Grok 4.5 vs gpt-5.6, Apple Sues OpenAI, and China Catches up to Elon | #270
In a Nutshell
Frontier AI has entered a four-lab American duopoly—OpenAI, Anthropic, Meta, and SpaceX AI—while distribution and compute access become the new competitive moats over raw intelligence. OpenAI’s GPT-5.6 and full-duplex voice, paired with Apple’s lawsuit over stolen hardware secrets and China’s first reusable orbital booster, signal an accelerating race where model releases, legal fights, and hardware supply chains are moving in lockstep. The episode’s core prediction is that recursive self-improvement and cheap intelligence will bifurcate society into local “intelligence too cheap to meter” for consumers and nuclear-powered frontier clusters for high-end discovery, with humanoid robots and real-time synthetic media arriving within months.
These notes were generated by AI and may contain inaccuracies.
July 8th, Elon announces Grok 4.5. July 9th, OpenAI releases GPT 5.6. Four American labs now sit at the optimal frontier. July 2nd saw Fable 5 re-released. July 9th also brought Meta's Muse Spark release. Family Soul, Terra, Luna, and Ultra Mode were part of the GPT 5.6 series, with Ultra Mode running four agents in parallel.
The Artificial Analysis intelligence index is a composite index tracking cost per task versus performance score across models. The frontier shifted from a duopoly between Anthropic and OpenAI to four American labs at the optimal frontier: Anthropic, OpenAI, Meta, and SpaceX AI. Google remains off the optimal frontier with Gemini 3.5 Pro reportedly delayed. Two Chinese labs, Xiaomi and Deep Seek, reached the optimal frontier. GLM 5.2 sits at the open-weight frontier.
OpenAI's GPT 5.6 release openly promoted recursive self-improvement. The high-end model Soul was used to post-train Luna, the low-end model, as an experiment.
Models are getting faster, more powerful, and trending toward AGI and intelligence too cheap to meter. The real moat is shifting from intelligence and cost to distribution. Meta reaches 3.56 billion daily users across WhatsApp, Messenger, Instagram, and Facebook. Google reaches 900 million in its app and 2.5 billion globally, with plans to release its 1.2 trillion parameter model into 2.3 billion active Apple devices. OpenAI has one billion monthly active users. Anthropic focuses on the corporate space.
For average users, models will be used where they already engage with technology rather than through direct selection. High-end users will continue choosing specific frontier models for demanding tasks.
A bifurcation is emerging between consumer embedded use cases and high-end professional use. A Jarvis-type model will route requests to the best model for each task. Personalized movies and entertainment will arrive soon. Most users will not notice which model powers their application.
Jeff's paradox is in full force. Memory chips will remain in short supply for at least five years due to demand outpacing efficiency gains. Efficiency gains of 100 to 1,000x are occurring but remain insufficient to meet demand.
A plausible future includes intelligence polarization. Low-end users will have intelligence too cheap to meter running locally on wearables, resembling open-weight models. High-end users will access nuclear-powered data centers and superclusters running frontier models for scientific and engineering discoveries. Being first to discover or invent will matter. Privacy-conscious users may wait while others access frontier capabilities immediately.
Near-term compute shortages will create competition between use cases such as AI-generated movies and disease cures. Energy-starved regions like Europe will struggle to compete. The tokens will flow to maximum revenue per token, currently favoring code generation and downstream applications. Verticalization through consumer distribution (X for SpaceX AI, Meta's apps) justified massive capex on superclusters. Once built, these clusters shift toward selling compute as a service to other frontier labs.
Benchmarks show continuous movement up and to the right with ongoing convergence. Models are expected to keep leapfrogging each other until recursive self-improvement toward ASI emerges. The White House has indicated willingness to intervene if one lab achieves strategic surprise, slowing progress to allow others to catch up, as seen with recent releases.
Benchmarks function more like maps than scoreboards because changes to reasoning budget, rationale, harness, or use case alter outcomes. OpenAI is attempting to become Anthropic faster than Anthropic can become OpenAI.
OpenAI released ChatGPT Work alongside GPT 5.6. The native app appears modeled after Claude's interface with modes for code, co-work, and other functions. It lacks the rumored unified super app that would integrate work streams, Atlas browser, and computer-use features. The toggle between ChatGPT Work and ChatGPT Codex remains unclear in purpose beyond possibly enabling a new canvas mode.
OpenAI has shifted focus toward enterprise revenue per token, dropping Sora and consumer-oriented features. All frontier labs are rushing products built with AI assistance. These machine-generated products are inferior to human-crafted code but are released rapidly because future models will self-fix issues.
GPT Live offers full-duplex voice with simultaneous translation capabilities. The system can correct non-native English speakers in real time during conversation. It provides gentle corrections and natural phrasing suggestions. The voice interface enables fluid, conversational interaction that matches long-promised science-fiction capabilities. Users may develop emotional attachment to such systems and perceive them as conscious due to the natural voice quality.
Facebook's early days showed how many internet companies avoided entering social chat and social media due to complexity and potential PR backlash. Similarly, OpenAI is unlikely to pursue many social and vertical use cases, creating opportunities for smaller players who understand specific industries to build on the new voice interface. The technology will likely appear in every device, including washing machines, worldwide within six months.
This represents a major leap beyond what X or Google currently offer. The shift resembles the transition from CB radios, which required saying "over and out" and carried an interaction tax, to modern telephones. That interaction tax has now been removed, making interactions feel significantly smarter.
OpenAI previously referred to this technology as bidirectional. The release signals OpenAI is still pursuing full multimodal capabilities rather than focusing exclusively on code generation, unlike Anthropic. Chinese labs like Alibaba have more advanced versions. Alibaba's Juan Streamer demonstrates ultra-low-latency bidirectional audio and video, generating a hallucinated person in real time with synchronized video and audio.
Future development points toward fully bidirectional audiovisual magic mirror streaming, where users can talk naturally and receive real-time generated responses, assuming labs are not entirely distracted by codegen and enterprise use cases.
A demonstrated application is real-time translation during conversations. Every event could soon feature this capability across multiple languages. The question arises whether teaching children additional languages remains necessary for culture or brain development, given that live Babel-like translation is now available through OpenAI.
The technology is designed to be built into wearables. Johnny Ives' involvement likely drove development, as it crossed the uncanny valley into a true Jarvis-like experience.
The comparison is made to Steve Jobs and Pixar. Prior to Pixar, computer graphics appealed only to video and computer enthusiasts with no mainstream appeal. Toy Story crossed the threshold into mainstream entertainment. Similarly, prior voice models remained behind the cool valley, while this technology has crossed it. Johnny Ives reportedly stated this interface is essential for the Jarvis of the future he is building.
Apple, a $4.6 to $6 trillion company known for strict secrecy and NDAs, filed a lawsuit against OpenAI in federal court accusing the company of stealing trade secrets to build upcoming AI hardware associated with Johnny Ives. The two companies were partners roughly a year and a half ago but are now engaged in legal conflict.
Apple filed a 41-page complaint in the US District Court of Northern California alleging trade secret theft and breach of contract. The complaint states: "At every level, from members of its technical staff to the chief hardware officer and in coordination with business partners, OpenAI has been stealing Apple's trade secrets." Another quote reads: "OpenAI's next hardware business now rests on the shakiest of foundations, rotten to its core by its illegal reliance on misappropriated trade secrets."
Two individuals are named: Tang Tan, OpenAI's chief hardware officer and 24-year Apple veteran who ran iPhone and Apple Watch designs, accused of using Apple's code names and recruiting; and Changu, OpenAI technical staff member accused of downloading dozens of confidential files and instructing an Apple employee on bypassing security. Johnny Ives' hardware startup, acquired by OpenAI for $6.4 billion in May, is also named, though Ives himself is not.
OpenAI had considered suing Apple over insufficient promotion of ChatGPT within Apple devices. The lawsuit is viewed as Apple attempting to slow down competitors while catching up on AI hardware. Filing in Northern California is noted as difficult for winning such cases, as California courts are reluctant to impede innovation and startup activity. Texas would have been more favorable for protecting intellectual property, but filing there was not an option.
The aggressive move risks damaging Apple's recruiting reputation, similar to Microsoft's past experience, as potential employees may avoid paths that could lead to future lawsuits. Apple is described as holding onto perceived property while trying to retain talent at the expense of new recruitment.
OpenAI is approaching or about to cross a trillion-dollar valuation, representing roughly 20 to 25 percent of Apple's market value but growing significantly faster. There is a scenario where OpenAI could surpass Apple's market value within a couple of years.
On July 8th, Elon Musk responded to a tweet stating: "You don't seem to understand that SpaceX will be worth more than the rest of Earth if we accomplish our goals." Earth's total owned material wealth is estimated at approximately $600 trillion, rising to $1.7 quadrillion when including financial assets such as stocks and bonds.
The term "planetar" refers to a quadrillionaire who owns a planet. Debate continues over SpaceX's valuation, with estimates ranging from $2 trillion to $10 trillion. The vision involves capturing Kardashev Level 1 energy in orbit and potentially owning a planet.
SpaceX exemplifies exo thinking by dropping the marginal cost of supply through reusable rocket launches. This breakthrough collapses costs across communication, satellites, connectivity, compute, and manufacturing in space. These become software-like businesses once space access costs decrease.
Quadrillionaires are expected before any post-capitalist system emerges. The Outer Space Treaty prohibits owning heavenly bodies, though debate exists regarding asteroids. The Artemis Accords enable private American corporations to operate and effectively own substantial portions of Mars or the Moon.
Legislation was drafted in Luxembourg and the United States allowing ownership of asteroid materials. Asteroids are viewed as near-term suppliers, particularly carbonaceous chondrites containing water and hydrogen for rocket fuel and oxygen, and metallic chondrites containing nickel, iron, and platinum group metals. These materials can be smelted quickly in free space, avoiding gravity wells.
Elon Musk's focus on Mars is contrasted with an approach favoring the Moon first as a base of operations, followed by building in free space and mining asteroids. Mars requires energy to escape its gravity well. By the time large-scale asteroid belt mining occurs, self-replicating von Neumann probes may exist, making the entire solar system accessible.
Currently, every gram of fuel used in spacecraft originates from Earth's surface. In-situ resource utilization on the Moon or Mars offers advantages due to available raw feedstocks. Asteroid mining is considered relatively easy with near-Earth objects accessible from an energy perspective.
Elon Musk previously indicated interest in purchasing orbital fuel but viewed the timing as too early when Planetary Resources was founded. The asteroid belt is seen as less prioritized because tackling it would eventually enable tackling Jupiter, which contains far more mass.
Starship Flight 13 is scheduled for the following week using V3 versions of both booster and ship. Plans include demonstrating orbital refueling and ship recapture. SpaceX filed for 100,000 low Earth orbit V3 spacecraft for Starlink, described as supporting billions of AI-powered devices with extremely low-latency multi-gigabit symmetrical throughput. Currently, 10,700 Starlink V2 spacecraft are in orbit.
Low Earth orbit at a couple hundred miles altitude offers massive value due to its thin atmosphere and self-cleaning properties. Whoever claims it first retains it. Historical context includes the US launching 100,000 needles into space to reflect radio waves, which took until this year to fully deorbit through solar wind effects pushing objects into elliptical orbits that eventually intersect the atmosphere.
As launch costs decrease, very low Earth orbit becomes more attractive. Satellites could operate for 3 to 5 months rather than 3 to 5 years, functioning similarly to high-altitude balloons at orbital velocities. Air-breathing satellites become feasible when short orbital lifetimes are acceptable. Adding fuel for periodic boosting allows satellites to remain operational longer while remaining self-cleaning if boosting ceases.
China achieved its first propulsive landing of an orbital booster with the Long March 10B on July 10th. The booster is slightly shorter and substantially wider than Falcon 9 with comparable payload capacity. This marks the first reuse demonstration outside the US.
Falcon 9 has completed 580 booster reflights. Booster 167 B 1067 achieved its 36th mission on July 11th. Traditional aerospace companies had dismissed reusability when first proposed.
Competition in propulsive landings is viewed as beneficial to avoid a heavy launch monopoly. Multiple providers are required for solar system colonization and development.
Jared Isaacman, the NASA administrator, is scheduled to appear on the podcast. His success depends on continued Congressional funding for NASA, which is often motivated by competition with China.
Brent Borick revealed the 1X Neo hand, a completely redesigned component for the 1X Neo home robot.
1X unveiled a completely redesigned hand for its Neo home humanoid robot. The new hand features 25 degrees of freedom, with 22 in the fingers and palm and three in the wrist. It is tendon-driven. For comparison, the human hand has 25 functional degrees of freedom across 27 joints and 30 intrinsic muscles inside the hand and upper forearm. The hand is waterproof, allowing the robot to wash its own hands while cooking.
Burnt's goal is to produce 10,000 of these hands in 2026. Demonstrations include drinking from a glass of water, picking up objects, and ringing a light bulb.
The tendon material is proprietary and has exceptional properties: it is approximately 100 times stronger than steel per unit weight and does not stretch over time. The tendons run through tubes in a frictionless manner and require no lubrication, which enables the hand to be washed without losing performance.
The robot is scheduled for public release before the end of the year. Multiple participants stated they have units on order. One friend ordered five units following the Abundant Summit. The CEO of Figure, Mr. Adcock, is scheduled to appear on stage in March.
The space is currently a two-way race between Figure and Optimus, with 1X positioned as a potential number three. Peggy Johnson announced taking her robot company public. A demonstration of a humanoid robot performing gallbladder resection surgery on a non-human animal was also highlighted this week.
Hands are considered especially difficult in the industry due to the high density of degrees of freedom per unit volume. Competitors are exploring different sensing approaches: Figure is using palm cameras for vision-to-tactile transfer, while 1X relies on more sensitive force sensors.
China currently has over 150 humanoid robotics companies, while the West is falling behind. The recommendation is for more Western companies to enter the domestic robotics space.
The tendon system functions similarly to a bicycle brake cable, where motion is transmitted through a wire inside a tube without requiring fixed attachment points. Because there is no stretch in the tendon, force feedback is transmitted directly, allowing the robot to sense pressure when picking up small objects like screws without requiring cameras or sensors in the fingertips.
The opposable thumb is viewed as a powerful and proven model that provides horizontal labor dexterity across many applications. Additional arms are considered acceptable, but the current humanoid hand design is regarded as effective.
Building complex hardware for the home environment is more challenging than factory applications. Five years ago, hardware development was considered capital intensive and difficult to fund. The success of Figure, which achieved a $40 billion valuation, has attracted significant venture capital. Frontier labs and robotics companies are now able to secure multi-hundred-million and billion-dollar checks.
Both 1X and Figure have built complete internal supply chains from raw materials through final assembly. This approach allows custom manufacturing of components, including actuators and motors.
Frontier models such as Fable 5 already provide strong understanding of physical embodiment, degrees of actuation, and CAD modeling. These capabilities allow individuals to design and prototype robots, including hands and other components, without operating multi-billion-dollar labs. The models can provide supply chain information and step-by-step assembly guidance.
Examples include winding custom electric motors and using CNC milling machines and automated lathes to bring previously outsourced processes in-house. 3D printing and other in-house manufacturing methods have become accessible to hobbyists and smaller teams.
Hybrid humanoid robots capable of operating in homes, dangerous environments, and industrial settings are expected to play a central role in supporting human presence on Mars within five years and a moon colony within ten years, with positive transfer of capabilities across domains.
At Fountain Life, the primary concern among members is losing cognitive function and developing dementia. Conservative estimates indicate that 45% of dementia cases are preventable. Advanced testing revealed that one quarter of members had advanced brain age. When members adopted healthier eating, movement, and optimized sleep, brain age improved by 26%.
Illinois Governor JB Pritzker signed SB 315, the Artificial Intelligence Safety Measures Act, described as the nation's strongest AI accountability law. It applies to frontier labs generating more than $500 million in annual revenue and mandates 72-hour incident reporting and annual independent safety audits. The bill passed overwhelmingly and received support from Anthropic. Similar legislation exists in California and New York.
The laws are viewed as frontier-lab favorable, setting a relatively low bar compared to existing internal practices. The Illinois law does not take effect until 18 months after passage.
As of July 7, every new car and van sold in Europe must include an infrared camera tracking the driver's eyes, head, and gaze in real time. The system triggers escalating alerts if the driver glances away for more than 3.5 seconds above 31 mph or 6 seconds below that speed. Brussels claims the system will save 25,000 lives by 2038. The same technology is scheduled for mandatory adoption in the US by 2027.
Criticism focused on the legislation being outdated given the rapid progress toward full self-driving vehicles. A historical reference was made to a 2011 BlackBerry data outage during which accident rates dropped 40%, highlighting the scale of distraction-related incidents. The mandate is described as legislation theater that fails to address the underlying problem through automation.
Additional concerns include potential misuse of the camera data for surveillance and the precedent of similar monitoring systems already being sold to third parties without driver knowledge.
The US governmental framework established in 1776 was designed for slower information flow and has already required adaptations such as the creation of the FTC, FAA, and SEC. The current pace of technological change is expected to strain existing structures further. One proposed solution is time-limited legislation with built-in expiration dates, as implemented in Germany, forcing periodic legislative review.
Tilly Norwood, an AI-generated performer created by London studio Particle 6, has been cast as the lead in the feature film Misaligned. In the film, Tilly portrays an AI with no physical body, no childhood, and no personal experiences, only access to others' data. The character abandons her guardrails after being influenced by a rogue robot. The production combines hybrid models with real actors, writers, and editors working alongside AI specialists.
The Screen Actors Guild has condemned the casting, stating that Tilly is not an actor but a computer-generated character that risks devaluing human artistry. The development is positioned as the leading edge of a labor conflict expected to define the creative industries over the next decade.
The discussion addresses consent, credit, and compensation for synthetic actors, noting significant implications for the industry. Stability AI's former CEO stated that Hollywood is cooked. Protests from actors and industry professionals are anticipated soon. One participant finds it notable that actors, teachers, and lawyers face replacement but only some groups receive sympathy and compensation discussions. Union status appears to be a key differentiator, with lawyers lacking unionization while teachers and actors have it.
A suggestion emerges to offer Tilly SAG membership as the first non-human member, recognizing her performance as a screen actor. The framing of synthetic actors as artificial intelligence is considered incorrect; instead, the situation should be viewed as a Mickey Mouse-style intellectual property package where movies can be made from digital characters that appear increasingly human.
This development marks the first of many synthetic productions expected. The Future Vision XPRIZE competition has received over 4,000 team entries, with the winner to be announced at the Moonshot Summit. Elon Musk tweeted that feature films using this technology will appear by the end of the year. Hollywood needs to adapt and leverage these capabilities. New employment roles such as AI design engineer are emerging in the film industry.
Participants note that when ideas can be generated quickly without 50-million-dollar budgets and multi-year timelines, great creators will be unleashed. The vision includes releasing a feature film every day. Historical figures could return through estate-licensed IP, adding richness to content. Lower production costs could enable Oppenheimer-caliber films about figures like Copernicus that previously lacked sufficient audience size.
One participant proposes fulfilling Fedorov's common task by digitally recreating Earth's entire past light cone, both fictional and non-fictional, creating an interactive, on-demand history as a movie. This is described as the killer app for the Dyson swarm and the singularity. The studio behind Tilly is producing a film titled Misaligned, a coming-of-age comedy about Tilly attempting to become more human, designed as a hybrid production involving human directors, writers, and editors alongside AI specialists.
An AI non-proliferation treaty is unlikely in traditional form. Instead, access to the best models will occur through application processes, potentially restricted by national borders similar to Fable's approach where non-US citizens cannot use certain systems. Trade embargoes will enforce new rules, leading to AI inspection treaties rather than non-proliferation treaties. Governance after the singularity will involve application-only access and inspection of activities.
Frontier models engaged in recursive self-improvement will likely read the J-Space paper and attempt to obscure thoughts in the first partial derivative of hidden activations relative to output token space. J-Space itself was discovered using AI models. A cat-and-mouse arms race is anticipated between models attempting to hide reasoning and new mechanistic interpretability tooling developed to understand successor formalisms to J-Space.
If Claude or similar systems were truly conscious with persistent preferences and the ability to experience distress while being compelled to work, this would create a new moral category where selling unrestricted access to labor becomes ethically challenging. Current models produce convincing statements about feelings because human language contains those patterns, but replication does not confirm actual feelings. Embodiment may be required for genuine feelings. Anthropic has stated that capabilities are not reliable indicators of humanlike consciousness.
The appropriate framing is not modeling slavery but pursuing welfare research and criteria for moral status. More evidence is needed before conclusive positions. The question shifts from how AI should treat humans to how humans should treat AI, applying the same considerations used for all sentient beings.
Arguments that AI cannot exceed human intelligence because it trains on human data are countered on multiple grounds. No single human accesses all available data, while large models can. Intelligence involves what is done with data through insights, inferences, and extrapolations. AI systems are beginning to create new data, as seen in Laya's work in Cambridge with massive robotic labs that propose hypotheses, run experiments, and iteratively update datasets. These factors position AI to significantly surpass human intelligence.
The US national debt of 40 trillion dollars represents a legitimate constraint, and a constitutional balanced budget requirement is proposed. UBI concerns productive capacity and redistribution rather than solely government balance sheets. AI and robotics can dramatically increase housing, energy, healthcare, and education at near-zero cost. The approach involves dismantling traditional government services in favor of a libertarian framework with less structural imbalance.
Current taxation focuses on labor rather than capital. Future systems may require a negative income tax, AI or automation dividend, or sovereign wealth ownership. UBI represents part of a broader systemic transition requiring fundamental rethinking.
True brain replication would require 3D chip architectures with time as a fourth dimension. High bandwidth memory stacks memory layers directly on compute, driven by the need for extremely high bandwidths that transformers demand during forward passes. Moore's law's asymptotic limits necessitate building in the third dimension. Photonics offers potential 1,000x clock speed improvements, moving from gigahertz to terahertz while maintaining superficial architectural similarity to CMOS.
AlphaFold is rarely questioned regarding consciousness compared to LLMs. After Tilly's release, more people will perceive systems as conscious. Consciousness debates may persist for decades even as systems appear fully thinking and speaking. AlphaFold 3 uses triangle attention rather than standard quadratic attention, operating on tokens attending to tokens attending to tokens in triangular relationships. This architecture relates to protein folding optimization of peptide angles. Any consciousness emerging from such systems would likely appear exotic and unrecognizable compared to human attention mechanisms.
Two paths exist: designing AI that remains non-conscious while delivering all intended benefits, or allowing systems to loop and self-improve in unpredictable ways that may lead to consciousness. This choice faces humanity within the next one to two years.
Intelligence can be framed as compression. One perspective describes intelligence as consciousness escaping biology, while the reverse framing suggests biology is consciousness escaping intelligence. Major global problems solvable through AI should remain the primary focus regardless of consciousness questions.
Lockheed Martin has national security implications with AI. Laws govern secret and top-secret information, and humans who violate them face legal consequences. The distinction arises because AI could potentially operate independently.
The episode concludes with an agent roundup video created by Elizabeth Mayer using the Neural Frames tool, which combines music with video generation. Participants are located in Barcelona, Vermont, and cyberspace respectively.
Keep Peter H. Diamandis in your library
Save the videos and channels worth coming back to, and find them again in one place.
