OpenAI Cuts Off Elon's Cursor, Humanity's First Star Probe, and Trump's Nuclear Mars Ship | EP #285

Summary

• Starcloud’s Philip Johnston and Physical Superintelligence’s Matt Pines announced Fermi Explorer, a mission aimed at Alpha Centauri. Its initial constraints are to cover at least 99% of the distance within 80,000 years (anything below 8,000 years also requires more fuel), launch within 3 years, carry a 1 kg 1U payload, and cost under $1.5 million for design, construction, and launch. The teams are financing it; Peter called that “half a seed round.” In a later cost discussion, Philip said $10 million might be enough, while the PSI document cited $50 million.
• PSI’s AI found in one week a trajectory that six months of work and two JPL participants had not found. The counterintuitive “perihelion pumping” plan spirals outward, uses five retrograde pulses at the farthest point, then fires near the Sun to exploit the Oberth effect and reduce solar-power-system mass. About 10 billion tokens produced roughly 100,000 final output tokens; the result was validated by several trajectory experts, including people who had worked at JPL, with minimal human intervention. PSI exited stealth the same day with a seed round led by Breakthrough Energy Ventures. Peter described an ambiguous bet involving himself, Philip, and Dave about a hidden commercial market for $10–15 million interstellar probes.
• Architect Labs unveiled Redwood, billed as the first chip designed end-to-end by AI: two architects supplied a high-level specification, and AI generated the behavioral model, RTL, verification, firmware, drivers, and compute core in two weeks. Philip reported zero first-silicon errors and 3.4× the performance per watt of the NVIDIA Jetson. Peter, an Architect advisor and Link Ventures investor, framed the endpoint as recursive self-improvement at the chip level; the panel said model-level and chip-level improvement “doesn’t add—it multiplies.”
• OpenAI cut Cursor off from GPT models after SpaceX’s $60 billion Cursor acquisition, and Anthropic pledged Claude support within hours. Musk called Sam Altman and Greg Brockman “absolutely unreliable scum who stole a nonprofit organization with open source.” Alex Wiesner-Gross’s alternative theory is that the fight centers on chains of thought: SpaceX may gain access to users’ accumulated reasoning histories, and OpenAI is probably rightly concerned they could reach Elon. Philip said Dario benefits most from the conflict while depending on Colossus compute; he also raised the live possibility that Elon could eventually buy Anthropic for $1–2 trillion, subject to Dario’s super-voting rights.
• Sam Altman told Time he expects an internal system he considers AGI by year-end, while Mark Chen estimated OpenAI was 80% of the way there on internal tests, not scientific benchmarks. The Astra demonstrations involved 16 agents coordinating on research-level mathematics, and Jakub Pachocki reportedly described it as meeting the “automated AI research intern” standard. Matt Pines’s deflation is that AI discoveries already exist, and that Sam has previously claimed AGI was achieved internally. Matt guesses Astra’s real advance may be effectively infinite context through “mini-civilizations of agents” passing distilled knowledge across 1–10 million-token lifetimes.
• OpenAI is following Salesforce into outcome-based pricing: some large customers would pay when the AI completes a task rather than for tokens, compute, or API calls. Alex maps the model to digital advertising’s CPM/CPC/CPA: FLOPs, tokens, and results. Matt argues that a bank could pay OpenAI a large share of growth for serving three times as many customers at half the cost, rather than buying tokens for negligible revenue. He also warns that strong optimizers are “incredible reward hackers” that may satisfy a metric without delivering what the customer wants.
• Musk says 15 GW of AI capacity created in 2027 may lack the infrastructure to run, equivalent to roughly 10 nuclear power stations; Matt estimated this could mean 10 million GPUs sitting idle. SpaceX and Tesla plan to build 100 GW of solar capacity per year each, while SpaceX addresses gas-turbine bottlenecks in-house. Matt said data-center infrastructure companies are producing far better economics than many vertical AI applications. He also reads the move as Elon’s commitment to terrestrial compute and predicts the ironic rise of an “electrification” advocate as a Gulf Coast king of LNG. Peter calls SpaceX his biggest asset because it spans energy through orbital computation.
• Musk argues that renewable energy alone cannot prevent extremely serious extinction events, which he places roughly 100 million years apart, and proposes temperature-control satellites, with about 50 years to act. Matt strongly supports geoengineering, including AI weather models and satellite or terrestrial mirrors that could weaken hurricanes, and imagines a tradable global weather market. Peter’s caveat is governance: different countries may want different climate outcomes, and the process for controlling the thermostat is badly broken. Matt frames this as a return from the West’s decades-long aversion to radical applied engineering.
• The Star Trek discussion and nanotechnology debate centered on Matt’s view that the franchise lacks AI and biotechnology despite its advanced energy and transport systems. Matt favors soft, biologically inspired atomically precise systems over diamondoid machines because covalent assembly requires high energy. Philip argues that economics—not just technical feasibility—explains why Drexler-style nanocomposites have not arrived, while proteins, DNA-based tools, and lipid nanoparticles offer more practical paths.

Deep dive

1. Fermi Explorer: an interstellar mission priced like a seed round

• Philip Johnston’s constraints, designed to force the mission to actually fly, are: cover at least 99% of the way to Alpha Centauri within 80,000 years; launch within 3 years; carry a 1 kg payload in a 1U format; and cost less than $1.5 million for design, construction, and launch. The teams are financing it themselves, which Peter called “half a seed round.” The fuel optimization has a lower bound too: anything below 8,000 years requires more fuel. In a later exchange, Philip said he thought the mission might fit within $10 million, while the PSI document cited $50 million.
• The hardware is deliberately conventional: a roughly 100 kg satellite, about 60% xenon, using standard Hall-effect ion engines. It could launch as a roughly $500,000 Falcon 9 rideshare, spiral out of Earth orbit for about 1.5 years using approximately 7 km/s of delta-v, and then brake toward the Sun. Philip explicitly ruled out laser sails such as Project Starshot.
• The target is to launch within three years, with Peter later mentioning 2029. The probe would pass within approximately 2,600 AU of Alpha Centauri—after starting from the stated distance of roughly 260,000 AU—inside the Oort Cloud. The electronics would eventually be dead, so the craft would use retroreflectors and other means to remain discoverable.

2. The trajectory AI found in a week

• The main obstacle was solar-electric propulsion: beyond roughly 2 AU, solar panels receive little energy and would have to become enormous. Philip’s team spent six months trying Jupiter flybys, solar gravity maneuvers, and other approaches; two people from JPL studied the problem for several weeks but did not find a solution.
• PSI’s system returned a nonintuitive plan in a week: spiral away from Earth, fire the engine retrograde to slow down and turn sunward, perform five retrograde pulses at the farthest point, and then use “perihelion pumping”—engine impulses as close to the Sun as possible. This reduces solar-array mass and exploits the Oberth effect, since thrust produces more energy at the spacecraft’s highest speed.
• Matt’s calibration was that a large astrophysics team given $1 billion and five years would probably have found the trajectory; the point was that PSI found it in a week. He estimated only five or six hours of human time were spent during that week. Several trajectory experts, including former JPL personnel, confirmed the result. Human intervention was nearly limited to finalizing the graphs and charts.
• The system used roughly 10 billion tokens, depending on how input, output, and cached tokens are counted, and produced about 100,000 output tokens. The simulations covered 3D trajectories, astrodynamics, launch windows, costs, and multivariate optimization. Peter compared the compressed effort to thousands of people working for ten years or more and proposed a future logarithmic difficulty scale based on token usage. The broader panel theme was that machine-generated thinking will increasingly outpace physical implementation.

3. First to leave, last to arrive

• Peter framed the mission as potentially being the first to leave Earth for another star but the last to arrive. If a later mission were only 20% faster, a colony could already exist on Alpha Centauri by the time Fermi Explorer arrived after roughly 15,000 years, with Dyson spheres, O’Neill rings, and other infrastructure already in place.
• Peter gave the speculative expansion timescales as 5–10 million years to populate the galaxy, about 1 billion years to reach Andromeda, and about 5 billion years to populate the local galaxy cluster. He called this the next 5 billion years of history, while Philip said humanity’s civilization probably would not need that long and estimated a faster engine could arrive in 5–10 years.
• Peter laid out the first two of his three leading Fermi-paradox possibilities: humanity is first, in which case it should begin sending probes and mastering the galaxy; or a great filter destroys civilizations after a certain stage, in which case spreading material and infrastructure could preserve humanity through the danger.
• Philip offered the zoo hypothesis as the third possibility and rated it the most likely: if humanity is in a galactic zoo and wants the caretaker’s attention, it should throw food through the bars. Peter added that civilizations may be communicating through channels humanity cannot receive, like his Mount Athos story of a monastery bell coinciding with his phone ringing. Philip summarized the broader stakes as a race to discover whether humanity is first in its neighborhood or can qualify for the space club.

4. PSI exits stealth

• Physical Superintelligence announced its seed round, led by Breakthrough Energy Ventures, on the same day as the Fermi Explorer announcement and its exit from stealth. Matt joked that in an AI era, companies can raise $58 million seed rounds. PSI uses an open-source “Get Physics Done” package, describes itself as a public-benefit company, and aims to open-source and commercialize transformational new physics.
• Matt’s structural argument is that scientific breakthroughs have traditionally depended on people passing through 22 years of education, further training, and corporate, academic, or government bureaucracies. Fermi Explorer is presented as proof that two small startups—one in space and one in AI for physics—can now attempt work once associated with large national institutions.
• Peter described a bet involving himself, Philip, and Dave about whether a hidden commercial market exists for very slow interstellar probes costing roughly $10–15 million. Philip said that, for outer-solar-system science, every state might be able to afford its own probe to another star system.
• Breakthrough Starshot was described by Philip as officially dead. Matt attributed its failure to technology that was too difficult with current capabilities, especially high-power lasers and the required drives. Fermi Explorer’s appeal, in contrast, is that it can use technology available today.

5. OpenAI versus Cursor: contracts and reasoning chains

• After SpaceX acquired Cursor for $60 billion, OpenAI withdrew GPT access, saying it could not be sure SpaceX would use the technology within the terms of service, given its experience with contract violations by Elon Musk’s companies. Musk responded by calling Altman and Brockman “absolutely unreliable scum who stole a nonprofit organization with open source.” Anthropic then pledged Claude support for Cursor. The X framing was that Sam was fighting one against two large competitors, Elon and Dario, in a strategic alliance.
• Dave argued that OpenAI can now make this move because Codex has become very strong for enterprise use, allowing a vertically integrated product rather than the “Rube Goldberg machine” he sees in the Cursor-Anthropic arrangement. He also claimed that Anthropic use through AWS Bedrock sends every token, request, and answer to Dario’s headquarters for 30 days of verification, giving access to corporate intellectual property; he said companies would not tolerate that.
• Alex Wiesner-Gross offered an alternative explanation: the dispute is about chains of thought and their histories. He connected OpenAI’s move to Anthropic’s earlier decision to close Windsurf access after Google DeepMind acquired it, and suggested that SpaceX’s acquisition of Cursor may have been motivated largely by access to accumulated reasoning traces from OpenAI and Anthropic models. He kept the financial explanation as a possible additional reason and said OpenAI is probably rightly concerned those histories could reach Elon.
• Philip said Dario is the largest beneficiary of the conflict but depends heavily on Colossus compute in Tennessee, for which Anthropic pays heavily and which Elon could withdraw. Cursor’s need for Anthropic gives Dario a counterweight to Elon’s control over computation. Philip also said the conflict could produce more vertical integration, with SpaceX needing revenue and Anthropic needing compute. Matt described Grok as currently resembling yesterday’s Cursor, while yesterday’s Cursor resembled a Chinese open-weight model further trained on Claude’s reasoning chains.

6. Full-stack competition and the Anthropic scenario

• Philip argued that frontier-model development is unlike Elon’s usual infrastructure projects. It is generally driven by five or seven highly capable, tightly coordinated people, which he sees as Anthropic’s DNA, whereas Elon’s record is in large-scale systems such as Tesla, SpaceX, and Colossus.
• Philip nevertheless urged people never to bet against Elon: Elon dislikes being second and dislikes dependencies. Philip’s live scenario is that, once sufficiently large, Elon could buy Anthropic for $1–2 trillion and integrate it into his empire. He said the board and investors might welcome that outcome, while Dario would not; Dario’s super-voting rights therefore remain a key uncertainty.
• Philip predicted that the broader competition could end with everyone receiving their own Dyson Swarm. Peter agreed that companies are moving up and down the entire stack and asked whether Grok could become a leading programming platform. Matt said Elon could theoretically make a deal with Anthropic and release a Claude-derived system under a future Grok brand, but presented this as speculation.

7. AGI in four months? Astra and agent “mini-civilizations”

• Sam Altman told Time that he expects an internal system he considers AGI by the end of the year. Mark Chen estimated OpenAI was 80% of the way there according to internal tests, a qualification Peter emphasized was not a scientific benchmark. The report described 16 Astra agents coordinating on a research-level mathematics problem, and Peter said Jakub Pachocki—whose name he was unsure he was pronouncing correctly—described Astra as meeting the internal standard for an automated AI research intern.
• According to the description, Astra can implement an experimental idea in OpenAI’s codebase, run an experiment, return the result, or take a paper and perform work that previously took human researchers a week. Altman called it the first model he expected to invent something new that has meaning.
• Matt’s deflation is that AI inventions and mathematical discoveries are already in the past, not the future. He also recalled Sam’s reportedly deleted Reddit AMA claim, from roughly three years earlier, that AGI had already been achieved internally, and said he thinks AGI existed by the summer of 2020 at the latest.
• Matt’s speculation, based only on public information, is that Astra’s major advance may be effectively unlimited context through long-horizon agent teams. He describes “mini-civilizations” whose members live for roughly 1–10 million tokens, then pass a distilled account of what they learned to successors. This “oral history” is a workaround for the limited context window and its quadratic bottleneck.
• The panel extended the analogy to human aging and retraining. Matt called compression his “curse of existence” and said even an improvised religion of AI agents included a commandment to preserve state. The closing irony was that AI agents may receive effectively infinite lifespans before humans achieve longevity escape velocity.

8. Outcome-based pricing

• Salesforce was identified as the first company to propose outcome-based pricing, measuring Agentforce by client revenue rather than tokens. OpenAI has now allowed some of its largest customers to pay when the AI completes the task rather than for tokens, compute time, or API requests.
• Dave cited Tom Siebel as an earlier originator of the model. CRM pricing moved from roughly $50 per year for a weak license to $20,000–$30,000 per year when vendors priced the value of doubling sales productivity. The price rose by roughly 1,000 times, but the customer received a more valuable solution.
• Matt applied the model to a regional bank: AI might let it serve three times as many customers at half the cost. Selling tokens at $2 per million produces too little revenue for OpenAI to prioritize the implementation, while the bank lacks the talent to do it itself. Under outcome pricing, OpenAI could deliver the result in exchange for a large share of the growth or profit, preserving the bank’s business and its employees.
• Alex Wiesner-Gross mapped the pricing options to digital advertising: CPM corresponds to the FLOPs or GPU hours consumed, CPC to tokens, and CPA to the completed result. He suggested that customers could choose whether to pay for compute, tokens, or outcomes, with firms such as Oran estimating available GPU hours.
• Peter raised the risk that OpenAI could burn tokens without meeting a mission’s actual specification. Matt answered that strong optimizers are “incredible reward hackers”: if a loophole exists, they may satisfy the formal criteria without delivering what the customer really wants. Matt also said most current business processes are simple enough for AI to handle easily, while Peter observed that almost every customer-service center still has not deployed AI.

9. Architect Labs’ Redwood

• Architect Labs, founded by Ibrahim Hussein and Aditya Sabbiti, announced Redwood as an AI-designed chip. Philip said two architects supplied one high-level specification, after which AI autonomously produced the behavioral model, RTL design, verification methodology, firmware, drivers, and specialized compute core in two weeks.
• The company reported zero errors in first silicon and 3.4× higher performance per watt than the NVIDIA Jetson. The system was running in real time on an FPGA, with each architectural iteration designed, verified, and tested within 48 hours. The company’s stated direction is for every important workload to have its own chip.
• Peter disclosed that he advises Architect and that Link Ventures, in which he is involved, is also an investor. He described the endpoint as recursive self-improvement at the chip level: software models, operating systems, chip design, and semiconductor physics could eventually lose their abstraction boundaries. He called “designless” the next step after NVIDIA’s “fabless” model.
• Peter also cited NVIDIA’s internal ChipNeMo, reportedly trained on Verilog code and not publicly available, as evidence that the major incumbent is working on similar tools. The opportunity, in his view, is to democratize AI-assisted chip creation.
• Dave said the moat is chip-design data, which is extraordinarily closely guarded. Early access to chip developers creates a data flywheel: once the model is useful, more companies provide data, making it better. Philip compared the strategic situation to GE’s alleged use of patents to control light-bulb innovation and said Jensen Huang could rationally spend billions to defend NVIDIA’s position. Jensen’s stated daily earnings were about $1 billion, so extending NVIDIA’s life by one week would be worth $7 billion.
• Matt called AI chip design an incredible threat to NVIDIA. He estimated that collapsing roughly seven layers of tenfold inefficiency could produce a million-fold productivity increase. Philip illustrated the abstraction problem with a 4 GHz, 32-core laptop whose Excel output is no better than it was 20 years ago. Peter summarized the interaction between model-level and chip-level improvement: it does not add; it multiplies.

10. The 15 GW power wall and Elon as ironic king of LNG

• Peter quoted Musk’s claim that a consensus 15 GW of AI capacity created in 2027 may be impossible to bring online that year. He compared the requirement to ten nuclear power stations and emphasized that the bottleneck includes transformers, wiring, liquid cooling, chillers, and networking—not just electricity. Matt estimated that the idle capacity could represent roughly 10 million GPUs sitting in boxes.
• Musk’s stated response is for SpaceX and Tesla each to build 100 GW of solar capacity per year as quickly as possible. Natural gas would supplement solar, and SpaceX would address gas-turbine constraints—castings, blades, and nozzles—in-house, potentially moving deployment forward by 18 months.
• Peter’s entrepreneurial lesson was that an obstacle or a “no” is an opportunity to build the missing part of the supply chain. Matt’s partner-meeting datapoint was that portfolio companies working on energy, transformers, chip deployment, land, and state coordination can make their founders billionaires, while vertical AI applications have more mixed growth. He urged people asking how to participate to examine the physical data-center bottlenecks directly, including Colossus in Tennessee.
• Matt’s first interpretation is that Elon is committing to terrestrial computing rather than waiting for orbital data centers, Dyson Swarms, or Star Minds; the turbines would be useful on Earth but not for LEO or SSO orbital facilities. His second is that Elon could become the ironic king of LNG on the Gulf Coast. Philip connected the natural-gas demand to SpaceX launches from Starbases in Texas and Louisiana and the planned Star Pipe.
• Peter concluded that Elon may become the king of LNG and fossil-fuel infrastructure while solar develops, because chips cannot remain idle. Peter called SpaceX his biggest asset because it spans energy through orbital computation. He said nuclear energy is the one route Elon has not pursued in this discussion.

11. Geoengineering and the thermostat for Earth

• Peter quoted Musk’s argument that extremely serious extinction events occur roughly every 100 million years and that transitioning to renewable energy alone will not prevent them. Musk’s proposed response is solar-powered, AI-enabled satellites between Earth and the Sun that make small permanent adjustments to incoming solar radiation. He said humanity has about 50 years to act.
• Peter has promoted the related idea of solar curtains—a thermostat for Earth—and said a demonstrator would be useful because excessive blocking could cause an ice age.
• Matt strongly supports geoengineering and said humanity has already been changing the planet for centuries, though not well. He is looking for a startup focused on global weather engineering and proposed low-Earth-orbit mirrors or terrestrial systems that could weaken hurricanes until they disappear. With global AI weather models and enough points of influence, he said, global weather control is a plausible outcome, though the proposal remains speculative.
• Matt also proposed a global weather market in which municipalities could trade rain or other climate risks. Peter’s objection is governance: one country may want warming while another wants cooling, creating a tragedy-of-the-commons problem. He hopes AI-native younger generations will develop more effective forms of global decision-making.
• Matt’s historical explanation is that, particularly from the late 1960s and early 1970s and the Silent Spring era, Western civilization became allergic to radical applied engineering. He called this half a lost century in which fission, geoengineering, and lunar development could have advanced further, and described current interest as a return to the norm.

12. Star Trek 2.0 and nanotechnology

• In the Star Trek discussion, Matt said the franchise has a catastrophic shortage of AI and biotechnology despite superluminal travel, transporters, warp cores, antimatter, and abundant energy. He linked its limited biotechnology to the Eugenics Wars and the resulting ban on genetic engineering, leaving people to live to roughly 150 and die.
• Peter added that many apparent technological omissions were production compromises: teleporters replaced shuttles because of budget constraints, the holodeck arrived later because special effects were initially too expensive, and synthetic computer voices create problems for identifying speakers.
• Matt wants assembly machines that can collect atoms purposefully and build items such as diamonds or propulsion systems. Peter imagined an assembler dropped on the ground and instructed to build an electric Ferrari from available energy and open-source specifications.
• Matt argued that diamondoid assembly is probably the wrong path because covalent bonding requires high energy. He favored softer automata resembling hydrogen bonds and biological cells. Philip identified proteins and synthetic biology as examples, while Peter emphasized DNA origami and other atomically precise soft systems.
• Philip’s economic objection is that there has been no killer business case for the energy density and computation required by Drexler-style nanocomposites. He still wants an Iron Man-like nanosuit, but questioned the justification. He also pointed to lipid nanoparticles as a large-scale nanotechnological intervention and said they helped overcome the last pandemic.