A paper just came out of Meta and the University of Washington, that looks like a real breakthrough.
It introduces Context Language Models—CLMs—which hand the model full responsibility for its own context: not appending to a transcript, but rewriting it on-the-fly, with ordinary shell commands. The model decides what from its own history is still worth keeping. As per early reports, the results back it up: less compute, better retention, context management that beats hand-tuned harnesses, etc. It also shows, almost in passing, that a model can be told in plain language to back up its context before every edit.
For my part, beyond the metrics I'm wondering how much such approach gradually increases a deeper risk that the benchmark never captures: I call it "epistemic closure".
It's quite different from hallucination, and closer to what Ancient Greeks define as syllogism. An apparently flawless reasoning that gradually collapses by curating its own premises. Form a belief, then the evidence against it looks a little less worth carrying forward each time you compact. No step in that chain is a lie. The whole thing can run on perfectly valid logic, each conclusion following cleanly from a context that has quietly stopped containing what would have challenged it. Worth noticing: the paper mentions some built-in backup process, still I'm not sure how much we really address the high-level, epistemic problem that those technologies quietly drift towards.
An instruction isn't an invariant.
I recently ran across these few lines, written long ago (1961) by theologian Hans Urs von Balthasar:
"In a world that no longer has enough confidence in itself to affirm the beautiful, the proofs of the truth have lost their cogency. In other words, syllogisms may still dutifully clatter away like rotary presses or computers which infallibly spew out an exact number of answers by the minute. (...) The very conclusions are no longer conclusive."
Written sixty-five years before anything like an LLM existed—today's "rotary presses"?—it's a strange thing to read now. Not to say that Balthasar actually predicted this, but precisely since he wasn't describing a machine that gets things wrong, but one that keeps getting things right, mechanically. The logic holds throughout. But what can close is the world it draws from, quietly folding in on itself a little more with every pass.
None of this makes CLM less of a breakthrough. But a breakthrough means opening something—a technology, a layer of an organization that used to resist it—and nothing has broken through so many jobs and organizations at once. Yet the oldest shape in logic can quietly close back on itself: a syllogism, consistent, undeniable, sealed.
I don't think we've ever broken through so much as in today's AI age. Nor have we been this exposed to such epistemic closure that doesn't need a single lie to take hold.
