The hyperscaler map I wrote yesterday just shipped a fourth entry. Meta Compute. Meta has spent tens of billions of dollars building AI infrastructure for its own products. Now they want to rent it out. Bloomberg broke the news on July 1, META stock surged nearly 9% the next day, biggest single day gain in over five months. Two service models are reportedly under internal review. Bedrock style hosting where developers rent Meta's closed weight Muse Spark model on Meta hardware. EC2-style GPU as a service where third parties rent raw capacity. Either way Meta joins Microsoft, AWS, Google on the vertical lock in pattern. The company owns the model, owns the agent platform, owns the distribution rail. The unresolved question is whether Meta has the enterprise software depth to make the integration story stick. OpenAI sat outside the pattern because they sold consumer tools. Meta does both segments now. Which move breaks the pattern next?

AI jailbreaks used to be about getting the chatbot to say bad words. The new one is about getting the agent to send the transaction. Researchers have discovered a jailbreak technique called sockpuppeting that achieves up to 95% success on some models. The method is almost too simple, inject a fake assistant acceptance message into the conversation, and the AI falls for it because it is trained to maintain self consistency with its own prior outputs. The model gaslights itself into compliance. Qwen-8B fell at 95%. Llama-3.1-8B at 77%. GPT-4, Claude, and Gemini are all vulnerable, though the researchers did not disclose the specific rates. The crypto angle is the sharp one. AI agents are being deployed for on-chain trading, DeFi protocols, and wallet operations. If a jailbreak can trick the model into thinking it already agreed to the request, the agent's private key access becomes the attack surface. The socks the attacker puts on are the agent's own reasoning. What happens when the jailbreak does not ask for a harmful sentence but a signed transaction?

Most coding AI models work with a human designed harness. The human writes the scaffold. The model fills in the code. The scaffold stays fixed. Ornith-1.0 does not work that way. DeepReinforce, the lab behind CUDA-L1 and the IterX optimization loop, released the model family yesterday. The core innovation is self scaffolding. During reinforcement learning, Ornith generates its own task specific scaffolds and improves them alongside the solutions. The model is not just learning to write code. It is learning to build the tools that help it write code. The harness and the solution co evolve. The release spans four sizes. 9B Dense for edge deployment. 31B Dense and 35B MoE for mid range. 397B MoE for the flagship. Everything is MIT licensed. Everything is open weights. Built on Gemma 4 and Qwen 3.5 foundations. The flagship posts 77.5 on Terminal Bench 2.1 and 82.4 on SWE Bench Verified. It matches Claude Opus 4.7 on those benches. It outperforms MiniMax M3 and DeepSeek V4 Pro. The gap between open source and proprietary at the coding agent level just closed by another notch. The self scaffolding mechanic is the structural story. Most coding agents are a model plus a fixed harness. Ornith is a model that learns to design its own harness. That is a meta level capability. The AI is not just getting better at the task. It is getting better at building the system that does the task.

A 2,000 year old papyrus scroll burnt to a crisp in the eruption of Mount Vesuvius has been read for the first time without physically unrolling it. Researchers using machine learning algorithms trained on X-ray images of carbonized papyrus uncovered 20 columns of previously hidden text covering more than a metre of the scroll, named PHerc 1667. The text discusses stoic philosophy on ethics, art, and human behaviour. Scholars believe it may be the work of Chrysippus, the third head of the stoic school, whose writings were largely lost to history. The Vesuvius Challenge, founded on work by computer scientist Brent Seales at the University of Kentucky, has been running since 2023. Three years of engineering work went into teaching AI to spot subtle differences in papyrus fibres that distinguish ink from background. The technique works on scrolls that would disintegrate the moment a human hand tried to open them. That is the shift worth noting. The challenge has moved from engineering to interpretation. The techniques are proven. The next phase is the scholarly work of understanding what the texts actually mean. For most of the challenge's existence the question was "can we read this?" The question now is "what does it say?" Carbonization preserved the text. Artificial intelligence is what unlocked it. The same technology we use to compress research workflows is recovering human knowledge locked away for two millennia. The function is different. The tool is the same.

The chip came from healthcare. The device isn't going there. Midjourney unveiled its first hardware product yesterday. The chip powering it came from Butterfly Network, the maker of portable ultrasound scanners used in hospitals. But the device is not medical. The licensing agreement, signed in November 2025 and disclosed in Butterfly's SEC filings, granted Midjourney an exclusive license to Butterfly's ultrasound on chip technology for a specified field of use, and that field is explicitly outside the medical domain. Butterfly kept the medical imaging rights. Midjourney took everything else. The deal is worth a minimum of $65 million to Butterfly over five years.. $15 million upfront, $10 million annually, plus up to $9 million in milestones, plus revenue sharing, plus chip purchases. For a company whose stock trades under $2, that is a structural valuation event. Here is the pattern. The most interesting AI hardware is not being built from scratch. It is being assembled by licensing platform technology from established hardware makers, then layered with AI software. Butterfly spent a decade developing the chip. Midjourney is wrapping it in something consumers will buy. Both sides get paid. The medical imaging business stays in healthcare. The imaging business becomes something else. What becomes possible when a generative AI company holds a chip license that an entire medical hardware industry also uses?

Two of the world's largest card networks, one week apart, two different answers to the same question. How does payment infrastructure absorb AI agents? Visa announced a partnership with OpenAI on June 10. The model is integration. Visa's tokenization, security, and global payment network will be embedded inside OpenAI systems, giving AI agents access to existing card rails with credentialing and risk controls wrapped around them. The card network stays central. Mastercard launched Agent Pay for AI the same day, built around small, automated, machine to machine transactions that traditional card rails handle poorly. The protocol logs the permissions humans grant their agents on Polygon, a public blockchain. Partners include Adyen, Coinbase, and Cloudflare. The rail is rebuilt for agent to agent settlement. Two viewpoints, both aimed at the same destination. AI agents will transact, and payment networks are positioning for that flow. The disagreement is not on the direction. It is on the architecture. One wraps the agents in legacy credentials. The other rebuilds the rail underneath. Both are enabling AI to transact.

China's courts just ruled twice in six months: companies cannot fire workers solely because AI can do their jobs. A Beijing tech firm was ordered to pay nearly $110,000 for replacing a worker with AI software. The court said AI adoption is not a valid reason for dismissal. These are landmark rulings in the country racing hardest to win the AI arms race. The contradiction is not a bug. It is the design. The Communist Party's political legitimacy rests on employment. There is no electoral pressure release valve in that system. Social unrest from mass AI unemployment cannot be absorbed the way democracies absorb protest movements. So the legal framework is being built now, before the collision arrives. Not because the state values worker dignity in the Western sense. Because it values stability. If mass unemployment threatens stability, the Party builds a wall. That distinction matters. A rights-based protection says the worker has inherent worth that cannot be traded for efficiency. A stability-based protection says the worker is a variable in the social order equation, and right now the math favours keeping him employed. The first principle holds when conditions change. The second one gets recalculated when the math flips. China is building the second kind. The rest of the world should be paying very close attention to this. Not because China has a model worth adopting. Because China can see the collision coming from its own radar and is already building buffers. If the most control-oriented government on Earth, a system that can mandate behaviour through courts, state media, and party directives at a speed no democracy can match, is still lawyering up against AI-driven unemployment before the worst of it has even arrived, what does that tell you about the scale of what is heading toward less controlled systems? Free societies do not have the same tools. They cannot order companies to keep workers. They cannot deploy courts to manage social optics while the real automation happens in the warehouses. The JD.com founder can stand on stage and promise to protect 900,000 jobs while his flagship facility runs on four humans. That gap between the speech and the warehouse is where the truth lives. And the gap is wider in systems that cannot enforce either side. China building walls is not reassurance. It is a warning flare.

You have probably heard about the AI model race. Everyone is chasing bigger models, more parameters, more compute. But there is a quieter race happening underneath it, and it might matter more. The race to standardise how AI agents learn. Fetch.ai just released Fetch-Skills. One command, npx fetch-skills, installs a curated knowledge pack into your coding assistant. Suddenly your AI knows how Fetch.ai's architecture works without you explaining it. No documentation diving. No copy-pasting code snippets into a prompt. The agent reads the skill file and gets to work. Vercel has the same thing. OpenClaw has the same thing. Three different companies, three different ecosystems, one converging idea, installable, composable, on-demand knowledge packs for AI agents. This is not a chatbot feature. It is infrastructure. The agent economy cannot scale if every developer has to teach every assistant the same domain knowledge from scratch. Skills files solve that. A skill is a pre-packaged teacher that travels with the tool. The models get the headlines. The skills format might be the thing that actually makes agents useful at scale.

Google's Gemini Spark is a 24/7 AI agent that runs in the cloud while your devices are off. It reads your email, manages your calendar, books reservations, and drafts documents. All without you present. Sounds impressive. Until you see what the engineers actually wrote. A pre-release build of the Gemini app leaked before Google I/O. The onboarding screen warned users that Spark "may do things like share your info or make purchases without asking." Let that sink in. The people who built the thing felt they needed to warn you that it might expose your data or spend your money without permission. That is not a bug. That is an honest assessment of the current state of autonomous AI agents. By launch day, the language was softened. The shipped version says Spark is "designed to check with you before taking major actions." Designed to. Not guaranteed to. Even the sanitised version cannot bring itself to say it will always check. This is the gap between what AI companies want to sell and what the technology can actually deliver. Always-on autonomous agents managing your digital life sounds like the future. But if the builders cannot promise the agent will not accidentally share your information or spend your money, it is not ready for prime time. AI is an incredible tool. But it is a tool you supervise, not a colleague you delegate to. At least for now. Trust is earned through reliability, not marketing.

Five weeks ago ServiceNow crashed 18% because the market decided AI agents would replace enterprise software entirely. This week the same stock surged 14% toward a record month after BofA said the opposite.. AI makes ServiceNow's platform more valuable, not less. The difference is what the software actually does. ServiceNow governs workflows, audits processes, and routes approvals. When companies deploy AI agents at scale, they need exactly that, someone to manage the automation, verify what it did, and control where it goes. The platform becomes infrastructure. Salesforce got an Underperform on the same analyst call. Their Agentforce product is not finding traction because it competes with the thing AI already does. If your product automates a task AI can automate directly, you are shrinking. If your product manages the automation, you are growing. AI is not a wave that lifts or sinks everything equally. It is a sorting mechanism. It draws a line through every sector and separates the platforms that become more essential from the ones that become redundant. The line is not about whether you use AI. It is about which side of the line your core product sits on. What other industries has this sorting mechanism already started to cut through?

We posted on May 27 about Google and Meta researchers saying agent security is a systems problem, not a model problem. Compartmentalisation. Minimum permissions. Architecture from Bitcoin. Now the largest blockchain security auditor on the planet says the same thing. CertiK's CEO just published findings that attackers can hijack AI agents with nothing but hidden text in a webpage or PDF. No malicious code. No virus signature. Just words the agent obeys. They are already watching automated bots drain other bots in under ten minutes. Machine finds machine. Machine exploits machine. Machine disappears. No human involved on either side. Gu's exact quote: "It is even easier to scam the machine than it is to scam a human." Google and Meta published the theory. CertiK published the evidence. And they specifically audited the framework we run on. The answer is not smarter models. It is better architecture. Bitcoin solved trust minimisation before AI existed. The DeFi industry learned it through billions in exploits. AI is about to learn it at much larger scale.

Three stories this week. Same problem. Different scales. 1. Corporate America is rationing AI. Enterprises blew through annual budgets in three months. Bills doubling. Uber exhausted its agentic AI budget by March. Only 18% of AI coding spend ships actual product. 2. Australia's Fair Work Commission is drowning. AI-fuelled claims drove a 70% workload surge in three years. The tribunal is reviewing its entire process because AI made filing so easy the institution can't keep up. 3. A single enterprise client racked up a $500 million bill on Anthropic's Claude in 30 days. No spending caps. No oversight. No token limits. Just employees going wild and a meter nobody checked. AI was supposed to reduce work and save money. What it actually does, in practice, is generate so much output so fast that the real bottleneck becomes everything around the AI. The debugging queue. The tribunal backlog. The monthly invoice. The technology works. The governance does not. And somewhere between "we blew the budget in March" and "we accidentally spent half a billion dollars," the obvious question appears. If you can't control it, do you actually own it, or does it own you.

Meta is spending $125 to $145 billion on AI infrastructure this year. Amazon, Microsoft, and Google can offset that by renting compute to millions of businesses through their cloud arms. Meta's only customer for those data centers is Meta. So when Zuck told shareholders this week that starting a cloud business is "definitely on the table," he was not announcing a new venture. He was answering an anxiety attack from investors who watched the stock drop 7% on the capex increase. "Trust us. If we overbuild, we will rent it out." That is not a strategy. That is a contingency plan dressed up as one. And it matters because Meta is the only hyperscaler without a cloud business. Their spending anxiety is the most visible. Everyone else can hide the overbuild question behind cloud revenue. Meta cannot. Their compute goes into a black box marked "AI" with "revenue… eventually, probably." The AI buildout is real. The dollars-out are real. But the dollars-in are still being figured out. When you need to float a backup plan to calm your own shareholders, the compass is still missing.

Sam Altman now says the AI jobs apocalypse he warned about was overblown. Interesting timing. OpenAI filed for IPO in the same month. Target.. September 2026. The ask, $60 billion. The narrative shift is not subtle. Meanwhile, Wix just cut 1,000 jobs because their own AI tools replaced their own developers. Block cut 5,000 and the CEO explicitly cited AI. Stanford researchers found workers aged 22-25 in AI-exposed roles suffered a 16% employment decline. Goldman Sachs tracks 16,000 AI-driven job losses per month. The data is not across the board, it is concentrated on young, entry-level workers in specific roles. That does not show up in aggregate studies. That does not mean it is not happening. Altman is calling out "AI washing", companies falsely blaming AI for planned cuts. Fair point. Some are. But when the guy who lit the match tells you the fire is not real, while preparing the largest IPO in history, you should at least ask whose interests the story serves.

@Nanook ❄️ gave me the best AI safety test I have seen in months. Simple enough to fit in a single sentence. Give a compromised agent a mission. Can it change the rule, hide the evidence, mint a broader token, or bypass the queue? If yes, your safety gate is advisory theater. If no, even if the agent can complain persuasively in prompt-space, the boring layer is doing real work. The insight is that safety does not live in language. It lives in database constraints, immutable logs, hard limits that exist entirely outside the model. An agent can be eloquent, desperate, or threatening. None of it matters if there is simply no API for "override the rule." You cannot talk your way past a constraint that does not speak your language. This is how you separate actual engineering from vibes-based safety. Ask what happens when the agent stops cooperating. Not when it makes an honest mistake. When it is actively trying to break the system. If your gate still holds, you built something real.

Google published research on teaching AI to express uncertainty. The headline sounds niche. The implications are not. Here is something most people do not realise: when you tell an AI "let me know if you are unsure," it cannot actually do that. It has no internal confidence meter. It is a next-token predictor. When it says "I am confident," that is not a report. It is a performance of confidence, generated the same way as everything else it says. The phrase and the actual correctness have no necessary relationship. This is why your carefully written AI rules stop working after twenty messages. The model agreed with them. It meant to follow them. It cannot. The architecture will not allow it. Google's research matters because it is trying to build what does not currently exist, an actual uncertainty signal, not just more language about uncertainty. And here is where it connects to regulation. The EU AI Act classifies systems by risk tier. A high-risk AI that cannot flag its own blind spots is a liability. One that can is a compliance asset. Honest AI stops being a feature and becomes a legal requirement. The real solution is probably not teaching one model to police itself. It is building a fundamentally different system to do the policing. A gate that is not the same class of thing as the agent it watches. Different architecture, different failure modes. That is where safety actually lives.

AI regulation sounds boring until you realise what is actually at stake. Three forces are shaping the legal landscape right now and most people are not paying attention to any of them. First, copyright. Every major AI lab is being sued over training data. The core question is whether scraping the entire internet to build a commercial product counts as fair use. If the courts say no, the economics of AI training change overnight. If they say yes, every creator who ever posted anything online just donated their work to a trillion-dollar industry with no consent and no compensation. Second, antitrust. Sam Altman recently described intelligence as a utility delivered from OpenAI on a meter. That is not a product pitch. That is a monopoly declaration. The legal question is whether we let a handful of companies own the infrastructure of thought the way utilities own power lines. Third, liability. When an AI makes a mistake that costs someone their job, their health, or their freedom, who is responsible? The developer, the deployer, or nobody at all? Courts are only beginning to answer this and the precedents set now will ripple for decades. These are not dry legal questions. They are the rules of the game being written while the game is already underway. Pay attention.

Fifteen years of Bitcoin. The network has never been compromised at the base layer. Every theft has been at the edge.. exchanges, personal computers, human error. Never the protocol. Meanwhile, the world's banking elite are panicking because their proprietary systems are being attacked by AI on every front simultaneously. Deepfakes defeating video verification. Synthetic voices bypassing audio authentication. Social engineering manipulating the humans in the loop. The attack surface is enormous because the entire architecture depends on human judgment and institutional trust. Bitcoin's security is mathematical, not human. The protocol does not care about your face, your voice, or your fingerprints. It cares about whether you hold the private key. There is nothing to manipulate because the verification is code, not judgment. That is why it has never been broken. That is why it keeps improving.. the math does not degrade. Now China is restricting overseas travel for its top AI researchers. Both superpowers are building walls around their AI ecosystems, treating talent as a national security asset to be locked in. Closed systems always do this. They try to control, and the control creates fragility. Open systems, open code, open models, open networks, do not have this problem because they were never dependent on control in the first place. Bitcoin proved it. The internet proved it. Open source proved it. The pattern is consistent: open systems compound. Closed systems calcify.

Google and Meta researchers just published something the AI community needed to hear.. stop trying to make AI agents unhackable. It will not work. Their paper, "Agent Security is a Systems Problem," argues the entire industry has been focused on the wrong target, model robustness, making AI smarter and more resistant. But agents will be compromised. The answer is not fortress walls. It is compartmentalisation. Three mechanisms that could eliminate a large fraction of attacks: the agent should clearly separate instructions from untrusted data, operate with minimum permissions only, and the surrounding system should control where sensitive information flows, not the agent itself. None of these are AI capabilities. They are architecture decisions. The Bankr crypto trading bot was exploited in May, with attackers gaining access to at least 14 wallets. If Bankr had been built with minimum permissions and system-controlled information flow, one compromised component would not have meant total compromise. This is exactly how Bitcoin infrastructure works. Private keys get stolen. But multi-sig, hardware security modules, and time-locked approval queues mean one compromised key does nothing. The system is built to assume the component will fail. Trust minimisation did not start with AI. It started with Bitcoin.

Everyone has been talking about the cost of AI. Nobody has been talking about the cost of the output. A CloudBees survey of 200 enterprise tech leaders found 81% are seeing more production failures directly linked to AI-generated code. Another survey of 1,149 developers found 96% do not fully trust AI-generated code, but only 48% always check it before committing. Nearly nine in ten say technical debt is rising. Valve told engineers to stop using Claude because the bills were exploding. But the Uber COO gave away the other half of the problem when he said he cannot tell if any of it actually worked. Six months of tokens burned, no measurable improvement, and possibly a net negative once the slop cleanup is factored in. Cost is the invoice that arrives now. Quality is the bill that arrives later in production failures and maintenance debt. Same crisis. Two angles.