Interview questions follow the work, and the work changed a great deal in 2026. Models shipped monthly, inference hardware split into chips for different phases, agents started running for hours and buying things, and AI answers began absorbing search traffic. None of that replaces the fundamentals of product sense, evaluation and metrics. It adds a layer of current material that a strong candidate can use to make sharper decisions.
This list groups seventeen topics by the round where they are most useful. Each links our concept page and, where one exists, a practice question with a public brief and follow-up rounds. Volatile facts on those pages are dated and sourced to the vendor, regulator or paper that published them.
Read diagram description
Seventeen topics in four clusters. Models and hardware: mixture of experts, AI accelerators, reading a model release, on-device models. Agents: the agent harness, long-running agents, agentic browsers, agentic commerce protocols. Trust and operations: delegated access for agents, frontier safety frameworks, agent traces. Product and strategy: answer-engine visibility, open-weight licences, enterprise adoption, generative media, physical AI, realtime voice agents.
Models and hardware
Mixture of experts. The largest open and hosted models now report two parameter counts: Alibaba's Qwen4-Max has 2.4 trillion total and 71 billion active per token. Total parameters set the memory bill, active parameters the compute. Read mixture of experts, then work a trillion parameters but only 32 billion active.
AI accelerators. Prefill is limited by arithmetic and sets time to first token; decode is limited by memory bandwidth and sets streaming speed. That split explains decode-focused hardware and paid fast tiers. Read AI accelerators for PMs and practise which products should pay for a fast tier.
Reading a model release. Six frontier releases landed between 9 and 18 November 2026, some with launch prices that expire weeks later. Learn to price a task rather than a token and to read a system card for constraints in reading a model release.
On-device models. Apple's on-device model fits about 3 billion parameters at 2 bits per weight. The memory arithmetic and the device-cloud split are in on-device models.
Agents
The agent harness. Two products on the same model can differ widely because of the loop, permissions, sandbox, memory and verifier around it. Read the agent harness, then work same model, but a competitor completes more tasks and users approve almost every permission prompt.
Long-running agents. Work that outlasts the context window needs a plan, a log and commits the next session can read. See long-running agents and an overnight agent that said done but skipped files.
Agentic browsers. Indirect prompt injection is the defining risk, and OpenAI retired its Atlas browser in August 2026. Read agentic browsers.
Agentic commerce protocols. Checkout, authorisation, agent identity and settlement each gained an open protocol, all young. Read agentic commerce protocols.
Trust and operations
Delegated access. An agent acting for a person should carry that person's narrowed permissions, with tokens bound to one service. Read delegated access for agents and work design an agent's access to customer CRM and email.
Frontier safety frameworks. Capability thresholds now shape refusals, vetted-access programmes and launch scope. Read frontier safety frameworks, then a cyber threshold before launch and a security copilot that now refuses legitimate work.
Agent traces. Structural traces give steps, errors, retries and cost per task without storing content, which matters for zero-retention customers. See agent traces and debug an agent for a zero-retention customer.
Product and strategy
Answer-engine visibility. When an AI summary appears, people click less. Segment the exposure by query type in answer-engine visibility and practise organic clicks fell after AI Overviews.
Open-weight licences. Licences range from Apache 2.0 to custom terms with user and revenue thresholds. Read open-weight models and licences and choose an open-weight model for a bank.
Enterprise adoption. Published adoption figures range from about a fifth of firms to two-thirds of enterprises, depending on the denominator. See enterprise AI adoption.
Generative media. Price kept clips, not generations, and plan for the model being retired. Read generative media products and work our video model provider is shutting down.
Physical AI. Robot foundation models moved from labs to pilots; the economics hinge on human interventions. Read physical AI for PMs and first task for a warehouse humanoid.
Realtime voice. Chained pipelines and speech-to-speech models trade auditability against latency. Read realtime voice agents and work move a bank voice agent to a chained pipeline.
If your product serves users in Indian or European languages, add two more. First, multilingual AI products, on token cost by language, evaluation beyond translated benchmarks and code-switching. Second, for growth and partnership roles, distribution inside AI assistants: plugins, connectors and the rules that govern them.
For the dated details behind the model and hardware topics, see six frontier releases in ten days and the 2026 inference hardware cycle.
Questions about this topic
Four clusters changed in 2026: models and hardware (mixture of experts, inference chips, reading model releases, on-device models), agents (harnesses, long-running agents, agentic browsers, commerce protocols), trust and operations (delegated access, frontier safety frameworks, agent traces) and product strategy (answer engines, open-weight licences, enterprise adoption, generative media, physical AI, realtime voice).