Three Warnings in Three Days, A Door Closes, Nine Billion Variants

The episode's claim that better data drove roughly 12x compute efficiency versus about 3.7x from architectures.

▸ Package each garden's verified corpus as a licensable training-and-eval dataset: schema, provenance column, freshness date, primary-source link per row. Ship BKG first as a sample pack plus a data card. Pitch it as the thing architectures cannot copy.

▸ agent: Agents assemble rows, attach citations, generate the data card, and flag unsourced entries for human review.

$ Dataset licensing and eval-set access fees; upsell to agent-priced MCP queries.

Salim's story of a company whose data subsidiary was valued on the balance sheet far above the operating business.

▸ Build a one-page Proprietary Data Valuation scorecard: uniqueness, verifiability, freshness, agent-readability, legal clarity. Run it on BKG contractor partners as a free diagnostic. It is both a lead magnet and the opening of a paid data-readiness engagement.

▸ agent: Agent ingests a partner's sample export and returns a scored, sourced readiness report unsupervised.

$ Free diagnostic converts to paid data-readiness audits and a cut of any licensing deal.

Alex's Bloomberg GPT cautionary tale: internal enterprise data had real value, but with a shelf life.

▸ Write the house doctrine that fine-tuned models decay but verified, re-verified ground truth compounds. Make freshness a product feature: every capsule carries a verified-on date and an auto-recheck cadence, and stale capsules visibly degrade rather than rot silently.

▸ agent: Nightly agent rechecks every source URL, diffs the claim, and downgrades or flags drifted capsules.

$ Freshness guarantees justify subscription pricing; stale-free data is what agents will pay for.

AlphaGenome precomputing every single-letter variant; Alex's point that bulk-solving a field wants to become a lookup database.

▸ Pick one narrow, enumerable garden domain and precompute the whole answer space as a cited table. Candidate: every common building-material and additive paired with its non-toxic status and primary source. Publish as browsable rows plus MCP endpoint, each row source-stamped.

▸ agent: Agents enumerate the space overnight, fetch primary sources, draft each row, and route unsourced rows to a human queue.

$ Agent-priced per-row lookups, plus a paid bulk export for specifiers and manufacturers.

Dave: give someone a hundred thousand genius agents and the hard part is specifying which problem to point them at.

▸ Productize problem specification. A short intake that turns a partner's vague goal into a runnable brief: success criterion, verifiable checks, agent count, budget, stop condition. Ship it as a template plus a live agent that drafts the brief from a transcript of one call.

▸ agent: Agent transcribes a call, drafts the brief, and returns it with named acceptance tests unsupervised.

$ Fixed-fee problem-framing sprints; the brief becomes the scope for paid execution work.

Noam Brown's arc from a costly ARC-AGI score to roughly twenty dollars, cited on-air as a ~25,000x collapse.

▸ Maintain a public capability-price ledger: for each task we care about, the cheapest verified way to do it today, dated, with a source. Update weekly. It doubles as our internal trigger list for when a garden feature becomes affordable.

▸ agent: Agent reprices the ledger weekly from published rates and flags any threshold crossing.

$ Cited, dated price curves are citation-magnet material and a paid briefing for operators.

The resignation post's reported reach, and the panel arguing on air over what is actually verifiable about it.

▸ Build a claim-provenance capsule format for contested AI-safety claims: the assertion, who said it, where, when, what is actually verifiable, what is not. Publish the first batch on this week's stories. Label hype as hype in the schema, not just the prose.

▸ agent: Agent drafts capsules from transcripts, resolves primary sources, and marks unresolvable claims unverified.

$ Provenance capsules are the citable unit agents pay to query; also earns inbound links.

ByteDance's founder personally leading a real-time world model; Peter asking what happens when most video comes from elsewhere.

▸ For The Call, lock a provenance-first posture now: every film ships with a manifest of sources, human authorship notes, and claim citations. Cheap video is commoditizing; a verifiable chain of custody is not. Add the manifest to the next film's release package.

▸ agent: Agent assembles each film's source manifest and citation ledger from the production files automatically.

$ Provenance manifests become a paid service for brands and studios facing synthetic-media doubt.

DeepSeek V4.1 Flash: MIT-licensed, $0.15 per million input tokens against far pricier Western models.

▸ Run a controlled bake-off across the cheapest credible design-capable models on one real garden task: generate a citation-capsule page in house aesthetics. Score by eye and by cost. Route routine page generation to whatever wins, keep judgment work on the premium model.

▸ agent: Agent runs the bake-off, renders every variant, tabulates cost per accepted page, and writes the recommendation.

$ Direct margin: same output volume at a fraction of current per-page generation cost.

The Ornn H100 index rising; Dave warning that assuming later compute access is a serious mistake.

▸ Write down the gardens' compute posture in one page: what we run where, what we would do if inference prices doubled, which workloads are movable to cheap open models. Not a procurement program. A one-page contingency so a price shock does not stall BKG delivery.

▸ agent: Agent inventories current model spend by workload and drafts the fallback mapping.

$ Avoided cost and continuity; protects margin on paid BKG work if rates spike.

The panel telling students in PhD and professional programs that momentum toward a credential is now the expensive choice.

▸ Ship an EdKG one-pager for people mid-credential: what to keep, what to abandon, how to convert domain access into a proprietary dataset before leaving. Cite the episode as a claim with a source, not as truth. Offer one free orientation call per week.

▸ agent: Agent drafts a personalized dataset-extraction plan from a participant's stated field and access.

$ Orientation calls feed paid cohort programs and sourcing for data-rich collaborators.

The FT report that a frontier lab withheld Claude Mythos 5.1 from the UK AI Security Institute.

▸ Third-party verification is becoming scarce exactly as it becomes valuable. Position the gardens' verification node as an auditable, publishable method: state the checks, the source hierarchy, the failure handling. Publish the method itself, so the moat is trust in the process, not secrecy.

▸ agent: Agent runs the documented check suite on every new claim and publishes a pass/fail ledger entry.

$ Verification-as-a-service for partners; method transparency drives citations and agent trust.