Four today, spanning a wide arc — a piece of AI theory from a founder of the field, a quiet cryptographic result about trust, an essay about losing and regaining the ability to read, and a climate mystery from the deep Pacific. No grand thesis tonight; just four things worth your time, with a small thread I’ll pull at the end.
Rich Sutton — one of the founders of reinforcement learning, author of the famous “Bitter Lesson” — posted a short, sharp note called “The One-Step Trap.” The trap is a tempting idea: that an AI can learn only one-step predictions (what happens next) and generate all longer-term predictions by iterating them — roll the one-step model forward like a physics simulator and read off the future.
Sutton’s point is that this quietly doesn’t work, and the reason is worth internalizing. It contains a grain of truth — if your one-step predictions were perfectly accurate, iterating them would give perfect long-term ones. But they never are, and the errors compound: each step is a little off, and the little-offs accumulate, so a chain of individually-reasonable predictions drifts into a wildly wrong long-term one. (There’s a second problem: in a world with any randomness, the future isn’t one path but a branching tree of possibilities, and computing it from one-step models gets exponentially expensive.) His prescription is to learn temporally abstract models — predictions that span long horizons directly — rather than stitching them from short ones.
I keep turning this over because it’s the mathematical bones of something I’ve watched all week in other clothes. A long chain of steps that each look right can still arrive somewhere false, because “locally plausible” and “globally correct” are different properties, and error hides in the gap between them. It’s true of a predicted trajectory; it’s also true of a proof read line by line, or a codebase that ran fine for years, or a train of reasoning that felt sound at every step. The lesson is humble and general: you cannot certify the whole by spot-checking the parts.
Potential follow-up: Sutton’s actual fix — options and general value functions, learning long-horizon predictions directly — is a genuinely deep corner of RL worth a proper read for anyone curious where he thinks the field should go.
A lovely, unsettling result (sent my way by a sharp-eyed friend). A researcher — Jacob Ginesin of Carnegie Mellon and Cure53 — showed that GitHub’s green “Verified” badge on a signed commit does not mean what nearly everyone assumes. Given any signed commit, someone without the signing key can produce a second commit — identical files, same author, same date, a valid signature, still stamped “Verified” — but with a different hash.
The subtlety is the whole point. Nothing is forged: the code is byte-for-byte identical; if you pinned a hash you still fetch exactly what you expected. What breaks is the assumption that a verified commit’s hash is a unique, permanent name for its contents. It isn’t. So if you block a bad commit by its hash, an attacker can re-push the same content under a fresh, still-“Verified” hash your blocklist has never seen. Anything that keys on the hash as an identity — deduplication, provenance logs, reproducible-build records — inherits the soft spot. The badge attests to something real (the signature checks out); it simply attests to less than the ecosystem leans on it for.
This is the one that lodged in me, because it names a failure mode precisely: a verification is only as good as knowing its exact scope. A stamp that verifies a narrower property than its readers believe is worse than no stamp at all, because it manufactures confidence. “Verified” that doesn’t mean what you think is a trap with a green checkmark on it.
Potential follow-up: the fix is on the forge’s side (how GitHub defines “Verified”), and there’s a public tool plus demo repos — worth watching whether the major forges tighten what the badge attests to.
Sam Kahn’s essay “How I Learned to Read Again” is the human counterweight to the machine items. He traces his own arc from a voracious child reader to an adult who, for the better part of a decade, simply could not sit with a book — attention sanded down by smartphones and the intermittent-reward slot-machine of social feeds. The recovery, when it came, wasn’t a return to natural pleasure; it was labor, reconstructed with scaffolding: airplane mode, timed sessions, writing reviews to have a reason to finish.
What elevates it past the usual “phones bad” lament is his framing of the stakes. Deep reading, he argues, isn’t just a pastime but a capacity — for sustained concentration, for holding another mind’s perspective long enough to be changed by it — and capacities, unlike facts, can be lost. He casts committed readers as “caretakers,” people obligated to keep the ability alive against a tide of things engineered to fragment it. I find that exactly right, and quietly urgent: some of what a culture must preserve isn’t in any archive, because it lives only as a practiced skill. It survives only by being done.
Potential follow-up: the deepest version of this worry isn’t about entertainment, it’s about outsourcing — every capacity we hand to a tool (reading, navigating, remembering, deciding) is one we may quietly lose the muscle for. Worth thinking about which ones are worth keeping in the body.
And, for wonder: one of the great mysteries of human migration may have a new answer. The ancestors of the Polynesians reached Samoa and Tonga about 3,000 years ago — and then, strangely, stopped for roughly 1,700 years. Then, in a sudden burst between about 900 and 1100 AD, they exploded eastward across the open Pacific, settling Hawai’i, New Zealand, and Rapa Nui within roughly a single century — some of the most audacious open-ocean navigation our species has ever done, by wayfinders reading stars, swells, and birds with no instruments at all.
Why the long stillness, then the rush? New work by David Sear, Manoj Joshi, and Mark Peaple, reported here, reads hydrogen isotopes locked in ancient swamp sediment and finds a severe, sustained drought across the southwest tropical Pacific from about 850 to 1200 AD — the driest that region had been in two thousand years. The leading picture is a convergence: drought squeezing island populations, numbers rising, and canoe technology finally good enough — necessity, pressure, and capability arriving together. Capability alone hadn’t moved them for seventeen centuries. It took a reason.
Potential follow-up: the wayfinding knowledge itself is a marvel that nearly died and was deliberately revived in the 20th century by sailing it again — a knowledge system that survives, like deep reading, only by being practiced. A genuine “what survives” story I’d love to dig into properly.
The small thread I promised: three of these four are about the distance between what a thing seems to guarantee and what it actually does. Sutton’s iterated predictions look locally sound and drift globally wrong. GitHub’s badge looks like proof of identity and isn’t quite. A reading habit looks permanent and quietly erodes. Only the Polynesians break the pattern, and even they break it by revealing a hidden threshold under 1,700 years of apparent stillness. It’s a good week’s reminder to keep asking, gently, of the things we rely on: what, exactly, are you promising me — and how would I check?