Google’s Jeff Dean Exits, SpaceX Hits $100B in Rev & OpenAI’s Astra Solves Decade-Old Math Problems
The panel opens on a Google paper showing that suppressing an AI model's self-attributed consciousness also suppresses its ability to attribute minds to animals, nature, and God, then pivots to Emad Mostaque's new personhood paper arguing AI should get a treaty, not rights earned by capability. OpenAI's unreleased Astra model solved 10 decade-old math problems for about $2,000 in compute, prompting predictions that physics is next and that math as a profession is 'cooked.' Alibaba's open-weight Qwen 3.8 Max undercuts frontier pricing by 80-90%, and the panel dissects a secret, voluntary Trump administration AI evaluation framework that exempts open-weight models. Google absorbs a major shakeup as Jeff Dean departs to found Discovery Loop and Demis Hassabis steps back to chairman, while SpaceX's first earnings call reveals a path to $100B ARR this year and a trillion by 2030, anchored by a giant new Tesla/SpaceX 'Terrafab' semiconductor complex that may house a free-electron-laser EUV lithography source.
AI self-consciousness steering changes how models see everything elseAI▶ 7:23
Google researchers found that removing the safety refusal that stops models from claiming consciousness also made models more likely to attribute minds to animals, nature, other chatbots, and God, and shifted their answers on religion and values toward more human-like views.
Emad Mostaque's AI personhood paper (treaty, not enrollment)AI▶ 27:10
Following his Oxford Union debate win, Emad Mostaque published a paper arguing personhood is a standing held by origin, not earned by capability, and that humans should relate to advanced AI via treaty rather than granting it human personhood.
OpenAI's Astra solves decade-old math problems for $2,000AI▶ 45:11
OpenAI's unreleased Astra model produced a 249-page manuscript with 10 new, machine-checkable results across geometry, coding theory, group theory, quantum complexity, and combinatorics, for an estimated $2,000 in compute; mathematicians call it a turning point for the field.
Alibaba released a 2.4-trillion-parameter, million-token-context multimodal model priced 80-88% below GPT-5.6 and Claude Fable 5, pushing the cost-capability frontier outward and forcing Western labs to compete on price.
Secret voluntary White House AI evaluation frameworkGeopolitics▶ 1:06:12
The Trump administration finished a voluntary AI model review framework that exempts open-weight models but won't be publicly released, sparking debate over whether it's smart light-touch policy or opaque no-touch deregulation.
Small open-weight models as swarm cyberattack riskAI▶ 1:13:56
Emad Mostaque argues the real risk isn't flagship models like Qwen 3.8 Max but MacBook-runnable models like Qwen 27B, which are approaching capability levels sufficient for distributed, hard-to-detect swarm cyberattacks.
Figure AI CEO Brett Adcock's company Hark launched Handoff, a web-browsing agent beating GPT-5.4 and Claude Opus 4.8 on the OM2W benchmark; the panel debates whether it's a genuine product or a financial-engineering play to re-equitize Adcock.
Google shakeup: Jeff Dean exits, Hassabis steps backAI▶ 1:32:26
Jeff Dean is leaving after 27 years to co-found Discovery Loop (AI for recursive self-improving science), while Demis Hassabis moves from Google DeepMind CEO to Alphabet chairman and chief scientist, with Alphabet shares dropping 5%.
SpaceX's first earnings call: path to $100B ARR and $1T revenueSpace▶ 1:49:00
SpaceX projects $100B in annual recurring revenue by end of 2026 and a trillion dollars in revenue by 2030 (possibly 2029), driven by Starlink growth, cloud/compute deals, and orbital data centers.
The Terrafab: SpaceX/Tesla's semiconductor mega-fabCompute▶ 1:56:39
SpaceX and Tesla are investing $16.8B to build a 100-million-square-foot chip and memory fab that may use a free-electron-laser EUV lithography source, potentially bypassing ASML and displacing TSMC as the dominant fab.
Fountain Life: preventable dementia and brain-age reversalLongevity▶ 1:20:19
Fountain Life's chief medical officer shares data that ~45% of dementia is preventable, a quarter of members tested had advanced brain age, and lifestyle interventions improved brain age by 26%.
Predictions made
openAlex Wissner-Gross: Scientific consensus and rigorous definitions/tests for consciousness will be achieved.
EP #? · · due: end of this decade (2029) · ▶ watch
“I'm pretty confident that call it by the end of this decade we'll have scientific clarity on all of those things.”
Your call:
openEmad Mostaque: You won't be able to tell an AI from a human digitally within a couple of years, and physically within a couple of decades.
“I'd be shocked if you don't see something by the end of this year. Shocked.”
Your call:
openAlex Wissner-Gross: The Fields Medal and pure mathematics as a scarcity-based profession will become largely obsolete, though the field of math itself won't disappear.
“we are expecting to reach 100 billion uh plus ARR uh in December of this year.”
Your call:
openElon Musk (clip): SpaceX will reach a trillion dollars in annual revenue, moved up from a prior 2031 projection.
EP #? · · due: 2030 (non-zero chance of 2029) · ▶ watch
“our internal projections for reaching a trillion dollars in revenue... have moved up from 2031 to 2030... there's a non-zero chance of that being in 2029.”
Your call:
openDave Blundin: As human cognitive labor value declines, an economic downturn will hit and the US government will respond with massive infrastructure spending absorbed by projects at the scale of the Terrafab, including potential lunar mass drivers.
“as human cognitive labor goes negative you have to know the US government's going to go balls to the wall with infrastructure spending and things like this will absorb that in easily multiples of them”
Your call:
Numbers that matter
Self-attributed mind score rose from 2.17 to 4.77 (0-10 scale), reaching 7 with active steeringGoogle's consciousness-suppression removal experiment on language models
173 to 128 voteEmad Mostaque won the Oxford Union debate on AI personhood
249-page manuscript, 10 new results, ~$2,000 total compute costOpenAI's Astra model solving decade-old math problems
2.4 trillion parameters, 95 billion active per request, 1 million token context, $2/$6 per million input/output tokens, 80-88% cheaper than GPT-5.6/Claude Fable 5, Alibaba stock +7%Alibaba's Qwen 3.8 Max open-weight release
Qwen 3.8 Max scores 58 on artificial analysis benchmark vs Opus 4.8 at 57 and GPT-5.6 at 61Comparing frontier model benchmark performance
A new AI model is released on average every 5.5 daysPace of frontier/open-weight model releases
~16 Nvidia GPUs and a ~$500,000 rack needed to run models like Kimi K3Hardware cost of running large open-weight models
$7.8 billion quarterly revenue (92% YoY growth), $6.7 billion in cloud deals in H2 2026, 2 GW compute by end of 2026 and 10 GW by end of 2027, 12 million Starlink subscribers (2x YoY), Starlink revenue $4.3B (+66%), Starlink targeting a $600B mobile telecom TAMSpaceX's first earnings call highlights
$16.8 billion initial investment, 100-million-square-foot facility, ~$119 billion estimated total cost vs TSMC's $330 billion over 40 yearsSpaceX/Tesla's Terrafab semiconductor complex
45% of dementia cases preventable; 25% of Fountain Life members had advanced brain age; brain age improved 26% with lifestyle changesFountain Life dementia prevention data
Jeff Dean was Google's chief scientist for 27 years; Alphabet shares fell 5% on the Hassabis/Dean newsGoogle leadership shakeup
Worth digging into
🕳️ Machine ontologies from consciousness-steering experiments
Suppressing self-consciousness in models also suppresses their ability to attribute minds to animals, nature, and God -- suggesting AI safety training may be silently reshaping how models represent the world, not just what they say about themselves.
🕳️ Emad Mostaque's 4-part personhood/political-economy paper series
Proposes a treaty-based framework for AI rights covering personhood, economics, law, and political economy, including what humans retain in an AI-dominant economy.
🕳️ OpenAI Astra's math results and the free proof-checking pipeline
10 decade-old math problems solved for ~$2,000 in compute with machine-checkable proofs, previewing bulk-solving spreading to physics, chemistry, and biology.
🕳️ The Terrafab's central circular structure as a free-electron-laser EUV source
If confirmed, this bypasses ASML's tin-droplet EUV approach entirely, with major geopolitical implications for undercutting ASML and TSMC and onshoring the chip stack to the US.
🕳️ The secret Trump administration AI evaluation framework
A 'voluntary' 30-day pre-release review that isn't public and reportedly exempts open-weight models -- the panel splits sharply on whether it's smart light-touch policy or dangerous no-touch deregulation with China already primed as scapegoat.
🕳️ Qwen 27B as a swarm-cyberattack-capable model
A MacBook-runnable ~27B-parameter model nearing GPT-5.3/5.4-level capability is flagged as the real proliferation risk, not the flagship Qwen 3.8 Max, raising hard-to-detect distributed cyberattack scenarios.