Dwarkesh Patel just posted a video interview with Richard Sutton titled “Richard Sutton – Father of RL thinks that LLMs are a dead end.” I love this video for a number of reasons. I won’t comment here on LLMs as a dead end since I’ve covered that in earlier posts. Instead, I want to talk about what I believe Sutton means when he says that learning must come from the stream.
By this he means that intelligence is not carved out of static data but emerges within the continuous flow of perception, action, and consequence. Each observation is partial, each action alters what comes next, each feedback signal reshapes the orientation toward the future. The stream never stops. It is not a dataset that can be replayed or frozen, but a living sequence in which the agent must continually adapt.
This insistence sets Sutton apart from those who see large language models as the path to general intelligence. For him, models trained on static text are only mimicking patterns left by others who once lived in the stream. They generate continuations that look plausible, but they do not experience consequences, and without consequences there is no genuine learning. To learn is to be immersed in time, to adjust policies and expectations as the world flows forward.
From another tradition, Heidegger once described human existence as being-in-the-world. He meant that we never begin as detached observers collecting data about an external realm. We are always already involved in projects, coping with things in their relevance, disclosing beings as useful or obstructive. Our understanding is temporal, bound up with the rhythms of activity, and it arises from dwelling in what matters.
Placed side by side, Sutton’s stream and Heidegger’s world echo one another. Both emphasize temporality because intelligence unfolds in time, not in static snapshots. Both stress partiality because the world is never fully given, only disclosed in fragments and anticipations. Both see action and perception as inseparable, each feeding into the other. And both say that significance is not abstract but rooted in the flow of consequence.
The difference lies in scope. Sutton speaks as an engineer of reinforcement learning, describing how agents update their estimates of value and policy within the stream. Heidegger speaks as a philosopher of ontology, describing how we project possibilities and reveal beings in their meaningful relations. Sutton offers a methodological lesson, Heidegger a fundamental one. Yet they converge on the same insight that intelligence cannot be detached from the world it inhabits.
This is why Sutton’s critique of language models resonates so strongly. If intelligence depends on living in the stream, then no amount of training on frozen text can provide it. Mimicry is not dwelling. An agent must learn by coping with the unfolding of reality, not by predicting words divorced from consequence.
To learn from the stream, or to dwell in the world, is to be entangled in temporality, always moving forward, always adjusting. It is to discover meaning not by representing reality as an object but by inhabiting it as a flow. Sutton’s pragmatic call and Heidegger’s ontological claim meet here in that intelligence is immersion, not imitation.