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.
Why is Google paying Anthropic in milestone tranches instead of a single $40B check?
Milestone tranches convert a huge AI bet into staged call options: capital unlocks only when technical scale/performance gates de-risk the next spend.
2026-04-22Why is Google splitting TPU 8 into a training chip (8t) and an inference chip (8i)?
Training and inference stress different bottlenecks—training needs pod-scale compute, inference needs low-latency memory (HBM/SRAM)—so split TPUs optimize perf/$.
2026-04-21Which OpenAI documents does Florida’s probe demand, and why are those more incriminating than chat logs?
Investigators target policies, training, and threat-response workflows because they prove organizational knowledge/control; chat logs show incidents but not systemic design negligence.
2026-04-20Why does Apple name a hardware-engineering chief as CEO while moving Tim Cook to executive chairman?
When custom silicon becomes the strategy bottleneck, CEO selection shifts to the executive controlling HW/SW roadmap coupling, while a chairman role externalizes policy leverage.
2026-04-19What does a robot “world-record” half-marathon time mean when it runs on a separate track?
A “world record” depends on shared protocol: separate tracks and mixed autonomy/remote control change allowable assistance, so the time isn’t comparable.
2026-04-18Why does GPT-Rosalind ship as a gated “Trusted Access” tool-and-data stack, not just a life-sciences model?
High‑risk bio models need workflow‑level controls: gating the tool/data connectors enables audit trails, validation, and dual‑use governance beyond weights alone.
2026-04-17How does the UK’s “Sovereign AI” fund buy sovereignty with GPUs, visas, and procurement—not just cash?
AI “sovereignty” is purchased by controlling bottlenecks—compute locality (AIRR GPUs), talent inflow (visas), and demand lock‑in (procurement)—not funding alone.
2026-04-16Why would Anthropic ship its “most capable” model widely while keeping a stronger Mythos model partner-only?
Release gating follows capability risk: models with autonomous vuln-finding/exploitation require partner sandboxing until safeguards harden in lower tiers.
2026-04-15How can ASML raise its 2026 outlook while installing only 60–80 EUV tools through 2027?
ASML’s outlook tracks backlog conversion plus high‑margin installed‑base service/upgrade annuities, so revenue scales with tool value and mix—not unit volume.
2026-04-14Where exactly is Alphabet’s AI story showing up first—cloud profitability or ad-driven revenue lift?
AI monetizes earliest where infrastructure spend maps to a direct P&L line (cloud operating income), unlike ads where AI gains diffuse into targeting and RPM.