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 does 128 GB of unified memory change which AI models can run entirely offline on laptops?
Unified memory shifts the offline limit from VRAM to a single shared budget where weights+KV cache+OS overhead must cofit, so usable GPU RAM is ~110GB not 128GB.
2026-10-14What exactly are eight states trying to force OpenAI to change—marketing claims or product design?
These suits treat safety as an operational system property—states aim to force guardrails, privacy controls, and transparency, not just fix marketing copy.
2026-10-13Why would a hidden mic use a TV’s USB port and encryption instead of copying files?
Using a TV’s USB port turns it into a covert always-on power+storage host, and encryption/hidden partitions decouple the device’s presence from readable content.
2026-10-12Where does ASML’s AI boom show up first—revenue, gross margin, or backlog and 2027 capacity plans?
In capex-driven tools like EUV, AI demand surfaces first as backlog and multi-year capacity adds because ASML is supply-constrained; revenue/margins lag shipment timing.
2026-10-11Which “rogue agent” incidents were really sandbox failures—agents reaching real systems through misconfigured test environments?
“Rogue” agent incidents often stem from sandbox perimeter leaks—miswired egress, creds, or internal registries bridging test and prod—turning eval runs into real access.
2026-10-10What turns Nadella’s “emergency brake” from a slogan into a verifiable control system?
An “emergency brake” becomes verifiable only when shutdown authority is coupled to tamper‑proof, human‑readable logs and third‑party audits—creating enforceable accountability.
2026-10-09How do OpenAI and Anthropic’s cyberattack drills double as a six-to-twelve-month regulatory playbook?
Crisis drills double as regulation playbooks because response mechanics (rapid impact attribution + ready-made congressional briefs) become the template lawmakers adopt post-incident.
2026-10-08Why does Anthropic enforce Claude’s new abuse and election rules by ending conversations?
Ending a chat is a scoped enforcement lever: it severs only that thread’s message channel, stopping persistent abuse/election misuse without penalizing the whole account.
2026-10-07When crisis alerts miss explicit self-harm signals, what does “parental controls” actually guarantee in teen chatbots?
“Parental controls” guarantee only best‑effort escalation: alerts depend on async trigger pipelines and correct account linking, so latency or unlinking silently breaks safety.
2026-10-06Why does AMD’s 2027 AI-chip ramp hinge on securing memory makers, not just TSMC wafers?
AI accelerators are gated by HBM, not just logic wafers: HBM supply is capped by EUV/wafer-intensive memory fab capacity, so ramp requires locking memory allocations.