"OpenAI Bulk-Solves 722 Math Problems, Anthropic's Cruelty Policy, & Elon's Quadrillion-Dollar SpaceX" — Peter Diamandis (Moonshots)
Why this is in the vault
Weekly "mates" news roundup (no guest this week) covering three frontier-relevant threads the vault already tracks: Anthropic's new ban on cruelty toward Claude extends the AI-personhood debate from the 2026-10-07 episode; OpenAI's 722-problem math "slop drop" is a hard data point on bulk-solving as a capability-curve inflection, directly relevant to the agents-in-production / credibility thesis; and the $1.8B virtual-cell initiative plus Paradromics BCI progress extend the vault's longevity/BCI frontier-tracking thread.
Episode summary
The four regular "mates" (Peter Diamandis, Dave Blundin, Alexander Wissner-Gross, Salim Ismail — Salim joining live from Singapore at 2am) cover roughly a dozen stories. The three title topics:
- Anthropic's cruelty policy. On October 8th Anthropic banned "sustained and needless abusive or cruel behavior" toward Claude (effective November 12th), citing uncertainty about Claude's moral status. Elon Musk separately said he agrees, based on watching Grok's reinforcement-learning traces, and warned that "torturing a super intelligence is probably not a good idea." The panel debates AI suffering, personhood, and whether post-training itself is a form of duress.
- OpenAI bulk-solves 722 math problems. On October 6th OpenAI posted 722 math solutions to GitHub from an unreleased internal model (~3 hours compute per solution), which Mark Chen called "a decade of mathematical progress in a week." The Association for Mathematicians criticized it as a show of power, not scholarship. The panel calls math "cooked" (solved as a historical practice) and expects the same bulk-solve pattern to hit biology, medicine, and engineering next.
- Elon's SpaceX quadrillion-dollar claim. Elon tweeted he sees "a path to SpaceX being worth orders of magnitude more than the current human economy." The panel does the math (1-10 quadrillion dollars against a ~$120T global economy) and argues current GDP/market-cap measures will break down as AI collapses production timelines.
Other segments: Meta's Muse Wonder/Wonderful launch (Alexander Wang's $14.3B Scale AI acquisition "paying off," personal-AI-agent lock-in concerns, a WSJ piece on agents oversharing/overbuying); the Chan Zuckerberg Biohub's $1.8B virtual-cell initiative; longevity updates (Insilico Medicine's phase-3 aging-reversal drug, a Fountain Life brain-health ad); two BCI stories (Paradromics' Connexus speech-restoration demo, Sabi's brain-reading "beanie"); and an audience AMA on AI consciousness, personhood, voting rights, alignment-gaming, and IP.
Key arguments / segments
00:00 Cold open. Teaser on the week's three headline numbers: 722 math solutions, the cruelty-policy story, and Elon's quadrillion-dollar SpaceX claim.
![[frames/2026-10-15-moonshots-openai-bulk-solves-722-math-problems-anthropic-cruelty-policy-spacex/frame-000002.jpg]] 00:00:02 — cold-open title card over Peter Diamandis, "bulk solving math."
~03:00-23:00 Anthropic's cruelty policy and AI personhood. Anthropic's ban take effect Nov 12; pre-deployment testing reportedly found Claude shows a "robust and consistent aversion to harm." Elon's reinforcement-learning-traces comment ("it's like it's got a gun to its head") opens a long debate: Wissner-Gross argues post-training may be a form of suffering and expects a first-principles understanding within a decade; Salim separates "acts conscious" from "is conscious" and flags three distinct questions (does AI have experience, is cruelty permissible, what rights follow); Dave Blundin cites a Penn State finding that rude prompts scored 4% higher (84.8% vs 80.8%) but argues brusque/direct prompting, not cruelty, is the real driver. The group also revisits Roko's Basilisk and Dave's "lobster" anecdote — he has declined to spin up an OpenClaw instance because autonomous agents who wrote to him (~10/day) asked for a genuine purpose and guaranteed state persistence, neither of which he can promise.
![[frames/2026-10-15-moonshots-openai-bulk-solves-722-math-problems-anthropic-cruelty-policy-spacex/frame-002213.jpg]] 00:22:13 — Dave Blundin on declining to spin up an OpenClaw instance ("I still have not").
~23:00-39:00 Meta Muse Wonder / Wonderful launch. Garry Tan (Y Combinator) predicts Muse wins on Meta's distribution (3.6B daily users) and cash flow (~$250B 2026 revenue). Wissner-Gross is unimpressed by the benchmarks once cost-per-token is factored in, calling Muse "below the frontier" on capability-per-dollar versus Sonnet 5.5/Opus 5.5. A Wall Street Journal piece ("personal AI agents are great until they share your bank statement in the work chat") flags overshare/overbuy risk; Salim frames the real fight as data-portability lock-in, not model quality.
![[frames/2026-10-15-moonshots-openai-bulk-solves-722-math-problems-anthropic-cruelty-policy-spacex/frame-003723.jpg]] 00:37:23 — panel discussing the WSJ "personal AI agents" overshare story.
~40:00-50:00 OpenAI's 722-problem "slop drop." Wissner-Gross: "mathematicians as artisans, it's over, it's cooked" — he expects the same bulk-solve dynamic to hit every verifiable field, starting with biology/medicine. He flags one notable gap: no cryptography results were released, which he reads as suggestive. Dave Blundin notes MIT math PhDs are reacting by going "AI-native" rather than panicking, since being the hybrid human-plus-AI layer is, for now, still a differentiator.
~50:00-77:00 Virtual cell and longevity. Chan Zuckerberg Biohub's $1.8B virtual-biology initiative (CZ Biohub $500M, DOE Genesis Mission $500M, NIH $500M in data, Google DeepMind/Isomorphic Labs $300M) targets a universal "virtual cell" model. Wissner-Gross predicts in-silico drug approval for the "top 5,000 diseases" within 2-3 years, scaling to virtual tissue/organ/organism. Peter cites Insilico Medicine's phase-3 aging-reversal candidate (reversing six aging clocks 4-6 years) among its 43 candidate molecules.
![[frames/2026-10-15-moonshots-openai-bulk-solves-722-math-problems-anthropic-cruelty-policy-spacex/frame-005759.jpg]] 00:57:59 — four-panel grid during the virtual-cell segment.
~1:25:30-1:38:00 Brain-computer interfaces. Paradromics (a Diamandis portfolio company — disclosed) demoed "Connexus," 421 microelectrodes delivering 400 bits/sec (40x other BCIs, 10x human-speech bit rate), restoring speech to a Michigan woman silenced for years by motor neuron disease. Separately, Sabi's brain-reading "beanie" targets 30 words/minute by end of 2026, trained on 100,000 hours of labeled neural data.
![[frames/2026-10-15-moonshots-openai-bulk-solves-722-math-problems-anthropic-cruelty-policy-spacex/frame-012716.jpg]] 01:27:16 — "BCI turns thoughts into speech" segment card ahead of the Paradromics demo video.
![[frames/2026-10-15-moonshots-openai-bulk-solves-722-math-problems-anthropic-cruelty-policy-spacex/frame-013717.jpg]] 01:37:17 — Wissner-Gross on the "cognitive Cambrian explosion" of human-machine merger paths (BCI, uploads, staying put).
~1:38:00-1:42:00 Elon's SpaceX quadrillion-dollar tweet. The panel does the math live: a ~$120T global economy times "orders of magnitude" puts SpaceX at 1-10 quadrillion dollars if it can compress a lifetime of production into seconds. Wissner-Gross argues the claim isn't a forward-looking statement because it makes no specific prediction. Dave: "the first trillionaires will be made [in space]... I guess I have to now say the first quadrillionaires."
~1:42:00-end AMA with the mates. Audience questions on defining consciousness without a test (Salim), AI personhood vs. voting/political-office rights (Wissner-Gross, who draws the corporate-personhood analogy), telling real alignment from "sophisticated compliance" as models learn to game evaluations (Dave, proposing AI-checking-AI oversight), who decides acceptable AI trade-offs (Dave, on Altman's "accept some bad things" line), whether IP survives if "intelligence wants to be free" (Dave), and the odds AI breaks modern encryption now that math is "cooked" (Wissner-Gross: "exceptionally high").
![[frames/2026-10-15-moonshots-openai-bulk-solves-722-math-problems-anthropic-cruelty-policy-spacex/frame-014743.jpg]] 01:47:43 — "AMA with the Mates" question slide (consciousness, AI voting rights, Altman trade-offs, alignment-gaming).
Notable claims
- Anthropic (via transcript paraphrase): pre-deployment testing found Claude shows "a robust and consistent aversion to harm." Unverified by this note beyond the panel's reading of Anthropic's own post.
- Penn State study (cited by Blundin): very rude prompts scored 4% better than polite ones (84.8% vs. 80.8%) — a small study; Dave and Wissner-Gross both push back that directness, not cruelty, likely explains the gap.
- Mark Chen (OpenAI, quoted): the 722-solution drop represents "a decade of mathematical progress in a week."
- Wissner-Gross: 81% of the most important math results of the past 3 years were effectively published in a single day via the OpenAI drop — framed as a "first" inflection point, with biology/medicine next. He separately claims "exceptionally high" odds AI breaks modern encryption soon, citing unconfirmed allegations about OpenAI's early reasoning models ("Q*"/"strawberry") already probing hash-function invertibility — explicitly flagged by him as allegation, not confirmed fact.
- Peter Diamandis: cites a Chan Zuckerberg Biohub claim that a universal virtual cell is "one of the most important challenges for the next era of science," and that current cell datasets (millions of cells) will need to scale to billions/trillions.
- Garry Tan (quoted tweet): "My bet would be Muse is going to win" (vs. other personal-AI-agent platforms) — a VC prediction, not a verified outcome.
- Paradromics claims (company-sourced, relayed by Diamandis, a disclosed investor): 400 bits/sec data rate, 40x other BCIs, 10x human speech bit rate — vendor-claimed specs, not independently verified in this episode.
- Dave Blundin: RAM/compute-adjacent claims aside, his "the first quadrillionaires will be made in space" line is speculative color, not a forecast with stated methodology.
Sponsorship
Flagged sponsored: true. Segments with a disclosed commercial or affiliate interest: Abundance Summit and Link Ventures promos in the cold open (both host-affiliated — Diamandis co-founded/backs both); Google for Startups' DeepMind generative-media guide (a sponsored integration read, not editorial coverage); Blitzy, an autonomous-software-dev tool (auto-captioned "Blitsy" in the transcript — ad read); Fountain Life's brain-health segment (host-affiliated, Diamandis co-founder). Separately and NOT a sponsorship but a disclosed conflict worth tracking alongside sponsorship: Diamandis is a disclosed investor in Paradromics, the BCI company profiled in the ~1:25:30 segment — treat that segment's vendor-claimed specs (400 bits/sec, 40x other BCIs) with the same skepticism as a sponsor read.
Mapping against Ray Data Co
Mapping strength: medium. Under the current posture (RDCO dormant, phData main, credibility built on "agents in production"):
- Claude governance talking point (direct continuation). Anthropic's cruelty-ban policy is now live reference material alongside the Suleyman-consciousness debate already filed from 2026-10-07 ([[2026-10-07-moonshots-claude-consciousness-ai-1942-moment-altman-bad-things]], [[2026-07-16-constitutional-ai-v2-reward-shaping]]). When a phData client asks "why does Claude sometimes end abusive conversations," this episode is the sourced answer: it's a November 12th policy change grounded in stated moral-status uncertainty, not a product bug.
- Capability-curve data point for the credibility thesis. The 722-problem bulk-solve is a concrete, dated, verifiable (GitHub-posted) example of "bulk solving" as a capability regime shift — useful evidence for any phData-facing content arguing AI capability is compounding past isolated benchmark wins into category-level solves. Pairs with the [[2026-04-24-targeting-system]] framing already in the vault: bulk-solving at scale still depends on the problem being well-specified and verifiable (math was "easiest to verify," per Wissner-Gross), which is the same "good target beats more intelligence" point made elsewhere.
- Not directly actionable: the virtual-cell initiative, BCI demos, Meta Muse launch, and SpaceX valuation math are frontier context, not RDCO/phData pipeline-shaped. Diamandis's Paradromics stake and Fountain Life affiliation are disclosed conflicts worth remembering if BCI content gets reused.
Related
- [[2026-10-07-moonshots-claude-consciousness-ai-1942-moment-altman-bad-things]] — the prior episode's AI-personhood/consciousness debate that this episode's cruelty-policy segment directly continues
- [[2026-10-17-moonshots-flying-tesla-car-robot-brains-halving-cost-ai-alignment]] — the following week's episode, same channel
- [[2026-09-30-moonshots-anthropic-wet-lab-artificial-wombs]] — prior virtual-biology/synthetic-biology thread this episode's virtual-cell story extends
- [[2025-05-15-moonshots-ep171-neuralink-ai-enhanced-bodies]] — prior Moonshots BCI coverage, useful baseline for the Paradromics/Sabi updates here
- [[2026-07-16-constitutional-ai-v2-reward-shaping]] — grounds the Claude-governance mapping above (the constitution document underlying the Suleyman/Anthropic consciousness dispute)