Technology explainers

Why Solar Cells Lose Charge Before It Becomes Electricity

A guide to electron-hole recombination, the different loss pathways inside photovoltaic materials, and why longer carrier lifetimes do not automatically guarantee a more efficient solar panel.

How Researchers Benchmark Mental-Health Chatbots

An accessible guide to scenario design, expert criteria, human and automated scoring, and why benchmark performance is not the same as clinical safety.

How the Developing Brain Removes Excess Neurons

A guide to why the developing brain produces more neurons and connections than it keeps, and how competition, activity and programmed cell death refine neural circuits.

How Binocular Suppression Lets Scientists Separate Seeing From Visual Input

Why showing different images to each eye helps researchers compare stable visual input with changing conscious perception, and where the method’s inferences remain limited.

How SQUID Magnetometers Measure Extremely Weak Magnetic Fields

A practical guide to superconducting loops, Josephson junctions, magnetic flux and the cryogenic systems behind SQUID magnetometry.

What Does It Mean When Lean Verifies a Mathematical Proof?

Lean checks whether a formally encoded proof follows from declared definitions and axioms. That provides powerful logical assurance, but human review is still needed to assess whether the encoded theorem matches the intended claim, whether assumptions are appropriate, and whether the result is novel.

How Do AI Sandboxes and Containment Controls Work?

AI containment relies on several independent barriers rather than a single sandbox. This explainer shows how network restrictions, disposable environments, short-lived credentials, monitoring and human approval reduce the impact of a failure.

Why Removing Training Data Is Different From a Runtime AI Guardrail

Training-time interventions change what a model learns, while runtime guardrails control what a deployed system may do. Each covers different failure modes, and neither is a complete substitute for the other.

How Approval Gates Keep AI Agents Within Authorized Boundaries

A practical guide to placing human approval, scoped credentials and audit logs around consequential actions taken by AI agents.

Why Offline AI Training Is Not the Same as an Air Gap

An AI agent can be denied ordinary internet access while its training infrastructure still uses external connections for monitoring and operations. Real containment therefore depends on layered controls around sandboxes, hosts, networks and shared hardware, not a single offline switch.

How an Electronic Nose Recognizes a Smell

An electronic nose combines several partially selective sensors with pattern-recognition software. It learns the response pattern associated with an odor category, but humidity, concentration, mixtures and sensor drift can make real-world recognition harder than a controlled demonstration.

How a Personalized mRNA Cancer Vaccine Is Made

Personalized mRNA cancer vaccines begin with sequencing a patient's tumor, selecting mutations likely to be visible to the immune system, and manufacturing a unique mRNA dose. The process targets remaining cancer cells, but it adds time, cost and complex quality control.

How Sequential Diagnostic AI Differs From a Medical Chatbot

Sequential diagnostic AI decides what information to request next and revises its hypotheses over several steps. That makes it closer to a simulated clinical workup than a chatbot answering a completed vignette, while still leaving major validation and safety questions.

How Antisense Oligonucleotides Turn RNA Into a Drug Target

Antisense medicines use short synthetic nucleic-acid strands to recognize specific RNA sequences and change what cells or pathogens produce.

How the FDA Regulates Apps Used With Prescription Drugs

A practical guide to how the FDA separates device-connected and standalone drug companion software, and when app output becomes official or promotional labeling.

How Do Self-Driving Laboratories Choose Their Next Experiment?

A self-driving laboratory connects experimental planning, instrument control, measurement and feedback. Its decisions can adapt to new results, but reliable progress requires clear objectives, calibrated measurements, traceable samples and controls that distinguish real effects from experimental artifacts.

How Do AI Companies Detect Model-Distillation Attacks?

Providers combine reasoning-access controls, request classifiers, cross-account telemetry and escalating limits to detect unauthorized extraction without treating every high-volume user as an attacker.

How Can Aviation Systems Respond to a Threat From Inside the Cockpit?

Cockpit doors stop unauthorized entry, but insider threats require overlapping personnel, procedural and operational safeguards.

How Can an AI Evaluation Affect Real Websites?

An AI test can create real effects when the model has live network access and tools that can submit forms or call services. Safe evaluation combines containment, narrow permissions, action checks and monitoring.

How Do Airlines Reduce Insider Threats in the Cockpit?

Airlines cannot eliminate every risk posed by authorised personnel through background checks alone. They combine continuing vetting, restricted access, crew coordination, standard alerts and independent safety investigations so that several barriers can prevent, reveal or contain a dangerous act.