Adapted from my presentation “Scaling AI Entertainment in the Car”. Numbers in the delivery scenario are illustrative estimates from the presentation, not a common-device benchmark.
The content space has changed
Traditional entertainment systems mostly distribute known assets: a music track, its cover, a searchable catalogue. There are already difficult problems in delivery, discovery and playback, but the unit of content is reasonably familiar. Generative AI changes that unit. A track can now have separate stems, remixes and alternative versions; an image can become a new composition or an outpainted cover. The experience becomes more flexible, while the number of assets and possible outcomes grows.
My presentation looks at that change through three connected questions: can we deliver the content efficiently, can we evaluate whether the generated result is useful, and can we select something relevant to the person’s current intention? A strong generator addresses only part of the system.
Delivery: compress the experience
In the presentation’s illustrative scenario, a conventional package contains a 4 MB track and a 0.5 MB cover. Add four stems, two versions and an expanded image, and the package becomes approximately 38 MB. This is an example of how richer experiences can increase delivery volume; it is not a universal estimate for AI-generated media.
Neural codecs offer a different representation for that content. An audio encoder turns a waveform into compact tokens; a compatible decoder reconstructs audio on the receiving device. SpectroStream explores this route for 48 kHz stereo audio. For images, Cool-Chic similarly explores a learned representation that is reconstructed by a decoder. The presentation combines approximate payload assumptions into a roughly 2.4 MB scenario, with decoder and learned-weight costs treated separately.
Compression moves work rather than making it disappear. The device must load the model, reconstruct the content, fit the memory budget and start playback at an acceptable time. My Android codec project measures this boundary directly: payload, startup latency, thread scaling and reconstruction quality are separate dimensions. A smaller file alone does not prove that the complete experience is better.
Quality: evaluate the result
The same model and prompt can produce different results across seeds. Choosing a provider or finding a good prompt does not remove this variability. Production therefore needs an evaluation layer: a way to compare candidates against the intended output and the constraints of the experience.
For audio, the presentation uses TuneJury as an example of learned evaluation between a track and its description. The same tracks can rank differently when the requested musical character changes. This suggests several uses: choose the best of multiple generations, guide optimisation, or support an expert’s iteration. A score still depends on what the evaluator has learned and what the prompt asks for; it should be validated against listening and the intended use.
For visual edits, EditReward illustrates a related question: how good is an outpainted image relative to its original and the editing instruction? A useful evaluator should reward a suitable extension, rather than just the most visually striking candidate. The engineering task is to make that judgement repeatable enough to support selection.
Same prompt. Different results.
Evaluation helps choose.
Relevance: understand the intention
Similarity is a useful retrieval signal. It is not a complete definition of relevance. The presentation’s example, “image for kid”, can retrieve a photograph of a child, a children’s illustration or a painting depicting a young person. All can be related to the words, yet they serve different intentions.
A practical selection pipeline starts with multimodal retrieval across text, audio and images. It then reranks candidates using the user’s intention, the current content and the product’s constraints. Domain evaluators can supply an additional quality signal for music or imagery. These layers answer different questions: what is nearby in representation space, what fits this situation, and what is useful enough to show.
Fast System-1 decision models are interesting here because many product decisions are bounded choices. They can rank a small set of candidates without generating a long explanation. My CLM work explores this for audio and visual decisions, while DeepReplay applies a similar separation of relevance selection and main-agent work to project context.
Putting the layers together
The next entertainment experience needs a chain of explicit decisions: generate or retrieve candidates, represent and deliver them efficiently, evaluate their quality, and select them in context. Each layer needs its own measurements. Bitrate cannot stand in for latency, similarity cannot stand in for intention, and a reward score cannot stand in for a complete user study.
The direction I find most useful is to prototype the full path early. Put the decoder on the actual device. Compare candidate rankings under different prompts. Test the evaluator’s failure cases. Measure where a model saves work and where it introduces new work. That is how an interesting generative capability becomes an engineering system that can be understood, improved and used.