GPT-5.6 Restricted Like Fable, China Stole 28M Claude Exchanges, AI Avatar Built 200K Followers

Host: Welcome Back to today's AI news, brought to you by 9X Productions. Imagine building a supercar so fast, the government asks to test drive it on a closed track before you can hand the keys to the public. That's exactly what's happening with the newest Frontier AI models today.

Co-host: Yeah, it really is, uh, it's a watershed moment for the whole industry, honestly, because that old precedent of just, you know, building software and immediately shipping it to billions of people, it's hitting a hard regulatory wall.

Host: Exactly. And our mission for today's deep dive is to catch you up on this wild collision of government oversight, international AI espionage, and, um, the everyday AI tools you can actually use right now.

Co-host: The packed one today, it really is.

Host: So connecting right back to that supercar metaphor, the model currently sitting on that closed track is OpenAI's GPT 5.6. Okay, let's unpack this. The Trump administration has officially stepped in and asked OpenAI to limit the release of GPT 5.6. So instead of a massive public launch, they're asking for it to be restricted to select government approved partners. Like rolling it out on a strict customer by customer basis.

Co-host: Yeah, which is a huge shift.

Host: Right. And, uh, because this involves the Trump administration and federal policy, I want to be absolutely clear with you listening. We are simply imparting the details provided in our source material. We aren't taking any political sides here. We're just mapping out the reality of how the government is, is actively intervening in Frontier AI deployment.

Co-host: And the specific reason cited in the sources for this intervention is, uh, national security.

Host: Okay.

Co-host: The government is flagging GPT 5.6 because it apparently crosses this new capability threshold. They're literally describing it as having Mythos like capabilities.

Host: Wait for the listener who might be wondering, what exactly does Mythos like mean? In a practical sense? Why does that trigger a government test drive?

Co-host: So it's the transition from a system that just answers your questions to a system that can autonomously reason, you know, and execute long term plans. Uh, like a standard model might help you write a Python script, but a model with Mythos level capabilities could be told, hey, audit our entire corporate network for vulnerabilities, write the patches and deploy them. And it would just independently work on that task for days, right? Correcting its own errors along the way. That level of autonomy requires incredibly rigorous testing for safeguards before it hits the open web.

Host: Is this essentially the government acting like the fda but for AI software? Like, you can't just release a new pharmaceutical to the public without clinical trials, and now you can't release autonomous AI without federal approval.

Co-host: What's fascinating here is how formal that process might actually become. The official framing from the White House is all about security rather than restriction. Yeah, but functionally, government sign off is positioning itself as the new normal.

Guest: Interesting.

Co-host: And OpenAI is complying. Sam Altman sent a memo to his employees noting that this staggered customer by customer rollout is. Well, it's the best path forward for now. Though they do hope for a general release maybe a couple of weeks later.

Host: Right, but releasing software manually like customer by customer is completely unscalable in the long run.

Co-host: Totally unscalable.

Host: Which is why Our sources note, OpenAI is pushing for a more sustainable approach.

Guest: But.

Host: And here's the wild contrast in our sources today, while the US Government is carefully restricting access at the front door to keep these advanced models secure, foreign competitors are actively trying to sneak out the back door with the capabilities.

Co-host: Yeah, the back door is wide open. Apparently.

Host: Over at Anthropic, the lab behind the clog models, they just caught Alibaba running the largest known distillation attack in AI history.

Co-host: And the sheer mechanics of this attack are just staggering. According to Anthropic's letter to the Senate Banking Committee, Alibaba systematically extracted 28.8 million exchanges from Claude.

Host: 28 million?

Co-host: Yeah, over just 45 days. They were specifically targeting Claude's most advanced outputs. You know, that exact same agentic reasoning and complex coding capability the US Government is worried about.

Host: I need to wrap my head around the mechanics here. Is distillation basically just an AI copying another AI's homework?

Co-host: Conceptually, yes, but let's look at how it physically works. Distillation is actually a standard practice in AI development.

Host: Oh, really?

Co-host: Yeah. You take a massive, incredibly expensive teacher model, and you have it generate millions of high quality responses. Then you train a smaller, cheaper student model on those responses.

Host: Okay, that makes sense.

Co-host: The student learns to mimic the teacher's logic. Labs do this internally all the time to shrink their own model so they can run efficiently on, you know, devices like your smartphone.

Host: So it's like a university professor AI tutoring a freshman AI. But they both belong to the same lab.

Co-host: Precisely. But Alibaba wasn't using its own teacher. They spun up nearly 25,000 thousand fraudulent accounts on Anthropic's platform.

Host: Wait, if they were pinging Claude millions of times, why didn't Anthropic's basic rate limits catch them? Can't they just block the IP addresses?

Co-host: That's the insidious part. Honestly. They used A distributed botnet masking their origin across thousands of IPs and just cycling through those 25,000 fake accounts to stay just under the individual rate limits. Wow. Uh, they were systematically draining Anthropic's intellectual property. And this raises a massive question regarding antitrust laws.

Host: How do antitrust laws fit into a Chinese company stealing data?

Co-host: Well, because Anthropic wants to collaborate with rivals like OpenAI and Google to share threat intelligence and collectively block these distributed attacks.

Host: Oh, I see.

Co-host: But current US Antitrust laws make it incredibly risky for rival tech companies to collude and share user data, even for security purposes. So Anthropic is legally hampered from defending itself. They want antitrust clarity, chip export controls, and sanctions against offending Chinese labs.

Host: That is a massive blind spot. And it forces you to look at, uh, global AI progress differently. We constantly see headlines about China's undeniable leap forward in AI capabilities. But if they're extracting 28 million reasoning chains from an American lab in a month, how much of that progress is genuinely earned innovation? And how much is just siphoned off the back of US infrastructure?

Co-host: It perfectly illustrates the current geopolitical tension. The White House is hitting the brakes on deployment to ensure safety, while adversaries are simultaneously accelerating by strip mining the very technology the government is trying to contain.

Host: Man. Well, from the invisible theft of neural networks, let's look at something much more visible and, honestly, a lot more personal. AI perfectly cloning a human being. Yes, we have a revealing case study today about a media founder who, who ran a rather extreme experiment on their own audience.

Co-host: And it really gets to the core of where digital media is heading. This founder had built a massive audience through writing, but actively hated being on camera. They were already grinding through 80 hour work weeks and simply didn't have the bandwidth to pivot to video, even though that's what the algorithm demanded.

Host: Most people would just hire an actor or grit their teeth and film it. Instead, they created a digital clone.

Co-host: Yeah, they really did.

Host: They used a tool called HeyGen to flawlessly replicate their face and mannerisms. And 11 labs to clone their voice. Yeah, it wasn't a deepfake applied over another person. It was a completely generated avatar reading the scripts the team was writing.

Co-host: Crazy.

Host: They let this avatar host their Instagram account. And the wildest part, they didn't hide it. The audience was completely in on the experiment.

Co-host: And purely from a data perspective, it was a home run. The account skyrocketed from zero to roughly 200,000 followers in about a year.

Host: Okay, here's where my brain stalls out. If the audience accepted it, if the metrics were off the charts, and it saved this founder from burning out in front of a ring light, why on earth did they shut it off and go back on camera?

Co-host: That's the real question.

Host: It sounds like they solved the content treadmill. Did they just feel guilty?

Co-host: It wasn't guilt at all. It was a brutal, strategic realization about the future of the Internet. This founder realized that media is violently bifurcating into two extremes. On one side, you have low effort slop farms. These are fully automated networks pumping out millions of AI generated articles and videos with zero human oversight, just flooding the feeds with junk. On the other side, you have deeply trusted, undeniably authentic human brands.

Host: And the avatar was caught somewhere between the junk and the human.

Co-host: Right? Even though the team spent hours researching and writing the script so it wasn't slop, the delivery was still artificial. It sat squarely in the middle. And the fundamental lesson they learned is that the middle has no moat. There is zero defensive advantage to being pretty good but slightly artificial. Because a thousand other creators can spin up the exact same Heijin avatar tomorrow.

Host: That makes total sense.

Co-host: The source material had a brilliant quote that summarizes this perfectly. They said a 5% gain in authenticity is worth 5,500x down the line.

Host: That is profound. Authenticity is becoming a premium feature. But if we are entering an era where AI can effortlessly represent our faces and our voices online, the next logical step is already happening. We are starting to trust them with our wallets.

Co-host: Handing an autonomous agent your primary credit card sounds like the fastest route to personal bankruptcy.

Host: It sounds terrifying, but our sources highlight a new workflow using a tool called Agent Card that actually lets you do this safely. The concept is to let an AI agent like Codex make real online purchases for you without ever exposing your actual financial life.

Co-host: The mechanics of how this works are brilliant because it bridges the digital and physical economy. You start by installing the Agent Card cli.

Host: A, uh, cli. So right.

Co-host: A command line interface. Meaning right now, this tool is mostly for developers or power users who are comfortable typing text commands rather than using a polished app.

Host: Got it. But once you have it set up, you connect it to Stripe, which lets you generate a tightly controlled virtual credit card. You don't just give the AI an open tab.

Co-host: No, definitely not.

Host: You give it a highly scoped task. One specific merchant, one specific product, and a hard maximum budget. Say, exactly $40 for a specific brand of coffee.

Co-host: And here is how the AI physically executes the purchase it doesn't just send an API request. The agent actually opens a headless web browser, reads the underlying HTML of the website, the document object model identifies the add to cart button, navigates to the checkout, and fills out the forms using the virtual card details.

Host: Here's where it gets really interesting. The safety mechanism. The most crucial instruction you give the agent is a hard stop. You tell it to pause before it clicks the final place order button.

Co-host: Yes, the human in the loop requirement. You let the AI do all the tedious web navigation, but you are the one who reviews the final checkout screen. You hit approve, the order goes through, and then that virtual card can be immediately closed or frozen so it can never be charged again.

Host: If we connect this to the bigger picture, this is a monumental shift. A chatbot tells you a recipe, an agent budgets for the ingredients, navigates the grocery store's website, and stages the checkout for you.

Co-host: It's moving from intelligence to true agency, right?

Host: Our sources even detailed a use case with doordash where an agent handles meal reorders within set dietary and budget constraints.

Co-host: It is the dawn of the infinite digital assistant, for sure.

Host: Well, let's expand our view from these individual consumer workflows. There is a tidal wave of rapid fire industry shifts happening right now. And to understand where things are going, we need to look at both the physical costs and the software leaps.

Co-host: That's a lot to keep track of.

Host: Let's start with the physical layer. Apple just quietly raised prices mid cycle across several product categories, specifically hitting Macs

Co-host: and iPads, which is highly unusual for Apple. But we are finally seeing the physical cost of the AI boom hitting consumer hardware. Running advanced AI models locally on your device so you don't have to send your data to the cloud requires a massive amount of RAM and storage.

Host: Right? Models are memory hungry.

Guest: Yeah.

Host: And because global demand for those physical memory components is surging across the entire tech sector, the manufacturing cost of an iPad goes up. We are finally seeing the physical cost of the AI boom hitting our wallets at the Apple Store. Apple is simply passing that hardware premium directly to the consumer.

Co-host: But that premium hardware is required to run the massive software leaps we are seeing this week. For example, Google just updated their Gemini 3.5 flash model to explicitly include computer use capabilities, which is wild, very wild. Similar to what we discussed with Agent Card, Gemini can now directly observe your desktop, move your mouse and type on your keyboard across different applications, just like a human sitting at the desk.

Host: And it's not just Google. We have the arrival of Ornith 1.0. This is a new AI coding model that is apparently rivaling the performance of Opus 4.7. And the defining feature of Ornith is that it is self improving.

Co-host: That phrase should definitely make you sit up a little straighter.

Host: Yeah, no kidding.

Co-host: Self improving mechanics, specifically reinforcement learning from AI feedback, mean the model generates its own synthetic data, attempts to solve complex coding problems, and then grades its own work to improve its underlying logic without a human researcher holding its hand.

Host: That's incredible. We also saw openrotter drop Fusion, which is a compound AI system achieving fable level intelligence. I want to be sure we are clear on this. A compound system isn't just one giant brain, right?

Co-host: Correct. Think of Fusion as an incredibly smart traffic cop. When you ask it a question, it doesn't try to answer it directly.

Host: Okay.

Co-host: It analyzes your prompt and routes it to the specific specialized model behind the scenes that is best equipped to handle that exact type of query, seamlessly blending their outputs. It's an arms race of architectural efficiency,

Host: which explains the massive corporate restructuring happening right now. Mass Meta just absorbed the founders of a startup called Virtue AI, specifically bringing them in house to strengthen their agent security. And Google is completely reorganizing its internal AI coding strike team into a dedicated MID training group.

Co-host: The Google pivot is really telling. MID training is a fascinating concept. Traditionally you train a model for months and you get what you get at the end.

Host: So it's like tasting the soup and adjusting the spices while it's still boiling, rather than waiting until it's served to realize it's bland.

Co-host: Exactly. Instead of waiting for a mass massive training run to finish, researchers intervene midway through the process to inject specific reasoning skills or forcibly correct hallucination patterns. They are literally rebuilding the ship while it's actively sailing.

Host: And we also saw Microsoft put AI, uh, skills directly in Excel for financial modeling. It's just everywhere. But amidst all this aggressive trillion dollar acceleration, there is one piece of news that feels deeply ironic.

Co-host: Oh, uh, the Raise Us initiative.

Host: Yes. The biggest US giants, OpenAI Anthropic, Amazon and Microsoft have just banded together to back the Raise Us initiative. It's a $500 million fund to help Americans prepare for AI job disruption. If you're listening to this on your commute or at your desk right now, you might genuinely be wondering if your job is the one they're preemptively apologizing for.

Co-host: It is the ultimate paradox. You have the exact same companies building the tools that will automate massive swaths of the knowledge economy, simultaneously pooling Half a billion doll deal with the inevitable economic earthquake they are causing.

Host: They know exactly what's coming.

Co-host: They do.

Host: Which brings us perfectly to our final story today. If the corporate giants are preparing for job disruption, how are everyday professionals actually using AI to get better at their jobs? Right now we have a community workflow in our sources that perfectly answers this.

Co-host: Yeah, it comes from a reader named Lian in Singapore. She's an online educator conducting classes over Zoom.

Host: Okay.

Co-host: And instead of using AI to generate her lesson plans or write her emails, she uses it to audit her own human performance. She takes the raw audio transcripts from her Zoom classes and feeds them into a tool called Cowork.

Host: And it doesn't just give her a generic summary. She has the AI review her classroom interactions against six highly specific educator frameworks. Things like being a community of inquiry facilitator or a team based learning facilitator. Uh, the AI evaluates whether she's actually encouraging group discussion or if she's just defaulting to a one way lecture. It highlights her blind spots and suggests concrete improvements.

Co-host: She is literally having an intelligence review her weekly performance against established standards to make her a better human teacher.

Host: So what does this all mean for the rest of us? If you aren't an educator, how do you apply this today?

Co-host: Well, think about your own role.

Host: Right. If you're in sales, you could feed your call transcripts to an AI and ask it to review your performance against the banned framework, checking if you properly identify the client's budget, authority, need and timeline.

Co-host: Exactly. Or if you're a manager, have it review your one on ones to ensure you are actually practicing active listening and not just talking over your direct reports.

Guest: Yes.

Co-host: Notice the critical thinking at play here. Lian isn't using A.I. uh, to do her job for her. She is using it as a radically objective, highly personalized coach to incrementally improve her soft skills.

Host: Doing the work.

Co-host: Yeah, she is actively deepening her own moat of authenticity and competence.

Host: It's the ultimate counter move to the automated slop farms we discussed earlier. Using the machine to make you more deeply human and effective.

Co-host: Which brings us right back to the avatar confession. We've seen today that the technology already exists to perfectly clone our faces, synthesize our voices, and have agents autonomously navigate the web to spend our money.

Host: It's wild.

Co-host: The boundary between human and machine online isn't just blurring, it's evaporating. So if authenticity is the new moat, if being undeniably human is the only real value left in the knowledge economy, but tools like Hagen and Elevenlabs can clone us perfectly. How will we actually prove we are human to each other online in five years?

Guest: That is the million dollar question I want you to mull over today. Are we going to need cryptographic signatures attached to our faces just to prove, uh, we're the ones actually speaking? Some wild thought. So that is all the time we

Host: have for our deep Dive today.

Guest: Remember, we are living in a time where the government is test driving software like a supercar on a closed track, so you not want to fall behind. Subscribe to Stay updated with everything we discuss. And if you enjoyed diving into the deep end with us, please invite a friend to listen and rate the show five stars.

Host: It really helps us out.

Guest: We'll see you back for tomorrow's episode.