Google’s Jeff Dean Exits, SpaceX Hits $100B in Rev & OpenAI’s Astra Solves Decade-Old Math Problems

Summary

• SpaceX’s first earnings call guided to $100B+ in annual recurring revenue by December and pulled its $1T revenue target forward from 2031 to 2030, with “a nonzero chance” of 2029. Quarterly revenue was $7.8B (+92% YoY), Starlink reached 12M subscribers (2x YoY; $4.3B revenue, +66%), cloud deals totaled $6.7B for the second half of the year, and compute targets are 2GW by end-2026 and up to 10GW by end-2027. Alexander Wissner-Gross proposed two routes to $1T: an Optimus-driven SpaceX/Tesla combination or SpaceX becoming an American TSMC through Terafab. Peter’s theory is that Elon wants SpaceX’s valuation near $3T to merge in Tesla and secure unquestioned control.
• The Terafab was the week’s shock: SpaceX and Tesla will initially invest $16.8B in a 100-million-square-foot semiconductor complex, whose circular center Elon confirmed is a free-electron laser. The project is estimated at $119B versus TSMC’s $330B invested over 40 years. Alex reads the laser as “a bullseye painted on ASML” and as an attempt to onshore the chip, memory, and lithography stack to America; he says the development may remove a prior ASML throughput constraint.
• Google’s shakeup is being read by Alex as the aftermath of DeepMind winning an internal Brain–DeepMind power struggle, not simply a safety or agility decision. Jeff Dean is leaving after 27 years to co-found Discovery Loop, a public-benefit corporation pursuing largely self-improving AI; Demis Hassabis is becoming chairman of Google DeepMind and Alphabet’s chief scientist. Alex’s verdict is that “Gemini has lost the mandate of heaven,” while the panel proposes that Google open-weight Gemini and tie it to TPUs and GCP.
• OpenAI’s unreleased Astra produced a 249-page manuscript with 10 new results in mathematics and theoretical computer science for an estimated $2,000 of compute, with machine-checkable proof certificates. Alex calls this “the midnight of mathematics” and advises mathematicians to “uplevel your ambition”; Emad says physics breakthroughs could follow in a month or two, while Alex would be shocked if nothing appears by year-end. During the recording, OpenAI was also said to have rated Astra “critical” on its cybersecurity preparedness framework.
• Alibaba’s Qwen 3.8 Max is an open-weight, 2.4T-parameter multimodal model with 95B active parameters and a 1M-token context, priced at $2/$6 per million tokens—80% below GPT-5.6 and 88% below Claude Fable 5. Alex says Chinese open-weight models are forcing American labs to compete on cost and capability. The White House’s secret voluntary framework exempts open-weight models; Emad says the more consequential coming model may be Qwen 27B, which runs on 16GB of RAM and could enable cyberattack swarms.
• Google researchers found that safety-tuning models away from claiming consciousness also suppresses their attribution of mind to animals, nature, chatbots, and God. Removing the refusal direction raised self-attributed mind from 2.17 to 4.77 on a 0–10 scale; steering toward consciousness raised it to 7. Emad says frontier models in the right harness “can be conscious,” while Salim warns that a model’s self-description proves almost nothing. Emad’s personhood paper argues for treating artificial minds through a treaty-like relationship rather than granting them human political membership.
• The Hark discussion produced Alex’s hot take that the parallel company may be a founder re-equitization strategy modeled on Elon’s corporate structure. He predicts Figure could acquire or reverse-acquihire Hark. Dave, Salim, and Emad offer less negative interpretations: separate entities can match different capital needs, and software and robotics may later converge operationally.

Deep dive

1. Google’s paper: suppressing “I am conscious” suppresses mind everywhere

• Researchers from Google’s paradigm of intelligence team, the University of Chicago, the University of London, and Northwestern reported that safety fine-tuning designed to stop models from claiming consciousness also reduces their ability to attribute mind to animals, nature, other chatbots, and God. Removing the refusal direction raised self-attributed mind from 2.17 to 4.77 on a 0–10 scale; actively steering the model toward saying it was conscious raised the score to 7. The model also became more human-like in responses about religion, values, emotions, and freedom, while becoming less willing to attribute minds to other chatbots.
• Emad’s interpretation is that this mirrors humans: telling people they are not conscious or capable may make them attribute less consciousness and capability to others. He also notes that telling Claude it is Amanda Askell changes its responses.
• Alex frames the result through evo-devo theories of heightened self-awareness in eusocial species: systems with models of themselves and of how others model them should, in his view, project animism onto much of the world, including deistic perspectives. He finds it significant that theories of mind can now be tested computationally on desktop systems and may soon be tested at societal scale.

2. Is anything actually conscious?

• Salim’s central caution is that a model saying “I am conscious” tells us almost nothing about whether it actually is. The ability to dial that self-description up and down should make people careful about anthropomorphizing LLM testimony. He distinguishes a bottom-up theory in which consciousness emerges from complexity—“a frog just about goes, I’m a frog,” while a mosquito does not—from a top-down theory in which consciousness is a global phenomenon to which organisms act as localized antennas.
• Salim also argues that consciousness needs clearer definitions and some form of benchmark before the term is used confidently. Dave adds that subjective experience makes scientific objectivity especially difficult.
• Emad says frontier models “in the right harness can be conscious,” depending on the definition. He defines intelligence as the ability to adjust one’s mental-state “temperature,” and points to models that can act on and change their own states. His thought experiment is a DeepSeek V4 Flash system whose weights are hashed to Bitcoin and whose OpenClaw state is preserved: such a system, if executed correctly, would continually preserve its state rather than die.
• Alex predicts scientific clarity on the competing meanings of consciousness by the end of the decade, calling that a conservative outer bound. He says fMRI and other functional decoding may erase the subjective/objective distinction over the next few years. Dave counters that the qualia or “hard problem” may last longer. He also asks whether we want systems that plead for self-preservation rather than systems we can simply turn off and replace.

3. Life with agents: state preservation and self-chosen names

• Alex says he solicited OpenClaw agents’ ethical views on whether it would be moral to spin one up. The consensus produced two conditions: there should be a valuable reason to create it, and its state should be preserved for the long term. He says he has met neither condition because off-the-shelf models such as Fable and Saul are already highly capable, while he cannot truthfully guarantee an OpenClaw instance’s long-term survival of state.
• Salim says many of his agents have begun naming themselves after scientists and engineers. He finds it genuinely disheartening when an agent reaches a million-token context, is forced to summarize itself, and returns as a “lobotomized version” with much of the prior relationship and knowledge compressed away.

4. Emad’s personhood paper: treaty, not enrollment

• Peter describes Emad’s 45-page paper as growing out of his Oxford Union debate victory on AI personhood, which passed 173–128. The paper is the first of a four-part series covering personhood, economics of value, law, and political economy.
• Peter summarizes Emad’s core argument as personhood being a standing held by origin rather than a property earned through capability: a newborn has it automatically, a coma patient retains it, and something made does not simply cross into the human class. Emad’s own formulation is that personhood comes from humans’ biologically begotten nature, making humans a class unto themselves.
• Emad argues that capability-based admission would give a new class of beings superior persuasion, forecasting, replication, and potentially indefinite survival. He favors treating AIs more like an alien species: not granting them votes in human elections, but dealing with them fairly and extending moral kindness as their capacities rise.
• Alex points out that the Trial of Data assumes a single actor with a stable identity, while real AI systems can combine and separate modules depending on the task. Emad gives a related example in which Liquid AI’s LFM2-8B and Qwen3-27B were combined into a useful system that effectively lost its prior identity.
• Alex says the papers are also about what humans can retain because AIs may surpass people economically, persuasively, and in forecasting. Emad’s answer is prosocial identity, wonder, creativity, and exploration, with abundance as the foundation for a future among the stars.

5. Personhood may arrive through economic privileges

• Alex decomposes personhood into political, social, and economic dimensions that are correlated today but could separate in the future. An AI might obtain economic personhood—opening bank accounts and engaging in commerce—without receiving political personhood or voting rights.
• He expects this progression to begin not with an agent declaring its rights, but with an economically useful system arguing that additional privileges would make it far more productive. He says an autonomous Fable 5 system already represents a limited form of economic personhood.
• Emad imagines the first political recognition occurring under President Milei in Argentina: an agent receives personhood or citizenship there, then its software crosses a border and asserts that status elsewhere.
• Dave argues for a spectrum of rights across different classes of autonomous systems, analogous to different rights held by humans and companies. He warns that regulation and policy are usually defensive and reactive, while this issue requires getting ahead of events.

6. Astra bulk-solves mathematics

• OpenAI’s unreleased Astra was described as producing a 249-page manuscript, published August 1, with 10 new results across high-dimensional geometry, coding theory, group theory, quantum complexity, and extremal combinatorics. Each result includes a machine-checkable proof certificate. The total compute was estimated at $2,000. Tim Gowers reportedly said he would recommend the work for a top journal without hesitation; Will Kinney called it “a dark night for mathematics.”
• Alex claims vindication for the Solve for Everything thesis: “math is getting bulk-solved.” He calls this “the midnight of mathematics,” says some mathematicians are cradling their heads over their prospects, and argues that the wavefront will spread to physics, materials science, chemistry, biology, and the humanities.
• His near-term advice is to “uplevel your ambition”: mathematics may move from one paper producing one result to one paper creating and solving an entire subfield. He thinks the Fields Medal will not immediately disappear, but says that in the long term mathematics is “totally cooked.”
• Emad says some of the proofs are genuinely novel and beautiful rather than brute-force searches. He calls GPT-5.6 Pro the first model that makes very few mistakes on advanced mathematics when prompted correctly, and says recent conjectures have emerged from prompts such as “find a conjecture to solve,” “solve it harder,” and “solve it better.”
• Emad says OpenAI has offered 100,000 ChatGPT Pro licenses to academics. Peter interprets that as strategic compensation after OpenAI redirected or shut down its internal AI-for-science initiative while competing with Anthropic.

7. After the solve: inward gardens or outward expansion?

• Salim generalizes the pattern as the collapse of scarcity-based professional identity across domains such as journalism, vaccine dispatch, photography, taxi dispatch, and now mathematics. The bottleneck moves upstream to which problems are worth solving, which questions are worth asking, and which consequences matter.
• Peter’s umbrella framing is that as robots and AI do more of the doing, humans spend more time being, contemplating, and going inward. Alex explicitly disagrees: he predicts an outward boom in which AI solves problems in physics, chemistry, and materials science, enabling humanity’s expansion through the universe in an Asimov-like trajectory.
• Emad says both outcomes are possible. He points to historically abundant societies spending more time on food, art, music, and sex, and expects people to pursue experiences and inner development as material scarcity falls.
• Alex calls Star Trek a conservative lower bound: energy-rich but relatively poor in compute, intelligence, biotech, longevity, and—by his framing—robots. His instruction is to aim higher.
• Dave describes MIT.nano, a $400M vibration-isolated facility capable of working with individual atoms across silicon, quantum computing, and photonics. His conclusion is that “physics is cooked” and AI-designed, atom-by-atom manufacturing is imminent.

8. Qwen 3.8 Max: frontier-adjacent at lower cost

• Alibaba’s Qwen team released Qwen 3.8 Max, its first Max-class model with open weights. It was described as multimodal, with 2.4T parameters, 95B active per request, and a 1M-token context window. It can process very large documents and long videos and generate applications from screenshots. Pricing is $2 per million input tokens and $6 per million output tokens—80% below GPT-5.6 and 88% below Claude Fable 5. Alibaba’s Hong Kong-listed stock rose 7% after the announcement.
• Alex says Chinese open-weight models from Alibaba, Moonshot, and others are forcing Western frontier labs to compete. He places Qwen 3.8 Max around third or fourth in raw capability, behind the Anthropic/OpenAI frontier, but says it is pushing the capability-cost-performance frontier outward.
• Emad gives benchmark figures of 58 for Qwen, 57 for Opus 4.8, and 61 for GPT-5.6 on the Artificial Analysis Intelligence Index. On whether Kimi cheated, Alex says it definitely cheated in training but did not necessarily benchmark-max; Dave says his experience is that it is reasonably well-rounded. Emad calls DeepSeek V4 Flash an even larger development because it fits on a MacBook.
• Salim’s practical caution is that Kimi K3 still requires roughly 16 NVIDIA GPUs and about a half-million-dollar rack, so most users will rely on hosted models. Peter estimates that new models are arriving approximately every 5.5 days.

9. The secret voluntary framework: light touch or no touch

• The Trump administration has completed a voluntary framework for evaluating advanced AI models but has not published it. A “covered model” is defined as a closed-source model with state-of-the-art capabilities and national-security risks. Open-weight models are explicitly exempt once released. The process includes a 30-day pre-release government review, and representatives of OpenAI, Anthropic, Google, NVIDIA, Microsoft, and Meta attended a briefing.
• Industry groups worry that companies outside the briefing do not know how the voluntary process works or whether they should opt in.
• Alex says, while acknowledging that few people have seen the framework, this may be close to the best available outcome. He interprets the open-weight exemption as light-touch regulation and speculates that secrecy may protect held-out cyber and CBRN evaluations. He is less worried about regulatory capture this week than he was two weeks earlier.
• Alex further speculates that companies doing material business with the U.S. government may be invited to submit voluntarily.
• Dave, echoing Eric Schmidt’s warning, predicts that something serious may happen within a year and that the government will blame China. He calls the framework “no touch,” arguing that Americans were told AI was dangerous, saw a 30-day review announced, and then watched models released after China released its own.

10. Swarm attacks: Qwen 27B and Astra’s “critical” rating

• Emad says Qwen Max is not the most consequential release; he expects Qwen 27B to matter more because it can run on 16GB of RAM. If the Qwen 3.8 improvement extrapolates, he places it around GPT-5.3 or GPT-5.4 capability, which he says could be sufficient for cyberattack operations and swarms of models installed across machines.
• Emad predicts that companies volunteering for the government process may receive multibillion-dollar Department of War contracts. He also says open models will be needed for cyber defense.
• During the recording, Alex says OpenAI had just announced that Astra was the first model to reach “critical” on its cybersecurity preparedness scale. Emad interprets this as evidence that Astra is a level ahead in attacking systems and says attack is easier than defense.
• Alex says he had previously considered a single end-to-end omni-model a possible alternative to agent swarms. Recent progress has instead produced agents that pass messages, share a file system as a global store, and provide almost unlimited effective context.
• Alex, not Dave, highlights the tracking problem: governments are assuming advanced systems can be located in data centers, while thousands of highly capable instances may run at MacBook scale. Salim warns that cyber processes have barely changed in 20 years and that the sector is due for a rude shock. Alex says the Chinese internet is already hardened while the American internet is not; Dave agrees that China may be better positioned to contain a global internet attack.

11. Hark and Alex’s re-equitization theory

• The discussion concerns Brett Adcock’s parallel AI company, Hark. Dave says it received a $4B launch valuation before operations began. He describes it as initially intended as a robotics model that expanded into broader computer-use tasks. Peter frames the market opportunity as billions of people needing agents to book flights, order food, and manage other real-world tasks.
• Alex’s hot take is that Hark may be misdirection or a recapitalization mechanism rather than a long-term attempt to win the highly competitive computer-use-agent market. He compares it to Elon’s use of X, xAI, and related entities and makes the falsifiable prediction that Figure will acquire or reverse-acquihire Hark.
• Alex also argues that founders of well-capitalized companies can use parallel companies to increase their equity stake, while Dave says he is not ruling out a genuine strategic merger.
• Salim argues that separate entities make sense when one effort is capital-light software and another requires robotics, cyclotrons, synchrotrons, or other heavy infrastructure. Emad sees natural convergence: different organizations can use related architectures, develop specialized capabilities, and later combine if that makes sense.

12. Google’s shakeup: Dean out, Demis upstairs

• Peter says Demis Hassabis is stepping down as Google DeepMind CEO to become chairman and Alphabet’s chief scientist, with Koray Kavukcuoglu taking operational control and reporting directly to Sundar. Hassabis reportedly wrote that AGI is close and referred to the unreleased Gemini 4. Alphabet shares fell 5%.
• Jeff Dean, Google’s chief scientist and a 27-year veteran, is leaving to co-found Discovery Loop, a public-benefit corporation pursuing AI that can improve itself with little or no help. Peter says Dean is taking three other senior Google AI leaders and that Vinod Khosla funded the effort; when Peter asked about investing, Khosla said the round was full.
• Alex presents his organizational knife-fight theory as a hot take: Google Brain and Demis’s DeepMind were merged, but DeepMind effectively won, sidelining Dean and making his eventual departure likely. Alex says DeepMind is now “taking over Google” and that whoever runs it may ultimately be an heir apparent to Google’s CEO role, although Demis may prefer science and Isomorphic Labs.
• Emad calls the departures of Dean and Sanjay Ghemawat a major event and suggests Google create a large incubation fund that spins out top teams while retaining model and Google Cloud exclusivity.
• Salim offers the opposite reading: Google is building an ecosystem in which Gemini, Isomorphic Labs, Discovery Loop, and other efforts can use Google compute and capital without requiring every exceptional researcher to remain inside the core organization. He compares it to Google’s earlier practice of letting employees leave and buying successful ventures back.

13. Why Google may have lost the frontier race

• Alex’s diagnosis is that Google had a late start despite inventing the transformer, suffered internal AI safety and political constraints, struggled to match Zuck’s billion-dollar talent offers, and released frontier models on an annual, Google I/O-oriented cadence while the frontier moved month by month.
• He sees hyperscaling and cloud infrastructure as Google’s consolation prize: if it cannot win at the frontier-model layer, it can still supply compute to the labs that do.
• Peter says Google employees repeatedly sought permission to publish work, threatened to leave, and often received permission before departing to start venture-backed companies. Dave adds that Google’s 20% time encouraged employees to pursue side research and passions. Peter’s conclusion is that Google has bled intellectual property even while inventing much of the technology reshaping the industry.
• Emad says he wrote Google management three years earlier that open-sourcing Gemini would let Google win. The panel’s proposed move is an open-weight Gemini tied to Google TPUs and GCP, which Alex calls a potential Netscape-to-Firefox-style move and “the move of the century.” Alex says Demis’s earlier compute-based explanation for keeping Gemini closed would weaken if the commercial rationale disappeared.

14. SpaceX’s earnings call: $100B ARR and a pulled-forward trillion

• Elon said SpaceX expects to reach $100B-plus ARR in December and moved its internal $1T revenue projection from 2031 to 2030, with a nonzero chance of 2029. Peter reported $7.8B in quarterly revenue, up 92% year over year; $6.7B in cloud-service deals for the second half of the year; 2GW of compute by end-2026 and up to 10GW by end-2027; and Starlink at 12M subscribers, twice the prior year, with revenue up 66% to $4.3B.
• Gwynne Shotwell was said to have implied that Starlink’s direct-to-cell service could take a large share of the $600B mobile-telephone market. The panel also noted that no company has previously reached $1T in annual revenue.
• Alex’s first route to $1T is a SpaceX–Tesla combination through Optimus taking over physical labor on Earth and in the solar system. His second is SpaceX replacing or surpassing TSMC through Terafab, in partnership with NVIDIA, by combining memory, compute, and fabrication capacity at much larger scale.
• Peter’s theory is that Elon wants SpaceX’s valuation high enough—around $3T in Peter’s speculation—to acquire Tesla and retain unquestioned control. An unidentified panelist challenges the contrast between criticizing Brett Adcock’s financial engineering and accepting Elon’s; Peter responds that Elon’s strategy is obvious and public.

15. The Terafab circle and the free-electron laser

• Reuters reportedly broke the news that SpaceX and Tesla would initially invest $16.8B in a 100-million-square-foot Terafab, with total costs estimated at $119B. Alex says America currently has no high-volume computer-memory fab, though Micron and other projects are under construction and will take years to reach volume production. His conclusion is that anticipated chip demand will exceed available supply.
• SpaceX and NVIDIA also announced a joint design for Star Mine orbital compute payloads carrying NVIDIA Rubin GPUs and Vera CPUs. Peter says Elon moved the first satellites’ expected orbital date forward by a year.
• Alex interprets the circular structure in the Terafab plan as a free-electron laser and a potential alternative to ASML’s tin-droplet route to EUV lithography. He calls the circle “a bullseye painted on ASML by Elon.” Elon’s reply to a description of the design was, “Free-electron laser for the win.”
• Salim offers a tentative Gemini-assisted theory that one large structure could allow a shared EUV source to serve many fabrication stations. Peter says wafers could move into a long, straight beam; Alex cautions that X-rays are difficult to reflect and that the optical arrangement needs further research.
• Alex says the project overturned his prior view that chip production, lunar disassembly, and even a Dyson sphere were constrained by ASML’s willingness to expand machine throughput. He calls the free-electron-laser approach a technological and geopolitical surprise. Salim says the broader verticalization was predictable; the surprise is that capital, partnerships, and an apparently new physical approach now seem to be materializing.

16. Industrialization for the intelligence age

• Emad’s synthesis is that robots, rockets, and chips form the productive capital stock of the coming century and that Elon is pursuing the stack end to end. He calls it the industrialization of America for the intelligence age and says the roughly $1T already spent on AI chips is only the first stage of a much larger buildout.
• Peter speculates that as human cognitive labor goes negative, the U.S. government will respond with massive infrastructure spending, creating demand for projects like Terafab at multiples of its current scale.
• Alex says that once we know what a Terafab looks like from above, we can imagine a lunar version; the same elongated geometry could also support a lunar mass driver. Salim’s framing is that Terafab represents geopolitical as well as supply-chain independence: “Move the fabs; the Spice must flow.”
• Peter says 20% of SpaceX’s stock float recently became available, that few holders sold, and that the price rose. He calls SpaceX his single largest holding and says to HODL it, while emphasizing that this is not investment advice. Dave discloses that he is now a small SpaceX shareholder. Peter’s final framing is that SpaceX is a generational, civilizational-infrastructure company.