The case for running models on the device is made in three clauses, and each represents a different kind of constraint (2:25, 2:34).
Cost, where a local model handling a task well removes a cloud API call. Availability, where the application works on a flight. And data, where requirements prevent information leaving the device at all.
The third is the one that changes what is buildable rather than what is affordable. Cost and offline capability are optimisations. A regulatory or contractual prohibition on transmitting data is a wall, and on-device inference is the only way through it.
The hardware caught up
The reason this is newly practical is unglamorous: improvements across compute targets, including new processor extensions integrated into the runtime (2:51).
Worth stating because on-device AI has been promised for several years and mostly disappointed. What changed is not a modelling breakthrough but silicon and the software that reaches it — instruction set extensions for the operations these models perform, and a runtime that actually uses them.
The reported result is around 30 per cent improvement in image editing functions including subject selection, sky selection and scene enhancement (23:17). Modest as a headline; decisive at the threshold where a feature becomes usable at all.
Small models need adaptation, not scale
The guidance about fine-tuning is precise about where it matters: adapting to a specific problem or domain is particularly important for smaller models, one billion parameters or below (7:57).
This is the practical inversion of the usual advice. Large models are general enough to handle a specific domain with prompting. Small models are not — their capability budget is spent, and getting them to perform on your problem means spending some of it on your problem specifically.
Which reframes what deploying a small model involves. It is not choosing a cheaper option from a menu. It is committing to an adaptation step that the large-model path lets you skip, in exchange for properties the large-model path cannot offer.
The examples are the argument
Two illustrations do more work than the specifications.
A game where characters respond rather than reciting pre-written dialogue trees (8:34). This is not primarily about cost — it is that dialogue generated per interaction cannot depend on a network round trip inside a game loop, so latency alone dictates the model runs locally.
And an acoustic monitor deployed into a forest to track a rare bird's migration (25:07). No connectivity, no power budget for transmission, and the processing has to happen where the microphone is.
Neither is a cheaper version of a cloud application. Both are applications that a cloud model cannot serve at any price, which is the honest case for this whole category.
关键数据
- ~30%
- reported performance improvement across image editing functions from runtime and silicon work 23:17
演讲章节
关键要点
- 01
Three arguments for local inference — cost, offline availability, and data that cannot leave the device — of which only the third changes what is buildable. 2:34
- 02
What made this practical is silicon and runtime work rather than modelling: processor extensions integrated into the inference runtime. 2:51
- 03
Around thirty per cent improvement is reported across image editing functions including subject and sky selection. 23:17
- 04
Domain adaptation matters most for the smallest models, whose capability budget is already spent and must be redirected toward your problem. 7:57
- 05
An acoustic monitor placed in a forest to track a rare species is the clearest case: no connectivity, no transmission budget, processing at the microphone. 25:07
提及的实体
相关演讲
Four of Google's model, product and search leads on what changes once agents run for hours rather than seconds, and the most quotable argument comes from Dean: the constraint is moving out of the model and into the tools around it. By Amdahl's law, an agent spending half its time in tools built for human-speed interaction cannot gain more than a doubling however fast the model becomes — which reframes a great deal of current infrastructure work as latency debt. Their internal response is concrete: rewriting Python tooling into Go, framed as a fully specified translation task rather than an open prompt, produced order-of-magnitude speedups overnight. Reid supplies the counterweight from Search, where acceptable latency turns out to scale with how much work is being taken off the user rather than being a fixed budget. Woodward's detail is the quietest and perhaps the most telling: teams that have stopped writing product documents for humans and now write context files for models to act on directly.
The anecdote that opens the panel does the work: a robot asked to pick up the extinct animal selected a dinosaur toy, with nothing in its training data connecting the phrase to the object. That transfer from language models into machines with hands is the premise of the current wave. What the practitioners then describe is where it stops. Physical intelligence is about exerting force and using a body to do it, which is knowledge about consequences — the one thing a corpus of internet images contains almost nothing about. The humanoid question gets an honest treatment: not that human shape is optimal, but that the world is already built for it, plus a development-loop argument about collecting data and deploying on the same hardware. The most useful passage is scepticism about the field's favourite shortcut: generated video looks realistic and does not hold up for dexterous manipulation, because looking right and being physically consistent are different properties.
The observation that organises this session is not about capability but about attention. Engineers increasingly file a ticket rather than opening an editor, and the code comes back — which changes what the surrounding tools are for. If the agent works while you do something else, the conversation between you is no longer a workspace; it is the mechanism by which the agent surfaces a question it cannot resolve alone. Interfaces built for continuous conversation optimise for flow, and interfaces built for interruption should optimise for the opposite. A runtime constraint follows immediately: an agent that starts a long-running job cannot block until it finishes, which turns out to be a workflow-engine problem rather than a model one. The panel's closing formulation — that deciding what to build is the hard skill and always was — reads as reassurance and functions as a warning, since that judgement is downstream of exactly the work now being delegated.
The most honest moment here is an aside about how the presenters have tracked their own projects: plans in documents, plans in spreadsheets, plans in bug comments, and once a plan written on a receipt. That describes the actual category being addressed — not software nobody has built, but the small internal tool every team improvises badly because building it properly was never worth the effort. The demo turns on a single question: the generated app is strictly single player, so what happens when you want to share it with the team? That boundary is where improvised tools historically died, because it is where accounts, shared storage and access rules begin. Here it is crossed in one step, with the access rules generated and deployed automatically — which is convenient, and is also the moment the application acquires obligations nobody reviewed.
The clearest statement at I/O of how an agent API differs from a chat API, and the reasoning behind each departure is stated rather than assumed. Three changes matter. Conversation state moves to the server: a call returns an identifier, and passing it back continues the thread, retiring the client-side history array. The data model abandons alternating user and model turns for discrete steps, on the argument that a trace containing reasoning, tool calls, environment responses and compaction was never really a conversation and modelling it as one distorted it. And agents receive their own persistent remote environment rather than acting on the caller's machine — addressable by identifier, and shareable, so a research agent's output files become an application builder's input without passing through the context window. Schmid is explicit that scaffolded environment files are deliberately not model input, which is what keeps large artefacts out of the context budget. Schaeff's first half covers the real-time voice path, where the notable property is speech-to-speech across ninety languages with transcription of both directions.
Pichai's framing of Google as the buffer between people and the raw internet is offered as continuity — search did it, browsers did it, agents do it more — and it is also the most contested claim in the industry, because a buffer decides what passes through. He reaches immediately for the counterweight, the connection people feel to creators they follow, which is precisely the tension the company is currently managing without resolving. Two answers are sharper than the format usually produces. On competition he describes participants running on different pre-training and release cadences rather than at different speeds in one race, which is a more honest account than the leaderboard framing and comes from someone with an interest in leaderboards. On security he acknowledges models improving at cyber work, which is the one domain where better capability does not obviously net out positive, since an attacker needs one vulnerability and a defender needs all of them.