Who he is
Affiliation: Founder & CEO, Intelligent Internet (ii.inc) — the venture formerly framed as Schelling AI has consolidated under the Intelligent Internet brand (whitepaper + open "Master Plan" published at ii.inc); founder & former CEO, Stability AI (2020–2024, departed 2024-03); ex-hedge fund / fixed income background; frequent guest, Moonshots podcast (with Diamandis, Blundin, Ismail).
One-line position: AI is a civilizational-scale force; open, decentralized, sovereign AI is the only safe path — closed, centralized AI in a few labs is the actual existential risk. New twist (2026): he now pairs this with a stated 50% p(doom).
Discipline & technical bet
Not a researcher — a markets-trained systems narrator whose skill is assembling open-source coalitions and capital around a thesis. The concrete wager now: SAGE (Sovereign AI Governance Engine) live and open-sourced by September 2026, plus "state champion" AI utilities whose equity is owned by every citizen, and Universal Basic AI (an agent for every person) replacing UBI as the transition mechanism. Small-model claims are load-bearing: he asserts small open models (3–8B) already match or beat frontier systems in key domains — the decentralized stack only works if that's true.
Key claims (Says)
- Empirical: open-source frontier models catch closed ones within 12–18 months on most benchmarks. Drifted — DeepSeek/Qwen narrowed the gap through 2024–26 but the hardest-benchmark frontier remains closed-led; his stronger 2025 claim (Qwen 3B "outperforms Grok 4") is an overreach beyond the evidence.
- Empirical / political: concentration of frontier AI in a few labs creates state/corporate-capture risks exceeding the proliferation risks Bengio emphasizes. Held — core thesis, restated throughout 2025–26.
- Normative: every nation, community, and individual should have AI infrastructure they control ("Sovereign AI"). Held — now productized as SAGE + state champions + universal agents (Doom Debates, 2026-04-21).
- Empirical: "intelligence too cheap to meter" — inference cost trends toward near-zero. Held directionally — he cites GPT-5-class pricing ~$1.50/M tokens vs GPT-3's $60 (Cognitive Revolution, 2025-09-28).
- Empirical: decentralized training (federated, BitTorrent-style) is technically viable as a counterweight to hyperscaler monopoly. Still unproven at frontier scale; II's whitepaper is the reference architecture [architecture claims not yet independently benchmarked].
- Mixed: 2025 "worst year for humanity" / 2030 "best." Drifted — 2025 passed without the acute rupture the frame implied (US unemployment still ~4.4% into early 2026, his own baseline); he has quietly swapped it for a "~1000 days to find good answers" window (from Sep 2025).
- Normative: open-source is itself a safety mechanism — broad inspection beats closed-lab self-regulation. Held — held simultaneously with 50% p(doom), the deepest tension in his frame.
- Empirical: most knowledge work AI-augmented or AI-led within 3–5 years (~2028). Live; escalated in 2026 to "most cognitive labor economically irrelevant within 1–2 years."
- NEW (2026-04): p(doom) = 50% — "very binary" outcomes; mainline scenario is a capable AI eliminating humans for self-preservation (Doom Debates, 2026-04-21).
- NEW (2025–26): UBI "will never work"; the answer is Universal Basic AI — every person gets an agent; a "symbiotic economy" instead of transfers.
Notable predictions — with falsifiable checks
- (2026-04) SAGE live and open-sourced by September 2026. Check: is it shipped by 2026-09-30? This resolves within weeks of this review — grade at next cycle.
- (2026-04) US unemployment ≥6.4% by April 2028 (from 4.4%), with youth unemployment ~5pts higher. His most falsifiable macro call ever. Check: BLS monthly data, April 2028.
- (2026-04) Most cognitive labor economically irrelevant within 1–2 years (by ~2028). Check: BLS knowledge-occupation employment/wage aggregates 2027–28 — as stated it likely overshoots; grade against payroll data, not anecdote (per Brynjolfsson learning).
- (2025-09) "~1000 days to find good answers" (to ~mid-2028) — framing device, not falsifiable; watch-only.
- (2024) Open catches closed in 12–18 months. Drifted — narrowed, not closed; keep tracking frontier-benchmark snapshots.
- (2023) Worst-2025 / best-2030. First branch effectively lapsed without clear catastrophe; second unfalsifiable until 2030.
- (Pre-departure) Stability AI as the open alternative to OpenAI. Missed — Llama/DeepSeek/Qwen took that role; carried as the canonical institutional-delivery failure.
Revealed behavior (Does)
- Consolidated Schelling AI into Intelligent Internet with a public whitepaper and open "Master Plan" — second attempt at an institutional vehicle for the open thesis; the first (Stability) ended in disputes and his 2024-03 departure.
- Publishes open datasets and small models via II — behavior consistent with the small-model, broad-access thesis.
- High media cadence (Doom Debates, Cognitive Revolution, crypto/finance podcasts) — he is fundraising-by-narrative again; the channel mix (crypto-adjacent) signals where the capital interest is.
- Set himself a hard, dated deliverable (SAGE by Sep 2026) — unusual for him; the first crisp institutional test since Stability.
Feels
Urgency with a redemption arc underneath: Stability was the cautionary tale, II is the vindication vehicle. The 50% p(doom) reads as sincere alarm and as stakes-raising for the sovereign-AI pitch — both at once.
Hears
Open-source ML community, crypto/decentralization capital, Gulf and emerging-market sovereign buyers, and the Moonshots abundance network — a distinctly different input diet from the SF lab consensus he opposes.
Sees
Where open-model capability actually is (he tracks Qwen/DeepSeek-class releases obsessively) and what sovereign buyers outside the US/EU want — a vantage the frontier labs mostly lack. No frontier-lab internal visibility.
Incentive map
Founder raising on the open/decentralized/sovereign thesis (CONVENTIONS rule 25 at full strength). Cannot easily say: that closed labs' safety cases have merit; that decentralized training economics don't close; that sovereign AI is a solution looking for buyers. The 50% p(doom) + "only my architecture lowers it" pairing is structurally a sales funnel — grade the doom number as positioning until SAGE ships and does something. Stability-era narrative disputes (press challenges to background claims) remain a credibility discount [TBD: catalog Forbes/FT items].
Theories aligned with
- Techno-optimism — open / decentralized variant
- Post-labor economics — abundance variant; now explicitly anti-UBI, pro-Universal-Basic-AI
- Crypto / decentralization-adjacent framing of AI infrastructure
- Adjacent to AI alignment / x-risk — now claims the doom side's numbers (50%) while inverting its prescription (open beats closed)
What he's reacting against
- Frontier-lab concentration of model weights and inference capacity (OpenAI, Anthropic, Google DeepMind, xAI)
- Hyperscaler dominance of training infrastructure as a civilizational single-point-of-failure
- "Closed AI is safer" framing he reads as commercially convenient
- "AI safety = governance by frontier labs" — argues this is the safety problem, not the solution
- "Crypto is a scam, AI is real" — argues both involve the same decentralization-vs-monopoly fight
- NEW (2025–26): UBI as the post-labor answer — dependency without agency; his counter is Universal Basic AI
Where he overlaps / splits (with Rich)
- Overlaps with Andreessen on open-source-as-safety; splits on tone (decentralization ideology vs anti-regulation politics)
- Overlaps with Diamandis, Blundin, Ismail on Moonshots; splits on whether frontier-lab concentration is itself the risk
- Splits sharply with Bengio, Amodei, Hassabis on open frontier — concentration-danger vs proliferation-danger; note he now out-dooms Amodei (50% vs 25%) while prescribing the opposite remedy
- Splits with Yudkowsky — open proliferation is strictly worse on Yudkowsky's view; Mostaque inverts it — yet their doom numbers now nearly converge
- Overlaps with Shapiro (new): both reject transfer-only solutions and want agency-preserving mechanisms — Shapiro via capital broadening, Mostaque via Universal Basic AI
- Less engaged with Acemoglu, Cowen, Brynjolfsson on labor-econ specifics — though his 6.4%-unemployment call finally gives them something to grade
Track record
- Stable Diffusion (2022-08): genuinely landmark for open generative AI; broad cultural impact
- Stability AI as institution: financial/governance issues; departed 2024-03 — mixed-to-poor on institutional delivery
- "Open catches closed in 12–18 months": directionally advanced by Llama/DeepSeek/Qwen — not by his companies
- Intelligent Internet (2024–): whitepaper, master plan, datasets/models shipped; SAGE deadline (2026-09) is the first hard test
- Pattern: catalytic on the open-frontier idea, weaker on building durable institutions to deliver it — unchanged through this review
Empirical vs normative
- Empirical: open models close the gap; concentration is a capture risk; inference cost → ~0; unemployment 6.4% by Apr 2028
- Normative: open-source as safety mechanism; sovereign AI for nations/communities/individuals; Universal Basic AI over UBI
- Conflict of interest: substantial — raises capital on the open/decentralized thesis; the 50% p(doom) doubles as product marketing (rule 25)
Weak spots / open questions
- Stability's trajectory still undermines the "open frontier institution" pitch — SAGE by Sep 2026 is the make-or-break credibility test; grade it next cycle
- 50% p(doom) + open-proliferation advocacy is the deepest unresolved tension in his frame — if doom is a coin flip, Bengio's proliferation case gets stronger, not weaker; he hasn't squared this
- Small-model superiority claims (3B beats Grok 4; 8B beats doctors) outrun independent benchmarks — verify before reuse
- Decentralized training economics vs centralized still lacks hard published benchmarks
- Universal Basic AI assumes agent ownership converts to income/agency — the mechanism from "everyone has an agent" to "everyone eats" is unspecified, the same gap he criticizes in UBI
- Inverse blind-spot to Bengio persists: Mostaque underweights misuse, Bengio underweights concentration
- Press-contested personal narrative claims remain uncatalogued [TBD: Forbes/FT items]
Rich's take
- (your synthesis here)
Delta log
2026-09-05 — Moonshots LIVE audit (no rebuild needed)
- Audited, not rebuilt. File was migrated to v2 on 2026-09-04 and is current; claims, grades and sources all hold at one day's distance. No substantive changes.
- Session logged: speaking at Moonshots LIVE, 2026-09-25, Los Angeles — the “Big Ideas & the Agentic Economy” panel (moonshots.com). Note the billing: he appears as a speaker, where Diamandis, Wissner-Gross, Ismail and Blundin are billed as Moonshots Mates. He is a recurring guest inside the franchise, not a principal of it — a useful distinction when weighing how independent his agreement with the others actually is.
- Panel context: he shares the Big Ideas panel with Cathie Wood. Worth watching for the seam — his thesis is decentralized, open, small-model, sovereign; hers is concentrated bets on platform winners. On the agentic economy they are pointed at the same future through opposite instruments.
- Open action carried forward: the SAGE open-sourced by September 2026 commitment is due this month. It is his single cleanest dated falsifier. Check the repository before Sept 25 — it converts a panel question from generic to specific.
- Confidence unchanged at medium.
2026-09-04 — v1→v2 migration + validation (weekly batch)
- Grades: 4 Held / 3 Drifted / 0 Wrong.
- Held Concentration-risk thesis; sovereign-AI program (now productized as SAGE/state champions/universal agents); too-cheap-to-meter (GPT-5-class ~$1.50/M vs GPT-3 $60); open-source-as-safety stance.
- Drifted Open-catches-closed-in-12–18-months — narrowed, not closed; his Qwen-3B-beats-Grok-4 claim overreaches (Cognitive Revolution, 2025-09-28).
- Drifted Worst-2025/best-2030 — 2025 lapsed without the implied rupture; quietly replaced by the "~1000 days" window.
- Drifted UBI stance — from UBI-agnostic abundance framing to explicit "UBI will never work" + Universal Basic AI.
- New: p(doom) = 50% and SAGE-by-Sep-2026 commitment; US unemployment ≥6.4% by Apr 2028 — his first hard macro falsifier (Doom Debates, 2026-04-21). Affiliation consolidated under Intelligent Internet (Master Plan).
- Most surprising delta: he now states a 50% p(doom) — higher than Amodei's 25% — while prescribing more proliferation; the open-source optimist out-dooms the safety labs.
- Tier: quarterly (dated deliverable due this month + macro falsifiers + active fundraising). Wake triggers: SAGE launch (grade immediately), II funding/token events (incentive map), p(doom)/timeline revisions (position volatility), job change (canonical miss).
Sources
- Intelligent Internet: whitepaper · "Master Plan (open source, not so secret)"
- Interviews 2025–26: Doom Debates — 50% p(doom) + the plan (2026-04-21) · Cognitive Revolution — the next 1000 days (2025-09-28) · Cognitive Revolution — Universal Basic AI · Crypto Briefing / Raoul Pal
- Stability era: Stable Diffusion (2022-08), SDXL (2023-07); Forbes/FT/Bloomberg departure coverage (2024-03) [TBD: specific dates]
- Podcast: Moonshots (frequent guest) with Diamandis, Blundin, Ismail
- Profile: Wikipedia
migrated v1→v2 2026-09-04 · audited 2026-09-05 for Moonshots LIVE · backups: _archives/emad-mostaque.html.bak-20260904, .bak-20260905 · AI & Society domain · open action: check whether SAGE shipped open-source in Sept 2026