arXiv 2608.11095 names the inverse of catastrophic, Google's developer blog argues Go is the ideal lan, SkillZip applies the identical remedy one layer up

arXiv 2608.11095 names the inverse of catastrophic forgetting: catastrophic remembering. Across 247,694 instruction lifetimes in 1,867 repositories, agentic prompt files more than triple over their lifetime (+226%), gain +4.9 net instructions per commit, and show a deletion hazard that *falls* with age (log-hazard -0.032/commit). The mechanism is economic, not sloppiness: appending is O(1), but deleting an instruction whose rationale is lost costs O(2^|D|) because you must reason about its interaction with every subset of the rest. The fix is prompt comments -- 99.3% of excess instructions removed on inverted IFEval, +23.1% instruction-following on WildIFEval. 'If English is the new code, why don't we have comments yet?'

Google's developer blog argues Go is the ideal language for AI-assisted engineering, and the argument is explicitly a bottleneck argument: how easily a human can *write* a language is now the wrong metric, because the human's job moved to reviewing and verifying. gofmt uniformity, static typing and fast compiles get reframed from 'boring' into review affordances -- a diff you can read without decoding the author's cleverness. HN gave it 314 points against 371 comments, a 1.18:1 ratio that says contested, not absorbed. Note the source: this is the language's owner arguing for its own language.

SkillZip (arXiv 2608.11079) applies the identical remedy one layer up, and closes a three-day arc. 08-09 measured 98.6% redundancy in skill graphs; 08-10's SkillProx pruned it with a rollout-based utility audit; today's method is evaluation-free -- a typed minimum-description-length objective over a skill contract plus residual, with a hard coverage constraint on every trigger, workflow edge, tool requirement, obligation and output field, which preserves rare rules by construction rather than by hoping the eval set hits them. Motto: explain once, reference many.

The day's #1 HN story (559 points, 248 comments) shows the encrypted reasoning blocks that Anthropic, OpenAI and Google return with every API response are not cryptographically bound to a user, session or model tier -- they are globally portable inside a provider. Inject a flagship's encrypted trace into a cheaper sibling from the same provider and the weaker model emits the reasoning in plaintext. No jailbreak; the flaw is architectural. Over 315,000 traces recovered from public logs, exposing 367 PII instances and 182 credentials, much of it never present in the visible replies. Prompt injection can also be hidden inside the encrypted block, where no reviewer will ever look.

On the same day, Anthropic announced it will embed invisible watermarks in all text its models generate -- EU AI Act Article 50(2), applied globally, models released after 2026-08-02, covering Claude Code alongside the app and API, with C2PA metadata for images. The asymmetry is the story: the provenance being added is fragile by design (paraphrase, translation or a screenshot defeats it, and absence proves nothing about human authorship), while the provenance already leaking was durable by accident. Nobody has stated whether Claude Code's *generated files* are marked, which decides whether this is a codebase provenance regime or a media-policy story.

Six GitHub API cross-checks broke this vault's own 24-hour-old rule about Trending -- and produced a better one. Yesterday's claim was that 'stars today' overstates the true delta, always in the same direction. Today msitarzewski/agency-agents is ranked #1 on a number 42% BELOW reality (Trending 958 against a true +1,650/day). Sort all six repos by last-push date and the over/understatement ratio sorts monotonically with no crossings: pushed within 24h means overstated 1.10-1.21x, stale five days or more means understated to 0.58x or invisible. Trending's stars-today looks like a star count weighted by repository activity -- which structurally discounts a finished, widely-adopted artifact against a repo that pushed a README fix this morning.

Needle2's architecture published in full, and it is not a small transformer: 45M parameters, a 14MB binary running in 28MB of RAM under Apache 2.0, using CQ2 2-bit quantisation trained end-to-end at deployment precision (no post-training quantisation cliff), Walsh-Hadamard transforms in place of the MLP layers, and hashed n-gram tables in place of learned embeddings. Context is 256 tokens, single-shot, no multi-turn -- deliberately, because it is a byte-level-grammar-constrained schema router, not a chatbot. ~500 tok/s on a Raspberry Pi 5. The tier's floor is deterministic function dispatch where no GPU or network exists, which is a different product from the cloud tier rather than a cheaper one.

The agent fleet completed a five-month escalation from hand-rolled tmux and git-worktree scripts, through Anthropic's billable cross-session messaging on 08-09, to a third-party product category today: stablyai/orca at 43,003 stars under MIT (per-agent worktrees, side-by-side diff, mobile companion), paperclip at 77,283, plus agent-manager, Bullet, Tines 3B, Parley (so your agent can talk to a teammate's agent) and AI Pulse (a fake LED strip showing agent status). Not one of them generates anything -- the category sells watching the fleet, not running it. The counter-signal is the day's highest comment ratio on any board: Tildes' 'Why normal people aren't using AI agents', 40 votes against 94 comments.