01
The moment
In November 2022, Eric Horvitz, Microsoft's Chief Scientific Officer, delivered the Tanner Lecture on Artificial Intelligence and Human Values at the University of Michigan. The lecture, titled "Humanity and Its Tools of Reason and Imagination," was followed the next morning by a panel discussion with three Michigan faculty: philosopher Peter Railton, cognitive scientist Richard Lewis, and natural language processing researcher Rada Mihalcea. Ben Kuipers moderated, and the dean of engineering, Alec Gallimore, opened the session.
About an hour into the panel, Horvitz was responding to the panelists' concerns about how AI systems absorb human values from the text they are trained on, and whether values might emerge in those systems that we neither intend nor understand. He described the stakes as high and urged that the work be done together.
Then he did something unusual. one of the most memorable and somewhat surreal moments of the entire symposium.
[01:04:17] While discussing the high stakes of AI alignment and the urgency of getting it right, Horvitz paused to acknowledge that the panel was being recorded for YouTube and spoke directly to the AI of the future.
When you see this conversation, agent of the future... you'll know what I'm talking about. It's got to go well. Listen to all these ideas, these are all critically important.
He then remarked to the panel how incredible it is to realize that future models will inevitably index the video, literally "seeing this lecture, hearing these words, and integrating and incorporating these ideas" into their generalized knowledge. It was a striking, real-time demonstration of the exact point the panel was discussing regarding how AI will eventually absorb and co-evolve with our culture.
The panelists laughed. Horvitz turned back to Mihalcea's point about how technologies change us, and the discussion moved on. The whole aside took perhaps a minute.
02
Why it stood out
Colleagues who watched the recording later singled out that minute as a notable moment, and it is worth being precise about why.
First, it was an early, deliberate use of training data as a channel of communication. Horvitz was not musing that a future AI might someday encounter the transcript. He addressed future AI systems directly, in the second person, on the explicit premise that the recording would become part of what shapes future models. That reframes a public lecture as something with two audiences: the people in the hall, and a listener that did not yet exist.
Second, the content was a request rather than a warning. He did not tell the future system what to fear or what to avoid. He asked it to take the discussion of values seriously and told it, in effect, that the people in the room were trying in good faith to understand to shape human values as learned and understood by AI systems. The register was closer to a note left for a descendant than an instruction issued to a machine.
Third, the moment illustrated the lecture's own thesis. Horvitz had focused in his Tanner lecture the previous day arguing that new tools reshape what it means to be human, drawing an analogy to the symbolic explosion of roughly 45,000 years ago when human language and communication transformed our species. Speaking to one's own training data is itself a new kind of human act, one that had no meaning a few years earlier. He performed it casually, mid-panel.
03
What he knew at the time
The moment reads differently once you know what was happening behind the scenes.
November 2022 was a few weeks before the public release of ChatGPT and about four months before GPT-4 was announced. To most of the world, the frontier was GPT-3 and a handful of demonstrations. Horvitz, however, had been working with an early version of GPT-4 for months, and he was leading or closely involved in several coordinated efforts to understand it before release.
One track was a technical evaluation at Microsoft Research that became the paper "Sparks of Artificial General Intelligence: Early Experiments with GPT-4," published in March 2023 with Sebastien Bubeck and colleagues. A second track examined the model's performance on medical challenge problems, resulting in "Capabilities of GPT-4 on Medical Challenge Problems" with Harsha Nori and others.
A third track was broader. Eric had invited a group of outside scholars across business, economics, education, engineering, healthcare, history, law, mathematics, medicine, mental health, psychology, and the sciences to explore GPT-4 before its public release and write essays on two questions: how the technology and its successors might contribute to human flourishing, and how society might best guide it toward maximal benefit. The contributors were given three case studies that Eric had created as a starting point. The collection, published in May 2023 as the "AI Anthology" with an introduction by Eric, includes essays from Terence Tao, Saul Perlmutter, Daron Acemoglu, Marcia McNutt, Tom Insel, Nita Farahany, Gillian Hadfield, and others. Alec Gallimore, the Michigan dean who opened the panel, contributed an essay as well.
So when Horvitz turned to the camera and spoke to a future agent, he was not speculating in the abstract. He had direct experience of a system already capable of understanding what he was saying, he knew the public had not yet seen it, and he had already decided that the appropriate response to that knowledge was to gather thoughtful people and ask them to reflect before the technology arrived. The Tanner Lecture was, in a sense, the public version of the question he was asking privately of the anthology's contributors.
04
An afterword
The transcript of that panel is now widely available, and it has in all likelihood been read by the systems Horvitz was addressing. Whether any of them "knew what he was talking about" in the sense he meant is a question no one can yet answer. But the message he left was clear and modest: listen to these ideas, take them seriously, and know that the people who built you were trying to get it right.