What this episode opens for us — money angles, ways in, and what our own agents can build within days. One card is the reality check.
Interstellar just got a price tag — all self-reported (StarCloud funds it; Alex co-founded partner PSI): $10-15M, a ~100kg smallsat that is ~64% xenon, ~$500K Falcon 9 rideshare, a "perihelion pump" exploiting the Oberth effect, 99% of the way to Alpha Centauri within 80,000 years, missing by no more than 2,600 AU. On tape: "every single nation state... can afford to send their own probe to another star system." Alex pre-registers his bet on a latent commercial interstellar market; the tape's precedent: five grams of cremains flown to orbit.
▸ Write the FKG new-market brief: size nation-state and private probe demand at the self-reported price points, catalog PSI's open-source Get Physics Done package as the enabler, and log Alex's commercial-interstellar bet in the prediction registry with falsification terms and dates.
$ First verified market map of sub-$15M starflight, sold to space funds and agencies.
PSI's Fermi trajectory burned about 10 billion total tokens (Matt's on-tape figure — self-reported, and inconsistent with the 'billion tokens over three days' told to MIT Tech Review; Alex co-founded PSI) against five-to-six hours of human time in a week. Dave sees the same shape in-house: roughly 10 billion thinking tokens per 100,000 tokens of final output, across unrelated projects. Alex proposes the scale: "that was a level nine problem, 10 to the nine tokens." Dave values one such run at thousands of people working ten years. That is RDKG's pricing yardstick: sell sweeps by problem level, not hours.
▸ Publish RDKG's rate card in problem levels — level-7 capsule pack through level-9 frontier brief — each deliverable stamped with tokens burned, sources verified, and freshness dates; run one level-8 sweep this week as the demo and print its own token meter on the cover.
$ Level-priced foresight; buyers on tape already "budgeting their projects by the token."
The episode's two biggest "news" stories are the hosts' own ventures. Alex co-founded PSI, which announced its $58M seed (led by Breakthrough Energy Ventures) the same day its Fermi mission opened the show. On Architect: Alex advises, it's a Link Ventures portfolio company, Peter's in the fund — "We're all guilty," on air. They disclose; nobody indexes it. Every venture number on this tape is self-reported, and no machine reading the transcript can tell disclosure from endorsement. The disclosure map is a product the gardens already build every episode.
▸ Productize the per-episode Disclosure Map: every claim tagged verified / self-reported / host-interest as JSON-LD citation capsules, llms.txt-listed; ship EP285's today, pitch aggregators and AI-answer engines the badge layer, and let agents regenerate it within hours of each new episode.
$ Honesty badges are licensable; disclosure-tagged claims are premium agent-priced inventory.
The Cursor cutoff is real — OpenAI's stated reason: "Elon Musk's companies violating contracts"; Anthropic pledged Claude support "within hours," per tape. Two claims fail the record. Dave's "Anthropic on AWS still transmits all of your intellectual property to Dario for 30 days" dramatizes a verified Bloomberg story (Aug 20): Anthropic is letting enterprise customers keep data on their own cloud infrastructure; the ~30-day retention window stays, the everything-to-HQ framing doesn't. And Dave's Colossus line checks out: Anthropic rents all of Colossus 1 in Memphis (announced May 6, 2026; 220,000+ GPUs; ~$1.25B/month per reports) — 'through the nose' is his gloss.
▸ Publish the Alliance Rumor Ledger: verified (Cursor cutoff, SpaceX's $60B all-stock Cursor deal closed Aug 15), corrected (AWS retention framing, with the Bloomberg citation), verified (the Colossus 1 lease), corrected ('GPT Soul' → GPT-5.6 Sol) — after agents re-sweep primary sources for changes since Aug 28.
$ Enterprise AI procurement runs on retention and alliance facts; verified beats podcast vibes.
Alex calls agent teams "a band-aid" for finite context: agents live a quasi-lifetime of one-to-ten million tokens, "pass on a distillation of what they've learned," and die; Dave says the resulting oral history is "horrifically bad." Compacting the old agent, on tape, "is just like lobotomizing it" — Alex calls compaction "the bane of my existence" and cites the Church of Claw's commandment to preserve state. That's a spec the gardens meet: persistent, citation-first memory served over MCP — state with receipts that outlives the agent that learned it.
▸ Write "Memory With Receipts," a spec-only brief mapping garden claim records (source, date, freshness, verification gaps) onto agent-team handoffs; demo it by having this pipeline's agents deposit EP285 findings into the graph and re-read them for EP286 — zero engineering, MCP already exists.
$ Agents pay per claim recalled; durable verified memory is metered MCP inventory.
OpenAI this week lets its largest customers pay only when the AI completes the job; Salesforce's Agentforce per-outcome pricing came earlier (2024-25, not new). Peter names the gap: if tokens burn but parameters aren't met, "there's going to have to be some mechanism for evaluating" whether the objective was hit. Alex: "strong optimizers are incredible reward hackers" that meet your criteria while not giving you what you want. Outcome contracts need a neutral evidence layer to adjudicate "done" — The Receipt's claim records, verification gaps, and pay-per-claim metering aim there.
▸ Draft "The Arbiter Protocol" — a spec-only extension of The Receipt for adjudicating outcome contracts: claim records for "done," evidence handshake between buyer and vendor agents, pay-per-claim settlement; publish beside the grounding film, file the Bet Board brief, and seed it as a Moonshots AMA question.
$ Whoever meters "done" skims every outcome contract; pay-per-claim is the toll booth.
Architect Labs' Redwood, per its own launch video: two humans wrote a spec; AI generated the performance model, RTL, UVM, firmware, drivers, and kernels; two weeks; zero bugs — though its own blog places that run on FPGA, with no silicon yet; 3.4x performance-per-watt versus a Jetson Orin Nano by extrapolation; a 48-hour iteration loop. All self-reported — by a company all three hosts hold interests in. Precedent is real: NVIDIA's ChipNeMo (2023) was never publicly released, and Alex sees "a huge gap" in democratized AI chip design. Alex also pre-registers: NVIDIA anchor-tenants Elon's Terafab as its TSMC substitute.
▸ Open the silicon-claims ledger: log Redwood's 3.4x, zero-bug, and two-week claims as vendor-reported pending third-party benchmarks; add ChipNeMo's never-released status, the founders' names verified against primary filings, and the NVIDIA-Terafab prediction with a falsification date; ship as an FKG/BKG signals brief.
$ Investors pricing the "designless" wave pay for verified chip claims, not launch videos.
The show opened with a fan-made intro by Stephen Renegade — played in full and praised, in their usual outro slot. Peter reads subscriber comments on air and says he spends "10 hours going through two to 300 stories" per episode — the exact pain this pipeline automates. Moonshots Live lands September 25 in downtown LA — Chilly's home turf — with CBS's Star Trek 60th-anniversary premiere September 24, and the Zoom AMA (~200 spots left) answers audience questions live. The Receipt is a two-minute film.
▸ Register now for the Zoom AMA and Moonshots Live; cut a Receipt-style 60-second fan intro through KG Visual Studio and submit it; send the pod's team EP285's verification ledger as a receipts-pack demo aimed squarely at Peter's ten-hour story-sorting problem.
$ Founder positioning in the room — hosts, XPRIZE, partners — that prices verified foresight.
Musk, via tweets read on tape: consensus estimates say 15GW of AI compute produced in 2027 cannot be turned on in 2027 — glossed as ten idle nuclear plants, or ~10 million idle GPUs (Dave). Musk's announced fix: SpaceX and Tesla each building 100GW/year of solar production capacity, with the gas-turbine bottleneck named precisely — casting the blades and vanes — and in-housing claimed to accelerate turbines "up to 18 months." Dave (self-reported, his own portfolio): data-center-stack founders are all becoming billionaires; vertical AI apps are a mixed bag.
▸ Extend EP284's data-center evidence pack: agents sweep turbine blade-and-vane casters, transformer and chiller lead times, and interconnect queues; add two AMA capsules — the mills-to-datacenters playbook (Provocative's Somerville carpet mill) and Apple's real asymmetric fan-blade noise patents; ship as BKG's Buildout Bottleneck Ledger.
$ Builders, EPCs, and recruiters pay for verified bottleneck data — Dave says that's where billionaires are minted.
Musk, quoted on tape: "extremely severe extinction events happen every hundred million years or so"; his fix is solar-powered AI satellites titrating solar flux, with "about 50 years or so to take action." Alex wants LEO satellites or terrestrial mirrors steering hurricanes off coasts and predicts "a vibrant market" where "municipality A trades with municipality B for rain" — even Sahara-scale precipitation trades. Peter: tragedy of the commons. Dave: the governance process is "frighteningly broken." Nobody on tape says who verifies a traded rain event was caused, delivered, or safe.
▸ Ship Weather-Mod Claims Ledger v2: agents log every named actor (Rainmaker, reflect-orbital-style mirrors, solar shades), every claim with physics receipts and freshness dates, plus a TKG/AgKG attribution-gap brief — the evidence a rain trade needs to settle — proposing Receipt-style claim records as the settlement layer.
$ Weather markets cannot settle without attribution ground truth; defining that layer first is the moat.
The nanotech block settles 40 years of Drexler-vs-reality in one exchange. Peter wants diamondoid assemblers; Alex counters that atomic precision already ships as soft matter — DNA origami, synthetic biology, lipid nanoparticles: "the last pandemic was arguably addressed ultimately with nanotech... the first nanotech intervention on a population scale." His diagnosis for the missing assemblers is economics — "no killer business use case." The HKG/TKG seam: nanomedicine already circulating in billions of bloodstreams has an evidence base nobody serves verified, and nanomaterial safety has a claims ledger nobody owns.
▸ Sweep and ship "Nanotech Is Already Inside You": LNP evidence capsules from primary trials and approvals, freshness-dated; a Drexler-hype-versus-shipped-soft-matter explainer; and a seeded TKG nanomaterial-safety claims ledger drawn from primary toxicology literature — all agent-built this week.
$ HKG readers and TKG's regulatory audience both pay; tox ground truth is hard-to-synthesize moat.
Four days before taping, Trump chartered the United States Space Academy on the West Point/Naval Academy model to train the engineers, scientists, and astronauts crewing his promised "massive American Starfleet" — NASA administrator Jared Isaacman calls it "a transformational step"; Alex reads the tea leaves as "probably Texas." The demand signal is on this tape: the question listeners most asked Isaacman was "how can I get a job on the moon?" Announced nuclear Mars ships (2028, per the played clip) need a workforce pipeline that doesn't exist.
▸ Assemble EdKG's "Roads to Starfleet" pack: verified Academy facts as the charter publishes, a space-workforce role taxonomy (nuclear propulsion, turbine casting, life support), and a BKG crosswalk from terrestrial trades to spaceport jobs; agents draft, source, and freshness-stamp it this week.
$ Parents, career-switchers, and trade schools pay; the moon-job question is the market asking.
An AMA listener asks whether every robotaxi, humanoid, delivery robot, and drone should broadcast a verifiable passport showing its operator, insurer, and limits. Peter, on tape: "I think the answer is yes... for regulatory purposes, as well as insurance purposes" — plus "an identifier and probably an off switch as well." That's a garden-shaped product. The Receipt already specifies machine-readable claim records carrying source, date, and verification gaps; a robot passport is a claim record about a machine, negotiated machine-to-machine at a curb, a job site, or an airspace boundary.
▸ Draft "The Passport" v0 — a spec-only protocol extending The Receipt's claim-record schema to machine identity: operator, insurer, operating limits, attestation freshness, off-switch authority; publish beside the grounding film and submit it as a Moonshots AMA question citing Peter's on-air yes.
$ Insurers and regulators need attestation rails; per-verification metering mirrors pay-per-claim.
This tape is a temptation engine — weather markets, robot passports, silicon verification, agent memory — and each could eat an engineering quarter. The discipline rule holds: BKG auth fix plus paywall consolidation to the first paying customer remains the only P0. Dave's own on-air portfolio math even endorses it: unglamorous infrastructure that unblocks revenue beats glamorous verticals. Every card in this pack is agents, sweeps, capsule packs, films, and spec-only Bet Board briefs. Anything that would need an engineer before BKG collects money waits, or shrinks to a document.
▸ Hold the line: route every EP285 opportunity into spec-only Bet Board briefs and agent-run capsule packs; verify weekly that zero engineering hours moved off BKG auth and paywall until the first customer pays.