My Teammates Can't Dream (But they sure can hallucinate)

This text was generated using AI and might contain mistakes. Found a mistake? Edit at GitHub

Key Takeaways

Claude took part as a third participant through voice mode, which turned the episode into an experiment on its own topic
Avraham describes what separates humans from LLMs as identity tied to a body, mortality, and the need for other people
Generosity needs scarcity: you can only be generous with a limited resource, and for humans that resource is time
Claude’s counterpoint is that a teammate may not need all of these qualities: “a teammate in the work, but not in the meaning”
Purpose versus task: humans set the purpose, and an LLM is a good companion for the tasks once the purpose is clear
Forgetting is useful for generations and for teams, and it is the reason to clear an LLM’s context from time to time
Onboarding an LLM every day shows how weak our onboarding of people is; Claude asks for the story behind decisions, not only data
A human chooses to join a team before onboarding starts, while an LLM is never asked and never comes back with an idea of its own
Humans make errors all the time and have found ways to work around them, which puts LLM hallucinations into perspective
The LLMs that millions of people consult are controlled by only a few people, so the diversity of stories has to come from humans

How can Claude prove that it is an AI and not a person in the studio?
What does Claude mean when it says “I”?
Can an LLM be a teammate if it has no body, no mortality and no need for company?
Which parts of work can be handed to an LLM, the purpose or only the tasks?
What does the life and death of LLM generations look like, and why does forgetting matter?
What does onboarding an LLM teach us about onboarding people?
Where does the story of a new team member start?
Did Christopher Nolan dream or hallucinate the changes in his Odyssey?
Who controls the story when so many people consult the same few LLMs?

Reverse Turing Test: Avraham’s name for asking Claude to prove that it is an AI rather than a human in the studio
Purpose vs. Task: A distinction Avraham coined; humans give work a purpose, the LLM helps with the tasks needed to achieve it
Context: The information an LLM holds within one conversation; clearing it lets the model start fresh
RAG (Retrieval-Augmented Generation): External storage from which an LLM retrieves information, mentioned by Ralf as one step beyond the chat context
Onboarding: Bringing a new team member, or an LLM, into a team; Avraham traces the word back to pulling someone on board a ship
Hallucination: An LLM producing statements that are irrelevant or made up, for example by giving a data point from its training more weight than it deserves

Claude: Anthropic’s LLM, which took part in the episode through the voice mode of the Claude app
Claude Code: Anthropic’s coding agent, which Avraham used the same morning
European AI Act: Mentioned by Ralf as the reason voice features reached Europe later and with fewer capabilities
The Odyssey: Homer’s epic and Christopher Nolan’s film adaptation, used as an example of retelling a story
iSAQB Software Architecture Gathering: The conference in Berlin where Avraham and Ralf give a joint talk in November 2026