Question archive
Each day, we pose a question inspired by the daily news brief and answer it using our database of indexed AI/ML articles. Browse and search past questions below.
How did one La7 copyright claim cascade into blocking NVIDIA’s own DLSS 5 announcement video?
YouTube’s Content ID keys ownership to the first registered/broadcast fingerprint, not provenance, so a rebroadcaster can “claim” the hash and auto-block the creator.
2026-04-05Why is a French LLM lab negotiating “AI memory” supply with Samsung like a chipmaker, not a cloud buyer?
As LLM labs build their own EU compute, memory becomes a binding scaling input, so they buy HBM supply like OEMs to lock capacity in a tight market.
2026-04-04Why buy an AI‑biology startup for $400M in stock instead of just hiring a few specialists?
Buying the startup captures a cohesive “capability bundle”—domain-tuned models plus the team and workflow—faster than hiring, since tacit know‑how isn’t separable.
2026-04-03Why would Anthropic pay $400M in stock for a biotech team instead of just partnering?
Buying a biotech team lets Anthropic tightly integrate and control biology-specific models inside Claude workflows, enabling cohesive drug+regulatory development that partnerships can’t enforce.
2026-04-02When is a Pentagon “supply‑chain risk” label effectively a penalty, even without contract termination?
In federal procurement, a “risk” designation penalizes via access control: it blocks eligibility across future bids, acting like debarment without ending a contract.
2026-04-01How can a harmless “source map” in a public npm release enable reconstruction of 500,000 lines of code?
Source maps embed a full “inverse” mapping from bundled/minified JS back to original per-file TypeScript, so shipping them makes the release a reconstruction blueprint.
2026-03-31What makes Iran’s threat to Apple, Google, and Meta a physical-security warning, not a sanctions-style pressure tactic?
It’s a physical-security alert because the IRGC moves from policy leverage to kinetic targeting: naming sites as “military targets” and urging evacuation signals imminent strikes, not sanctions.
2026-03-30Why is Microsoft turning Copilot into a model referee instead of betting on one “best” LLM?
Copilot’s “referee” setup boosts reliability by separating generation from verification: one LLM drafts, a different LLM critiques, so errors aren’t shared.
2026-03-29How did a public dispute with the U.S. Department of Defense translate into more Claude Pro subscriptions?
When demand is driven by attention funnels, a high-profile regulatory dispute functions as earned-media acquisition, converting awareness into Pro signups.
2026-03-28How does class certification let Nvidia’s crypto-revenue fight proceed without any fraud finding yet?
Class certification is a Rule 23(b)(3) gatekeeping decision about shared factual/legal issues, not liability—bundling claims lets litigation advance before any fraud finding.