Evaluate emerging technology with a falsifiable question
AI and machine learning, serverless computing, WebAssembly and blockchain were the four themes of this article in 2024. They remain useful subjects for evaluation, but none should be adopted simply because a trend list predicts inevitable growth.
This November 2026 review keeps those subjects and replaces sweeping forecasts with questions about correctness, workload, costs and operating responsibility. The examples are evaluation scenarios, not claims of shipped projects or commercial results.
AI and machine learning: define acceptable error
A model can support classification, retrieval or language tasks, but a useful integration needs a measurable goal. For a support-drafting tool, decide what information may be sent to the model, how answers are checked and who approves a response.
Compare the model with a simpler baseline. Evaluate representative inputs, ambiguous cases and deliberate misuse. Record quality, latency and cost. A plausible answer is not sufficient evidence of correctness.
Keep authorisation outside the model. Treat retrieved text and model output as untrusted data; do not let a generated instruction grant itself permission to change records. Provide a fallback when the service is unavailable and make uncertainty visible where it affects the user’s decision.
Using a hosted API is different from training a model. State which responsibility the team owns. The mobile engineering article discusses additional constraints when inference is part of a device experience.
Serverless: less server administration, continuing system design
Serverless services can move parts of infrastructure operation to a provider. Application teams still own data access, security, failure handling and their deployment configuration.
Evaluate a representative workload rather than assuming usage-based billing is always cheaper. Include invocation frequency, memory, execution duration, storage, networking and dependent services. Check time limits and concurrency behaviour against the task.
Retries are particularly important for state-changing operations. Use stable event identifiers and a durable record of completed work when duplicate processing could cause harm. AWS Lambda’s best practices recommend idempotent code; apply the principle to the business operation rather than treating a retry setting as a correctness guarantee.
For long-running or stateful work, consider whether a worker, queue or conventional service is easier to operate. Write down why the deployment model fits before adopting it.
WebAssembly: benchmark the expensive part
WebAssembly is a compilation target that can be used alongside JavaScript. It can make sense for a compute-heavy component or reusing a suitable existing library. It does not automatically speed up a page dominated by network requests, layout or oversized images.
Identify the bottleneck first. Measure startup and download cost as well as execution time. Include the cost of moving data across boundaries and test on representative devices. A small benchmark of an algorithm is not proof of a faster complete user journey.
MDN’s WebAssembly concepts guide explains its relationship with JavaScript. Start with one isolated component and keep a fallback or clear compatibility policy where necessary.
Blockchain: establish why shared authority is necessary
Ask whether independent parties need to verify a shared history without relying on a single operator. If one organisation controls all writers and readers, a conventional database may satisfy the requirement with less complexity.
Consider key management, transaction costs, contract upgrades and data visibility before selecting a network. Immutability does not establish that an input was truthful. A public ledger is not an automatic privacy mechanism, and application code can still contain vulnerabilities.
Evaluate security and privacy through a project-specific threat model and review of the chosen protocol. Decentralisation alone does not provide either.
Use a short decision record
For each experiment, record the problem, baseline, proposed technology, success criteria, budget, risks and a stop condition. If the trial does not beat the baseline on the relevant measure, keep the simpler approach.
Curiosity is useful; evidence makes it actionable. The stack-selection guide offers a broader framework for turning those experiments into maintainable product decisions.

