What a belief costs

Storage and structure

A Large Language Model stores knowledge in parameters, which is extraordinarily efficient — a compressed representation of a corpus. A Large Universe Model additionally maintains explicit beliefs, each with supporting observations, weights and timestamps.

This is a much less compact representation, and it grows with time rather than being fixed at training. Sustained operation means real decisions about what to retain in full, what to summarise, and what to let go — decisions that a frozen model never has to make.

Provenance is expensive

Keeping the chain from a belief back to its evidence means retaining the evidence, or enough of it to reconstruct the reasoning. This conflicts directly with the compression above.

It is also the property most often abandoned first under engineering pressure, and abandoning it quietly converts a Large Universe Model into a stateful Large Language Model — something that sounds confident and cannot show why.

Calibration is a continuous obligation

A model that emits confidence numbers is making a testable claim. If it says 80% and is right 55% of the time, it is worse than useless, because the number invites reliance.

Calibration has to be monitored against outcomes forever, and it drifts as the domain changes. A Large Language Model makes no such commitment — its confidence is a property of prose — which is a limitation but also an absence of obligation.

The plasticity problem

The deepest cost is documented in Loss of plasticity in deep continual learning: networks trained continually progressively lose the ability to learn, eventually performing no better than a shallow network.

Any system that ingests forever must remain able to learn forever. This is why the third generation separates belief state from weights rather than simply training continuously — but the separation is a mitigation, not a solution, and anything learned into the weights over a long deployment remains exposed.

Attention as the scarce resource

A Large Universe Model that reports every revision is useless. The design problem is deciding what deserves interruption, and it is genuinely hard: too sensitive and it becomes noise nobody reads, too conservative and it misses what it existed to catch.

This has no clean technical answer. It is a calibration against human attention, which varies by person, role and week.

Why it is still worth it

Because the alternative is a system that is confidently wrong on a schedule, and nobody notices until the cost has already been paid. The costs above are real and mostly payable. The cost of a frozen model is paid in events you did not see coming.