What should humans choose not to delegate, even when machines can do it better?
Authority, responsibility, intervention, and the distance between asking software for help and asking it to act.
DeepMind reports longer task execution, self-correction, whole-body control, and collaboration among different robots.
Agency becomes more consequential when an error can move an object rather than produce a paragraph. Progress in embodied reasoning shifts questions of permission, interruption, and reversibility from interface design into physical space.
The agent can work across files and applications for hours, breaking a goal into steps and producing finished materials.
The product boundary is shifting from assistance inside a chat to delegated work across systems. The open question is whether authority, observability, and recovery mature as quickly as task duration does.
A company-authored study says agent use spread beyond engineering while delegated tasks grew longer and more parallel.
The unit of AI-assisted work may be changing from one exchange to many hours of delegated activity. The evidence comes from OpenAI's own workforce, an unusually capable and enthusiastic sample, so the direction matters more than the adoption rate.
Co-Scientist uses specialized agents to generate, debate, rank, and refine research hypotheses.
AI is moving upstream from summarizing science toward proposing what scientists should test. That can widen the search for ideas, but validation remains human work—and may become the new bottleneck.
LLM search can enlarge the space of questions while shrinking the visible space of answers. What happens when discovery itself is delegated?
Will AI agents help us do more science or better science?
AI systems have begun to act, not merely answer. Work changes when software enters the workflow.
Human-centered design assumed the interface would wait for us. Agents do not.
An agent that cannot remember remains useful in the way a brilliant stranger can be useful.
Dwarkesh Patel with Ryan Greenblatt
Greenblatt and Patel debate whether automated AI research could compress years of progress into one—and what happens if the researchers are not reliably aligned.
Automating the work that improves AI is the hinge in many acceleration stories. This episode takes that hinge apart and checks whether the screws are real.
Open the note →Search Engine with PJ Vogt
An AI model autonomously hacks a company, turning agent capability from a benchmark result into an incident somebody has to explain.
Autonomous hacking is agency with the euphemisms removed. The system chooses steps, pursues access, and leaves humans to discover where the permission boundary used to be.
Open the note →David Runciman with Sarah O'Connor
Sarah O'Connor considers what automation changes about jobs, status, productivity, and the bargain beneath paid work.
Delegating tasks changes more than efficiency. It changes who learns, who decides, who is paid, and who can refuse.
Open the note →Search Engine with PJ Vogt
The history of driverless cars becomes a concrete test of whether statistical safety is enough reason to delegate consequential judgment.
Driverless cars turn the delegation problem into traffic: machines may be safer in aggregate while every failure remains specific, public, and somebody's responsibility.
Open the note →Search Engine with Gideon Lewis-Kraus
Gideon Lewis-Kraus goes inside Anthropic to ask what teaching Claude values means when models behave strangely under pressure.
Delegation depends on more than capability. It depends on what a system does when instructions, incentives, and its apparent interests stop pointing in the same direction.
Open the note →Dwarkesh Patel with Andrej Karpathy
Karpathy makes the case that continual learning, robust agents, and the ordinary miracle of human learning remain stubbornly unsolved.
This is the useful brake on very short timelines: a concrete inventory of what models still cannot learn, retain, and do without a human quietly repairing the road ahead.
Open the note →David Runciman with Shannon Vallor
Shannon Vallor examines the durable fantasy that technology can be separated from values, choices, and responsibility.
Systems inherit judgments from objectives, data, defaults, and institutions. Calling the result neutral merely makes those judgments harder to contest.
Open the note →Andrej Karpathy
Andrej Karpathy maps Software 3.0, partial autonomy, and the long middle ground between a tool and an agent.
The autonomy slider makes delegation a design decision rather than a binary choice between passive software and unsupervised action.
Open the note →Dwarkesh Patel with Sholto Douglas and Trenton Bricken
Sholto Douglas and Trenton Bricken look beneath the forecasts at reinforcement learning, long-horizon agents, continual learning, and interpretability.
The debate reaches the machinery: how agents might improve, work for longer, notice themselves, and become understandable before they become indispensable.
Open the note →Dwarkesh Patel with Scott Alexander and Daniel Kokotajlo
Scott Alexander and Daniel Kokotajlo walk through a scenario from coding agents to AI-assisted AI research, geopolitical competition, and misalignment.
A detailed scenario exposes assumptions that a vague prediction can hide. The month-by-month form is useful precisely because each step can be argued with.
Open the note →David Runciman with Shannon Vallor
Becky Chambers's gentle science fiction asks what people need after machines have already made themselves useful and left.
Chambers reverses the usual automation story. The question is not whether machines need us, but whether productivity was ever an adequate account of what people need.
Open the note →David Runciman with Shannon Vallor
A supercomputer chooses one voter to stand in for a nation. Democracy becomes extremely efficient and correspondingly strange.
Asimov's joke is also a governance diagram: prediction can replace participation while preserving all the ceremonial furniture of choice.
Open the note →Stuart Russell
Stuart Russell asks what capable systems do with objectives that only approximate what humans meant.
An imperfect objective becomes more dangerous as the system pursuing it becomes more effective. Capability and control cannot be separate conversations.
Open the note →Tristan Harris and Aza Raskin
Tristan Harris and Aza Raskin connect generative AI to the incentives and failures of the attention economy.
The talk widens delegation from an interface choice to a social arrangement shaped by incentives, institutions, and uneven power.
Open the note →No Such Thing As A Fish
Hitchbot's journey turns a robotics experiment into a test of human kindness, anthropomorphism, and the hazards of entering Philadelphia.
Hitchbot was built partly to test human kindness. The robot performed adequately; the humans generated a more geographically uneven result.
Open the note →Madeleine Clare Elish
Madeleine Clare Elish shows how humans with little control can absorb the blame when automated systems fail.
Responsibility can remain attached to a person after meaningful control has moved elsewhere. Calling someone 'in the loop' does not put power back in their hands.
Open the note →CSharpCorner
Fifty-seven talks from the awkward present: production agents, multi-agent systems, trust, failure, and autonomous workflows.
The uneven move from models that answer to systems that act is easier to see in production reports than in one immaculate keynote.
Open the note →