Ephemeris · Issue 090 Sunday · 19 Jul 2026 · Zürich
Ephemeris · Issue 090 01 / 04 · Thinking Machines
Model labs · Open weights

The lab's first model ships to be forked.

Mira Murati's Thinking Machines released Inkling, an open-weight mixture-of-experts — 975B parameters total, 41B active per token — pretrained from scratch on 45 trillion tokens of text, image, audio and video. It reads all four natively, flags its own uncertainty instead of guessing, and lets you dial thinking effort up or down. It's a base to fine-tune on Tinker rather than a finished frontier model.

975BTotal params
41BActive / token
45TTraining tokens
4×Modalities in
Ephemeris · Issue 090 02 / 04 · GitHub Engineering
Agentic coding · Practice

The code got cheap. Owning it didn't.

GitHub's Dalia Abuadas splits two costs that used to move together: producing code, now trivial for an agent, and owning it, still exactly as expensive. Her heuristic for a scoped request: stop debating in the abstract, have the agent generate a constrained patch, and read the diff as evidence. Does it touch the files you expected, does it resist testing, can a person confidently own it?

The most expensive part of a small feature request used to be writing the code. Now it's usually the meeting about whether to write it. Dalia Abuadas · GitHub
Cost to produce
Collapsed
Cost to own
Unchanged
Ephemeris · Issue 090 03 / 04 · via @TochkiNadAI
Local inference · Playbook

Tune every layer of your local rig.

A living deep-dive from a year of single-workstation experiments, refreshed in late June. It works through every layer of local inference: RAM speed and CPU governor/EPP, llama.cpp flags, the ik_llama.cpp fork's MoE kernels, and keeping trusted agent endpoints behind Tailscale or WireGuard rather than an open port.

$ llama-bench -m qwen3-coder-next.gguf
  # stock, default governor
  tok/sec ......... 18.4
$ export EPP=performance  # + perf governor
$ ik_llama.cpp --moe-kernels on --n-gpu-layers 99
  tok/sec ......... 40.6   # +121% on the same box
Ephemeris · Issue 090 04 / 04 · Cloudflare
Security · WordPress

Two WordPress holes. Patch now.

WordPress disclosed two high-severity flaws: a critical unauthenticated RCE reachable through the REST API's batch endpoint when no persistent object cache is set, and a SQL injection from crafted input. Cloudflare shipped blocking WAF rules across every plan tier, and core is force-updating affected sites — but confirm your build is patched instead of assuming it happened.

CVE-2026-63030 Unauthenticated RCE · REST API batch endpoint Critical
CVE-2026-60137 SQL injection · crafted query input High

PATCH → 7.0.2 · 6.9.5 · 6.8.6 (or 7.1 Beta 2)

End of issue 090 Back to top ↑

That's Sunday.

Four picks for 19 July 2026: an open-weight base to fine-tune, a working rule for when to accept a change, a local-inference runbook, and a WordPress advisory worth acting on before coffee.

Fed today's picksThinking Machines · GitHub Engineering · carteakey (via @TochkiNadAI) · Cloudflare
Also readThe Batch · Import AI · Don't Worry About the Vase · OpenAI · Anthropic · Sentry · Vercel · PostHog · Next.js · Fly.io · JeffBullas · @seeallochnaya · @denissexy · @rvnikita_blog · @ProductsAndStartups
RubricTools to adopt this week · creative & dev tooling · agentic coding · privacy & security when signal is high · research with a practical kernel
IssueNo. 090 · Sunday 19 July 2026 · Zürich · shipped 08:00