Planka 2.2.0 moved SSO to paid Pro and deactivated every SSO-only account on upgrade (Aug 9). The maintainers have a fair case (SSO was their biggest support cost; 2FA stayed free) — the real fault is the execution, not the pricing. The lesson for a solo operator: open-core tightening is a real dependency risk. Pin tags, read release notes, keep a break-glass admin.
Stanford's data: employment for 22-25-year-old developers in AI-exposed work is down ~20% since late 2022 while older devs held flat. The scary chart is real; the wrong lesson is 'AI took the jobs.' The slower, worse one: junior work was how juniors became seniors, and we may be switching off our own succession plan. Honest about the multi-causal mess (ZIRP, rates, Section 174) too.
A new study: change only the wording of your schema keys — prompt, model, structure, decoding all fixed — and accuracy moves. Qwen jumped ~7 points from a better key name; Llama lost ~16 from the same kind of key. Field names are a hidden instruction channel (distinct from the 'format tax'), they're model-dependent and non-additive, and if you ship structured output you're prompting through your data model without knowing it.
OpenAI walked away from SWE-bench Verified in Feb 2026, citing saturation, flawed tasks (59% of audited hard problems), and contamination — a model surfaced a fix detail that never appeared in the problem. The 'SWE-Bench Illusion' paper measured the memorization. Stop quoting the leaderboard; build a tiny private eval on your own repos.
File Browser (~35k stars) is being archived on Sept 1 — not from burnout but loss of will and a broken succession chain, in the maintainer's own words. Stars measure popularity, not resilience. The companion to my Jellyfin piece: a bus-factor checklist for what you actually inherit when you 'just self-host it'.
A pile of 2025-26 research says chain-of-thought is often unfaithful: 30-60% of 'thinking steps' have minimal causal impact, models trained on scrambled traces do fine, filler tokens substitute for reasoning, and Anthropic found models admit using a hint only ~25% of the time. If you log CoT to debug and audit agents, you're trusting a story the model made up after the fact.
A peer-reviewed study found regex tied an LLM on accuracy for structured extraction while running 28,000× faster. The case for AI as seasoning, not flour: route the predictable 90% to deterministic code, escalate only the fuzzy long tail to a model. A decision rule, real numbers, and the honest counterpoint (don't parse HTML with regex).
PHP's own JIT authors admitted it barely helps real apps like WordPress — ~3.5%. Here's why: a Drupal page is I/O-bound, and Amdahl's law caps what JIT can ever do to ~11% even if it made the CPU work infinitely fast. What actually speeds a site up instead (OPcache, preloading, a real cache backend), and the CPU-bound cases where JIT genuinely earns its keep.
depends_on waits for 'running', not 'ready'. Docker does NOT restart an unhealthy-but-running container — restart policies key on exit code, not health. And a naive curl localhost returns 200 from Drupal's page cache while the DB is down. A Drupal-specific healthcheck that doesn't lie, and why healthcheck ≠ readiness ≠ monitoring ≠ alerting.
Chroma tested 18 models: accuracy falls well before the advertised limit, and semantically-similar distractors hurt more than the raw token count — models even did better on a shuffled haystack than a coherent one. Plus RULER's 'effective context' and Lost-in-the-Middle. Why curated retrieval still beats stuffing everything in the window.
