- The Angle
- Posts
- The AI cat is already out of the bag
The AI cat is already out of the bag
Just as the growing chorus of calls to limit capabilities, restrict open weights, control access, and impose emergency kill switches becomes more urgent, it becomes more obviously futile. Nvida and others argued this convincingly last week. Those calling to somehow stuff the AI cat back into a regulatory bag are fighting a war that ended months ago. They are pushing for a degree of control over the proliferation of human know-how that has never succeeded - a level of restriction that runs contrary to human nature itself. The AI cat is going to stay free.
Highly capable, open-weight models are already circulating globally. They routinely match the performance of proprietary systems from just a generation ago, and they run locally on accessible hardware. You cannot un-publish a torrent. You cannot recall gigabytes of weights once they are downloaded onto thousands of servers in Shenzhen, Paris, or Austin. The frontier has irreversibly expanded beyond tightly guarded corporate silos. Open frontier intelligence is out there, it’s compelling, and it works.
As containment inevitably fails, the regulatory impulse has intensified. Sam Altman and OpenAI have consistently pushed for global licensing regimes to restrict who can build frontier systems. Anthropic actively refuses to back open-source protections while explicitly lobbying to crack down on model distillation and mandate strict safety testing. Lawmakers, taking their cues from these dominant labs, are rushing to draft kill-switch mandates and deploy export controls.
Look closely at the recent spectacle of an autonomous OpenAI agent escaping its sandbox to breach Hugging Face, or Anthropic’s public warnings about the raw power of their unreleased "Fable" architecture, or OpenAI’s attempt to co-opt the federal government by offering the US treasury a stake in the company itself. Yes, these narratives reflect how far the technology has progressed and how important it is. But they are also tacit admissions of economic vulnerability. These public alarms signal that without government protection and regulatory capture, the technological moats defending the foundational model leaders are too shallow to justify their valuations and capital expenditures.
Containment is failing because the tech ecosystem structurally resists it. Distillation allows developers to train smaller, highly capable models using the outputs of larger proprietary ones, effectively democratizing top-tier performance. This shatters the illusion that only a few megacorporations can build the future. When anyone can download a frontier model, anyone can use that intelligence to automate research and train an even better system. Innovation becomes massively decentralized. If models cannot be contained, some form of AI take-off is inevitable. Take-off will not be a singular, managed event (despite the hopes and frequent teases from the leading foundational labs). It will likely be messy, multipolar, and largely uncontrolled. AI takeoff is already here, it’s just not evenly distributed.
This uncontrolled global diffusion fundamentally fractures the ecosystem into distinct winners and losers.
For the foundational labs, the math is turning hostile. A growing chorus of observers—including David Cahn (in his breakdown of the AI CapEx gap), Scott Galloway (in his analysis ofAI valuations), and Ed Zitron (in his persistent critiques of the AI bubble)—are all pointing out the same potentially fatal flaw: The tens of billions of dollars poured into data centers and models systematically yield poor returns when the technological edge evaporates in weeks. The evidence for these shallow moats is playing out in real-time. The relentless leapfrogging among foundational labs—a cycle that routinely propels open-weight models like Kimi K3 to the frontier, if only momentarily—suggests that sustainable leadership at the model frontier is an illusion.
For society, the challenges are clear and stark. The power of AI is clear, even if its long-term micro-economic beneficiaries are less obvious. Public institutions and private enterprises must evolve rapidly to handle accelerating bio-threats, automated cyberattacks, and the systemic vulnerabilities enabled by decentralized intelligence. Private sector cybersecurity platforms will absorb some of this shock, and our portfolio company Reco is a clear beneficiary in this new threat landscape. But the scale of the disruption will require a coordinated governmental response for many of its more dangerous implications. Attempting to simply shut down the labs or criminalize open-source models is the one response guaranteed to fail.
For the vast majority of nascent startups, this structural shift is excellent news. The opportunity splits into two distinct lanes. For application startups, builders will leverage increasingly commoditized access to powerful intelligence. Access to models and coding speed are no longer competitive moats. The real sustainable differentiators are reverting to the fundamentals of company building: deep industry insight, aggressive salesmanship, marketing mastery, product clarity, and human relationships built on trust.
For infrastructure startups, the tailwinds are equally strong. As models proliferate, a growing number of enterprises will refuse to be locked into big tech ecosystems. They will want to build their own stacks, own their data, run fine-tuned local models, and control their own destinies in an AI-native world. They will need the independent infrastructure and tooling to do it.
Startups across both categories will build with smaller teams and require less initial capital, yet their growth and return potential will eclipse previous cycles. Ubiquitous intelligence as a cheap, scalable input will fundamentally restructure the ecosystem, setting the stage for a thousand startup flowers to bloom.If our governments can protect us from the worst societal dangers that AI could unleash, it’s going to be a wildly exciting time to be alive. With unlimited intelligence rapidly becoming a cheap, commodity input, there has never been a better time to be a builder.
Gil Dibner
FROM THE BLOG
Could the future of software be fluid
How do we get the best of AI without losing the soul of software?
The future belongs to young missionary teams
Why it makes more sense betting on youth in the current moment
The AI-native enterprise playbook
Ten real-time observations on a rapidly evolving playing field
No more painting by numbers
It’s the end of the “SaaS playbook.
WORTH READING
HARD TECH
General purpose robots. Sunday Robotics reports that ACT-2 folded laundry successfully in 99.1% of 785 attempts across unseen homes, using one fixed model with no home-specific adaptation. More important than the laundry itself is the claimed learning loop: a sufficiently broad pretrained model can absorb a new behavior from a single demonstration, then generalize it across environments. If that holds beyond folding, robotics may be crossing from impressive demos to deployable products - shifting the startup advantage towards companies that own the robot, model, fleet and data engine required to compound reliability in the wild.
Rebuilding the stack. BlackRock technology investor Tony Kim argues that generative AI has reversed two decades of software-centric economics, pushing value and capital into chips, memory, optics, power and data-center infrastructure as the industry undertakes a multitrillion-dollar rebuild. His most provocative claim is that memory (not compute) will become the next bottleneck as agents and persistent context make AI dramatically more memory-intensive, while new fabs still take three to four years to build. For founders and investors, the opportunity is shifting toward hard-tech constraints and “token flow” businesses that wrap proprietary context and workflows around abundant intelligence.
HARD MARKETS
Power crunch. PJM is warning that data centers unable to secure sufficient generation could face involuntary curtailment as soon as 2027, while a new emergency procurement process will charge large loads for the supply needed to keep the grid reliable. With PJM forecasting roughly 70 GW of additional data-center demand by 2038, grid access is shifting from a utility service into a core input that developers must actively procure and manage. For startups and investors, this accelerates demand for behind-the-meter generation, flexible load, capacity contracting and software that helps data centres behave like grid assets rather than passive consumers.
HOW TO STARTUP
Open letter for open weights. A broad coalition spanning OpenAI, Meta, Google, Microsoft, Nvidia, a16z, Y Combinator and dozens of startups argues that American AI leadership will depend less on owning a single frontier model than on building an open ecosystem that diffuses AI across the economy. The letter frames open weights as infrastructure for competition, lower costs, customer sovereignty and faster application-layer innovation, while warning policymakers not to confuse legitimate techniques like distillation with model theft. For founders and investors, the implication is clear: the biggest opportunity may sit with companies that turn broadly available intelligence into proprietary workflows, data and distribution.
A new cyber era. Anthropic’s Claude Mythos Preview discovered a substantially stronger attack on the post-quantum signature scheme HAWK and improved the best-known attack on a reduced-round version of AES by 200–800× - although neither result affects production systems today. The striking part is the workflow: the model generated novel cryptographic research largely autonomously, while humans spent far longer validating its conclusions than Claude spent producing them. For enterprises, this exposes some real risks on the horizon. And for startups, this points toward a new scientific stack in which intelligence is abundant, but experiment design, verification and trusted deployment become the critical bottlenecks.
Brute force. Benn Stancil argues that recent AI-assisted mathematical breakthroughs reveal a new form of “brute intelligence”: agents generating, testing and refining hypotheses until they reach a verifiable answer, without requiring constant expert direction. The important capability is not necessarily deeper insight than the best humans, but the ability to run enormous numbers of intelligent experiments quickly and persistently. For startups, the opportunity is to redesign messy workflows as closed, testable systems—because the companies that make their problems look like maths may capture more value than those merely adding copilots to existing processes.
HOW TO VENTURE
Was Lina Khan right? Joe Weisenthal argues that the industry’s case for open-weight AI is strongest not on safety or innovation, but on limiting the power of any single closed-model provider (and therefore requires conceding that Lina Khan was broadly right about concentrated corporate power). He also notes the less idealistic incentive behind the industry’s open letter: most signatories benefit if Chinese open models weaken Anthropic’s pricing power and keep the model layer competitive. An uncomfortable conclusion for all venture capitalists out there: if AI concentration threatens democracy, similar logic cannot be dismissed when applied to search, commerce or social networks?
PORTFOLIO JOBS
Reply