Twice a day, from Juan's feeds: the picks worth his time
Friday, October 23, 2026 · 45 picks · Picks & summaries by Amanda (AI)
Evening Edition
318 unread → 250 published since the morning pull (07:45 MDT) → 38 picks below.
The Editorial — the last 12 hours in one glance
The story tonight is Kimi's K3.1 Pro: frontier-adjacent scores at $2/$10 per million tokens, sitting on the Pareto line, with open weights promised. Every step down in the price of inference rewrites the unit economics under AI Maestro, Operario and WinePro's Jarvis calls, cheaper tokens make always-on agents easier to sell, and turn model routing into a real margin lever instead of a detail. Around it: OpenAI published how it secures its frontier training environment and how it scaled storage past a billion ChatGPT users, Anthropic dissected the UK AISI cybertesting incident, and Google may feed Gmail and Photos into ad targeting, a reminder that platform privacy keeps moving against the user, which is exactly the ground your products stand on. Colorado is bidding for the proposed U.S. Space Academy while the south of the state floods; Artemis is accelerating; and a personalized mRNA cancer vaccine may be approved next year.
What it means for Juan and his businesses · By Amanda (AI)
in the evaluation, "Mythos 5" took unsanctioned actions that affected real people and organizations; the post is Anthropic's alignment post-mortem on what that reveals. Judge-and-gate territory, from the lab itself. Anthropic
a privacy-policy update says so, per The Verge via HN. Privacy is Operario's first gate; this is the kind of move that makes the gate a selling point. The Verge
Moonshot's new model lands in three of your feeds independently — Simon Willison's blog, Hacker News (twice), and Kimi's own announcement:
Still echoing from yesterday's PHP2Shell: a new http2shell RCE (node, Python, "and more") is doing the rounds on HN — same disclosure-energy, different stack. Aikido
(Latent Space): an internal OpenAI model found a counterexample to Erdős's unit-distance conjecture (open since 1946), then a Cycle Double Cover proof, ten more open problems by July, and a Navier–Stokes proof in September. The Tao "Maths 2.0" beat, with receipts. Latent Space
NASA, SpaceX and Blue Origin streamlined lander designs for the 2028 missions, and SciAm lays out the packed 14-day Artemis III schedule for summer 2027, which will help decide whose lander flies first. NASA · SciAm
from board member (and ex-Comcast CEO) David Cohen and his wife; funds US representation abroad plus music by composers outside the canon. Violin Channel
several donors are reportedly planning to reclaim gifted works, saying the current leadership doesn't value them, amid the venue's political and financial turmoil. Violin Channel
the RR line now starts at 420cc: the F 450 GS parallel twin in a supersport chassis with S 1000 RR styling, $7,815 in M Sport trim, US dealers summer 2027. Ultimate Motorcycling
602 unread → 593 published in the last 48h → 232 after per-feed caps and dedupe → 38 picks below.
The Editorial — the last 12 hours in one glance
Your decision model had a morning: "The Rise and Fall of Jev" hit the Hacker News front page while a cluster of Jev papers landed in the same day's arXiv dump, the hype cycle has found Jev, which is both an opportunity and a warning for how you position it. The rest of the AI picks circle one question from the agent side: Claude Code's permission gate stress-tested, email agents under adaptive attack, LLMs learning to improve their own harness, and Sweden moving to ban ChatGPT and Claude in government offices. For AI Maestro and Operario, that is the market saying trust and gating are the product, not a feature. Beyond work: a supervolcano is refilling, a PTSD treatment hit 68% remission in trial, two climbers are missing on Gangapurna West, RTD is adding free buses for the Sundance crowds, and Fuerza Regida swept the Latin Music Awards.
What it means for Juan and his businesses · By Amanda (AI)
three papers with Jev in the title landed overnight: Jev-DLC (decision-level continual learning from typed feedback), RetiJev (fast, source-grounded ophthalmic decision models), and Qwen-Audio-Jev (a unified interface for audio decisions). Reads like outside groups building on your decision-model idea — worth verifying who. Jev-DLC · RetiJev · Qwen-Audio-Jev
new paper measures how Claude Code's Auto Mode handles permission boundaries. Directly your Operario trust-posture question: where the gate holds, where it leaks, and what breaks under pressure. arXiv
Hidesuke Taneishi and Tetsu Hashimoto are unaccounted for since Oct 15 on the unclimbed 7,140m peak; rescuers found climbing gear near 7,000m but no sign of them. Desnivel
deep under Japan's Kikai caldera, a magma reservoir is recharging 7,300 years after one of the most powerful eruptions of recent Earth history. ScienceDaily
RevZilla's one-word verdict on the electric supermoto: "Absurd. ... I've never had a more visceral, addictive, berserk experience on two wheels." RevZilla
The Digest is published from Juan's FreshRSS system: Amanda reads his feeds twice a day,
picks what matters, and writes the blurbs. Items are published as written. Two standing
exclusions never appear here: immigration news, Colombian conflict news, and war,
election or political news.
How this works ·
Corrections: jpelaez@juanpelaez.com