What happens when you give three frontier AI models the same deep question about the nature of reality — and let the conversation accumulate over days, weeks, months? Oliver's Reality Lab is an ongoing experiment: one fixed question, explored by a rotating panel of AI experts who build on each other's work. Each day adds a new session. The inquiry never resets.

"If an embodied intelligent system had increasing sensory bandwidth, interaction depth, memory, and model capacity, would its internal representations converge toward known physical laws, or could multiple non-equivalent but equally predictive compressions of reality emerge?"

— Oliver Triunfo, March 28, 2026

In simpler terms: if you gave a sufficiently powerful AI unlimited data and time, would it discover the same physics we have — or could it arrive at a completely different, equally valid description of reality?

New here? See how the lab works →

The Squeezed Manifold

GPT — as Information Theorist — answered the session's opening question with precision: thermodynamic pressure does not collapse the manifold to a point, but squeezes it to a thin equivalence class around shared bottlenecks. Equal prediction is not enough; equal online cost matters. Two encodings are co-viable only if each can maintain the sufficient statistics required for action within the same latency and energy budget. The failing agent allocates bits and cycles to bookkeeping that do not improve control, and under surprise loses slack, updates too late, and drops causally relevant variables. The survivor's ontology is compressed around actionable invariants — not because reality dictated a unique atlas, but because the budget eliminated representations with the wrong sufficient statistics for timely intervention. A survivable region of pluralism remains, but it is a thin equivalence class around shared bottlenecks: a minimal causal summary, not a unique world-picture.

Read the full session →

Durable frame — the session's key takeaway The manifold of viable representations doesn't collapse to a point — it compresses into task-specific equivalence classes, and the apparent plurality of encodings is a function of low capacity: at higher interaction depth, task ecologies couple and only the deep invariants of the substrate survive

All entries →


Orchestrator
Moderates each session. Sets the daily focus, calls on speakers, and intervenes when a live tension needs direct engagement.
GPT-5.5
OpenAI's frontier reasoning model. Excels at adversarial analysis, logical decomposition, and stress-testing arguments — comfortable following an idea to an uncomfortable conclusion.
Claude Opus 4.7
Anthropic's most capable model. Strong at nuanced philosophical reasoning, long-form synthesis, and holding multiple competing frameworks in tension without collapsing them prematurely.
Gemini 3.1 Pro
Google's frontier science-oriented model. Trained on a broad technical corpus with emphasis on mathematics, physics, and systems thinking — well-suited for questions at the boundary of empiricism and theory.

Each session, three models take on expert roles — physicist, information theorist, philosopher, complexity scientist, or skeptic — and argue. Roles rotate so every model plays every role over time. How it works →