← All notes

Technology & everyday life · Working notes

Beyond the Scroll

Building digital experiences worth our attention. A constructive look at what our interactions can tell a system, what they leave uncertain, and how better questions can inspire better technology.

A song arrives at the right moment. A short explanation makes a difficult idea understandable. A small business appears just when someone needs its service. Sometimes a few seconds on a screen lead to something that stays with us long after we put the phone down.

Those moments make digital discovery valuable. Behind them are creators making things worth sharing, engineers building systems at extraordinary scale, and people whose interests are always changing. The opportunity ahead is to understand those people with greater care.

My interest in this question comes from making creative work and exploring practical technology. Music, visual storytelling and everyday service work repeatedly bring me back to the same concern. What makes an interaction useful to the person on the other side?

This essay develops that question through illustrative examples, published engineering work and proposals for investigation. The examples are thought experiments. The design ideas are invitations to test, with room for findings that challenge our expectations.

The value of discovering something unexpected

Imagine a student in a small town finding a teacher who explains a subject in a familiar language. Imagine an independent musician reaching someone who would never have searched for their name. Imagine a useful local service being discovered by a person who had assumed it was unavailable nearby.

Each encounter expands someone’s possibilities. Discovery helps knowledge travel beyond geography, gives creative work another chance to be heard and helps people find resources outside their existing circle. A thoughtful recommendation can reduce the distance between a need and something that meets it.

Yet usefulness takes different forms. Someone seeking a repair tutorial wants a clear answer quickly. Someone exploring unfamiliar music may welcome a little surprise. Someone researching a purchase may need several competing perspectives before deciding. A design that serves one situation beautifully may require adjustment for another.

That makes context a productive starting point. Before choosing a success metric, ask what the person came to do and what a worthwhile outcome would look like. A completed task, a remembered idea and a newly discovered artist deserve to be considered in their own terms.

What an interaction can mean

Consider a cooking video with a song playing in the background. A viewer watches it twice and saves it. Perhaps the recipe is useful. Perhaps the presentation is enjoyable. Perhaps the soundtrack is the main attraction. Several explanations fit the same visible behaviour.

This ambiguity becomes a valuable engineering question. How much confidence should a system place in each explanation? What further evidence would help? A deliberate visit to the song’s page might provide additional information. A saved ingredient list might point in another direction. Even those actions need interpretation.

The central insight from my earlier research exploration is a question about this attribution. When a piece of media combines a topic, a creator, a visual style and a soundtrack, how accurately can engagement with the whole tell us about preference for each part? A useful investigation would measure these preferences separately and report where they agree or diverge.

Now imagine encountering the same soundtrack in several unrelated videos. A listener might discover a favourite, become tired of it, or barely notice it. These are competing possibilities worth testing. Recognition, enjoyment and the desire to hear more should each have their own measurement.

For students, this is a lesson in the meaning of a label. A dataset column called “liked” records an action under particular circumstances. Documenting those circumstances can be as important as selecting a more sophisticated model.

The engineering behind the moment

Meta’s 2023 account of Instagram Explore describes retrieval followed by successive ranking stages and final reranking. This architecture narrows a large collection into a manageable set before more expensive evaluation. It also describes using different interaction signals and filtering history inputs. That publication is a concrete engineering example with its own date and scope.

Read Meta’s engineering account of Explore.

A simple way to reason about the design is to distinguish three questions. Retrieval asks which items deserve consideration. Ranking asks how promising those candidates appear for this person and situation. Evaluation asks whether the resulting experience actually served the intended purpose. A weakness at any stage can shape what the next stage gets to work with.

Suppose we are designing a classroom demonstration with 100 educational clips. The retrieval step selects 20 candidates. The ranker orders them and displays five. The learner chooses one. We can observe that choice, but the learner had no opportunity to choose the other 95 clips during that interaction. That missing opportunity matters when we later interpret the data.

The next version of our classroom system might learn from the previous version’s recommendations. Our original decisions would then influence the examples used for future decisions. Research on algorithmic confounding examines this kind of feedback and its possible effects on diversity and utility. Its findings motivate careful evaluation in each new setting.

Chaney, Stewart and Engelhardt, research on algorithmic confounding in recommendation systems.

1. SelectA system chooses what can be seen.
2. EncounterA person responds in a particular context.
3. InterpretThe response becomes a learning signal.
4. UpdateFuture choices reflect that evidence.
A conceptual feedback cycle. Each update leads to another selection, so the system also influences the evidence it collects.

This creates a rewarding design challenge. We can ask whether our evaluation gives unfamiliar material a fair opportunity, whether observations include enough context and whether a different selection policy would produce different conclusions. These questions belong in the experiment plan from the beginning.

Measuring an experience that was worth having

Begin with a human outcome and work backwards. For a tutorial, the outcome might be completing a task correctly. For music discovery, it might be voluntarily returning to a song later. For a business enquiry, it might be finding the right service with fewer unanswered questions.

Each outcome needs an operational definition. “Helpful” might mean a learner can complete a short exercise after watching. “Discovery” might mean someone independently returns to an unfamiliar creator within a defined period. Definitions should be chosen before inspecting the results, so they remain consistent when a favourite idea performs poorly.

Questions that make measurement more useful
ObservationA useful follow-up question
A video was watched to the endDid it answer the viewer’s question or leave them confused?
A song appeared repeatedlyDid the listener choose to return to it later?
A post received many clicksDid visitors find what the post led them to expect?
A campaign generated enquiriesHow many were relevant, and what additional value did the campaign create?
A new creator received exposureDid that opportunity lead to sustained audience interest?

Technical teams already study competing objectives and biased feedback. Google’s 2019 paper on video recommendations describes a multi-objective ranking system and approaches to selection bias. It offers students an example of translating a complicated product problem into modelling choices.

Read Google Research’s paper on multitask ranking.

For a small business, the same discipline can begin in a spreadsheet. Record the campaign objective, dates, audience, spend and definition of each outcome. Comparing a profile visit with a video play produces a misleading sense of precision. A useful comparison requires matching the question being answered.

We should also ask what would have happened without a campaign or product change. A carefully designed comparison group helps estimate additional impact. Random assignment can strengthen that comparison where it is feasible and appropriate. When it is unavailable, report alternative explanations such as timing, audience differences and other promotion.

Every measurement has a cost. Surveys ask for time, detailed tracking raises privacy questions and longer observation delays decisions. A practical team chooses a small set of meaningful measures, documents their limitations and improves them as evidence accumulates.

Where human value and business value meet

Consider the different participants in a useful discovery. A viewer finds something relevant. A creator reaches an interested audience. A business receives an appropriate enquiry. A platform becomes a place people trust enough to return to. This is a promising business hypothesis because the value is shared across the experience.

The alignment requires work. A viewer may prefer a quick answer, while a product dashboard rewards a longer session. An advertiser may value immediate clicks, while the business ultimately needs satisfied customers. A creator may seek discovery today and a durable audience over months. Product decisions should make these different time horizons visible.

Imagine two designs for an educational application. One encourages a student to keep opening lessons. The other helps the student choose an appropriate lesson, finish a useful exercise and leave knowing what to practise. A business could test whether the second experience earns stronger trust and more voluntary returns. The answer requires observation across time.

For a platform, useful questions include whether creators find the rules understandable, whether advertisers can connect activity to outcomes, and whether viewers feel able to steer their experience. Confidence in those relationships may support commercial value. Any claim about revenue improvement should follow measurement.

Companies such as Meta, Google and OpenAI could draw inspiration from these questions in different product settings. For a feed, the concern may be the next item presented. For search, it may be finding a relevant source. For an AI assistant, it may be whether an answer helps someone complete a task with appropriate confidence. Each setting needs its own evidence and evaluation.

Design opportunities worth exploring

Let people express the context of an interest

A person researching a temporary project could benefit from a way to signal that context. A possible design might let them mark an interest as something they are exploring this week. Engineers would need to test whether the extra control is understandable and whether people use it. The resulting preference should be editable and easy to withdraw.

Give discovery room to surprise

A learning app might reserve some opportunities for relevant material outside a student’s recent pattern. This could help reveal interests that past behaviour has never had a chance to express. The exploration rate becomes a testable choice. Too much unfamiliar material could make the experience feel unfocused, so satisfaction should be evaluated alongside variety.

Make feedback easy to interpret

When a person dismisses a recommendation, an optional reason such as “already know this” could be more informative than dismissal alone. Explanations should remain brief and voluntary. The design challenge is obtaining useful context without turning ordinary browsing into a questionnaire.

Evaluate the experience across time

A single session offers a narrow view. A proposed evaluation could combine immediate relevance with later voluntary return, task completion and repeated dissatisfaction. Teams would need to examine how results differ for newcomers, frequent users and people whose interests change. Averaging all of them together could hide important trade-offs.

Keep improvement proportionate

More context does not automatically justify collecting more personal information. Start with the minimum data needed for the question. Consider aggregate measures, voluntary feedback and short retention periods. Build a way to inspect mistaken decisions and compare alternatives before expanding a feature.

A practical challenge for engineering students

Build a small synthetic discovery system for educational videos. The aim is to practise experimental reasoning. You can implement it with a spreadsheet or a short program and explain every assumption in a notebook.

  1. Create a synthetic catalogue. Define 100 videos across five topics and 200 simulated learners. Give each learner a hidden topic-preference vector. Make those preferences part of the simulation specification.
  2. Define a response model. Specify how topic match, item quality and randomness affect a simulated choice. Store the probability for every displayed item. Run a separate scenario if you want repetition to influence responses.
  3. Compare two selection policies. Use a baseline that favours previous engagement. Compare it with a policy that reserves some exposure for relevant, less-seen items. Keep the response model and starting conditions consistent.
  4. Log opportunity as well as action. Record which items were shown, their positions and whether they were selected. Retain the hidden preferences for evaluation rather than letting the baseline read them.
  5. Measure several outcomes. Track topic match, catalogue coverage, exposure concentration and the number of relevant new items discovered. Report results over multiple random seeds.
  6. Challenge the result. Change the initial popularity distribution, strength of preferences and exploration rate. Check whether the apparent benefit survives these changes.

Write the hypothesis before running the comparison. One possibility is that modest exploration increases relevant discovery while preserving acceptable topic match. Specify what counts as a meaningful improvement and show uncertainty across repeated runs. A finding that the proposal fails under some conditions can be an excellent project result.

The simulation demonstrates consequences of its assumptions. Applying those conclusions to people requires additional evidence. In a report, explain which behaviours you invented, which parameters you tested and which outcomes you never measured. That habit makes your work easier for another student to reproduce and improve.

What we can apply today

As readers, we can pause after an interesting post and ask what we gained. Was it a useful explanation, enjoyment, a new perspective or a reason to investigate further? Describing the benefit helps us make more deliberate choices about our attention.

As creators, we can give each piece a clear purpose. A tutorial should help someone do something. A song can offer an experience worth returning to. Clear presentation and transparent collaboration help the audience understand what is being offered.

As businesses, we can define the result we need before choosing a channel. Enquiry quality, completed bookings and customer satisfaction often require different measurements. A useful review asks which information would change the next decision.

As engineers, we can document how observations were produced. Which options did a person see? What did the system withhold from consideration? What uncertainty remains about the response? Those questions help turn a dashboard into an investigation.

As students, we can practise explaining a model’s goal in ordinary language. If we cannot describe who benefits and how we would know, the project needs another conversation before another layer of complexity.

A digital experience worth returning to

The most memorable digital moments often feel small. Someone understands a lesson, finds a useful service, hears a beautiful song or discovers that another person sees the world in a way they recognise.

Building for those moments asks us to take both people and measurement seriously. Curiosity gives us questions. Engineering gives us ways to test them. Care helps us decide which outcomes deserve our effort.

There is room here for creators, researchers, businesses and the next generation of developers to contribute. A worthwhile place to begin is a question simple enough to carry into every product discussion. What will someone be glad they discovered?

Appendix — A recent example from Meta

In its 16 November 2026 update on Instagram recommendations in India, Meta describes work undertaken over the preceding year. It reports stronger AI-based assessment before recommendation and adjustments for Indian audiences, including improved filtering of sexually suggestive and violent content.

Meta says recommended content is reported roughly half as often as a year earlier. The announcement provides no detailed denominator or false-positive analysis. This is a company-reported outcome. It also describes feedback controls and additional protections for teens.

For students, the announcement offers a useful evaluation exercise. What would distinguish better detection from changed reporting behaviour? How would a team measure mistaken exclusions, viewer satisfaction and effects on creator discovery? Which comparison period and exposure denominator would make the result interpretable?

The update illustrates an ongoing effort to refine recommendations. The relationship between improvements, user feedback and measured outcomes remains a productive subject for further study.

Meta’s original announcement, 16 November 2026.

Sources and further reading

← More from the notebookContinue the conversation on Instagram ↗