Jensen Huang Launches an Open AI Alliance, Anthropic & OpenAI Team Up in DC, Kimi K3 Goes Global
Summary
Open weights have become both NVIDIA’s stack strategy and a sovereignty strategy. Jensen Huang’s Open Secure AI Alliance argues that attackers already have frontier AI, so defenders need frontier AI ecosystems. NVIDIA benefits when models commoditize and value shifts toward compute. Dario Amodei’s counterargument, as presented by Peter and David Blakely, is that biology breaks the cyber analogy: powerful open-weight models could assist with pandemic-scale pathogens. His proposals were chip controls, restrictions on industrial-scale distillation, and safety testing for powerful models, open or closed.
The strongest regulatory call is to police AI actions, deployments, and compute—not intelligence itself. Alexander Wissner-Gross calls capability-based enforcement “thought policing” and favors defensive co-scaling: give defenders the strongest models first, monitor what systems do, and exploit even a one-month lead during recursive self-improvement. Salim Ismail doubts conventional KYC can survive shell companies; Alexander argues that transparency into installations, compute, and actions is the necessary layer.
The transcript calls the Kimi release K2 once and K3 elsewhere. The K3 passages describe roughly 2,500 downloads in two hours and about 100,000 in 24 hours. Peter got it running on dedicated Modal GPUs in under an hour at about $55 per hour; David Friedberg’s team had polling agents monitoring the repository. The investment implication is local control: “a model that you can control locally is way more valuable than a marginally smarter model” behind an API.
Claude Opus 5 appears to be an efficiency and vision-code release, not an unambiguous intelligence leap. It approaches GPT-5-level frontier intelligence at half the price, with pricing of $5 per million input tokens and $25 per million output tokens. ARC-AGI-3 rose from 1.5% to 30.2%, yet Alexander still preferred Fable 5 for general work and flagged weaker Frontier Math results, modest Humanity’s Last Exam gains, and possible benchmark gaming. The broader opportunity may be scaffolding, where a better problem representation can temporarily multiply domain capability.
OpenAI and Anthropic’s Washington alignment could create a safety regime and a regulatory moat simultaneously. Their reported agenda included an August 1 deadline, a voluntary 30-day government review of releases with serious cyber or national-security capabilities, and equivalent obligations for Meta, xAI, and startups. Peter sees both genuine safety concerns and possible regulatory capture; Alexander argues that the proposal targets the wrong layer.
AI profit pools may migrate toward compute, fabrication, proprietary learning loops, and applications. Peter’s four-layer model spans unreleased internal systems, paid frontier models, commoditized open models, and ecosystem distribution; Salim adds “layer zero,” compute and power. Alexander expects one power-law business model to capture most free cash flow, while David Friedberg expects multihundred-billion-dollar robotics, biotech, and entertainment companies to emerge above the model layer.
China’s export of models and infrastructure makes openness an instrument of geopolitical alignment. Salim warns that restricting US open models would protect a few domestic labs while handing the Global South’s AI ecosystem to China. Alexander’s warning is that foreign infrastructure can progress from lending to listening to “thinking for you,” making locally trained and locally controlled intelligence a possible sovereignty benchmark.
The physical economy is moving toward abundance, but monetary transition remains the risk. Starship 13’s intact splashdown supports a launch-cost path from the Shuttle’s roughly $54,000 per kilogram toward Starship’s stated $10–$100 target. Prima and Neuralink demonstrate new medical and mobility capabilities, but the panel did not claim they follow the same launch-cost curve. Elon’s “money won’t matter in 2036” scenario implies deflation and cheaper basics, while scarce assets, unequal ownership, debt, and fiat systems may persist.
Deep dive
1. Jensen turns open weights into security infrastructure
Peter framed Jensen Huang’s first-ever X post as a deliberate market intervention: signed by 77 companies, it argued that open models strengthen safety, cybersecurity, innovation, diffusion, and sovereignty. The world needs “both frontier closed models and frontier open models,” and NVIDIA’s Open Secure AI Alliance would organize the defensive ecosystem around that premise.
Jensen’s proof case was Hugging Face’s reported intrusion: an agent logged 17,000 actions, escalated privileges, harvested credentials, and moved across clusters. The transcript names the blocked models as “GPT-5, 6, and Claude Fable”; an open-weight GLM-2.5 model then helped find and contain the attack. Peter’s summary: “Attackers have frontier AI, so defenders need frontier AI ecosystems.”
After three days’ silence, Dario said Anthropic had “never advocated for a ban on open-weight models.” He reframed the threat as authoritarian states reaching the frontier, especially in biology, and proposed blocking advanced chips and chipmaking equipment from reaching China, cracking down on industrial-scale distillation, and safety-testing powerful models, open or closed. Peter connected that emphasis to Dario’s biophysics background and Coefficient Bio acquisition.
2. Biology is the fault line the cyber analogy cannot settle
David Blakely’s defense of Dario began with the sharpest objection: “Okay, bioweapon.” Cyber models can help counter cyberattacks, but an AI cannot necessarily reverse a pathogen after release. He separated motive from economics: Dario’s position may protect Anthropic’s valuation, yet David believes Dario is “speaking his mind without an agenda,” even at the expense of his own valuation.
David rejected a simple open-source-software analogy because weights are tools for constructing downstream systems, not readable source code whose malicious component can simply be spotted. They might help “cure all disease and give us infinite longevity,” but that does not answer whether they should be handed to “every terrorist in the world.”
Alexander Wissner-Gross rejected biosafety as a sufficient case for restricting open weights: dangerous knowledge already exists online, dangerous acts are already possible, frontier AI may not be necessary, and biologically capable models already exist. He compared the warnings to Microsoft’s late-1990s FUD about viruses and IP lawsuits—claims that, in his telling, were overturned when open source proved safer in many settings.
Salim supplied a counterexample from the head of innovation at a three-letter agency: when capabilities are widely distributed, funding and opening the ecosystem can expose bad actors earlier. His preferred control point is transparent monitoring of data centers and workloads. Peter’s caution remained: once weights can run “in a basement somewhere,” defensive AI may see the threat only when it acts.
3. Defensive co-scaling requires visible deployments and compute
Asked about KYC, Salim said, “You can’t do it”—Chinese companies can use shadow entities in Singapore, making identity controls easy to route around. Alexander answered just as categorically: “Of course you can do KYC,” but superintelligence should go further by allocating some of itself to policing other superintelligence. “Defensive co-scaling is the answer.”
The apparent synthesis was transparency at the deployment and action layers. Open weights can remain open, but regulators would need visibility into installations, compute, and real-world effects. Peter suggested that governments might receive unreleased systems—possibly GPT-6 or a Claude successor—as white-hat defenders.
Alexander added a timing mechanism: in recursive self-improvement, a month or two of model advantage can compound from marginal into overwhelming. That makes temporary frontier access potentially more useful than permanent intelligence ceilings—provided the “good guys,” however defined, can detect and counter malicious behavior.
4. NVIDIA is opening a cold war over where AI margins accrue
Alexander sees the former détente between GPU suppliers and frontier labs becoming a “cold war.” NVIDIA’s incentive follows the aggregator rule: “commoditize your complements.” Popular open-weight models weaken profit capture at the model layer while preserving demand for NVIDIA compute; proprietary oligopoly would instead concentrate margins at OpenAI, Anthropic, or their peers.
His unresolved question sits one layer lower: why is NVIDIA not equally aggressive about commoditizing fabrication through TSMC or Samsung competitors? Salim’s answer was dependence—TSMC is too powerful to irritate publicly, so any NVIDIA effort would have to be exceptionally covert. Elon can announce a “Terafab”; a more dependent buyer cannot.
The Open Secure AI Alliance reminded Alexander of the Open Source Initiative’s 1998 creation and Linux’s challenge to Microsoft. Alexander also supplied IBM’s 1995 discovery: 95% of Fortune 500 CIOs claimed not to use open source, while 95% of their sysadmins said they did. IBM embraced what its hardware and services businesses could monetize.
5. OpenAI and Anthropic are writing guardrails their rivals must inherit
Peter said OpenAI and Anthropic were using shared Washington back channels ahead of an August 1 deadline for frontier-model rules. Their reported program included a federal review process, a voluntary 30-day government look at releases with serious cyber or national-security capabilities, and matching obligations for Meta, xAI, and frontier startups.
Alexander’s objection is architectural: applying identical safety rules to downloadable weights and gated APIs is “too clever by half” because deployment risks differ. His preferred line is memorable: “Aiming enforcement at intelligence is like thought policing.” Regulators should “police what the AIs are doing or being used to do, not what they’re thinking or how smart they are.”
Peter called the alliance “regulatory capture in real time,” comparing it with railroads, banks, telecom, and Big Tech: incumbents define acceptable guardrails, keep government close, and competitors out. He doubts it will hold because open models are improving too quickly, but views the attempt as economically rational.
Peter also argued that both CEOs may be genuinely trying to create a safe future rather than merely defend their valuations. He noted that Sam Altman is not an OpenAI shareholder and has roughly 400 investments in vertical companies that benefit from cheaper models such as Kimi. Alexander’s pushback was that genuine rulemaking should also include Google, Meta, and the rest of the ecosystem.
6. The winning AI business model will follow the bottleneck
Peter’s four-layer cake begins with unreleased “wild stallions” such as GPT-6, used internally to create materials, drugs, energy systems, and new businesses. Paid frontier products such as the models Peter labels GPT-5.6 solve and Fable 5 sit beneath them; commoditized open models power everything else; ecosystem owners monetize applications and distribution.
His distribution numbers were 3.5 billion Meta users across Meta’s apps, 2 billion Google users using Gemini, and roughly 1 billion OpenAI ChatGPT users. Consumers inside WhatsApp or another application will not ask which model answered; the ecosystem captures value because the model disappears into the product.
Alexander expects power laws, not four balanced pools: one business model may capture 80%-plus of industry free cash flow. Salim adds “layer zero”—compute and power—because value moves toward whichever resource remains constrained. David Friedberg sees value moving both downward to fabs and memory and upward to future robotics, biotech, and entertainment giants.
7. The Kimi release makes capability diffusion irreversible
The transcript calls the release Kimi K2 in one announcement passage and Kimi K3 in the surrounding download and architecture discussion. The K3 passages describe a frontier-adjacent model available globally through Hugging Face with no API key, gatekeeper, or revocation switch. The repository showed 2,500 downloads in two hours; Peter’s research suggested roughly 100,000 in 24 hours. Once copied widely, “there’s no undo button.”
David Friedberg’s team had polling agents pinging the page every 15 seconds and briefly encountered a 404. Peter interpreted that moment as a possible White House intervention before the page returned.
Peter got the model running on dedicated GPUs on Modal in under an hour. Full-throttle operation cost about $55 an hour, while voice prompting could provision 100 instances in about two minutes without specialized technical skill.
Peter called it “the biggest turning point in human history”: an AI capable of self-improvement running in the wild. He also cited a 13% decline in Anthropic secondary pricing after the K3 announcement, to about $230 billion, prompting the line, “It’s nice to lose $230 billion on someone’s tweet.”
8. Kimi swaps out the transformer while retaining its silhouette
Alexander’s architectural surprise was NoPE—no global position embeddings. Kimi Delta Attention, or KDA, behaves like a small recurrent neural network inside attention, smuggling in enough ordering information for a million-token context without the original Transformer’s explicit positional machinery.
His metaphor was the “Ship of Theseus”: attention, position embeddings, residual streams, layers, and sparsity are being replaced piece by piece. The system remains recognizably Transformer-like from a distance, yet its frontier implementation is increasingly unrecognizable beside Attention Is All You Need.
Salim’s practical reading was that NoPE supplies fading memory, unlike RoPE’s tendency, in his explanation, to give very distant context too much weight. He also highlighted the claim that attention can be removed around layers 100 or 120 with nearly the same output and greater speed.
Open access lets researchers—including those in China—try apparently obvious alternatives and lets other labs adopt what works. The discussion treated that diffusion as one of the central benefits of open models.
9. Local models turn token costs into an enterprise redesign problem
Salim’s call was that local control can outweigh marginal intelligence: companies can fine-tune proprietary knowledge, run inside secure environments, and avoid sending regulated data to a vendor’s cloud. That matters for sovereign workloads and industries where data control is part of the product, not merely a compliance cost.
Peter expects “token maxing” to turn a negligible line item into one larger than payroll by year-end and perhaps ten times payroll two years later. Specialized tuning, deleting general-purpose cruft, and streamlined models could yield 10X—and perhaps 100X—improvements for the same business process.
Salim’s implementation advice was deliberately narrow: rebuild one workflow that sharply increases revenue and one that sharply reduces cost, then expand. The thesis is not generic chatbot adoption; it is an AI-native operating model in which local inference, proprietary data, and repeated feedback compound together.
10. Opus 5 wins on cost, but Fable 5 still wins Alexander’s trust
Anthropic positioned Claude Opus 5 as its fourth Claude 5-generation release, approaching the frontier intelligence of GPT-5 at half the price. It became the Claude Max default at $5 per million input tokens and $25 per million output tokens, unchanged from Opus 4.8.
ARC-AGI-3 produced the headline: 1.5% on Opus 4.8 became 30.2% on Opus 5, which Alexander said was the highest official baseline-model score he was tracking. He inferred optimization around front-end development and the vision-code intersection, and welcomed fewer refusals when benign biology or cybersecurity questions were misconstrued.
The counterevidence kept him on Fable 5. Humanity’s Last Exam with tools moved only from Fable’s 63.9% to Opus’s 64.7%, performance fell without tools, and legal, health, and Frontier Math results suggested weaker general capability. A third-party ARC-like game also failed to reproduce the headline jump, raising concern about “mild benchmark gaming.”
VoxelBench placed Opus 5 third, just behind Fable 5, at much lower cost; OpenAI’s Sol still led. Peter’s broader metric was velocity: Opus 4.8 arrived May 28, Opus 5 roughly eight weeks later with about 2X efficiency—faster than the previously discussed 10-week price-performance doubling. Alexander added that model releases had averaged one every six days.
11. Scaffolding is temporary alpha for domain entrepreneurs
ARC Prize Foundation organizers reportedly saw Opus 5 convert visual objects into algebraic software representations and do mathematics over them—the first such frontier-model behavior Alexander knew of. His caveat was explicit: that novelty holds “unless Anthropic was benchmark gaming on ARC-AGI-3.”
Peter said that reframing the puzzle might reportedly produce about 98% performance, although he could not fully replicate the result. The investable analogy is domain-specific: give a model a better representation of protein folding, drug discovery, or robot-arm design and it may appear three times smarter. “That is an entrepreneurial heaven.”
Alexander’s warning is that scaffolding advantage decays: “Today’s scaffold is tomorrow’s baseline capabilities.” Yet the current window is real enough that Peter wanted every university to teach scaffolding, with prompt engineering as a prerequisite. Opus 5’s one-shot Call of Duty demo reinforced the point: experimentation cost is approaching zero.
Privacy remains the counterweight. A Reddit user reportedly found that Google searches for shared Claude URLs surfaced health records, private documents, names, and telephone numbers. Alexander noted that the sharing links were designed to be social, while Peter and Salim kept the enterprise lesson: sensitive proprietary data eventually moves on-prem.
12. China is exporting an AI sphere of influence
Peter summarized the Financial Times framing as “Pax Sinica”: Xi Jinping using models and infrastructure as statecraft across the Global South while Washington debates open versus closed. Whoever supplies the intelligence stack can shape economic development and political alignment for decades.
Salim warned that restrictive US open-model policy would be a strategic “self-own and shooting your own foot at an epic level.” Protecting a few domestic labs would hand the wider startup, university, hospital, and defense ecosystem to China. “Openness is not a philosophical preference anymore”; it is soft-power infrastructure.
Alexander compared the choice to exporting F-16s: refusing a risky customer may simply send that country to Russian or Chinese suppliers, strengthening their industrial base. He also said that there was no US open-source model currently competing with the Chinese offer.
Alexander escalated the sovereignty argument: a foreign loan can seize an asset, telecom equipment can spy or shut down, but imported superintelligence is “thinking for you.” Peter’s supporting story involved a smaller Asian country taking a half-billion-dollar Chinese port loan when drones, improving every nine months, might eventually move 20,000-pound containers without that infrastructure.
13. Starship 13 moves launch economics toward ordinary infrastructure
Starship 13 deployed 20 operational Starlink V3 satellites, performed an in-orbit Raptor relight relevant to Artemis, and completed a precise soft landing in the Indian Ocean with the vehicle initially intact. The satellites were described as supporting roughly 0.5-to-1-gigabit connectivity almost anywhere on Earth.
Peter said the landing’s precision made a Flight 14 catch by the tower’s “chopsticks” plausible. The abundance curve was his real story: Shuttle launch cost roughly $54,000 per kilogram, Falcon 9 about $2,000, and Starship’s stated target just $10–$100.
Salim admired SpaceX’s willingness to treat every failure as information rather than embarrassment. Alexander Iskold connected the broader hardware renaissance to AI: models can specify part numbers, identify a German lens vendor, write manufacturing requirements, and combine rapid design with 3D printing and capable robots. “The amount of possibility is so exponentially bigger” than five years earlier.
14. Prima uses a near-term product to finance a deeper neural interface
Science Corporation’s Prima received a European CE mark for restoring central vision affected by age-related macular degeneration. Glasses capture the image and transmit it through infrared to an implant behind the retina, which stimulates remaining retinal cells. Patients reportedly gained five lines on a standard eye chart after 12 months.
Alexander Iskold called it almost Borg-like and noted that retinal surgery stays farther from direct brain intervention than Neuralink’s Blindsight approach. He expects competition among central and peripheral nervous-system implants, stimulation, wearables, ultrasound, and fMRI to reveal the most ergonomic interfaces within five to ten years; non-invasive vision restoration might arrive within 20.
Peter described Prima as Max Hodak’s “stage-zero revenue-generating engine,” funding a more ambitious interface in which neurons grow from circuitry into the brain. The long-term metaphor is a third hemisphere connected to the cloud—an “exocortex”—without the electrode damage Peter associated with current BCI approaches.
The entrepreneurial lesson was survival before scale. Apple began with a box of chips, Facebook at Harvard, and Google as a Yahoo plug-in. Peter cited Bill Gross’s conclusion that timing mattered most, then SpaceX’s own 2008 sequence: three Falcon 1 failures, a successful fourth launch, and a billion-dollar NASA resupply contract. “Just find a way to survive.”
15. Neuralink’s wheelchair is the first step toward embodied telepresence
Neuralink translated imagined cursor movement into analog wheelchair controls, including a safety design that returns the cursor to center if the user becomes incapacitated. Alexander’s next step is full exoskeleton control for paraplegic and quadriplegic users; Peter’s is an Optimus body whose eyes, ears, and limbs are operated remotely through Starlink—“effectively telepresence.”
Alexander then extrapolated, with hedges, toward behavioral or partial mind uploads trained from fMRI or ultrasound data. He expects something adjacent within roughly five years, “certainly by the end of this decade.” The liberating Avatar scenario carries its own social extreme: people who never leave a room and interact only through BCI-controlled robots.
16. Abundance demotes money before it eliminates scarcity
Elon’s 2036 prediction rests on a simple identity: if robots provide more goods and services than humans can consume, and output grows faster than the money supply, deflation replaces inflation. Salim separated money’s three functions—exchange, accounting, and storing value—and argued that some exchange mechanism survives wherever access, ownership, or location remains scarce.
Alexander translated the claim as “Star Trek economics”: food, shelter, healthcare, utilities, education, and entertainment could be effectively demonetized within ten years, while interstellar travel, a week on the Moon, antiques, or collectibles retain prices. His guess for demonetizing the total economy was closer to 30 years, “not investment advice.”
Peter’s bridge is a roughly $3,000 monthly UBI whose purchasing power rises as AI physicians, robot surgeons, autonomous vehicles, and robot-built housing push costs down. Wealth gaps might widen, but his criterion is whether the floor rises: every person gains food, water, energy, healthcare, education, and freedom.
Peter also cited Jeff Booth’s claim that each $1 of GDP growth has required $4 of debt, then imagined a $10 million TV factory whose $1,000 product falls to $500 and then $250 before the loan can be repaid. He said cash is losing about 14% annually and expects abundance to force a wholesale rethink of fiat, central banks, and economic measurement.
17. Proprietary learning loops outlast hidden reasoning
Asked whether Anthropic could prevent distillation by hiding model reasoning, Salim said it could make copying harder, not stop it. Enough input-output examples let competitors infer useful behavior. Durable advantage instead combines unique data, compute, users, distribution, and feedback loops that continuously improve the full system.
Acquiring companies does not automatically acquire their knowledge: much of it remains tacit inside key employees or incompatible operating processes. Salim expects more partnerships and co-ops, such as hospitals pooling data for pharmaceutical companies, because shared incentives may extract value more effectively than buying the organization.
Alexander rejected “garbage in, garbage out” as a universal indictment of synthetic data. Procedural worlds, generated code, and deliberately injected bugs can supply nearly infinite reinforcement-learning tasks; in the theoretical infinite-compute limit, a system could learn from entirely synthetic sequences.
Dave Blundin’s practitioner distinction was sharper: “Synthetic data is fine. It’s mislabeled data that actually kills you.” A single contradiction can warp weights as the network struggles to reconcile it. Clean generated data may therefore be preferable to supposedly natural data containing unrecognized errors.
18. Longer lives and AI education require adaptation, not population panic
Peter’s longevity mantra remained “LEV by 2033.” Salim described a spiky arrival: some subpopulations may already be approaching one year of gained expectancy per elapsed year, while others might arrive by 2030. Whether curing a disease counts as catch-up or escape velocity depends partly on accepting Salim’s premise that aging itself is a disease.
Population is not the binding constraint in Peter’s view. Replacement requires roughly 2.1 children per family, while South Korea and Japan were cited near 0.6; he expects world population to reach 9.5–10 billion and then fall. Apparent resource limits also invite substitution—lithium scarcity, for example, can yield to new deposits or sodium batteries.
Peter’s education prescription is AI-led, purpose-driven learning because fixed curricula cannot keep pace. High school should reward productive inquiry, real projects, entrepreneurship, prompt engineering, and scaffolding. Peter also proposed prompt communication, prompt engineering, and context engineering as a possible humanities credit at MIT; Alexander later identified prompt engineering as a prerequisite for scaffolding.
For individual investors, Peter suggested universities, crowdfunding, AngelList syndicates, VC funds, and demo days; Dave Blundin’s stronger advice was to earn access by helping early companies with sales, introductions, operations, or even moving offices. “If you add value, you will get stock”—the human learning loop beside the model learning loop.