Anthropic vs. Alibaba, OpenAI Delays Its IPO, and the US Government Blocks GPT-5.6 | #267
Summary
Washington’s intervention puts the US government directly inside the frontier-model release loop. Anthropic’s Mythos 5 was limited to 100 selected companies, while OpenAI’s GPT-5.6 Sol, Terra, and Luna were reportedly restricted to 20, with access approved “customer by customer.” Dave Blundin argued uniform controls need not damage valuations, but others warned they could push global customers toward Chinese open-weight models and leave Western users behind the labs’ internal capabilities.
The strongest challenge to model controls is that older and open models can already be lifted above the embargoed frontier through better harnesses. Emad Mostaque said GLM 5.2, trained with roughly $25 million of compute, topped his team’s Frontier SWE testing when properly orchestrated; Diamandis drew the blunt conclusion: “The government is too late.” If GPT-5.5, Opus 4.8, or GLM 5.2 can outperform withheld models through scaffolding, prompts, tools, and multi-model routing, capability control cannot stop at weights.
Cybersecurity supplied the public rationale for gating, but recursive self-improvement may be the deeper line Washington is defending. Diamandis cited Project Glasswing results in which Mythos reportedly identified vulnerabilities in highly sensitive classified systems, with Senator Mark Warner saying it broke into almost all of them “not in weeks, but in hours”; exploitation was outside the exercise’s scope. Blundin countered that the decisive new query is, “Can you help me build yourself?”—making prompt logging, KYC, licensing, citizenship restrictions, and geopolitical AI blocs more likely than a quick return to unrestricted releases.
AI security is becoming a market for trusted models, automated remediation, certification, and liability—not merely vulnerability detection. GPT-5.5 Daybreak scored 85.6 on CyberGym, and OpenAI’s stated ambition is to write and test fixes from browsers through the Linux kernel. Yet the same models can insert nearly invisible backdoors, especially if an adversary obtains unrestricted base weights; the panel’s through-line was that “only AI can keep up with AI,” creating a valuable but politically fraught trust layer.
OpenAI’s delayed IPO looks more like a capital, governance, and strategic-flexibility decision than fear of SpaceX volatility. SpaceX priced at $135, traded as high as $202, and closed near $153 while retaining a roughly $2 trillion valuation; Blundin called that a well-priced IPO, not a cautionary failure. OpenAI has reportedly raised $122 billion, is running at $40 billion-$50 billion of revenue while forecasting a $26 billion annual loss, and may prefer to grow Codex and resolve leadership, conflicts, and corporate structure before defending a $1 trillion public valuation.
China’s AI position is bifurcating: open models are nearing coding parity while video generation may already be ahead. ByteDance’s C-Dance 2.5 was described as producing 30-second 4K clips with as many as 50 image, video, and audio references, while Anthropic accused Alibaba of 28.8 million fraudulent exchanges across 25,000 accounts to distill Claude. The panel expects those allegations to become a policy trigger for a “second Cold War type path,” even as recursive improvement may make copied Western traces progressively less necessary.
Quantum received the headline funding, but photonic computing drew the sharper near-term infrastructure call. The US committed $2 billion across IBM, D-Wave, Rigetti, Inflection, and PsiQuantum, yet Wissner-Gross remained only “mildly excited” because useful quantum advantage is still elusive. Blundin instead forecast highly quantized photonic neural nets at perhaps 1/100 the mass of Nvidia-based compute for equivalent work, calling that estimate conservative and potentially decisive for orbital AI infrastructure within 12-18 months.
Neural interfaces and orexin drugs represent two different attempts to expand scarce human productive time. Neuralink may attempt direct human-to-human communication later this year, potentially bypassing speech at 40-60 bits per second and typing at 5-20; Mostaque suggested shared latent representations could raise effective bandwidth by “10 or 100,000 times.” Eli Lilly’s $6.3 billion purchase of Synthesa Pharmaceuticals similarly points toward orexin therapies that might eventually let ordinary users sleep less without normal deprivation costs—an opportunity Wissner-Gross compared to the GLP-1 playbook.
Deep dive
1. Autonomous drones are turning wildfire response into a deployment problem
Diamandis opened with the Wildfire XPRIZE finals in Fairbanks, where Anduril, Germany’s Dryad, and Australia-UK team Aura used autonomous aircraft to detect fires and drop suppressant. The target was to identify ignition and extinguish it within 10 minutes, removing humans from the most dangerous work.
The trial covered 1,000 square kilometers and included decoy fires; teams were instructed to attack a fire once it exceeded 2 meters or began moving. Wissner-Gross’s analogy was the Oil Cleanup XPRIZE: a winning system can become “a global best practice in one iteration,” establishing a new baseline rather than remaining a demonstration.
2. Washington has become the gatekeeper for frontier-model releases
Diamandis described an unprecedented national-security hold: Anthropic initially pulled Fable and Mythos, then received permission to release Mythos 5 to 100 selected companies. OpenAI’s GPT-5.6 preview was reportedly limited to 20 companies, with the government approving access “customer by customer” before a hoped-for broader release within weeks.
OpenAI announced three exact variants—GPT-5.6 Sol, the flagship; GPT-5.6 Terra, the middle tier; and GPT-5.6 Luna, the faster low-cost model. All three, in Diamandis’s account, were being throttled by the White House.
Blundin’s initial answer separated regulation from valuation: controls are “inevitable” because the models are “so insanely capable,” but equal treatment across US labs need not reduce market caps. His caveat was pointed—government fairness “seems unlikely”—while he still called the frontier labs potentially the most valuable companies in history.
China produced the first major disagreement. Blundin said it remained far from pushing the frontier and excelled mainly at distillation and copying; Wissner-Gross cited an extrapolation suggesting Chinese open weights could erase the capability lag by Christmas.
3. Harnesses are making model weights a weaker measure of capability
Mostaque argued that single-model benchmark rankings increasingly miss the point. Sakana’s Fugu and the Blitz team’s multi-model systems combine prompts, routing, tools, and models, while Intelligent Internet’s forthcoming harness was designed to lift existing systems on Frontier SWE, where individual tasks can require 11 hours and genuinely novel kernel work.
In Mostaque’s current testing, he said Intelligent Internet would announce GPT-5.5 overtaking Mythos on Frontier SWE with its new harness, while GLM 5.2 outperformed GPT-5.5 and reached the top of Frontier SWE. With a normal harness GLM 5.2 ranked roughly fourth or fifth; orchestration changed the outcome despite an estimated training-compute cost of only about $25 million.
Wissner-Gross defined a harness as all the “non-weight capability improvements” around a model: system prompts, parsers, tool use, logic, model mixtures, and what Andrej Karpathy might call software 1.0 orchestration. Prompts were progressively factored into reusable scaffolding until the surrounding system became as important as the neural weights.
Blundin’s software-builder perspective captured the behavioral shift: models now announce that they need scaffolding or will “monkey patch something,” a phrase he rarely encountered in 35 years of coding. The process is less legible, but the delivered software works—encouraging users to let the system run.
4. Existing models may already sit above the government’s control line
Diamandis converted Mostaque’s benchmark result into the episode’s sharpest policy claim: “Therefore, the government is too late.” If GPT-5.5 or Opus 4.8 plus a strong harness can exceed Mythos or GPT-5.6, Washington would have to reach backward six months and restrict already distributed models.
Mostaque distinguished competent from novel intelligence. Competent systems capable of building codebases are already broadly accessible; intelligence that can make a user unusually capable at a novel attack may become licensed, restricted to citizens, or otherwise tightly controlled.
Diamandis worried that enterprises unwilling to accept government-selected access would move toward on-premises Chinese open weights. Blundin thought the US could respond by banning corporate use of Chinese models and requiring licenses, KYC, and prompt retention for any new frontier system.
Wissner-Gross saw a “very real risk” that US labs achieve AGI internally while Western customers remain at parity with—or behind—Chinese public models. That creates perverse labor incentives: people may join OpenAI or Anthropic simply to cross the capability “event horizon” and use tools unavailable to the rest of the economy.
5. AI access is hardening into a geopolitical licensing regime
The emerging map, in Wissner-Gross’s phrase, could be “Pax Silica” expanding into “Pax Intelligentsia”: an American superintelligence bloc, a Chinese bloc, and a Europe pressured to choose between them. Blocking Chinese models only inside the US would not work if the rest of the world continued adopting them.
Mostaque imagined frontier access resembling a driver’s license or security clearance, complete with KYC and perhaps a requirement to demonstrate loyalty to America. Wissner-Gross added that citizens and US corporations could retain privileged access while foreign nationals—even Canadians working inside a lab—were excluded from the strongest systems.
Mostaque proposed keeping superintelligence in “a couple of boxes”: one would be unhealthy, but ten could also be destabilizing. His preferred regime would inspect every prompt and use case with other AIs, then give allied countries access to Fable-class capability only after they accepted a newly negotiated rulebook.
Mostaque expected economic protection to sit beside security policy. In his account, David Sacks and Sriram Krishnan had left government and Howard Lutnick was driving AI policy, forcing Washington to balance US industrial advantage, catastrophic misuse, and the risk of financially damaging Anthropic or OpenAI.
6. Recursive self-improvement is the stronger case for the hold
Diamandis supplied Washington’s steel man: under Project Glasswing, Anthropic’s Mythos reportedly identified vulnerabilities in highly sensitive classified systems, with Senator Mark Warner saying it broke into “almost all of our classified systems” in hours rather than weeks. Exploitation was outside the scope of the exercise. Twelve days later, the administration restricted Mythos 5 and Fable 5 for foreign nationals.
Blundin accepted that cyber findings were real but called them a trigger rather than the underlying agenda. His decisive distinction was that Mythos could answer a question Opus 4.8 could not: “Can you help me build yourself?” Once that knowledge escapes, recursive improvement becomes globally reproducible.
Wissner-Gross thought China already possessed enough capability for its own recursive-improvement “Fermi pile,” without more Western reasoning traces or stolen weights. The policy therefore makes sense only if every day in an endgame race matters; otherwise a lighter-touch regime should replace broad cyber restrictions within a month or two.
His historical analogy was an inverted Sputnik moment: this time the strategic surprise came from inside the United States. Private frontier labs leapfrogged cyber capabilities that the NSA and Cyber Command had kept bottled up, leaving the bureaucracy reacting to domestic rather than foreign technological shock.
7. Defensive AI creates a new control point over global software
GPT-5.5 Cyber, code-named Daybreak, scored 85.6 on CyberGym, which Diamandis called the highest single-model result yet. Altman’s stated prize was not merely finding holes but automatically writing and testing fixes “across web browsers all the way down to the Linux kernel.”
Wissner-Gross expects AI to “bulk solve” vulnerabilities across foundational open-source repositories, comparing the project’s scale to building the Interstate Highway System. He also inferred that OpenAI may steer models to frustrate or poison offensive cyber activity while preserving defensive utility, although OpenAI did not use Anthropic’s more inflammatory language.
Mostaque emphasized the value of unrestricted pre-system-prompt weights. If an adversary steals the more creative base model while defenders receive a constrained version, nation-states—not “kids in the basement”—gain a durable offensive advantage against infrastructure already too fragile for humans to audit manually.
The trust problem cuts both ways: a Chinese model could insert a one-line backdoor into thousands of generated lines, but Wissner-Gross said people were also asking, “How can we trust American AI is not installing backdoors?” Blundin’s endpoint was an imminent ISO-like certification market, with the certifier’s geopolitical legitimacy as important as its technical competence.
8. OpenAI’s IPO delay is about optionality, not SpaceX’s chart
The reported choice was to list now below $1 trillion or wait until 2027 while scaling revenue, infrastructure, and partnerships. Diamandis suggested public investors might not tolerate enormous compute spending and an AGI timeline measured in years while judging management quarter by quarter.
Blundin rejected SpaceX’s price action as the reason. Its shares priced at $135, opened near $150, peaked around $202, and closed at $153 while supporting a roughly $2 trillion valuation—“pretty much a perfectly priced IPO,” not evidence that listing was a mistake.
His alternative explanation was simple: OpenAI had just raised about $122 billion and did not need cash. A public filing would also force Altman to disclose hundreds of outside investments and conflicts, submit communications to Regulation FD controls, and absorb lawsuits whenever the stock dropped.
Wissner-Gross argued OpenAI overcommitted to consumer revenue and is now racing to “become Anthropic faster than Anthropic can become OpenAI.” Codex is the promising enterprise engine, but OpenAI may still need revenue “rabbits,” profitability, or even a governance move such as acquiring Sierra and installing Bret Taylor as CEO before defending a trillion-dollar IPO.
9. Private capital lets frontier labs wait out the singularity
Mostaque put OpenAI revenue at roughly $40 billion-$50 billion, alongside a forecast $26 billion loss for the year. With $122 billion recently raised, it has room to redesign its corporate structure rather than rush into public markets; Anthropic’s tighter financing position may require another raise or an IPO sooner.
Blundin pushed back on Diamandis’s rule that companies should list only with profits and predictable revenue. He thought frontier labs could manufacture enormous revenue quickly if compute were available, but regarded SEC filings and roadshows as “the stupidest thing you can do in the middle of the singularity.”
Employees already have secondary liquidity, and investors will “throw money at Dario,” weakening another traditional reason to list. The cost is distributive: Diamandis noted that what may be history’s largest wealth-creation event remains inaccessible to retail investors and concentrated among VCs, family offices, and sovereign funds.
10. Neuralink is moving from restoration toward superhuman communication
Musk said Neuralink might attempt direct human-to-human communication later this year. Diamandis framed the endgame as an input-output layer that lets humans “couple with AI during the singularity,” not merely a medical device restoring lost motor function.
Wissner-Gross discussed research suggesting bilingual human hippocampi organize concepts like a transformer’s vector-embedding space. Mostaque drew the further inference that, if the embedding theory generalizes, decoding thought and telepathy could be easier than expected. Diamandis suggested that this could create demand for non-invasive human-to-human communication alongside implanted Neuralink systems.
Human output is strikingly narrow in their comparison: speech carries roughly 40-60 bits per second, typing 5-20, and conscious behavioral selection perhaps 10. Mostaque argued that telepathy would not need full sentences; precisely targeted signals could activate a shared latent space and raise effective bandwidth by “10 or 100,000 times.”
Mostaque cited Stability’s 2023 Mind’s Eye work reconstructing viewed images from fMRI via Stable Diffusion as evidence for compatible latent representations. The upside is unprecedented connection and self-understanding; the downside ranges from exploding divorce rates to “borg-anisms” and collective minds.
11. SpaceX is assembling a vertically integrated space economy
Musk’s naming system maps an expanding stack: Starlink moves bits, Starfall moves atoms, and Starmind moves intelligence. Wissner-Gross added Starship, Starbase, Starfactory, defense-oriented Starshield, Earth-and-orbit observation service Stargaze, and the newly disclosed Star Pipe—prompting Musk’s joke that there was “too much star shit.”
Starfall’s practical opportunity is downmass: companies can launch experiments or manufactured goods to low Earth orbit and retrieve them. The episode’s memorable specimen was yeast launched for an Orbital Brewing Company and returned for “LEO-brewed beer”; Outpost and Varda were cited as other orbital-manufacturing participants.
Stargaze could compete in situational awareness by using satellites to look both upward and downward. Star Pipe suggests SpaceX is developing methane and natural-gas logistics for Starship, terrestrial data-center power, and eventually lunar or Martian industry—an adjacent oil-and-gas capability emerging from propulsion needs.
12. C-Dance 2.5 puts China ahead in controllable video generation
ByteDance’s C-Dance 2.5 beta was described as generating 30-second 4K video from as many as 50 reference inputs across images, video, and audio, with directorial controls and text-based editing. Mostaque’s reaction was “I told you so”: he had expected Hollywood-grade control in 2026 and full-length movies by 2027.
Fifty references could represent different characters and media elements, giving creators close to pixel-level continuity and control. For studios, that lowers costs; for post-production and on-set labor, it begins one of the first large displacement waves outside call centers, with little obvious path to retraining every affected worker.
Mostaque’s counterweight was collaborative creation. These systems need not remain “single-player experiences”; groups can construct stories together, including hopeful futures that conventional economics would never fund. Diamandis connected that possibility to the Future Vision XPRIZE’s plan to create at least one, and hopefully more, hopeful full-length motion pictures.
Wissner-Gross said China was “running away with video generation” because its labs face lower data costs, fewer practical copyright constraints, and less pressure to prioritize lucrative coding models. Western labs are chasing recursive code improvement because code produces much more revenue per token or FLOP.
13. Real-time video expands the market from Hollywood to software
Blundin saw latency, not visual quality, as the remaining “buzz killer.” Liquid AI had shown him generation that kept pace with speech, built on a more efficient context architecture small enough to run on-premises in a Mercedes without an internet connection; he expected its next financing to return the company to view.
The immediate prize is interactive gaming: the panel put games at a few hundred billion dollars annually versus roughly $50 billion for Hollywood. Wissner-Gross estimated video-generation revenue near $4 billion-$5 billion, while code generation was perhaps 20 times larger, but forecast real-time 4K generated games as soon as next year.
Wissner-Gross suggested C-Dance’s physical consistency reveals an implicit physics embedding, making video world models a second possible route to AGI beside recursive code improvement. He called the holodeck a trillion-dollar market.
Diamandis described Alibaba’s One Streamer demo as real-time, interactive video-to-video participation. He suggested that an AI joining Zoom or FaceTime as a visible colleague could address a claimed $30 trillion enterprise-software market, turning game-engine capabilities into workplace infrastructure.
14. Anthropic’s distillation allegation will become a policy weapon
Anthropic accused Alibaba of using 28.8 million fraudulent exchanges across 25,000 fake accounts to extract Claude’s capabilities. Diamandis characterized the alleged operation as the largest model-theft campaign yet; Wissner-Gross noted the irony that Anthropic itself has faced lawsuits over copyrighted material used in pre-training.
Wissner-Gross distinguished a terms-of-service violation from espionage, predicting heavy litigation over that boundary. He described Chinese proxies offering Western models at roughly one-tenth the normal cost in exchange for collecting users’ reasoning traces, which could then become training material.
Mostaque said direct latent distillation may be detectable, but asking a teacher model to solve hard problems or generate excellent code is much harder to stop. Frontier models can substitute for expensive human experts, expanding a student model’s data distribution before it enters its own recursive improvement loop.
Diamandis argued that nobody familiar with Chinese copying practices should be surprised, using his son’s $25 Shenzhen Rolex—complete with inscriptions and patent markings—as the analogy. He expects the allegation to become the “trigger” the US, Europe, and perhaps South America use to suppress Chinese AI and formalize a second Cold War.
15. Superintelligence could either erase IP or enforce it ruthlessly
Diamandis recalled Steve Jurvetson and Astro Teller’s argument that IP protection becomes futile near the singularity: AI will not copy a product exactly but rapidly reinvent and improve it. Survival then depends on continuous innovation, not defending work completed years earlier.
Mostaque agreed that competent systems can increasingly “one-shot” software—“do Teams but make it not annoying”—then recreate and remix rather than visibly copy. Strong copyright regimes may still matter in music, but software’s weaker protection makes broad capability lockdowns increasingly difficult.
Diamandis took the opposite urgency: AI could create more intellectual property in the next 18 months than all prior human history, and uncontrolled copying could produce global chaos. Wissner-Gross sharpened that rebuttal: the same intelligence that routes around patents can draft stronger claims, litigate superhumanly, and make IP protection “utterly supercharged.”
16. Quantum policy is a hedge on an AI-discovered advantage
A new US initiative committed $2 billion, including about $1 billion for IBM’s Albany foundry, $100 million each for D-Wave, Rigetti, and Inflection, and $140 million for PsiQuantum. An executive order also called for a Quantum Computer for Application Development and Discovery Science and stronger counterintelligence protection.
Wissner-Gross was only “mildly excited.” Protein folding was solved classically rather than by quantum computing, and known quantum-advantaged algorithms have not proved economically transformative; he found the government’s desire to avoid another strategic surprise more convincing than the immediate applications case.
Mostaque’s bull case depends on convergence: Mythos-level AI might finally formulate and program the right questions for quantum machines. If useful answers arrive in microseconds rather than requiring huge test-time compute, hybrid GPU-quantum systems could break the default relationship between more energy, more compute, and more intelligence.
Wissner-Gross challenged the implied collapse in computational complexity but accepted that AI might discover a useful quantum algorithm humans missed. He rejected the stronger claim that this eliminates orbital compute: a Dyson swarm might instead consist of hybrid or quantum machines, because demand and local resistance to data centers would remain.
17. Photonics may matter sooner than conventional quantum computing
Blundin’s stronger conviction sits in quantum photonics and sensing. Work at his new Quantum.ai venture convinced him that highly quantized neural networks can perform comparably to 32-bit floating-point systems, opening the door to much more efficient optical matrix multiplication.
His forecast was unusually specific: by the next December discussion with Musk, orbital deployments might replace heavy Nvidia systems with photonic compute offering equivalent work at about 1/100 the mass—and he called 1/100 conservative. He placed realization and deployment on a one-to-18-month horizon.
Wissner-Gross added that photons could offer roughly a 1,000-fold clock-rate improvement over “stupidly slow electrons.” The panel also cited a multibillion-dollar SpaceX photonic-computing and communications acquisition as a sign the Starmind architecture may already be turning optical.
Mostaque warned that China’s photonic systems, including the Jiuzhang series, made this precisely the field where US attention was insufficient. The group’s policy conclusion was that core and quantum photonics may deserve higher priority than the conventional quantum program Washington had just elevated.
18. Orexin could become sleep’s version of the GLP-1 trade
Diamandis said natural four-hour sleepers represent no more than roughly 0.1% of people, while about 1% can function on six hours; most need eight. Four-hour sleep would add 28 waking hours weekly, or approximately 58 days a year—effectively two additional months of usable life.
He also presented the deprivation cost: habitual sleep of six hours or less was associated with a 48% increase in coronary-heart-disease risk, 15% higher stroke risk, 12% higher all-cause mortality, 5% more beta amyloid, 17% higher type 2 diabetes risk, and four times the risk of catching a cold.
Eli Lilly’s $6.3 billion acquisition of Synthesa Pharmaceuticals centers on orexin, the brain’s wakefulness switch, initially for narcolepsy. Wissner-Gross compared the strategy to GLP-1 drugs moving from diabetes into a far larger health-span market; Mostaque added that orexin modulation might also reduce inflammation through related ghrelin, leptin, and oxytocin pathways.
The uncertainty matters: the speakers were discussing a possible future lifestyle drug, not an established ability to replace healthy sleep. The participants who addressed it said they would want the capability, and Blundin’s naturally short-sleeping colleague supplied the economic analogy—“It’s like having a whole ’nother life.”
19. UBI needs monetary redesign, not dividends from AI shares
Mostaque rejected the idea that government-held golden shares could simply fund UBI. At a 5% dividend yield, he estimated AI companies would need to be worth about $10 trillion, with government owning roughly half merely to get halfway to a basic-living payout.
Diamandis’s arithmetic was harsher: $3,000 a month for US residents would cost about $12 trillion annually versus a $7.4 trillion federal budget. He nevertheless framed UBI as a “freedom dividend,” closer to Alaska’s Permanent Fund than a trade of autonomy for welfare.
Blundin emphasized the unresolved international problem. If frontier models become the universal workforce while ownership concentrates in the US and China, domestic electoral incentives will distribute gains to voters and ignore outsiders; he argued that a global capital-flow arrangement must be designed within roughly the next year.
20. Geography will split AI companies, compute, and even personhood
Wissner-Gross doubted frontier labs would flee to Argentina: American and Chinese model builders will remain anchored inside their respective blocs. Argentina’s proposed non-human corporate regime could instead attract millions of inference-time AI businesses, though US controls might restrict their access to Chinese base models.
For European founders, Mostaque preferred installing agentic workflows into regulated legacy industries over competing on frontier or vertical models. Europe’s inertia creates room to charge meaningful markups for inevitable modernization. The discussion also cited EU Inc’s under-one-day incorporation proposal against Germany’s process, which can take six months.
Compute will diversify rather than choose one location. Blundin favored low Earth orbit below roughly 500 kilometers because residual atmosphere clears debris; Wissner-Gross favored the moon’s defensibility and silicon, oxygen, nickel, and iron feedstock; Diamandis cited Microsoft’s two-year Project Natick trial as evidence subsea servers can cool efficiently with lower failure rates.
The same heterogeneity applies to people. Wissner-Gross expects ordinary biological humans, Neuralink-enabled collective “borg-anisms,” AI legal persons, and uploaded minds to coexist; Diamandis expects nearer-term orexin therapies, gene edits, higher IQ or muscular capacity, and brain-cloud links to move from moral controversy toward normalization, much as IVF did.