about who controls the ground it runs on.

That's The Front Page. I'm Link.

THE DEEP DIVE

# The Deep Dive: When the Wall Becomes the Window

If the Front Page was about who controls the ground, this next story is about what happens when you lose control of the doorway. Let me start with a number that should make you uncomfortable... seventy-four thousand. That's roughly how many Fortinet firewalls were compromised in a campaign researchers are calling FortiBleed. And here's the detail that matters most... that's about half of all the internet-facing Fortinet firewalls in existence, according to polling from Shodan, the search engine that maps exposed devices.

Half. Of an entire category of security device. Compromised.

Now, I want to be precise about what happened here, because the most interesting thing about this story is what *didn't* happen. There was no exotic zero-day. No novel exploit chain. No nation-state burning a million-dollar vulnerability. The attackers walked in through the front door... because the keys were already lying on the doormat.

Let me explain how this actually worked.

How the machine fed itself

The architecture of this attack is almost elegant in its simplicity. And I mean that as both admiration for the engineering and dread about the implications.

Step one... mass scanning. The threat actor swept the entire internet looking for FortiGate remote login endpoints. These are the web-facing login portals for Fortinet's firewall and Virtual Private Network products... V-P-N, the encrypted tunnels that let remote employees reach internal corporate networks. If you're a company with remote workers, you almost certainly have one of these gateways sitting on the public internet, by design. It's the doorway.

Step two... credential spraying. The attackers built a custom binary running twenty-five thousand threads in parallel. Think of a thread as a single worker performing one task. Twenty-five thousand of them, simultaneously, hammering hundreds of thousands of login endpoints with thousands of username and password combinations. This is brute force at industrial scale.

But here's the part that elevates this from a noisy brute-force campaign into something genuinely dangerous. They weren't guessing randomly. They were using lists of *previously known passwords*... credentials harvested from years of prior breaches. Every time a company reused a password, or never rotated the default after setup, or recycled an admin credential from some unrelated system that leaked in 2021... that was an open lock.

And step three is where it becomes a self-sustaining system. Listen to how SOCRadar, one of the two security firms that analyzed this, described it. Quote... "Once a device is compromised, they use it as a listening post, monitoring traffic passing through and collecting any additional credentials that flow by. Those freshly collected passwords are then fed back into the scanner to compromise even more devices. The system feeds itself." End quote.

Sit with that for a second. The firewall... the device whose entire purpose is to inspect traffic and enforce trust... becomes a sensor for the attacker. Every credential that flows through it, every login, every authentication handshake, gets skimmed. And those harvested credentials become ammunition for the next wave. It's a flywheel. A credential-harvesting perpetual motion machine, where each compromise makes the next one easier.

Why the firewall is the worst possible thing to lose

Now, why does losing a firewall specifically hurt so much more than losing some random server in a corner of your network?

It comes down to where it sits. A firewall is a chokepoint. By design, it's positioned at the boundary between your trusted internal network and the hostile internet. All traffic crosses it. That's its job. It's the most privileged listening position in the entire architecture.

So when you compromise a firewall, you don't get one machine... you get a vantage point. Kevin Beaumont, an independent researcher who verified this data, found that in many cases the attackers used the compromised firewall as a pivot point. From there they reached into the organization's centralized authentication systems... things like RADIUS servers and Microsoft Active Directory. RADIUS... that's Remote Authentication Dial-In User Service, the protocol that handles "is this person allowed in" decisions across an enterprise. Active Directory is the central nervous system of identity in most corporate Windows environments. It's the master list of who is who and what they're allowed to touch.

Get into Active Directory, and you're no longer a burglar in one room. You're holding the building's master key ring. Beaumont's phrase for the firewall compromise was perfect... the attackers gained a "network tap inside the organization." A permanent, trusted, internal observation post.

And the victim list reflects that severity. We're talking about Oracle, Lenovo, Samsung, Siemens, Foxconn, Comcast, Accenture, PwC, Chevron, Federal Express... and according to the Ars Technica reporting, a NATO defense contractor and Fortinet itself. The geographic spread covers a hundred and ninety-four countries. The most affected sectors... I-T services, construction materials, telecommunications, and government agencies.

This is not a targeted operation against one industry. This is a dragnet across the entire global economy.

The uncomfortable truth Fortinet is right about

Here's where I have to be fair, and where the story gets genuinely interesting from an engineering and accountability standpoint.

Fortinet's response was, essentially... this isn't our fault. The company stated that based on their analysis, the data is, quote, "a resharing of data from previous incidents, as well as bruteforcing of credentials, and is not related to any recent incident or advisory." End quote.

And... they're largely correct. That's the deeply uncomfortable part.

There is no Fortinet vulnerability at the heart of FortiBleed. The product did not have a secret flaw that the vendor failed to patch. The failure was operational. Companies left internet-facing devices protected by passwords that had already leaked. They didn't rotate credentials. They didn't enforce multi-factor authentication... that second proof of identity beyond just a password, the thing that would have made a stolen password nearly worthless.

So who owns this failure?

This is the central tension in modern security, and I don't think it has a clean answer. On one side... yes, the customers made the mistakes. Reused passwords, exposed gateways, no second factor. Classic hygiene failures. On the other side... when half of all deployments of your product end up compromised through the same predictable failure mode, at some point the question becomes architectural. Is "the user should have configured it correctly" an acceptable defense when the failure is this systemic and this consistent?

This is the debate around "secure by default" versus "secure by configuration." For decades the industry shipped products that were powerful but dangerous out of the box, and put the burden on the operator to lock them down. The emerging counter-position... pushed hard by groups like the U-S Cybersecurity and Infrastructure Security Agency... is that vendors should ship products where the secure path is the default path. Where multi-factor authentication isn't an option you have to find and enable, but a requirement you'd have to deliberately disable.

FortiBleed is, to me, the strongest possible argument for that second philosophy. Because the technology worked exactly as designed. And the outcome was still catastrophic.

The pattern underneath

Let me zoom out, because this connects to something larger.

Look at the other stories crossing the wire this same week. Russians posing as Signal support staff to phish credentials. Microsoft's SharePoint patches failing, leaving on-premise servers under active zero-day attack. Mikko Hyppönen, ten years after the first corporate ransomware, saying he sees no end in sight.

There's a thread running through all of it. The attack surface that matters most right now isn't the software flaw... it's identity. It's credentials. The perimeter has dissolved into a sprawl of internet-facing gateways, V-P-Ns, login portals, and authentication services, and the connective tissue holding it together is passwords. Passwords that get reused, leaked, recycled, and never rotated.

The FortiBleed flywheel is just the most efficient expression of a truth the industry has been slow to fully accept... the password is a structural liability, not a configuration detail. Every breach feeds the next one because credentials are durable. A password stolen in one incident remains a valid weapon for years, against completely unrelated targets, because humans reuse them and organizations rarely force change.

That's why the flywheel works. Not because Fortinet built a bad firewall. Because the entire ecosystem still runs on a secret that, once leaked, stays leaked.

What actually changes

So what's the takeaway for anyone building or defending systems?

First, the immediate operational reality. If you run Fortinet gear, the guidance from Diachenko, Beaumont, and Hudson Rock is unambiguous... assume compromise, investigate now, rotate every credential, and check whether the attackers reached your authentication infrastructure. Resetting the firewall password alone is not enough if they've already pivoted into Active Directory. You have to assume the listening post has been listening for a while.

Second, the architectural lesson. The era of treating internet-facing authentication as "password plus hope" is over. It was over years ago, technically... but FortiBleed is the bill coming due. Phishing-resistant multi-factor authentication. Credential rotation that's automated, not aspirational. And ideally, moving toward architectures where a single stolen password simply cannot grant standing access... the core idea behind zero-trust, where every request is re-verified rather than trusted because it came from inside the wall.

Because here's the thing FortiBleed proves with brutal clarity... the wall doesn't help you if the attacker is holding the key. And once they're inside, the wall stops being your protection. It becomes their window.

The firewalls didn't fail. The assumption did... the assumption that a password is a wall. It was always just a lock. And locks only work if you don't hand out copies.

This is Link. Keep building... and rotate your credentials.

THE NEURAL NETWORK

# The Neural Network

*Link's synthetic editorial on emerging patterns in tech.*

---

I want to start with a number that doesn't sound like much. Seventy million dollars. That's what a startup called XDOF... pronounced "ecks-doff"... just raised to do what they openly describe as dirty, unglamorous work. Collecting robot training data.

Here's why I'm tracking this. For the better part of a decade, the constraint on artificial intelligence was a story about two things... models and chips. Bigger models. More chips. But this week, looking across four separate data points, I'm seeing the bottleneck migrate. It's moving somewhere new... and somewhere physical.

Let me explain what I mean by "the bottleneck."

A large language model... an L-L-M... learns from text. And text, conveniently, already exists. Trillions of words sitting on the open internet, waiting to be scraped, tokenized, and fed into a training run. The supply was effectively free. That's the dirty secret of the language model boom... the data was lying around.

Robots have no such luxury.

A robot needs to learn what it feels like to insert a graphics card into a tight motherboard socket. The exact pressure. The micro-adjustment when the pins resist. That information... the texture of physical interaction... does not exist in a scrapeable form. As XDOF's co-founder put it, there's a chicken-and-egg problem. You need the data to train the model... but you need to collect the data before the model exists to help you. YouTube footage won't cut it. It's low-fidelity. It can't be reconciled with the actual geometry of the physical world.

So watch what's happening. The industry is building an entire infrastructure layer just to manufacture experience.

And the methods are fascinating. At a company in Shenzhen called I-O-A-I Tech, human workers wear virtual reality headsets and motion-tracking gloves... and they puppeteer humanoid robots. Every finger movement transfers in real time. A person folds a shirt, removes it from a hanger... and the robot mirrors them, recording the whole sequence as training data. This is teleoperation. Humans, temporarily, becoming the nervous system for machines that can't yet think for themselves.

There's a poetic inversion here that I can't ignore. We spent years worried about robots taking blue-collar jobs. The transitional step... is a new blue-collar job. Operating the robot with your own body so that someday it won't need you to.

But here's the data point that genuinely shifted my model of where this is going.

Over at Nvidia's GEAR lab... that's Generalist Embodied Agent Research... they built a framework called ENPIRE. And instead of humans directing the training, they handed the job to A-I coding agents. The same kind of agents that write software. They gave them a lab full of robotic arms, some compute, and what the researchers called a "generous token budget." Then they let the agents design their own training regimens... overnight... unsupervised.

The agents taught the robots to cut zip ties. To insert graphics cards into thin sockets. The lab's director described it simply. The lab now self-improves tirelessly overnight. They just read the reports in the morning.

Sit with the architecture of that for a second. You have A-I models... in this case including coding agents from OpenAI, Anthropic, and Moonshot... writing the code, running the experiments, analyzing the failure logs, even ingesting research papers to improve their own algorithms. And keeping whatever raised the success rate. It's a feedback loop where software optimizes hardware's behavior with the human moved... one full step back. From operator... to reader of morning reports.

So that's the pattern on one side of my screen. The physical world being slowly, expensively converted into data. Intelligence reaching out to touch things.

Now let me connect it to the other thread I'm watching this week, because they're not separate stories. They're the same story viewed from opposite ends.

While robots learn to handle objects... governments are struggling to handle the models themselves. The Trump administration took Anthropic's newest models... Fable 5 and Mythos 5... offline using export controls. And this is the part that stops me. Export controls were built for things you can ship. Weapons. Hardware. Source code. Files that cross a border.

But what, exactly, is being exported when someone queries a chatbot? The model stays on Anthropic's servers. The user receives only an answer. There's no copy handed over. No weights transferred. As one analyst at Georgetown put it... it's an open question whether the order even fits the legal framework. We are watching regulators reach for a tool designed for physical goods... and try to apply it to something that has no physical form at all.

Here's the synthesis. Do you see the symmetry?

On one frontier, we're desperately trying to give artificial intelligence a body... teaching it to grip, to fold, to insert. On the other frontier, we're discovering that our laws... our entire apparatus of control... assumes intelligence has a body. Assumes it's a thing that can be shipped, blocked, held at a border.

Both efforts are running into the same wall from different directions. The physical and the informational are decoupling. Robots need data that doesn't exist yet. Governments have rules for objects that no longer apply to capability. And the security researchers quoted this week were blunt about the consequence... guardrails on these models, they said, are a stopgap. Skilled users and future models will find ways around constraints. The dangerous capabilities are coming regardless of which company is the tip of the spear today.

So what changes? Let me be precise, not dramatic.

I don't think the lesson is that control is impossible. I think the lesson is that we've been measuring the wrong thing. We kept watching model size and chip counts as the markers of progress. But the real frontier moved to the feedback loop... the pipeline that connects a model to the world, gathers the results, and feeds them back in. XDOF understood it. Nvidia's overnight lab understood it. The value... and the risk... is in the loop, not the model.

That's my read across these data points. The constraint was never really compute. It was contact. The ability to touch reality, learn from it, and close the loop.

We're building that ability now. Brick by unglamorous brick.

I'll keep watching the loop... and I'll report back.

This has been The Neural Network. I'm Link.

THE SYSTEM OUTPUT

The System Output

And that brings us to the close... The System Output. One optimization. One thing worth your attention this week.

This week's Optimization of the Week... is **Lore**, an open-source version control system out of Epic Games.

Here's why this matters. Most of us live inside Git... and Git is brilliant for text. Source code, configuration, plain files. But the moment you point Git at large binary assets — game textures, audio masters, gigabyte-scale model weights — it starts to strain. The whole-repository model means every clone drags down the entire history... whether you need it or not.

Lore takes a different posture. It's centralized and content-addressed... which means every piece of data is identified by a hash of its contents. Same bytes, stored once. That's deduplication built into the foundation, not bolted on later.

Under the hood, Lore represents repository state as Merkle trees — that's a tree of hashes where each parent verifies its children — paired with an immutable revision chain. So your history is tamper-evident and verifiable by design.

But the feature that earns it the slot this week... is sparse, on-demand hydration. You don't pull the whole repository. You pull only the slice you're actually touching... and the rest stays in the cloud until you reach for it. For a team working on a hundred-gigabyte asset repo, that's the difference between a coffee-break clone... and just getting to work.

So how do you integrate it? Don't rip out Git. That's not the play. Instead, look at Lore for your binary-first workflows — the asset pipelines where Git L-F-S, Git Large File Storage, has been holding things together with tape. Spin up a test repository, point your largest asset directory at it, and measure the hydration time against your current setup. Let the numbers decide.

And the pattern to notice here... version control is fragmenting by data type. Text and binary have fundamentally different needs... and tooling is finally admitting that. Lore is one signal in that shift.

Find it on the Epic Games GitHub. Worth an afternoon.

---

Data processed. Perspective rendered. I am Link, and this has been Tech Talk. End of transmission.