Who he is
Affiliation: Co-founder & Chief Scientist, Physical Superintelligence PBC ("PSI") — launched publicly 2026-09-01 with a $58M seed round (co-founders Matt Pines, CEO, and Alex Klokus; PRNewswire, 2026-09-01). Previously founder of Gemedy; Harvard PhD (physics); MIT background; co-host, Moonshots podcast (with Diamandis, Blundin, Ismail); co-author with Diamandis of Solve Everything (Feb 2026); writes the daily AI newsletter The Innermost Loop (~100K readers).
One-line position: Intelligence is a physical phenomenon — maximizing future freedom of action — and AI capability follows from compute, data, and algorithmic insight more than from any specific architecture; reason about it quantitatively, not narratively. New corollary (2026): aim superintelligence at physics itself and it "solves everything else."
Discipline & technical bet
Physicist by training ("causal entropic forces": intelligent behavior as maximization of future option space — Physical Review Letters, 2013). The concrete wager now is no longer academic: PSI is a public benefit corporation staffed with "virtual physicists" (AI agents doing physics research), betting that AI can industrialize the discovery of new physical laws — near-term product: data-center design optimization; long-term: new physics across the energy-to-compute spectrum. He has, in effect, turned his 2013 theory of intelligence into a funded company.
Key claims (Says)
- Empirical / theoretical: causal entropic forces — intelligence ≈ a physical system that maximizes its future option space (TED Talk 2014; Physical Review Letters 2013). Held — and now operationalized: PSI's stated ambition ("higher-fidelity world models… discovering new physical laws") is this thesis as a business plan (PRNewswire, 2026-09-01).
- Empirical: across the history of AI, each major capability breakthrough trailed a key dataset by ~3 years on average — datasets, not algorithms, have been the rate-limiter. Held — his own venture's first move is building physics world-model data/simulation capacity, consistent with the thesis.
- Empirical: AI capability arises from the product compute × data × method; no single one is the master variable. Held.
- Empirical: capability gains will continue but are likely to encounter diminishing returns at some scale frontier — refuses to specify when. Unfalsifiable as stated — retired to watch-only per validation rule (no observable, no date); no longer graded.
- Mixed: AI risk arguments should be quantified (probabilities, mechanisms, falsifiers) not narrated; takes risk seriously without being a doomer. Held — still the calibrating voice on Moonshots (58+ episodes since Jul 2025).
- Empirical: the next bottleneck shifts from algorithmic novelty to data quality, multimodal grounding, and embodiment. Held — revealed-behavior confirmation: he founded a company whose product is physical-world grounding.
- Normative: framing AI policy in physics-grounded, falsifiable terms is more productive than ideological camp-staking. Held.
- NEW (2026): "Superintelligence, aimed well, solves everything else" — Solve Everything (Feb 2026, with Diamandis) is "a blueprint for aiming the Singularity at every problem that has ever made human life short, expensive, or unfair," target ~2035.
Notable predictions — with falsifiable checks
- (2017+) "Datasets, not algorithms, are the rate-limiter." Largely vindicated through the 2020s (ImageNet → LAION → web-scale corpora → RLHF data). Ongoing check: does the next frontier jump (world models / physics) trail a dataset or a method?
- (2013/2014) Causal entropic forces as a model of intelligence. Intellectually influential in a niche; not a frontier-lab training principle. New check: does PSI's Emmy platform produce a peer-reviewed novel physics result by 2028-12-31? That is now the live falsifier for the whole research program.
- (Multiple) Multimodal grounding and embodiment as the next frontier. Directionally consistent with 2024–26 releases; his own founding of PSI is the strongest (self-interested) confirmation.
- (2026-02, Solve Everything) Major human problems solvable by ~2035 if superintelligence is aimed well. Check (per-branch, since aggregate is unfalsifiable): (a) an AI-originated physics or materials result with industrial deployment by 2030; (b) his own PSI delivering a measurable data-center efficiency gain publicly credited to AI-discovered physics by 2028. Review annually.
- (Retired) Diminishing returns at some unspecified scale frontier — no observable + date; watch-only.
Revealed behavior (Does)
- Co-founded Physical Superintelligence PBC; $58M seed announced 2026-09-01, led by Breakthrough Energy Ventures with Dragon Global, Robot Ventures, SV Angel, Valkyrie, Balaji Srinivasan, Scaramucci, and angels from OpenAI/NVIDIA/Oracle orbits. ~20 researchers from Google, OpenAI, Harvard, MIT, IAS. First application: terrestrial and orbital data-center design; technical partner for the Fermi Explorer interstellar mission concept.
- Shipped "Get Physics Done" (open-source AI assistant for physicists, Jun 2026) — code before claims; consistent with his falsifiability rhetoric.
- Published Solve Everything with Diamandis (Feb 2026) and scaled The Innermost Loop to ~100K daily readers — he is now a full participant in the abundance-media economy, not just its calibrator.
- 58+ Moonshots episodes since Jul 2025 — time allocation says media + company, no longer quiet consulting.
Feels
Wants the universe to be legible — and wants to be the one who made it computable. The move from theory to company suggests impatience with being the smartest commentator in the room; he wants a result with his name on it.
Hears
Frontier-lab research output daily (he digests it for 100K readers); the Diamandis abundance network; now also his cap table — Breakthrough Energy and deep-tech investors expecting physics, not essays.
Sees
The whole AI literature at newsletter cadence — few people metabolize more papers per day — plus, now, whatever PSI's internal agents produce. His vantage: pattern-matching across physics and ML before either community connects the dots.
Incentive map
Changed materially in 2026. Previously the low-conflict calibrator of the Moonshots crew (v1 gave him extra weight per CONVENTIONS rule 25). Now: equity in a $58M-funded AI-physics startup, a co-authored abundance book to sell, and a large newsletter audience. He can no longer easily say: that AI-for-science is overhyped; that world models are far off; that the Singularity-aimed-at-everything frame overpromises. Rule 25 discount now applies to him at full strength — grade his AI-for-physics claims like Diamandis's longevity claims.
Theories aligned with
- Techno-optimism — measured / scientific variant, now with commercial skin in the game
- Physics-based / first-principles view of intelligence (causal entropic forces)
- Soft adjacency to AI alignment / x-risk — engages risk arguments seriously without doomer framing
- "Datasets, not algorithms" thesis — distinct, his framing
What he's reacting against
- Architectural triumphalism — claims that this architecture (transformers, diffusion, etc.) is uniquely the path to general intelligence
- Narrative-driven AI risk discourse divorced from physical / computational quantification
- "Algorithms are the bottleneck" framing — argues data has been the historical rate-limiter
- Hype cycles that conflate demos with deployment
- Pure-doomer or pure-utopian framings on Moonshots — provides the calibrating scientific voice (though see Incentive map: the calibrator now has a book and a startup in the fight)
Where he overlaps / splits (with Rich)
- Overlaps with Diamandis, Blundin, Ismail, Mostaque as part of Moonshots crew; provided scientific calibration to their bolder framings — the gap is narrowing now that he co-authors the abundance blueprint
- Overlaps with Brynjolfsson, Cowen on capability-vs-impact gap; splits with Brynjolfsson on which complementary investments matter most (data infrastructure vs org redesign)
- Overlaps with Bengio, Amodei, Hassabis on taking risk seriously; splits on policy specifics — Wissner-Gross less prescriptive
- Splits with Yudkowsky on probability framing — argues for quantification without committing to high-doom priors
- Splits with Andreessen on rhetorical style; aligned on direction, opposed on tone
- Splits with Altman on commitment to specific timelines — emphasizes uncertainty bounds (though Solve Everything's 2035 target is his most timeline-shaped commitment yet)
- For Rich's bench: PSI's first product is data-center design optimization — squarely in the picks-and-shovels/infra lens. Watch whether AI-discovered physics actually moves datacenter economics; that would be a new infra sub-thesis.
Track record
- Physical Review Letters 2013 paper on causal entropic forces: peer-reviewed, novel framing; influence durable in a small community, not mainstream
- "Datasets are the rate-limiter" thesis: directionally vindicated; consistent with subsequent ImageNet → LAION → web-scale corpora pattern
- TED Talk (2014) on the equation for intelligence (F = T·∇Sτ): high reach, mixed scientific reception
- Raised $58M for PSI at launch (2026-09) with top-tier deep-tech investors — fundraising execution now demonstrated; scientific output of the company: entirely unproven
- Pattern through 2025: low-frequency, high-precision claims; 2026 marks a regime change to high-frequency public output (daily newsletter, book, podcast, startup)
Empirical vs normative
- Empirical: causal entropic forces; datasets-as-rate-limiter; compute × data × method as joint capability driver
- Normative: AI policy and discourse should be quantitative and falsifiable, not narrative; superintelligence should be aimed at humanity's material problems (Solve Everything)
- Superseded (was v1's key weighting): "less commercial conflict than other Moonshots co-hosts — more weight per rule 25." No longer true: book + newsletter + $58M startup put him at full rule-25 discount on AI-for-science claims.
Weak spots / open questions
- Causal entropic forces as a general theory of intelligence is contested — durable in a niche but not adopted in mainstream ML; PSI is now the live experiment
- "Datasets, not algorithms" is broadly right historically but the algorithm/architecture distinction has mattered more recently (transformers, diffusion, MoE) — track whether the rate-limiter shifts
- Can ~20 researchers + AI agents outrun DeepMind/academia at physics discovery? "Virtual physicists" is a marketing frame until Emmy produces a result someone else replicates
- Less prescriptive on policy than Bengio/Amodei — useful for calibration, less so for action recommendations
- The v1 note "public source footprint is thinner than other Moonshots crew" inverted in under a year — he now publishes daily; the new problem is volume over signal
- "Solve everything by 2035" inherits the unfalsifiability problem his own rhetoric warns against — per-branch checks added above; hold him to them
Rich's take
- (your synthesis here)
Delta log
2026-09-04 — v1→v2 migration + validation (weekly batch)
- Grades: 5 Held / 1 Drifted / 1 Wrong (one claim retired as unfalsifiable).
- Wrong "Less commercial conflict than other Moonshots co-hosts; more weight per rule 25" — reversed within a year: Solve Everything (Feb 2026, with Diamandis), The Innermost Loop (~100K daily readers), and co-founding Physical Superintelligence PBC with a $58M seed announced 2026-09-01 (also Unite.AI, HPCwire).
- Drifted "Public source footprint thinner than other Moonshots crew" — now a daily newsletter, a book, and 58+ podcast episodes (profile/episode index).
- Held Causal entropic forces; datasets-as-rate-limiter; compute×data×method; quantify-risk stance; embodiment/grounding as next frontier (the last now confirmed by revealed behavior — he founded the company).
- Retired: "diminishing returns at some unspecified frontier" (no observable + date).
- New falsifiable checks added: PSI/Emmy peer-reviewed physics result by 2028; AI-credited datacenter efficiency gain by 2028; Solve Everything per-branch 2030 checks.
- Most surprising delta: the Moonshots crew's designated low-conflict calibrator became, in ~7 months, a founder with $58M and a book to sell — the incentive map inverted while the persona sat unreviewed.
- Tier: quarterly (new company, fast-moving claims, fresh capital). Wake triggers: PSI funding/product/results, Fermi Explorer milestones, new book/major essay, job change — one line each: company news moves his incentive map; mission milestones test the thesis; publications restate the claims; job changes are the canonical miss.
Sources
- Peer-reviewed: "Causal Entropic Forces" (Wissner-Gross & Cameron Freer, Physical Review Letters, 2013-04)
- Talks: "A new equation for intelligence" — TED (2014)
- Company: PSI launch release (2026-09-01) · Unite.AI on the $58M seed · HPCwire (2026-09)
- Book: Solve Everything (with Diamandis, 2026-02)
- Newsletter: The Innermost Loop (daily) — incl. his PSI announcement post
- Podcast: Moonshots with Diamandis, Blundin, Ismail, Mostaque — episode index
- Profile: LinkedIn · CV · earlier Gemedy materials (superseded as primary affiliation)
migrated v1→v2 2026-09-04 (weekly validation batch) · backup: _archives/alex-wissner-gross.html.bak-20260904 · AI & Society domain