Berkeley lectures on AI, robot goals, and non-human worlds
Each episode is a talk presented by University of California Television, and the speakers are clearly addressing a room, not a studio mic. The AI talks put four-year-olds beside large language models and ask how a robot should treat human correction, and two episodes on communities turn to animist politics and the mounds of Cahokia.
PRESS PLAY
UC Berkeley (Audio) collects recorded talks presented by University of California Television, and it keeps the sound of a room. Episodes open with a UCTV appeal for donations or ratings and close by pointing to uctv.tv. In between, the speakers hold the floor. The AI talks deal in ideas and consequences more than products. In 'Imitation and Innovation in AI: What 4-Year-Olds Can Do and AI Can’t (Yet)', the speaker proposes a different way to think about large language and image models and sets them beside childhood and human intelligence. The argument turns on one question: is such a system an agent or a cultural technology? On this account, reinforcement learning from human feedback does not send a model out to find the truth. It only gives the model another, less transparent way of taking information from other people. 'AI Agents That Do What We Want' comes at the problem from robotics. Its speaker remembers graduate school at Carnegie Mellon's Robotics Institute, when robots mostly knocked things over, and moves on to the question of what objectives a robot should have. The demonstration is concrete: a robot pushed to stay closer to a table does exactly that, because it takes the correction as evidence about what people want. It also stays uncertain about its reward model, and when the possible models start to disagree in a new situation, the talk treats that as the warning sign. By the end, the speaker has moved past AI to a broader question about decisions that change our own future values. 'Data Dignity and the Inversion of AI' starts inside Microsoft's year in AI and the release of ChatGPT, then gets less comfortable. The speaker distrusts a future in which everyone lives on universal basic income paid through a single channel, warning that whatever organization carries a whole society's support will be seized by the worst people. A friendly jab at Sam Altman's eyeball-scanning crypto project makes it in too. 'How to Create AI to Solve Real-World Problems' opens with a robot sent to explore another planet with no mission control, which needs both preparation and the resourcefulness of someone stranded on a desert island. Then the feed leaves computing behind. 'Exploring Communities: Humans and Non-Humans Together' presents a political anthropology beyond the human, in which cosmopolitics becomes diplomacy among humans, plants, animals and more or less vindictive spirits, and it closes on legal rights for a milieu as something local people have to fight for. 'Exploring Modern Communities' takes Professor Descola's lectures in a different direction, setting Cahokia's Monk's Mound beside the modern landfill of East St. Louis while warning against idealizing Cahokia. A listener who wants the reasoning rather than the headline will find it here, and should be ready to go from reward models to spirits in the same feed.
Made for: Suits a listener who wants the thinking behind AI debates set out as a full argument rather than a headline, and who is curious enough to follow the same feed into the anthropology of humans, animals and spirits.
What sets it apart: Instead of hosts trading takes, it carries university talks built as full arguments, from a robot nudged toward a table to diplomacy with spirits. AI theory and the anthropology of non-human communities share one feed.
Programs from the University of California, Berkeley.
A robot meets lava, and your tidy theory of what it wants suddenly needs a meeting.
Press play if you want
Short reviews from the PlayNext desk, based on the episodes we processed.
A developmental view of AI asks what children can do that current systems still cannot.
The review asks what kind of intelligence large language and image models actually show, then shifts the focus to what changes when people tune those systems through reinforcement learning from human feedback.
A robot pausing at lava turns uncertainty about human objectives into an engineering requirement.
AI agents that do what people want can sound like a tidy software problem, right up until a robot reaches unfamiliar ground and its instructions stop being enough.
A demanding, thoughtful episode on layered worlds, living territories, and rivers as legal subjects.
Territory becomes more than a human boundary in this single-episode review, as the discussion follows animist cosmologies where people share layered worlds with plants, animals, and spirits.
A focused lecture on cities as world-making systems, from Cahokia's mounds to an East St. Louis landfill.
A lecture about assembling humans and non-humans opens into a sharper look at how cities arrange people, other beings, and the ground they live on.
From economic pressures to investment consequences.
AI capability meets the work of building.
Daily AI news through the lens of trust, infrastructure, and real-world deployment.
At UC Berkeley, the speaker argues that AI models are better understood as "cultural technologies" than as individual agents.
At UC Berkeley, the talk "AI Agents That Do What We Want" argues that robots should infer "intended reward" instead of optimizing a fixed reward function.
A speaker discusses the impact of ChatGPT, AI's evolution, and lessons from mentor Marvin Minsky.
The lecture links Eduardo Viveiros de Castro’s perspectivism, the Matsiganga myth from the Peruvian Amazon, and animist cosmopolities across Amazonia, Melanesia, and Southeast Asia.
An academic lecture discussing the assembly of humans and non-humans in extra-modern collectives, challenging the nature-culture divide.
A planetary robot needs both preparation and the ability to adapt when unfamiliar problems arise.
Stuart Russell argues that general-purpose AI could expand access to healthcare and education, with a stated estimate of $13.5 quadrillion in net present value.
The UC Berkeley lecture questions whether anthropology can describe other worlds using European concepts such as “nature,” “society,” and “culture.” Wodabe Fulani “doodal” joins a human lineage with its zebu herd, while Tuva “Al Kodan” includes people, animals, buildings, and land.