Grading the earlier braced priors: Google’s AI Overviews spent the fortnight telling America to put glue on pizza (sourced from an eleven-year-old Reddit shitpost), eat “one small rock per day” (sourced from The Onion), and assorted other confidently-cited absurdities, screenshots at meme velocity, a partial feature rollback within days, and the year’s cleanest public demonstration of an earlier pre-registration’s full mechanism: retrieval-augmented generation grounds confabulation in sources without ranking source credibility (the retrieval was accurate! The Reddit post really says glue! The summarization faithfully summarized a joke, because irony-detection at web scale is an unsolved eval and the corpus is load-bearing satire all the way down). The file’s structural note, beyond the comedy: this is the data-provenance war’s consumer surface, Google licensed Reddit’s corpus for $60M this year (the API-pricing wars’ actual endgame, exactly as filed), and the glue incident is what happens when training-data economics meet retrieval trust-scoring nobody built (the eval discipline now has its canonical missing-test-case: “is the source kidding”). Search’s fortress renovation continues regardless, the convenience gradient remains undefeated, but the incident bought every skeptic’s talking point a year of shelf life, and the file notes the deeper asymmetry: Google’s error rate is likely tiny in percentage terms and infinite in screenshot terms (the consensus-hallucination doctrine: brand damage propagates at meme speed, denominators don’t, the Peloton lesson, now for epistemics).
The ScarJo coda: Scarlett Johansson published a statement that OpenAI had approached her to voice the assistant, she’d declined, and the shipped “Sky” voice’s resemblance was, her word, “eerily similar”; OpenAI paused the voice while insisting on independent provenance (a different actress, documented casting). The file’s read: whatever the casting paper trail shows, tweeting “her” at launch converted resemblance into intent in the court that matters (the material-words doctrine: the CEO’s one-word tweet is the exhibit), and the likeness-wars thread gains its most famous plaintiff-shaped datapoint without a suit even filed. The affect layer has a rights layer; the industry keeps discovering its layers via incident.
Life patch note: the sabbatical is over. As of mid-May I’m at a major observability company, platform and DevOps for the products half this archive’s incidents were debugged with, hybrid between Bengaluru and Hyderabad. Six months of coursework and open source told me what I wanted next: scale again, and telemetry specifically, there’s a poetry the streak appreciates in fourteen years of writing about observability ending in shipping it. The remit: an internal developer platform across product lines, OpenTelemetry adoption inside and out, the CI/CD spine, compliance automation at FedRAMP grade. The M.Tech continues at night; the badge photo is terrible; the graphs, for once, are literally my job.
The sports ledger: the Champions League final is set, Real Madrid-Dortmund (Madrid’s remorseless European machine versus Dortmund’s underdog defensive architecture), and the WSL’s record-attendance spring continues breaking broadcast telemetry — the parity thread compounding into the sport’s best growth story.
TIL: adversarial-content retrieval scoring, the emerging discipline of ranking sources by satire-probability, community-context, and claim-verifiability before summarization (Reddit’s corpus arrives with vote-context that the glue incident’s pipeline apparently discarded, the signal existed and was dropped in ingestion). The fix, as always, was in the metadata nobody shipped (fund the plumbing, again, still).