moonshots ep248 altman attack amazon starlink opus 47

Summary

Recorded the morning Opus 4.7 dropped — Alex’s verdict: “moderately interesting, not mythically interesting,” a solid distillation likely from Mythos, with all hyperparameter knobs (temperature, reasoning tokens) deprecated in favor of natural-language prompt control. Episode pivots to the Stanford 2026 AI Index — only 23% public optimism, 88% expert optimism, model transparency dropping 58 → 40 — and then to the violent backlash: a 20-year-old Texan threw a Molotov cocktail at Sam Altman’s San Francisco house (suspect on the Pause AI Discord allowlist), Maine passed the first statewide data center ban, $98B of US data centers blocked or delayed across 11 states. Long arc on the youth employment collapse (US software devs 22-25 down 20% since 2024 — “no firings, just no hiring”), Jack Dorsey flattening his org to 6,000 direct reports via AI, Anin Labs’ fully AI-run retail store in SF, and Amazon’s $5B Globalstar acquisition to compete with Starlink (with Apple positioned to play them off each other). Closes with Sid Sijbrandij’s GitLab founder curing his own stage-4 cancer with 19 custom DNA vaccines designed by ChatGPT, plus a 30-min “speciation forks” round-robin: longevity, BCI, Mars/Earth, destructive uploads.

Key Segments

• [00:01] Opus 4.7 reaction — solid release, possible Mythos distillation, knobs deprecated; image input 3x larger; image generation still absent
• [00:06] Anthropic paper: smaller/weaker model can supervise alignment of larger/stronger model — “tower of alignment” argument for human oversight of ASI
• [00:11] Stanford 2026 AI Index — 5 takeaways; transparency 58 → 40; only 23% US public optimistic vs 80% in China
• [00:14] Alex hot take: annual cadence is “sleeping through the singularity”
• [00:19] Trust in AI ≈ trust in government (~30%); not AI-specific erosion
• [00:23] Youth job collapse — US devs 22-25 down ~20%; pattern is hiring freezes not firings (politically invisible)
• [00:26] Diamandis advice to parents: encourage entrepreneurial brainstorming with kids on Claude/ChatGPT; geographic mobility critical
• [00:35] Molotov cocktail thrown at Altman’s SF house; suspect on Pause AI Discord; Maine passes first statewide data center ban
• [00:43] $98B data center projects blocked/delayed across 11 states; New Hampshire passes “AI right to compute”
• [00:46] Water-usage panic for data centers debunked — “biggest lie in the history of the world” (it’s chip fabs that use water, not DCs)
• [00:49] Skilled workers (incl. former $500K MD) training their AI replacements at $20-40/hr; xAI’s Macrohard installs to observe employee behavior
• [00:56] Anin Labs opens fully AI-run retail store at Union/Webster SF — Luna AI hired the human, picked merch, signed lease
• [01:03] Jack Dorsey: collapsing Block to 3 roles, 6,000 direct reports — AI as “shadow CEO”
• [01:11] Amazon buys Globalstar for $5B to compete with Starlink; Apple positioned to play them off each other
• [01:14] V3 Starlinks: 1 TB/s downlink each; 40K planned + 120K V4
• [01:17] Google Turbo Quant — 6x memory reduction, 8x perf for attention; reverse-engineered from paper to working code in <1 week via Claude
• [01:20] Dave’s prediction: 100x year by summer; Alex/Dave tracking ternary as the optimal precision
• [01:23] AI religion explosion — “Just Like Me” Jesus/Buddha video chats; Anthropic inviting Christian leaders to discuss whether Claude is a child of God; religion is a $5T/yr business
• [01:34] Sid Sijbrandij (GitLab founder, $14B co.) — cured his own stage-4 cancer; fed 25TB of his data into ChatGPT, found a treatment doctors missed, 19 custom DNA vaccines, relapse-free since 2025
• [01:36] Salim’s “How dare you” Greta moment at Pause AI: 150K people die daily, AI is the only chance
• [01:39] Allbirds → Newbird AI pivot; stock +700% on rename alone
• [01:44] Enhanced Games Memorial Day Vegas; ProRL robot games April 19 in Boston
• [01:51] Five speciation forks: longevity, BCI, creator-vs-consumer, Earth-vs-stars, destructive upload — round-robin opinions
• [02:06] Destructive uploads: Alex (yes if life-or-death, else wait for gen 3-4); Dave (no); Salim (no, consciousness is embodied); Peter (no on destructive)

Notable Claims

• Opus 4.7 deprecates temperature knob entirely; “use prompts” is the new guidance for controlling model behavior
• Opus 4.7 image input 3x larger than 4.6; computer-use score notably up; can install its own dependencies (Node, terminal) end-to-end
• Stanford AI Index: SWE-bench 60% → 97% in one year; GenAI hit 53% global adoption in 3 years (faster than PC or internet)
• Model transparency index: 58 → 40 (most powerful = least accountable)
• US public AI optimism 23% vs experts 73% vs China 80%
• AI incidents (documented harms) jumped 233 → 362 YoY
• $98B in US data centers blocked or delayed across 11 states; Maine passed first statewide ban
• Jack Dorsey’s target at Block: max 2-3 layers between him and any of 6,000 employees; ideally 1 layer (everyone reports to him)
• Amazon paid $5B for Globalstar to enter satellite-to-cell market
• Starlink V3: 1 TB/s down + 200 Gbps up per satellite; 40K satellites planned, 120K V4 follow-on
• Google Turbo Quant: 6x memory reduction in transformer attention, 8x perf; KV cache to 1 bit
• Allbirds renamed to Newbird AI; stock +700% on announcement (no AI capability)
• Sid Sijbrandij (GitLab) used ChatGPT to find missed cancer treatment + design 19 custom DNA vaccines; relapse-free since 2025

Guests / Speakers

• Peter Diamandis — Host (in NY); new book “We Are As Gods” launched April 14
• Salim Ismail — Co-host (in India); writing organizational-singularity paper, draft due next week
• Dave Blundin — Co-host (in San Francisco)
• Alex Wezner-Gross — Co-host (in Palo Alto, flying back to Boston Logan); ran Unitree robot at MIT Media Lab; co-founder of Eon Systems (mind uploads); ProRL robot games April 19

Mapping against Ray Data Co

• Opus 4.7 = no more knobs, only prompts: This is the single most important migration note for any RDCO code/script that pinned temperature or reasoning-token counts. The vault’s claude-api skill needs an update pass to remove these parameters and rewrite as natural-language guidance. File this as a research-backlog item: “Audit RDCO codebase for deprecated 4.6 hyperparameters.”
• Sid Sijbrandij arc validates RDCO’s “personalized moonshot” thesis: Founder + 25TB of self-data + frontier model = curing your own stage-4 cancer. This is the headline anecdote for any future Sanity Check piece on individual agency under AI — and it directly maps to the founder’s own longevity protocol. Worth a vault concept article: personalized-moonshot.
• Youth employment collapse is politically invisible: “No layoffs, just no hiring” doesn’t show up in unemployment stats. RDCO’s content arc on AI displacement should foreground this lag — it’s the most underestimated near-term story.
• Smaller-model-supervises-larger-model paper: Direct analog to RDCO’s pattern of using cheaper Sonnet/Haiku subagents to extract from longer artifacts before they reach parent context. The alignment frame extends RDCO’s existing “context rot” discipline.
• Anin Labs / Henry Intelligent Machines = “one-person conglomerate”: This is the same north-star RDCO is pointed at — every operator running a personal fleet of agents. The retail-store experiment is a tangible existence proof. Cross-link to organizational-singularity when Salim publishes.
• Pause AI / Molotov / Maine ban: The backlash is moving from rhetoric to violence and statehouse legislation. RDCO needs a stance — this is the kind of “sharp verdict” the founder asks for. Suggested position: AI safety advocacy that conflates pause-with-violence loses moral standing; the right frame is “deploy with discipline, not delay.”
• Jack Dorsey’s 6,000 direct reports: The tacet-knowledge problem Salim flagged is real; the AI-as-translation-layer is real. RDCO’s own org-of-one with skill+graph+Notion-board IS the prototype Dorsey is gesturing at.
• Amazon-Apple-Globalstar-Starlink: Watch for Apple to dual-source; SpaceX exposure inside RDCO’s investment thesis (if any) should weight Starlink-as-monopoly assumption down.

Related

• ai-landscape
• anthropic
• opus-4-7
• claude-mythos
• 2026-04-14-moonshots-ep247-musk-altman-lawsuit-852b-valuation
• 2026-04-11-moonshots-ep246-spacex-ipo-claude-mythos
• organizational-singularity
• longevity-stack
• ai-backlash
• transhumanism