Daily Picks
A daily selection from the Strange Future Lab archive, rewritten as compact briefs with source trails.
Microsoft Research arXiv paper (Apr 2026, HN frontpage ~400-600 pts in last 24h); frontier models (Claude 4.6, Gemini 3.1, GPT-5.4) corrupt ~25% of content in simulated professional editing across 52 domains (code, crystallography, music notation); errors are sparse/severe, compound with length/distractors; agentic tools don't fix it.
LayerX Security disclosure/blog (May 7 2026, HN Anthropic mentions ~last 24-48h); affects Claude Chrome extension (v1.0.69/1.0.70 partial patch); PoC with zero-perm extension steals Google Drive "Top Secret" files, sends emails, bypasses approvals via DOM manipulation/approval looping, trusts origin not context.
The Atlantic feature (May 3 2026, renewed discussion/HN surveillance signals last 24-48h); concrete examples—MorphCast Zoom emotion scoring ("amused/determined/impatient"), Cogito tone analysis for MetLife call centers, Burger King "Patty" headset friendliness judge, biosensor chairs (Framery), Slack sentiment (Aware/Microsoft), interview valence scanners (Imentiv); EU ban (2023, prompted relocations to Bay Area) but market to triple to $9B by 2030.
Specific prompt ("turn entire photo into single pastel/crayon child drawing on white paper, simplify details like 10yo drew it, cute/sevimli atmosphere, don't use original colors") exploding on X/Twitter May 9 2026 (multiple posts 2k+ likes, viral in Turkish/English feeds with before-after grids of pets, people, memes).
Fortune-reported experiment (published ~May 10 2026, fresh news cycle); AI-generated identical résumés for "Emily" vs. "James"—hers more often labeled "weak," reviewers 22% more likely to question trust for female version; male version 97% approval in some cohorts (Gen Z men notably harsher).
SCMP reporting (recent, cited in last 48h searches); high-volume sellers on Xianyu advertise low-latency shadow APIs/proxies for restricted Western models; concrete builder workaround economy.