AI / REPORTFive Questions to Ask Before Choosing an Edge AI Chip
At the edge the metric is performance per watt, and the compiler decides adoption. Quantisation, on chip memory and support lifetime outweigh peak throughput.
Category: AI
News & analysisAI / REPORT
AI / REPORTOpen Source Licenses Face Their First Real Test in the AI Era
Training, weights and attribution all sit outside what traditional licences anticipated. Model specific terms are appearing, and provenance records matter more than optimism.
AI / REPORTOn-Device AI Changes What a Privacy Claim Means
Local inference removes a real risk category, but storage, telemetry and model extraction still matter. Verifying the claim takes more than reading a policy.
AI / REPORTTwo Reasons AI Benchmarks Go Stale So Quickly
Benchmarks decay, get contaminated and measure the wrong things. Practical evaluation now means private sets, dynamic questions and workload specific tests.
AI / REPORTQuantum Error Correction Moves From Physics to Engineering
Logical qubits have moved error correction from theory to engineering. The meaningful metrics are error rates and overhead, not the total count of physical qubits.
AI / REPORTThree Audit Trails Every AI Agent Deployment Needs
Agents take actions, and most deployments cannot reconstruct why. The fix is unglamorous: tool permissions, approval gates, logged trajectories and rehearsed shutdowns.
AI / REPORTSmall Language Models Are Finding Work the Cloud Cannot Reach
Models that fit on a device are solving latency, privacy and connectivity problems that hosted systems handle badly. The engineering trade-offs are concrete and measurable.
AI / REPORTEurope’s AI Rules Reach Their First Transparency Deadline
Disclosure and documentation duties now apply well beyond frontier labs. The hard work is classification, inventories and records that stand up to scrutiny.
AI / REPORTOpen-Weight Models Are Changing the Economics of AI Inference
Downloadable models have turned inference into a cost calculation rather than a vendor relationship. The trade-offs involve licences, hardware and evaluation discipline.
AI / REPORTAI Chips Face a Memory Wall That More Compute Cannot Solve
The bottleneck in AI hardware is no longer arithmetic. It is the cost of moving data, and the workarounds are reshaping chip design, software and data centre planning.