Google's Chief Scientist Says Infinitely Fast AI Won't Help You.
TL;DR
The real AI bottleneck is the human-shaped web, not model speed — Nate Jones uses Jeff Dean’s GTC point that even an “infinitely fast” model might only deliver a 2–3x productivity gain because compilers, APIs, auth flows, file systems, pagination, and dashboards were all built for humans, not agents.
Agents are already moving at superhuman speed, and infrastructure is lagging badly — He cites agents operating at 10–50x human reasoning speed, FANG companies saying 20–40% of code is AI-written, Anthropic claiming Claude Code writes 80% of its own code, and NVIDIA’s Billy Dally saying inference now consumes 90% of data center power and is heading toward 10,000–20,000 tokens per second per user.
“Just add MCP” is not an agent-native strategy — Wrapping a human-friendly API in MCP may make it technically accessible, but Jones argues agents still burn wall-clock time fighting pagination, authentication, and other human-speed assumptions, so the experience remains fundamentally non-native.
The rebuild is happening in three layers: faster tools, agent-native primitives, then a rewritten web stack — He walks from TypeScript 7 moving to Go for 10x+ speedups and Rust being easier for agents to verify, to OpenAI’s persistent containers and hosted shells, to deeper primitives like BranchFS and shared KV caches that assume no eyes, no hands, and no coffee breaks.
The durable career value shifts upward into human judgment and orchestration — Jones says the future belongs to four or five roles: the AI tool-using generalist, the pipeline/infrastructure builder, the relationship-heavy business closer, the “grown-up in the room” leader who knows when to brake, and possibly a rare high-end creative in the Steve Jobs mold.
He frames this as a promotion, not a demotion — His core message is that humans are no longer the primary users of computing systems, but that doesn’t make us obsolete; it pushes us toward the hardest layer of the stack: directing, constraining, designing, and humanizing an agentic economy arriving in the next 12–24 months.
The Breakdown
The web was built for eyeballs and hands
Jones opens with a big claim: for 50 years, humans were the center of computing, and that’s no longer true. He makes it concrete with familiar examples — spreadsheets, dashboards, logins, API pagination — all of it engineered around how fast people can read, click, and think, which he says was brilliant design until roughly now.
Jeff Dean’s “infinitely fast AI” warning
The turning point is Jeff Dean at GTC: if an agent is already operating 50x faster than a human, then startup time, tool switching, and interface friction suddenly matter a lot. Dean’s punchline, which gives the video its title, is that even infinitely fast models might only improve productivity 2–3x because the slowdown lives in our tools, not the model.
We spent a trillion dollars on intelligence — and bottlenecked it on tool calls
Jones leans into the irony here. AI can now reason and code at serious speed — he mentions 20–40% AI-written code at major tech firms, Anthropic’s claim that Claude Code writes 80% of its own code, and Billy Dally’s note that inference is already 90% of data center power — but so much wall-clock time is still spent waiting on file systems, APIs, CRMs, ERPs, and test suites built for humans.
Layer one: existing tools get rebuilt for speed
The first layer of change is straightforward optimization, especially in developer tooling. He points to the JavaScript ecosystem moving toward Rust, Go, and Zig, and TypeScript 7 being rewritten in Go for a 10x+ speedup; he also highlights Lee Robinson building a 38,000-line Rust image compressor with coding agents, where Rust’s strict compiler becomes a kind of built-in verifier for AI-generated code.
Layer two: tools give way to agent-native primitives
Then he gets more radical: maybe the right answer isn’t better human tools, but systems designed for consumers with no eyes, no hands, and no coffee breaks. His examples include OpenAI’s persistent containers and hosted shells, BranchFS for sub-third-of-a-second branch creation, and research showing multi-agent coordination through a shared KV cache that cuts latency by 3–4x versus text-based handoffs.
Layer three: the human scaffolding gets pinched off
This is where he invokes the “bitter lesson”: general computational methods eventually beat hand-built human scaffolding. Citing Aaron Levie, he argues every new model generation makes yesterday’s framework overhead feel heavier, so simply optimizing current tools is structurally the wrong move because faster models keep exposing more drag in human-centric systems.
So what happens to the humans?
Jones shifts from infrastructure to careers and lays out four main roles, maybe five. The future team, in his framing, needs a tool-using generalist who gets things moving, a pipeline/infrastructure builder, a human relationship-builder who can still close business over dinner, a mature leader who knows when to hit the brakes, and perhaps a rare creative visionary in the Steve Jobs mold.
His final framing: this is a promotion
He’s careful not to cast this as human obsolescence. The emotional landing is that humans are being moved up the stack: less raw execution, more judgment, direction, taste, and restraint in a world where agents handle more of the doing — and he says those role shifts are already visible now, not hypothetical.