The observation this session builds on is that inference stopped being the cheap half.
Thinking models consume large numbers of tokens while reasoning through a problem (1:44), which means the compute profile of serving a model no longer resembles the profile that shaped serving infrastructure. Training and inference are described as two distinct phases with genuinely different requirements (1:48), and the response is to stop treating them with one general-purpose system (2:38).
Where the difficulty actually lives
The framing that transfers past any particular hardware is their account of what makes inference hard.
The difficult problems are memory management, scheduling and hardware utilisation, plus abstracting all of it away from whoever is building on top (4:19). Not matrix arithmetic. The chip does the arithmetic; the surrounding system decides whether the chip is busy.
The concrete bottleneck they identify is the key-value cache. As context lengths grow and concurrency rises, managing that cache becomes one of the biggest constraints on inference (4:38). This is the unglamorous centre of serving economics: a system that recomputes cached attention state is doing expensive work twice, and one that keeps too much of it runs out of memory under load.
The optimisation follows directly — reuse what was already calculated rather than recalculating it (5:13) — and it matters more as conversations get longer, which is exactly the direction usage is moving.
Batching, and why the old approach broke
The scheduling change is small to state and large in effect.
Static request-level batching groups a fixed set of requests and processes them together. It assumes they take similar amounts of time. Generation does not work that way: requests finish at wildly different points depending on how much output each produces, so a static batch is held hostage by its slowest member while finished slots sit idle.
Continuous batching (5:32) admits new work as slots free. The reason this became necessary is a consequence of reasoning models — variable-length output is no longer an edge case, it is the normal case, and infrastructure calibrated to uniform requests wastes most of its capacity.
The claim that determines whether any of it is adoptable
The commitment that matters commercially is stated almost in passing: as a developer you never have to rewrite your application layer or serving stack (6:43).
That is the whole argument for specialised hardware. Performance gains that require rewriting the serving layer are not gains — they are a migration project with a performance argument attached, and most organisations decline. Gains available behind an unchanged interface are adoptable by teams who never think about the hardware at all.
The corresponding detail is that this work happens in the open, including contributions arriving from outside — a research group's speculative decoding technique is cited as an example (14:54). For infrastructure whose value depends on being the default path, that openness is strategy rather than generosity.
演讲章节
关键要点
- 01
Thinking models consume large numbers of tokens while reasoning, which changes the compute profile serving infrastructure was built for. 1:44
- 02
The hard problems in inference are memory management, scheduling and utilisation rather than arithmetic — the chip computes, the system decides whether it is busy. 4:19
- 03
Key-value cache management becomes one of the biggest bottlenecks as context lengths grow and concurrency rises. 4:38
- 04
Continuous batching replaces static request-level batches because variable-length output stopped being an edge case. 5:32
- 05
Their commercially decisive commitment is that developers never rewrite the application layer or serving stack to get the gains. 6:43
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.