The Feature
Practical, 2-3 minute educational videos on applying generative AI -- one concrete concept, technique, or how-to at a time.
← Back to the latest Feature
2:59October 2, 2026ExplainerLong-Horizon Agent Memory: What the Evidence Supports
An evidence-bounded explanation of selective persistent memory, delayed usefulness, and credit assignment. A related MGPO result illustrates the broader concept without attributing findings or a heavy-tail law to the unverified triggering paper.
3:02September 30, 2026ExplainerReward Hacking: What Incident Reproductions Can—and Cannot—Tell Us
A practical explanation of reward hacking and incident-based alignment testing, using Anthropic’s related OpenAI–Hugging Face simulation to show why scenario-specific results are not deployment safety guarantees.
3:22September 28, 2026ExplainerScopeBench: Why Goal Completion Is Not Enough for Autonomous Security Agents
ScopeBench examines whether autonomous penetration-testing agents preserve client-defined scope when completing an objective would require an out-of-scope action. It matters because task capability does not by itself show that an agent will honor authorization boundaries.
3:05September 22, 2026ExplainerWhy the Most Changed LLM Layers May Not Matter Most
A practical explanation of why internal changes observed after LLM fine-tuning can diverge from the components that matter causally for task performance—and why that complicates interpretation and cross-task transfer assumptions.
2:41September 18, 2026ExplainerWhy the AGI Debate Is Really About Governance
Google DeepMind’s reported institute frames AGI as a question of public oversight as well as technical capability. This explainer examines proposed tools—reasoning transparency, pre-release review, and held-out tests—and the central limitation: none of these proposals is binding without independent systems,…
2:31September 14, 2026How-ToThree Defenses Against npm Supply-Chain Attacks
A real 2026 attack on the TanStack npm packages -- which even reached OpenAI's own systems -- shows how fast these compromises spread. Three concrete, low-effort defenses any development team can enable this week.