Just two years ago, human beings wrote software code manually. We typed every single logical instruction by hand. Yeah, line by tedious line. Right. Today, we are building autonomous AI managers instead. Those managers hire other AI workers to do it. Welcome to our latest deep dive. I am truly glad you are here with us. Okay, let's unpack this. Today, we explore a profound technological shift. We are moving from human prompts to continuous agentic loops. It is a massive leap forward.
It really is. We will trace the staggering hardware costs of this autonomy. We will see how it enters Hollywood and the beauty industry. Finally, we examine a plan to shield the Internet's foundation. We are crossing into completely uncharted territory today. The fundamental architecture of digital creation is transforming. It is happening much faster than anyone originally predicted. We should start with how work actually gets done now. AI is operating autonomously in continuous, invisible
loops. Beat? At Meta's At Scale conference, we heard a fascinating announcement. Right, from Boris Cherny. He created Cloud Code. Exactly. He officially declared the era of agents prompting other agents. This changes the baseline of software development entirely. The evolution of software engineering is actually quite striking. First, we relied on humans writing handwritten code. Which, you know, took thousands of hours. Then we transitioned to asking AI agents to write
code. That required complex prompt engineering and constant human oversight. Prompt engineering is rapidly becoming an obsolete skill. Yeah. It really is. We used to spend hours perfectly phrasing our technical requests. Now the AI agent optimizes its own internal prompt structure. Human engineers just set up managers to handle tasks. Those managers then prompt other sub -agents to complete specific jobs. These autonomous processes
are formally known as agentic loops. They rely entirely on complex, non -deterministic logic. Let me define that specific technical term for you quickly. An AI making its own choices instead of following fixed rules. Right. It doesn't follow rigid tracks like a train anymore. It is more like a self -driving car encountering a detour. It calculates a brand new route dynamically in real time. It evaluates the physical obstacles.
Right. And adjusts its path continuously. Czerny noted this is already happening professionally. He has agents constantly running in his system's background. Just constantly working. Exactly. They scan his entire code base for outdated legacy code. They unify duplicated abstractions and improve code architecture endlessly. They do this tedious maintenance work without any human prompting. I still wrestle with prompt drift myself. My AI forgets instructions during a long
chat session. Goal. It loses the original context after 20 or 30 messages. How exactly do they keep these advanced models on track? Well, developers now use clever engineering tricks like the Ralph Loop. It is a brilliant mechanism for maintaining digital focus. How does it work? It forces the AI to output its reasoning aloud first. By reading its own logic step by step, the model self -corrects. It feeds its previous output back as the next input. So it prevents the autonomous model from
hallucinating a wrong turn. It constantly grounds the AI in its own verified logic. Exactly. That is crucial for what engineers call hill -climbing problems. Meaning an AI making incremental software improvements endlessly. Right. It optimizes complex architecture as long as your budget allows. Anthropic recently showed a truly powerful demo of this concept. Their engineers used AI to build software entirely from scratch. Not just asking it to
build a whole app lazily. Exactly. They guide the agents through the complex development process iteratively. There is an emerging tool called SkyBridge doing exactly this. It handles the complex development loop entirely on its own. Meaning it spins up a virtual network and tests the written code? Yeah, it establishes secure testing tunnels to evaluate external webhooks safely. It manages view rendering and client
compatibility automatically behind scenes. So human engineers just focus on designing the core product features. It's like stacking Lego blocks of data. That is a great analogy. But how does a non -deterministic loop actually know when to stop? If it runs in the background endlessly, what kills it? It seems like it would just optimize code forever. Well, developers program a separate independent subagent to monitor the process. This subagent compares mathematical confidence
scores against a predefined threshold. Ah, I see. Yeah, it constantly evaluates the current output against the primary goal. When the score passes the success metric, it cuts power. It acts as an objective independent auditor for the working agent. So an AI manager simply decides when the job is finally done. Yeah, it becomes a completely self -regulating digital ecosystem. The human becomes a supervisor rather than a
direct creator. Infinite loops are undeniably amazing in theory, but they require infinite guardrails to remain perfectly safe. Two secs silence. They also require practically unlimited computing power to function properly. This brings us to the immense financial cost of autonomy. These continuous loops burn through processing tokens insanely fast. Every single loop cycle costs real money to execute. There is practically
no theoretical compute ceiling here at all. This raises an urgent question about evaluation and software testing. Before you run infinite loops, you need strong safety nets. You cannot afford to let an agent hallucinate endlessly. There is a new tool called AgentX for this. It evaluates these agents and creates comprehensive test suites. Before they fail in production, right? Exactly. It runs intense evaluations before agents go live. It simulates thousands of random user interactions
simultaneously. This helps pinpoint catastrophic edge case failures. incredibly early. But better software testing does not solve the hardware bottleneck. If loops just burn tokens indefinitely, what happens next? Aren't we just building a machine designed to bankrupt developers? It sounds like a severe financial drain for smaller teams. Whoa, imagine scaling to a billion queries. The hardware compute costs become absolutely astronomical very quickly. This is precisely why Grok recently
raised $650 million. They were challenging the current graphical processing paradigms directly now. NVIDIA also completed a $20 billion deal recently. Right. A massive not -acquire -hire deal. They absorbed key engineering talent without triggering antitrust monopoly laws. That deal also absorbed highly specific chip architecture intellectual property. Traditional AI models
rely heavily on graphics processing units. Those chips process thousands of math problems simultaneously in parallel, but they suffer from severe memory bandwidth limitations during generation. Why is Grok's alternative hardware so crucial for infinite loops? They engineered what they call language processing units instead. These specific chips process sequential data without traditional memory bottlenecks. So they eliminate the memory bandwidth limitations that plague traditional
graphics cards. Exactly. They skip the heavy graphics processing overhead entirely during operation. This makes running text -based AI models significantly faster and cheaper. Grok is spreading this specific architecture across 13 data centers. They clearly see an enormous, unending demand for continuous compute. Will these severe hardware bottlenecks eventually kill the infinite loop dream? Are we hitting the physical limits of our digital ambitions? It certainly will, unless alternative clouds
challenge the chip monopolies. Companies like Grok must successfully provide affordable, scalable computing power. They need to drive the cost of inference down dramatically. Otherwise, only tech giants can afford to run these models. Basically, endless AI needs endless chips or the whole thing stalls. Physical reality firmly limits our theoretical digital potential. To fuel these highly capable models, they need enormous memory. They also require constant ongoing observation of complex
human behavior. beat. This collision directly impacts our foundational expectations of digital privacy. Google's Gemini 3 .5 Pro nears a June launch. It features deeply advanced reasoning capabilities for complex tasks. It also features a staggering 2 million token context window. Let me define that specific context window for you quickly. The amount of text the AI can remember in one conversation. A 2 million token window
is practically an entire personal archive. Yeah, and pro and ultra tier users get this advanced access first. To capture proper human context, new software tracks everything constantly. A tool called ReadyWhen monitors your daily workplace decisions continuously. It indexes your daily Slack messages, your email, and meetings. Right. It builds a comprehensive vector database of your professional life. It drafts your necessary next steps automatically. based on behavior.
Here's where it gets really interesting. Meta paused an internal AI training program recently. Employee personal data was exposed during the initial training process. The internal program tracked daily keystrokes and mouse movements continuously. They wanted to train models on how employees actually work. They wanted the AI to learn our natural human hesitations. Anthropic updated Claude's official privacy policy in a similar vein. They now strictly require identity
checks for specific flagged users. This verification process includes scanning passports and official driver's licenses. It even requires uploading selfies and complex face geometry data. They map the physical topology of your face for security. If we connect this to the bigger picture, it feels invasive. Models clearly need deep, intimate access to be truly helpful. But that deep access creates enormous, terrifying new privacy vulnerabilities. There is a playful new tool online called In
the Weights. It lets you check if large models learned your actual name. You can literally see where you rank in their training data. How does that tool actually work mathematically behind the scenes? Well, it queries the model's latent space for specific token associations. It measures how strongly your name... connects to specific data vectors. If your name appears frequently, the mathematical connection is stronger. Are we trading our digital anonymity just for better
autocomplete? Why do they need something as intimate as face geometry? As AI capabilities grow dangerously powerful, accountability becomes a break. Companies cannot allow anonymous users to run autonomous infinite loops. Because a malicious infinite loop could launch devastating cyber attacks continuously. Exactly. They need to know exactly who deployed the destructive code. Verified human accountability is the ultimate emergency stop button. Bigger memory means bigger risks. Forcing companies
to demand your actual face. Right. And the glorious anonymity of the early Internet disappears. All this highly personalized data isn't staying confined to dashboards. Beat. It is actively moving into our creative and physical realities. It is rapidly moving from the back office to the movie set. That is great analogy. Google is currently investing $75 million into A24. That is the brilliant studio behind everything everywhere all at once. Right. They are famous for their incredibly specific,
surreal visual aesthetics. Google wants to build powerful new AI filmmaking tools together. Because algorithms are excellent in generation, but they lack actual taste. Yeah, they tend to create sterile, plastic -looking visual content by default. So Google is using direct feedback from real Hollywood artists. They want to ensure the tools serve the true creative process. The beauty industry has historically relied entirely on physical products. Now they are transforming into massive
digital technology companies, essentially. L 'Oreal recently partnered with OpenAI for a highly visual project. They revealed an incredible Maybelline virtual makeup. try -on tool recently. Which they showcased at VivaTech 2026, right? Yeah. You will soon use this tool directly inside the chat GPT interface. They understand that digital appearance is becoming increasingly important today. The Maybelline tool analyzes the unique
geometry of your face perfectly. It applies digital makeup that reacts perfectly to virtual lighting. You can test makeup virtually before buying the physical product. There's also a marketing tool called Align 2 .0. It captures exact micro -brand details for modern targeted marketing campaign. It learns the specific typography and hex codes of a brand. It generates perfectly on -brand social assets and digital advertisements easily. It is kind of incredibly efficient for large
creative teams. Does Google paying A24 mean AI officially replaces human artists? Or does it prove algorithms desperately need our aesthetic judgment? It seems like they're admitting their models lack artistic soul. Developers finally realize algorithms lack true human aesthetic taste entirely. They're essentially buying access to artists to train aesthetics. They need human creativity to guide the enormous computing power. Right. Without human taste, the generated output
remains completely sterile. They're paying artists to help build the exact tools they'll use, sponsor. If we are going to rely on AI for our movies and makeup, we desperately need to secure the open source code running underneath. Two sec silence. Open source software projects are the fundamental bedrock of commercial tech. They run our servers, our phones, and our global banking networks. But they are mostly run by unpaid,
deeply passionate volunteers. This widespread decentralization creates huge, terrifying security gaps worldwide. OpenAI just announced a major new program called Patch the Planet. Which is a geeky nod to the 1995 movie hackers. OpenAI officially teamed up with a cybersecurity firm called Trail of Bits. They are actively defending major open source projects using codec security. This is a fascinating approach to a deeply decentralized problem. The AI scans vast open source repositories
for deeply hidden vulnerabilities. It understands the underlying logic of the complex software architecture. It flags potential security flaws that human reviewers easily miss. Do you remember the massive log4j security crisis a few years ago? A tiny unnoticed piece of open source code broke half the internet. So what does this all mean? That is exactly the kind of nightmare scenario Trail of Bits targets. Human reviewers cannot possibly audit millions of lines of volunteer
code. The AI acts as an untiring, incredibly thorough security guard. But the human engineers at Trail of Bits carefully review the AI findings. Right. They handle all the difficult heavy lifting for the community. They translate the AI's complex math into human readable security patches. They actively work with volunteer projects to deploy these essential fixes. important software tests and build reusable workflows securely. And they do all this before project maintainers ever see
the bug. It is an incredible technological irony when you really think about it. We deploy autonomous AI systems to shield our heavily decentralized internet. We use it to fix dangerous vulnerabilities we created ourselves. It is a wild cycle for sure. Does relying on open AI to find bugs create a single point of failure? Is it bad for internet security to rely on one corporate entity? Human engineers at Trail of Bits act as a vital buffer. They carefully review abstract syntax trees before
pushing patches live. They run the suggested fixes in isolated sandbox environments first. This ensures the AI model does not accidentally introduce new backdoors. We cannot trust a machine to rewrite our security protocols blindly. Exactly. The machine acts as an incredibly powerful digital magnifying glass. But the human engineer must remain the ultimate decision maker. Humans still verify the AI's math before the fix goes live. Right. And that expert human oversight is absolutely
critical currently. We have journeyed through an incredible, profound technological shift today. We moved from manually writing simple code to swarms of agents. Now, autonomous digital managers build our complex software and our art. But this incredible autonomy requires absolutely unimaginable computing power to sustain. It demands a profound, unprecedented surrender of our personal biometric privacy. And it requires a completely new security
paradigm for protecting the Internet. We are trading fundamental control for unprecedented convenience on an enormous scale. Thank you so much for joining this deep dive today. We always appreciate you spending your valuable time with us. I want to leave you with one final lingering thought. Two years ago, humans literally wrote our basic software manually. Now imagine a powerful AI loop running endlessly in the background. It subtly rewrites its own complex architecture,
continuously and silently. Human developers just sit back and lazily review the final patch. Beat. At what point does the software officially stop belonging to us and start belonging entirely to the continuously evolving autonomous swarm?