A follow-up to Alien Intelligence
The Room We Get Only Once
What carries over, what can start again, and what makes us stop and ask.
At six on a rainy Saturday, forty people are coming to a neighborhood supper. Last month the first guests stood outside under an awning while the volunteer with the key was still on her way. You forget who brought the cornbread. You remember to have somebody at the door.
Give an AI last month’s schedule, the room dimensions, and the list of volunteers. It can preserve the exact minute the first guest arrived, try three ways to set the room, and put a volunteer’s name in every box. It can still assign the door to someone who has never agreed to take the key.
In Alien Intelligence I used two arrows pointing in opposite directions to describe how a person and an AI can attend to the same material. I’ve been looking for places where the difference shows up in what they do. This room gives us three. Hold the job steady and watch what each carries into the next hour.
What survives the night
Ask the person who ran the last supper a week later what happened. She may remember the guests caught in the rain and the rule she made afterward. Ask whether they waited six minutes or sixteen, and she may have to check. Melanie Sekeres and her colleagues found something related when they showed people short films. Across a week, people lost peripheral details faster than the core of what happened. A brief reminder could bring some of the details back. Read the memory study.
An assistant with access to the old event log can take the opposite route. It can recover an exact timestamp with a search, yet in a fresh session it may repeat last month’s staffing plan. The evidence is sitting in a file. For it to change the next plan, the assistant has to retrieve the right record, extract the lesson, and bring it to bear. A correction in one chat can change the rest of that chat without becoming part of the next one.
In a recent engineering study, Anthropic gave its agents a folder that persisted across fresh sessions. On a series of database questions, some models used it to turn their mistakes into checked rules they could use later. The lesson could travel because it had somewhere to live and a way back into the work. Read Anthropic’s study of agent memory.
I would give the supper two records. In one I’d keep the arrival times and the message from the person with the key. In the other I’d give the next organizer a rule to follow. Assign a keyholder before opening and confirm they can get there before the guests. Save only the log and the next assistant has to discover the lesson again. Save only the rule and you may keep following it after the venue installs a keypad. With both in hand, someone can ask why the rule exists and when it should change.
What can start again
You can imagine opening the side door instead of the front. You can even come back the following Saturday and try it. The first evening still happened, and the people who waited in the rain will bring that experience with them. We’re quite good at imagining different versions of the past. Living one version changes the conditions under which we can try another. Read the study of imagined alternatives to personal events.
Give three AI runs the same saved plan and three copies of a test schedule. In the first copy, change the arrival pattern so that most guests come at once. In the second, remove a volunteer. In the third, make the front entrance unavailable. Each run starts with the plan as it stood before any of the incidents, and each leaves a separate record of what it tried. You can rerun the same case with a revised plan and see which part of the response changed.
Go offers a clean example of this difference. AlphaGo Zero used tree search to explore alternative continuations from the same board position before choosing a move. A human player can study several lines too. A computer can copy the board state and run out the alternatives repeatedly under the same rules. The room is harder than the board. For instance, a test schedule can account for one missing volunteer, but it takes a visit to find out how loud the front room gets when the rain hits the tin awning. Read the AlphaGo Zero paper.
I’d use the three runs to find the steps we can’t afford to improvise. If all three plans depend on somebody fetching the spare key, we know what to check before Saturday. If all three agree because they inherit the same obsolete note about where the key is kept, the agreement adds very little. Repetition lets us change the assumptions one at a time. A visit lets us find out which ones were wrong.
What makes you stop
There’s a name beside “front door” on the schedule. There’s no reply from the volunteer. At five fifteen, a person who has stood under that awning has a reason to put down the schedule and make a call. In experiments led by Kobe Desender, people were more likely to request another look at the evidence when they felt less confident, even under conditions where their answers were equally accurate. Uncertainty can change what people do before they answer. Read the study of confidence and information seeking.
A model can ask too. Whether that behavior is encouraged depends in part on what counts as success. Adam Tauman Kalai and his colleagues showed how evaluations that award a point for a correct answer and give the same score to a wrong answer and an abstention encourage guessing. Sometimes the schedule with every box filled gets the point. On Saturday, a definite answer with no confirmation can leave forty people standing in the rain. Read the paper on evaluation incentives and hallucinations.
Before the doors open
At five fifteen I want the AI to say, “Maya is listed for the front door. She hasn’t confirmed. I’ll ask her now.” If Maya confirms, the schedule can carry that fact into the next shift. If she can’t make it, the assistant can try the backup we rehearsed. If nobody can confirm, I need to know while there’s still time to walk over with my key.
After the supper, we can save the record and revise the plan. Next month the AI can bring back exactly what happened and rehearse from a clean copy again. When a real guest reaches for the handle, there’s one door, one key, and one Saturday night.