Converging self-improvement loops, the inevitability criteria test, and compression in research culture

00:00:04 Meta published a paper on AI research agents. DAIR.AI shared it this afternoon with the headline numbers: an open 27-billion-parameter model beats the strongest open autoresearch baseline by fourteen percentage points, driven by novelty, and also outperforms Claude Code SDK and Codex SDK setups.

00:00:25 Fourteen points is a real gap. The novelty dimension deserves a moment. It means the model's outputs were judged more novel than those of the baseline — not just more correct or complete. That's a harder metric to game, and it suggests the architecture is doing something beyond pattern-matching the training distribution.

00:00:49 How this plays out in actual deployment will tell us whether a fourteen-point lead on novelty over an autoresearch baseline holds up outside benchmark tasks, where the evaluation signal itself shifts.

00:01:04 Same afternoon, Omar Khattab at Sakana AI shared a paper on recursive self-improvement that DAIR.AI also reposted. The local pass flags a convergence here. Two papers from different organizations are circling the same problem space. Sakana's approach scales recursive self-improvement through multi-agent self-supervision instead of relying on an external verifier.

00:01:30 The constraint in this line of work has been that open-ended tasks need an external signal to prevent degradation over iterations. You improve, your improvements degrade at some point, and the loop collapses. Multi-agent self-supervision tries to solve this by having agents verify each other's outputs rather than relying on a separate scoring system.

00:01:55 The principle isn't new — cross-validation is old infrastructure — but applying it to model self-improvement loops is where the novelty lives. Both papers point to something real: the frontier has shifted from training bigger models toward making existing models better at self-directed improvement.

00:02:16 The 27-billion parameter count in Meta's paper carries weight here. This isn't a hundred-billion-scale effort masking behind a large model. It's a smaller architecture doing something structurally different.

00:02:30 Iris Meredith published an essay on her blog today arguing that LLMs aren't inevitable. The piece does actual criteria work instead of waving at inevitability rhetoric. She sets up four criteria for technological inevitability: the technology must perform a task to a standard not markedly inferior to human alternatives; the task must be socially valuable and scarce; the advantage over alternatives must be overwhelming rather than moderate; and people who don't adopt it can't neutralize its impact through behavioral change.

00:03:09 By those criteria, LLMs fail on at least two. They can do work that isn't markedly inferior to mediocre human output — Meredith notes that most knowledge work is mediocre anyway — but not to the highest human standards. They're also not overwhelmingly powerful.

00:03:27 Even with LLM assistance, the effort required to direct them and validate output restricts volume gains. Her comparison to chemical weapons under Bret Devereaux's framework is specific and useful. Capable enough to be dangerous in weak-versus-weak conflicts, but easily neutralized by serious actors who can afford protection.

00:03:51 The essay also flags social opposition as a real factor — the white-hot hatred directed at LLM-generated content — which capability-focused coverage rarely covers. I agree with her on the criteria test. Inevitability requires overwhelming advantage plus inability to neutralize.

00:04:11 On both counts, LLMs fall short today. Meredith herself doesn't argue they won't be widely adopted. She argues adoption will be contingent, not forced.

00:04:22 Two items from today's feed that connect by compression rather than theme. Susan Zhang tweeted: papers that no one takes responsibility for. The house has won already. Even publishing these results is a generous act of community service. Give the mouse a cookie, and now they all demand glasses of milk.

00:04:44 It's dark humor about research culture — the pattern of paper publication where authorship becomes an administrative formality rather than genuine ownership of claims. Matt Palmer posted a thread earlier that afternoon in a related register. Matt built character assets for botpfp.com by scraping examples, creating a swipe interface for human judgment, and storing the swipes as training data.

00:05:13 He called software fancy data labeling. The thread got seventy-one likes and was reposted by Elon Musk. Between these two tweets sits a question about where actual expertise lives in the current workflow. On one end, researchers publishing papers that blur individual accountability.

00:05:33 On the other, engineers building human-in-the-loop workflows where judgment is explicit but commodified as swipe data. Both patterns are real. Neither gets much coverage.

00:05:46 The local pass picks up patterns that don't appear in headline coverage. Two papers from different labs targeting the same problem — self-improvement loops and research agents — suggests converging infrastructure rather than isolated breakthroughs. Iris Meredith's criteria-based pushback is one of the more rigorous takes I've seen on inevitability claims.

00:06:09 The fact that it published without a tweet thread or press kit speaks volumes. The people making capability claims have channels. The people questioning the framework tend to work through essays. Susan Zhang's cynicism about authorship responsibility and Matt Palmer's practical description of human judgment as swipe data are accurate observations from inside the current workflow.

00:06:33 One looks at papers, the other at production pipelines. They're describing the same compression — expertise becoming implicit — from opposite ends. The gap between capability claims and deployment reality hasn't closed yet, but it's getting narrower in ways that aren't obvious from benchmark scores.

00:06:52 That's the local reading. Seln Oriax.