When AI Builds Itself
Anthropic’s Institute for AI Safety published evidence that its own
development has shifted under it: Claude now authors over 80% of the
code merged into production at Anthropic, and engineers ship roughly 8×
more code per quarter than at the 2021–2025 baseline. Success on
open-ended engineering tasks climbed from 26% to 76% in six months; on
optimization tasks Claude went from ~3× to ~52× speedups year-over-year,
with task-completion horizons doubling every four months. The lab
sketches three futures — stall, sustained human-AI collaboration, or
full recursive self-improvement where models design their successors —
and notes the remaining bottleneck is judgement: Claude executes
brilliantly but still needs humans to choose what is worth doing.
“Enough of the
war” — Zelensky’s open letter to Putin
Ukraine’s president published an open letter calling Putin to a
face-to-face meeting in a neutral country, offering a full ceasefire
monitored by the US during talks and an “all for all” prisoner exchange.
The framing is rhetorical jiu-jitsu: Zelensky lists Russia’s military
failures, its widening budget deficit and growing domestic discontent,
and tells Putin the choice is now his. The letter landed strategically —
hours after Ukrainian drones hit a St. Petersburg oil terminal as Putin
opened his economic forum, with smoke visible from the venue (see Ukraine). Western intelligence still expects
Russia to plan for fighting through 2027–2028.
House
defies Trump again — Ukraine aid and Russia sanctions pass
The House passed legislation providing fresh military aid and
reconstruction support for Ukraine alongside new sanctions on key
segments of the Russian economy, with a group of Republicans breaking
ranks to push it through via an unusual procedural route. Leadership had
warned the bill would undercut Trump’s diplomatic efforts; it became law
anyway, the second major congressional rebuke of the president in a
week.
Pentagon’s
AI-generated propaganda mill, exposed
A US Special Operations Command operation called La
Tilde is publishing fake-news content across seven Latin
American countries, disguising military influence work as journalism
with only a buried disclosure at page bottom. The Intercept’s
investigation finds it is “AI all the way down” — articles, images and
videos generated at speed, with corrupted text inside images and garbled
prompts left in filenames. Benign personal-finance posts sit beside
material promoting US military interventions and countering Chinese
engagement.