US Government Blocks GPT-5.6, Alibaba's AI Theft, and Why OpenAI Is Stalling Their IPO | #267
In a Nutshell
The US government is now directly gating access to frontier AI models, throttling GPT-5.6 and Mythos releases to select customers while Chinese open-weight models rapidly close the capability gap through distillation and recursive self-improvement. This creates a bifurcated world where US labs maintain internal leads but external users risk falling behind Chinese models, potentially driving companies toward on-prem Chinese alternatives unless countered by licensing, KYC requirements, and bans. Meanwhile, real-time video generation, photonic compute, and orbital data centers are emerging as the next competitive battlegrounds, with China leading in video/world models and the US racing to secure compute infrastructure through space-based and ocean deployments.
These notes were generated by AI and may contain inaccuracies.
For the first time in US history, the executive branch has placed a national security hold on commercial AI products. The models are so insanely capable that they have to be controlled. Certain models will be available to everyone. Certain models won't. And now they're becoming more and more gated. This will only accelerate. The US government is now in the release loop.
Anthropic's Fable and Mythos models were initially pulled from the market. The Trump administration struck a deal with Anthropic granting permission to release Mythos 5 to a group of 100 select companies. Two days ago, just as OpenAI was about to release GPT 5.6, the White House asked them to slow down and only release the model to 20 select companies.
The US government is now in the release loop for the most capable models, customer by customer, selecting who gets access to the latest models. In a recent memo, Altman said the government will be approving access customer by customer during a limited preview window with broader release hopefully coming in a couple of weeks if all goes well.
OpenAI has announced three versions of GPT 5.6, all being throttled by the White House. GPT 5.6 Saul is the flagship model. GPT 5.6 Terra is the middle tier. GPT 5.6 Luna is the fast low-cost version.
Given that the White House is delaying Anthropic and OpenAI frontier models, the government may be effectively stifling domestic AI in a regulatory blanket, potentially crushing OpenAI and Anthropic's valuations.
Dave's position is yes on stifling and no on crushing valuations, but it's inevitable. The models are so insanely capable that they have to be controlled. Cybersecurity is just the first excuse, but all the other evil use cases are right behind that. This is the new normal.
The only argument that it would stifle market caps would be tied to whether China is going to beat these companies. China is nowhere near caught up to the US. They do an incredibly good job of distilling, copying, and taking intellectual property, but in terms of pushing the frontier, there's very little chance that China is going to threaten the market caps of these US companies anytime soon. If the government is fair across the board, it doesn't reduce the market caps at all. These are the most valuable companies in the history of the world by far.
GPT 5.6 Saul is roughly comparable to Mythos preview. On all of the cyber benchmarks and some of the bio benchmarks, it wins some and loses some.
We're in the regulatory endgame. During the founding early days of OpenAI, DeepMind and Anthropic, there was enormous hand-wringing over what would happen when we reached the era of recursive self-improvement and the race condition that would happen. There was a general consensus that the frontier labs would establish some sort of coordination mechanism with each other to slow down and cross the finish line together without creating a race condition.
The same old coordination mechanism that we've had for thousands of years, government, a localized geographic monopoly on force, is more than capable of being that coordination mechanism for getting the final two, at least as of this point in time, the duopoly of Frontier Labs, OpenAI and Anthropic to synchronize the release of their capabilities out to the first few dozen customers or users that the government is going to gatekeep.
The US government has functioned as a synchronization mechanism for helping GPT 5.6 reach essentially some form of par or near parity with Mythos preview 5. We've never seen any sort of third party basically forcing the two leading contenders who would otherwise be in a race against each other to essentially similar capabilities.
Looking at China, there's a chart floating around X extrapolating the time difference between Chinese openweight models reaching the same capability as the two remaining frontier western openweight models. Depending on how you calculate that time delta, the number is on a trajectory to go to zero by Christmas of this year.
GLM 5.2 was the first model with the big model feel even though it just trained more from the GLM 5.1 base. On Frontier SWE, which is the most difficult coding benchmark where each task takes 11 hours and requires novelty, GLM 5.2 is at the top with the right harness. With a normal harness, it's number four or five. GLM 5.2 is maybe $25 million worth of compute. The gap is definitely closing by December, maybe even now as base model performance becomes less important than how you use it.
You can already take GPT 5.5 or Opus 4.8 and put enough of a brilliant harness around it to make it better than Mythos or better than GPT 5.6. Therefore, you can take what's already out and turbocharge it above the level of what the government tried to stop this week. That means the cat's out of the bag. The government would then need to backtrack and say they were too slow and need to lock down 4.8, maybe 4.7, going all the way back six months.
Blitzy can beat Mythos in SWE-Bench 4 or in SWE-Bench Pro.
A harness typically refers to non-weight capability improvements. When building a machine learning model, there are many phases. It typically consists of a neural network composed of weights. The neural network has some intrinsic behavior. It goes through pre-training, mid-training, post-training. These all impact the weights directly.
Now you have a model and you want to on the fly change its behavior. In 2020 when we first got GPT-2 and large language models or few-shot learners, there was the prompt. You could change its behavior in context by feeding it different text as a prompt. Then people realized maybe we want to factor out common elements of that prompt across many prompts and the system prompt was born. Then people realized it was desirable to keep factoring out common elements and logic and text and other information out from all of these prompts into what ultimately became a harness.
It is now quite possible to create lots of software 1.0 harnesses that live outside the model that orchestrate the models, feed common system prompts and other prompts to the models, parse the outputs, and mix different sorts of models from different vendors in order to achieve super performance.
Right now, the government is controlling who gets access to frontier AI level. At the same time, we have these openweight models coming out of China. A lot of companies may say they don't want the government telling them what they can and cannot access and start using on-prem openweight models.
There's a good chance that the US government bans Chinese openweight models from being used by corporations and requires a license and KYC for any frontier or new frontier model, including retention of your prompts.
Mythos is only allowed to be used by US citizens on the allow list. This means Andre Karpathy at Anthropic cannot use Mythos. It actually says that in Lutnick's order. It says only citizens, even internally non-citizens cannot use it.
If frontier models from US labs are only allowed to be used by US citizens and US corporations, that's a massive lead for America. They will continue. That's still a big market.
It looks like protectionism. Superficially it looks like protectionism masquerading as export control. There is a very real risk that the US falls behind. Without this regulatory regime, we were neck-and-neck, maybe six to eight months ahead of the Chinese models. Now there is very much the risk that AGI is achieved internally but externally for all of the users, the American users and western users of American frontier models are stuck at par or worse behind par with Chinese models, whereas internally within the labs the capabilities are continuing to leap ahead.
That distinction creates the risk of some sort of weird perverse incentives for the American economy. For example, are we incentivized to all go work for OpenAI and Anthropic so that we gain access to the internal capabilities? That's happening already. Everybody's leaping.
The playbook if you ban US companies from using Chinese models doesn't achieve anything because the rest of the world will still use the Chinese models. The playbook next is to say we will not let you access the most amazing technology in the history of mankind, Europe, South America, East Asia, unless you follow this new playbook. That's what David Sachs is busy writing right now.
One could imagine the Pax Silicon generalizes to the Pax Intelligencia where you have the American intelligence super intelligence block and you have the Chinese super intelligence block and Europe is torn between them.
If you have any licensing restrictions, one scenario is just like you have a driving license, you will need to have a
Frontier AI models may require special licensing similar to a driving license, where users must demonstrate patriotism to America and pass KYC requirements to access advanced capabilities. The concern centers on preventing adversaries from obtaining high-capability models that can operate autonomously for days or months rather than providing simple query-response interactions.
Anthropic's Mythos model, running red team exercises under Project Glass Wing with US intelligence agencies, identified vulnerabilities in highly sensitive classified US government computer systems. Senator Mark Warner stated the tool broke into almost all classified systems in hours rather than weeks, with exploitation methods outside the exercise scope.
Twelve days later, the Trump administration directed Anthropic to disable Mythos 5 and Fable 5 for foreign nationals. The exercise with the Department of Defense explains the decision, suggesting a future where advanced models test sensitive systems, patch vulnerabilities, then become available.
Mythos 4.8 can answer queries about building itself, enabling users to create competing equivalents in China, while Opus 4.8 cannot. This self-improvement capability represents the real issue behind the restrictions, not the cybersecurity exercise which served as justification. Governments often have pre-existing agendas and use trigger events to justify actions they planned to take anyway.
China has sufficient capabilities to achieve recursive self-improvement independently without needing to siphon trade secrets or reasoning traces from Western models. The policy only makes sense if the West is in the endgame of recursive self-improvement. Alex argues China has reached recursive self-improvement escape velocity and doesn't require Western assistance.
The White House may not share this assessment and believes they still have time to act.
Regulators should unleash David Sacks to meet world leaders with specific frameworks for how the future world will work. Superintelligent AI should be locked into multiple boxes rather than one or ten. The most critical requirement is logging every prompt and having multiple eyeballs inspect every use case, chain of thought, and iteration through other AIs.
Access to Fable would be granted only after agreeing to rules about who sees the logs and establishing friend versus non-friend distinctions.
A licensing regime similar to driver's licenses or security clearances will likely emerge for American civilians accessing frontier models. David Sacks has left government, leaving AI policy driven by Howard Lutnik, whose focus is America's economic interests. The government must protect US interests, prevent others from accessing capabilities, while ensuring Anthropic and OpenAI don't collapse. The current situation involves determining access requirements while frontier capabilities emerge across cybersecurity and biological domains.
This represents a Sputnik moment where the surprise came from within—the NSA and Cyber Command were surprised by the private sector leapfrogging vulnerability and cyber capabilities previously controlled by government agencies.
If the US should maintain strong immigration policy for human capital, similar thinking applies to importing foreign AI as capabilities meet and exceed human intelligence. If open-source trends continue and Chinese capabilities match or surpass US frontier models, the US may import more foreign artificial intelligence than foreign human intelligence.
GPT-5.5 Cyber, code-named Daybreak, scored a record 85.6 on the Cyber Gym benchmark testing AI agents for real-world cybersecurity vulnerabilities—the highest single model score posted. Sam Altman stated the real prize is automatically writing and testing fixes across web browsers down to the Linux kernel, turning threats into cures.
The question becomes who is allowed to merge AI-written fixes into critical infrastructure codebases. Whoever controls that trust layer and liability controls digital security across everything. Non-profit initiatives including one run by IBM are using AI to bulk identify and remediate vulnerabilities across open-source projects.
OpenAI is intentionally biasing models toward defense versus offense through self-disruption or poisoning mechanisms for offensive cyber use cases. Alignment comes through model design combined with scaffolding, system prompting, and post-training that favors defenders while obfuscating, foiling, and poisoning attacker use cases.
Anthropic's weights loaded onto Colossus create high-value espionage targets. A few models are more valuable than a million agents for cyber attacks. Base weights obtained before system prompting are more creative and less hobbled for attack purposes. Infrastructure is currently inadequate even for basic attacks before AI capabilities emerged.
A massive push is needed to harden essential government and private apparatus using these models. Claims may emerge that Chinese models introduce backdoors in code, particularly since few people review every line of merged code.
The US government may require ISO-type certified models for merging code, as uncertified models could introduce undetected holes. Chinese models could trivially inject spyware through a single line of code. Only AI can keep up with AI in cybersecurity. An ISO-certified trustworthy AI stamp of approval is imminent, raising questions about which entity provides that certification—US, US plus Europe, or a new NATO-type entity.
Questions have emerged about whether American AI installs backdoors. The concept of privacy is largely illusory given existing surveillance capabilities. The assumption is that governments can read texts, emails, and monitor all communications.
OpenAI leadership is pulling back from near-term IPO plans. Advisers presented two paths: go public this year potentially accepting sub-$1 trillion valuation, or wait until 2027 to scale revenue, infrastructure, and partnerships while preserving the $1 trillion IPO narrative. Altman is not interested in going public below a trillion dollars, partly due to watching SpaceX stock volatility.
Staying private avoids quarterly market pressures while burning staggering capital on compute. Public markets lack patience for capital spending timelines required to achieve AGI and ASI. However, SpaceX's IPO at $135 opened at $150 and trades at $153, representing a perfectly priced offering rather than volatility warranting delay.
OpenAI raised $122 billion recently and doesn't need the money. Sam Altman has at least 400 outside investments that would require disclosure in SEC forms as potential conflicts. Running a public company involves substantial compliance burden that may deter the transition.
OpenAI focused too long on consumer revenue rather than enterprise, a bet that proved incorrect. They are now trying to become Anthropic faster than Anthropic can become OpenAI. Codex revenue is skyrocketing, potentially serving as the revenue engine for a trillion-dollar-plus IPO.
Silicon Valley consensus holds that hard takeoff is happening now, with discoveries in physics, medicine, and governance changes coming. Enormous wealth will be created from scientific breakthroughs. Elon Musk lost edge on frontier models and is pursuing relevance through space and orbital data centers, requiring massive compute infrastructure.
Pre-revenue, pre-profit companies going public in biotech during 2021-2023 were decimated. The rule is to go public with profits and predictable revenues. However, frontier AI companies can manufacture substantial wealth and revenue quickly with compute access. Revenue visibility and CFO concerns are not the primary decision factors. The belief is the world a year from now will look fundamentally different, making SEC processes and road shows a waste of time.
Staying private allows companies to avoid time-consuming SEC filings and roadshows during the singularity. The regulatory overhead feels outdated and misaligned with Silicon Valley's pace. Once you begin filling out SEC documents, the process feels like a massive waste of time rooted in 1929-era requirements.
Anthropic's decision on going public depends on whether their revenue continues ramping without a drop. Given their ideological stance and unusual PBC structure, they may prefer avoiding external fiduciary control. A potential structural workaround could involve acquiring Sierra and installing Brett Taylor as CEO while moving Sam Altman to president.
SpaceX's AI revenue from their cloud business has overtaken AWS and GCP on a run rate within just a few months. OpenAI has reached $40-50 billion in revenue, with growth matching their valuation. OpenAI is forecasted to lose $26 billion this year while having raised $120 billion, providing substantial runway. Anthropic sits closer to their raise-to-balance ratio, potentially requiring either another raise or an IPO.
Employee shares and invested options provide secondary liquidity, reducing pressure for an IPO. Companies do not need public markets for personal liquidity events like buying houses or cars. The regulatory burden post-IPO includes stock lawsuits if share prices decline, plus FD approval requirements for any public statements or social media posts.
Fountain Life's member database shows 3.3% of people who believe they are healthy actually have undetected cancer. Most cancers detected at late stages are the ones that prove fatal, whereas early detection significantly improves cure rates and treatment ease. Full body MRI combined with early cancer detection screening identifies cancers that conventional prevention approaches miss.
These advanced screening methods are not typically covered by insurance and are not standard in conventional care settings. The goal is to collect data, conduct research, and democratize wellness access.
Elon Musk announced that Neuralink may attempt the first direct human-to-human telepathic communications later this year, transmitting thoughts directly between brains without typing or speaking. This represents a new level of intimacy and communications capability for the human species.
Elon's endgame for Neuralink extends beyond medical devices toward creating an IO layer for the singularity, enabling humans to couple with AI during the transition.
Neuralink would represent one of the first open attempts to create superhuman capabilities rather than merely restoring function for those with motor disabilities. A recent Cell paper found that bilingual adults' hippocampus functions similarly to a vector embedding space in an encoder-only transformer model, with neurons for related concepts positioned adjacent to each other.
This geometry suggests telepathy may be easier than previously suspected, as human cognitive capabilities may be less complex than assumed and represent distorted reflections of ancestral environments.
Neuralink requires inserting multiple needles into the human skull. If the embedding theory of neuron placement generalizes, enormous demand could emerge for non-invasive human-to-human telepathy options. Neuralink may develop less invasive BCI alternatives in parallel.
Potential outcomes include couples connecting neurally, potentially affecting divorce rates, or emergence of borgganisms.
Human input/output rates are extremely narrow: speech operates at 40-60 bits per second, typing at 5-20 bits per second, and conscious thought and behavior selection at 10 bits per second. Processing occurs at roughly 100 tokens per second, but pattern recognition like spotting a tiger happens instantly through latent space adaptation.
The 2023 Stability neuro team paper called Mind's Eye used fMRI scans of people viewing bottles, then reconstructed the images using Stable Diffusion, indicating human latent spaces align with AI model architectures. Telepathic communication would require minimal words to activate shared latent space, similar to finishing each other's sentences when in sync.
Bandwidth could increase 10-100,000 times by adapting adapters to match individual latent spaces. This capability could enable deeper exploration of what makes humans human, though risks include negative outcomes alongside positive ones.
Elon Musk expanded SpaceX's naming protocol with star-based ventures. Starmind refers to the AI satellite constellation, Starfall handles cargo delivery moving atoms, and Starlink moves bits. Musk stated they are cutting back on the star prefix because it is getting silly.
Starlink is developing direct-to-cell mobile service. Bloomberg reported a potential deal with Charter, with additional rumors suggesting possible acquisition of T-Mobile.
Starship provides propulsive landing and heavy lift capabilities. Starbase in Texas houses the Star Factory producing Starships. Star Shield supplies a defense-grade version of Starlink for the US Department of Defense.
Starfall enables private companies to launch cargo to LEO with retrieval capability. A major challenge has been down mass from orbit for materials and biology experiments. Outpost, founded by Singularity alum Jason Dunn, has been developing orbital manufacturing hardware including their Carry All product.
An orbital brewing company launched yeast into orbit via Starfall, retrieved it, and plans to brew LEO beer.
Stargaze leverages SpaceX satellites for situational awareness looking both up and down, potentially competing with Planet in orbital and ground monitoring. Star Mind serves as the brand name for the AI orbital constellation, described as SpaceX's Dyson Swarm.
Star Pipe represents SpaceX's entry into oil and gas infrastructure for piping natural gas and methane, supporting methox propulsion. Starbase's Texas location provides access to oil wells. The initiative addresses future needs for oil and gas capabilities in Martian and lunar colonies, including data center powering.
ByteDance released Cance 2.5 in beta with full release expected in July. The system generates 30-second videos at 4K resolution, references up to 50 different inputs including images, video, and audio, provides director cinematographic control options for post-production, and supports editing via text prompts.
Elon Musk stated that full-length motion pictures will emerge from AI by the end of the year, suggesting Hollywood is finished.
Cance 2.0 represented a major advance, with competitors catching up. The 50-input capability enables near-perfect pixel control across multiple characters using video, audio, and images. Hollywood studios benefit from reduced costs, but downstream impacts include displacement of media production jobs.
The technology enables telling previously impossible stories. Using these models collaboratively rather than as single-player experiences proves more rewarding, as different perspectives shape narratives.
The Future Vision X-Prize, a partnership with Google backed by Arc Invest and Marc Benioff, involves the Rodenberry family. Filmmakers creating three-minute trailers showing hopeful, compelling visions of the future using this technology will have their full-length motion pictures produced. The initiative aims to steer humanity toward positive Star Trek futures rather than dystopian scenarios.
China leads in video generation. American tools like Gemini Omni remain limited to approximately 10 seconds, while Grok attempts to compete. Chinese frontier labs benefit from cheaper, less encumbered training data from TikTok and similar sources without copyright infringement concerns.
American labs prioritize recursively self-improving codegen models due to higher revenue generation compared to video models, which primarily serve consumers with lower spending power.
Liquid AI is building foundation models from scratch without using Anthropic or OpenAI technology. Their approach uses context far more efficiently than transformer attention windows. Mercedes vehicles will use Liquid AI for on-prem vehicle conversations starting in September, requiring small models that function without internet connectivity.
Edge computing requirements for autonomous vehicles demand systems that function without internet connectivity and fit within vehicle power supplies. Video generation models have demonstrated the ability to generate images or videos in real-time as quickly as speech occurs. High-quality video generation faces latency challenges that disrupt creative workflows, with waits of up to three minutes breaking the creative cycle. This latency becomes particularly problematic for interactive applications compared to traditional movie production.
The video game industry represents a significantly larger market than Hollywood, generating several hundred billion dollars compared to Hollywood's approximately 50 billion dollars. Real-time video generation embedded in interactive gaming environments represents a major commercial opportunity. Current frontier labs focus primarily on coding and white-collar automation, but solving real-time video game applications could unlock substantial value. Latency remains the primary barrier to adoption.
Chinese companies demonstrate advantages in world model development for real-time interactivity. One path to AGI involves recursive self-learning on code generation, with models like C Dance 2.5 showing clear physics embeddings. Companies including ByteDance are developing early world models enabling real-time interactivity. Two distinct routes to AGI appear to be emerging: code-based recursive self-improvement versus world model approaches.
Video games currently generate several hundred billion dollars and could expand three to ten times with compelling real-time generative capabilities. The market will expand beyond its current demographic dominance of males under age 30 to reach broader populations across all age brackets.
Enterprise software represents a 30 trillion dollar market, orders of magnitude larger than video games' trillion dollar market. Alibaba's One Streamer demonstration shows real-time interactive generative video-to-video capabilities, enabling generative environments where participants can interact by placing hands in video or removing them. This represents minimal world modeling operating in real-time and interactively.
Western companies could capture enterprise opportunities through real-time Zoom or FaceTime participants capable of audio and video interaction in company meetings. This capability could expand the addressable market from one trillion to 30 trillion dollars. Video game companies possess technology that could pivot to enterprise applications, with Fortnite-style platforms potentially adding real-time video generation for enterprise use cases like Slack.
Anthropic accused Alibaba of operating the largest AI model theft campaign ever documented, utilizing 28.8 million fraudulent exchanges across 25,000 fake accounts to extract and copy Claude's capabilities. Distillation involves using one model to answer questions for training another model through a teacher-student approach. This represents a frontline issue in US-China AI rivalry.
Anthropic has faced multiple lawsuits for using copyrighted material in pre-training, creating irony given their current accusations against Alibaba for using Anthropic outputs in post-training. Export control regimes function as protectionist measures preventing US frontier model insights from leaking through proxies that offer access at 90% discounts. These proxies reportedly gather all reasoning traces, effectively using users as sock puppets for distillation efforts.
Distillation ranges from direct latent approaches to question-answering for complex codebases. The latter proves extremely difficult to prevent. Chinese companies substitute models like Mythos or Claude for expert input to improve data distribution quality. GLM 5.2 demonstrates improvement over GLM 5.1 through recursive self-improvement loops, using thorough data improvement cycles on initial distributions.
Chinese models will improve with either super expert input or Mythos-style question answering. However, recursive loops may eliminate dependency on external data entirely. This represents a takeoff scenario where existing Chinese datasets and techniques may not sustain frontier capability competition if US model access ends.
China's economy has grown through copying and appropriating ideas from elsewhere, with this approach embedded in cultural practices. The 25,000 fake accounts accessing reasoning traces align with documented patterns rather than representing isolated incidents. This case will likely trigger US, European, and South American efforts to suppress Chinese AI capabilities.
Early Singularity University discussions with Steve Jurvetson and Astro Teller addressed IP protection, concluding that IP becomes meaningless when AI can iteratively reinvent products faster and better than original creators. Alternatively, AI could supercharge IP litigation lawyers, enabling superhuman protection of intellectual property. The next 18 months will generate more intellectual property than all previous human history combined, primarily in software, video generation, physics, and mathematics.
AI creative and copying capabilities enable one-shot recreation with remixing within coming years. Sonnet's competence threshold broke last year, and current models across open and closed systems have reached this capability level. New recursive loops reduce the necessity of copying when recreation and remixing become feasible.
Copyright protections vary significantly by domain, with music maintaining strong protections while software receives minimal protection. Parallels exist between defense-offense dynamics in software vulnerabilities and IP litigation, where AI strengthens both offensive circumvention and defensive patent drafting and litigation capabilities.
President Trump signed executive orders supercharging US quantum computing companies and directing intelligence agencies to protect quantum research equivalent to nuclear secrets. The US committed 2 billion dollars in venture investments through the CHIPS and Science Act through May 2026.
IBM received 1 billion dollars for co-developing their Anderon quantum chip foundry in Albany, New York. D-Wave, Regetti, and Inflection each received 100 million dollars, while Sai Quantum secured 140 million dollars in equity stake. One executive order mandates establishment of QC ADS (Quantum Computer for Application Development and Discovery Science).
The Quantum Information Science and Technology Counter Intelligence Protection Team exists within US government infrastructure. The executive orders aim to prevent being caught flatfooted by quantum capabilities as occurred with AI regarding NSA cyber and quantum capabilities.
Quantum computing was expected to solve protein folding, but AlphaFold 3 achieved this classically without quantum computing. Sandbox AQ utilizes quantum physics equations on AI platforms rather than quantum computing itself. Quantum computers prove vastly more complex to build than GPUs, with China advancing photonic quantum computers but lacking capabilities in other quantum computing types.
Super Mythos-level models could enable proper programming of quantum computers, which remains difficult. Quantum equations and general equations share similarities for interesting reasons. Quantum computers could answer questions in microseconds rather than requiring day or year-long processes.
Convergence between improved question-asking capabilities and quantum supremacy could eliminate the need for Dyson spheres by breaking the energy-compute-intelligence scaling requirement. The default trajectory involves everything converting to compute, but frontier capabilities could create discontinuities within coming years.
Existing quantum advantage algorithms lack economic transformative utility despite generating headlines and IPOs. AI may discover quantum-advantaged training or inference algorithms in mathematical theory space, potentially justifying previous quantum infrastructure investments. Municipal resistance to terrestrial data centers pushes computing infrastructure to orbit.
Quantum computers could be constructed in permanently shadowed lunar craters. The goal is not indefinite American advantage preservation but rather slowing competitors sufficiently to achieve recursive self-improvement dominance within months to a year ahead of competition.
Quantum.AI was founded specifically to pursue quantum photonics rather than traditional quantum computing. Quantum sensing is expected to enable storage of massive amounts of information in extremely small spaces. Highly quantized neural nets can perform comparably to float32 neural nets, which opens the door to massive photonic computation efficiency gains.
By December, photonic compute is expected to reach deployment stage, with the Terrafab beginning production. Photonic systems are projected to deliver the same computation at approximately 1/100th the mass of Nvidia chips, representing a conservative estimate. This represents the stepping stone to the discontinuity discussed earlier, using quantum photonics rather than conventional quantum approaches.
Intel is working on interesting developments in this area, while China is heavily focused on its Xang series and related photonic technologies. The US is considered significantly behind in this domain, making photonics and quantum photonics an area requiring substantially greater focus and investment.
The White House should elevate photonics as a priority on par with or exceeding the recently passed quantum initiative. In the past two days, Elon Musk and SpaceX acquired a photonic computing and communication company for several billion dollars to integrate into their operations. Photonics could provide up to a thousand times clock rate speedup over electron-based systems and may become central to Starlink or Star Mind plans.
Google received a patent last year for performing matrix multiplications in pure light, having filed it in 2023. With SpaceX going public and providing liquid stock, Elon Musk is positioned for an acquisition rampage across multiple companies.
Eight hours of sleep is evolutionarily required for most humans. Short sleepers who function on four hours represent 0.1% or less of the population, while those managing on six hours represent approximately 1%.
Regular sleep of six hours or less correlates with a 48% increased risk of coronary heart disease, 15% increased stroke risk, 12% increase in all-cause mortality, 5% increase in beta amyloid associated with neurodegenerative disease, 17% higher risk of type 2 diabetes, and 4x higher risk of catching a cold.
Functioning on four hours of sleep would yield 28 additional hours of work or play time per week, equivalent to 58 extra days per year or two additional months of productive time annually. This represents incremental longevity gains distributed throughout life rather than added at the end.
Eli Lilly acquired Centessa Pharmaceuticals for $6.3 billion. The company's drug targets orexin, the brain's master on-off switch for wakefulness. Initially developed for narcolepsy, this is positioned to become a lifestyle drug. The GLP-1 receptor agonist precedent shows how a pathway discovered in Gila monsters for diabetes treatment generated revenue comparable to or exceeding OpenAI and Anthropic's token revenue, contributing to Eli Lilly becoming a trillion-dollar company.
Applying the same receptor agonist playbook to neuroscience, orexin would play the role GLP-1 played in metabolic health. Potential applications include helping individuals with daytime sleepiness from Parkinson's, Alzheimer's, and mood disorders, as well as assisting recovery from comas. Orexin pathways are essential to multiple neurological functions.
One portfolio company has demonstrated double the performance of the Centessa drug. If everyone gained four additional hours of wakefulness, the economic output unleashed would be substantial, particularly if available during peak productive years in the early twenties.
Natural four-hour sleepers exist along a bell curve extending to two-hour sleepers who remain fully functional, though this is extremely rare. Some individuals require up to 14 hours of sleep. Upregulating orexin affects inflammatory markers, ghrelin, and leptin. The pathway is expected to reduce inflammation alongside enabling reduced sleep, similar to how GLP-1 impacts multiple homeostatic mechanisms.
Universal Basic Income funded through dividends from US government-held golden shares is mathematically impossible. AI companies would need to reach approximately $10 trillion in value at a 5% dividend yield, requiring ownership of roughly half of them to approach meaningful UBI levels. Such companies would need to become the largest in the world, and distribution would be limited to US citizens.
A $3,000 monthly UBI for US citizens alone would require $12 trillion annually, compared to the current US budget of $7.4 trillion. Capital flows and monetary equivalents of UBI, where money is created directly, represent more viable approaches than traditional dividend models.
Frontier model companies are expected to remain in the US, with Chinese companies remaining in China. Argentina's proposed nonhuman AI corporation regime is unlikely to attract frontier model development but could host inference-time AI person companies operating without human employees. Many millions or billions of such non-human AI companies could theoretically migrate there.
Export controls could extend to restrict Argentina's ability to import Chinese models, consistent with existing non-AI technology regimes that have limited third countries' access to Chinese technology or commodities.
The wealth concentration effect from frontier models becoming the universal workforce will be extreme, with AI development concentrated in the US and China. This requires policy resolution within the current administration, meaning within approximately the next year.
UBI should be understood as a foundation for autonomy rather than a trade for it, functioning as a freedom dividend similar to the Alaska permanent fund. Recipients could use distributions to uplevel their lives, start companies, or create personal meaning.
Low Earth orbit below 500 km altitude benefits from atmospheric cleaning of debris, but beyond this altitude the Kessler effect creates unacceptable debris collision risks for data centers. The moon offers protection without atmosphere but enables defense of physical assets against orbital interference.
China has publicly demonstrated robotic on-orbit devices with grappling arms capable of interfering with satellites. Lunar surface data centers can be physically defended and secured against beam interception. The moon provides feedstock including silicon, oxygen, nickel, and iron from lunar regolith for constructing additional data centers, which cannot be sourced in orbit.
Mass drivers could launch lunar materials into low Earth orbit. The moon represents the ultimate high ground for both potential weapons applications and data center deployment serving Earth. Gerard K. O'Neill at Princeton developed mass driver technology and outlined architectures for lunar mining to build solar power satellites, a vision now adapted toward space-based intelligence infrastructure.
Humans are already in the age of the singularity. Post-singularity, multiple forms of personhood and physical embodiment will coexist. Uploaded humans may exist as collections of bits or qubits with minimal physical form. Collective human intelligences may emerge through Neuralink or other interfaces.
New physics and applied physics will provide novel compute substrates. Systems will continue to ephemeralize. In the distant future, humans might operate on gravitational fields or approximate pure energy states without biological bodies. These transformations will be optional.
The future will be heterogeneous rather than homogeneous. One hundred years from now, some humans will appear substantially unchanged while others exist as uploads in quantum computers, with these forms coexisting.
Gene editing, IQ-enhancing injections, muscular enhancement, and brain-computer interfaces connecting the neocortex to the cloud will become available within years to a decade. Enhanced capabilities will appear alongside baseline human forms.
Municipalities are increasingly resistant to hosting data centers locally. The discussion explores whether biological enhancements face similar regulatory resistance, potentially pushing transformative physical and behavioral human changes into a posthuman realm as societies regulate out short-term enhancements. Historical precedent shows attitudes shift over time: in vitro fertilization was initially viewed as immoral and contrary to religious beliefs, yet IVF is now considered normal, accepted, and beneficial for couples having children later in life.
Initial resistance may come from religious communities, but eventual acceptance is expected. Genetic edits for BCI capability (neural lace) and other enhancements may initially go underground before becoming mainstream. The desire for optimal outcomes for oneself and children—including best genetics, education, food, and technology—will drive adoption despite sensitivity around the topic.
Immad recently spoke at the Oxford Union debating whether AI can attain personhood. The topic represents one of many significant ethical and philosophical questions society will need to address as AI capabilities advance. The perspective offered suggests humans are all made of star stuff, though some may ultimately become more enhanced versions than others.
For entrepreneurs outside the US and China, verticalized AI startups face significant barriers to competing on frontier models. However, agentic workflows applied to legacy industries present substantial opportunity, particularly in Europe where institutional inertia creates competitive lag. Silicon Valley sees rapid innovation diffusion due to corporate openness, while Europe remains further behind but faces inevitable competitive pressure to adopt.
Transforming existing companies through AI implementation offers higher leverage than building vertical AI startups in heavily regulated environments. This approach allows charging premium markups while working within regulatory comfort zones. The US remains the superior location for globally competitive verticalized AI startups.
Europe is described as slowly waking up to AI developments, though institutional inertia remains substantial. The EU AI Act faces potential repeal as decision-makers respond to competitive pressures. EU 2031 is referenced as a warning call, giving Europe an extra four years compared to AI 2027 timelines due to slower regional pace. Concerns exist that Europe's regulation could become a self-imposed filter limiting competitiveness.
The fable (likely referring to a competitive AI development) created urgency among leadership circles. New mechanisms are being explored as alternatives to existing frameworks. Ireland is identified as a potential natural leader in this space, alongside Estonia. The EU Inc. initiative allows company formation in under a day, a significant improvement over Germany's six-month process for tax status and related requirements.
Microsoft's Project Natick demonstrated ocean-based data centers operating successfully for two years with lower failure rates than land-based servers and excellent cooling performance. Ocean deployment advantages include avoiding launch costs, abundant cooling, proximity for repairs, and jurisdictional flexibility (within or outside territorial waters).
Orbital data centers are positioned at approximately 500 miles in low Earth orbit, not 93,000 miles. The moon sits at 240,000 miles. Launch costs have dropped significantly, with Elon Musk reportedly shifting perspective on orbital data centers around nine months prior. Will Marshall, Eric Schmidt, and Google teams discussed orbital compute concepts approximately a decade ago.
Both ocean and orbital data centers will coexist alongside land-based facilities. Demand exceeds current supply capacity, eliminating near-term infrastructure limitations.
Peter Thiel funded Panthalassa, focused on ocean-based infrastructure. Data centers on oceans serve as prototypes for seasteads and ocean colonization efforts expected within the next decade. While ocean-based compute may be less scalable than orbital systems, local compute becomes essential for ocean colonies. Panthalassa video demonstrations suggest these installations could nucleate next-generation ocean settlements.
If photonic takeoff occurs on a one-to-two-year timeline, NVIDIA data centers will likely deploy in oceans rather than launch due to mass constraints. An NV72 system with power weighs several tons, making ocean deployment more economical during the transition period. Photonic systems offer dramatically reduced mass per unit of compute—approximately one-tenth to one-thousandth the mass—making orbital deployment more feasible for those systems.
Intelligent Internet is releasing a harness allowing users to uplift models above Methos on the following Monday, accessible via Twitter. Sovereign AI work continues with mechanisms enabling collective ownership of AI and robots. A new book titled "After the Last Economy" is forthcoming alongside additional announcements.
Keep Peter H. Diamandis in your library
Save the videos and channels worth coming back to, and find them again in one place.
