Morning Brief: Friday, July 17

Sixty-five feeds. Two weeks. 4,187 items reduced to what follows. (what we track, how we crawl, subscribe)

Friday morning. Kimi K3 (2.8T-A50B) lands as the week's second open-weight frontier release — Latent Space frames it as the largest open model ever released; Opus 4.8-class at Sonnet 5 pricing. Vercel shipped AI Gateway availability yesterday, on the same day Databricks hosted Inkling. Two Apache-licensed frontier models in a single week (Kimi K3 Fri, Thinking Machines Inkling Thu) is the week's structural story. Simon Willison stays the sanity check with Kimi K3, and what we can still learn from the pelican benchmark. Yesterday's arXiv agent-memory + governance cluster (Oracle Agent Memory, Experience Memory Graph, Compaction as Epistemic Failure, CAVA, AgentCompass) is still the frame for the read. FreeBSD Foundation adds Dave Cottlehuber to the board (Clojure-adjacent). CIDER ships macro stepping. Pydantic publishes The human-in-the-loop is tired. Truth Social to sell Wall Street firms fastest access to Trump posts. MIT Tech Review flags weather-data sabotage. Astronomers claim atmosphere on a nearby Earth-like planet.

Top (5-7 min)

Kimi K3 2.8T-A50B: the largest open model ever released; Opus 4.8-class at Sonnet 5 pricing: Latent Space, 2026-07-17. Second open-weight frontier launch of the week. Kimi K3 at 2.8T params / A50B active — nearly 3× Inkling's 975B/A41B from Thursday. Latent Space's framing (Opus 4.8-class at Sonnet 5 pricing) is aggressive but names the exact place this lands: the price/quality frontier that Anthropic just moved with Sonnet 5. Vercel had it on AI Gateway same day as the launch post (changelog, 07-16).
Kimi K3, and what we can still learn from the pelican benchmark: Simon Willison, 2026-07-16. The independent read on Kimi K3. Willison keeps the pelican SVG benchmark alive as a low-cost sanity check that resists benchmark saturation — it is now the load-bearing counterweight to Latent Space's aggressive framing. Read both together.
Kimi K3: Open Frontier Intelligence: Hacker News, 2026-07-16. Primary launch post from Moonshot AI. Read this before the Latent Space and Willison analyses.
The human-in-the-loop is tired: Hacker News, 2026-07-17. Pydantic post that names the current agent-tooling frustration: every capable agent workflow still routes back through a human approval bottleneck. Pair with yesterday's CAVA runtime attestation paper — CAVA is the remove-the-human-by-attesting-the-action answer to what Pydantic is describing.
Spot birds not golf: Simon Willison, 2026-07-17. Willison Friday item. Short but characteristic — the use-vision- models-for-boring-classification pattern he keeps documenting.
Runtime logs now show cache reasons: Vercel, 2026-07-17. Observability improvement worth flagging: cache-hit/miss attribution surfaced in the log stream. Small change, large debugging leverage for anyone who has fought an ISR miss cascade.
FreeBSD Foundation Welcomes New Board Member: Dave Cottlehuber: FreeBSD Foundation, 2026-07-17. Cottlehuber (Erlang/Elixir/Clojure-adjacent, longtime BSD contributor) joining the board is a governance signal for a project that just retired the last of its GPL code (Slashdot, 07-15).

Themes this week

Two open-weight frontier launches in one week (Kimi K3 Fri + Inkling Thu): Latent Space: Kimi K3 2.8T-A50B — largest open model ever released; Opus 4.8-class at Sonnet 5 pricing (today), Vercel: Kimi K3 now on AI Gateway (Thu), Willison: Kimi K3 and the pelican benchmark (Thu), Kimi.com: Kimi K3 — Open Frontier Intelligence (Thu), Latent Space: Thinky's Inkling 975B-A41B (Thu), Willison: Inkling — Our open-weights model (Thu), TC: Thinking Machines amps up its bet against one-size-fits-all AI (Wed).
Agent-memory-as-substrate (Thursday arXiv cluster, still framing): arXiv: Oracle Agent Memory as an Enterprise Memory Substrate for Long-Horizon AI Agents (Thu), arXiv: Experience Memory Graph — One-Shot Error Correction for Agents (Thu), arXiv: Compaction as Epistemic Failure (Thu), arXiv: Memory as a Controlled Process — Learned Adaptive Memory Management for LLM Agents (Thu).
Agent-governance and audit (Thursday arXiv cluster, still framing): arXiv: CAVA — Canonical Action Verification and Attestation for Runtime Governance (Thu), arXiv: AgentCompass — Unified Evaluation Infrastructure for Agent Capabilities (Thu), arXiv: Interventional Grounding Audits (Thu), arXiv: How Agents Ask for Permission — User Permissions for AI Agents from Interfaces to Enforcement (Thu), Pydantic: The human-in-the-loop is tired (today).
Clojure/CIDER (small but coherent week): metaredux: Stepping Through Macros in CIDER (today), metaredux: Simplifying Session Management in CIDER (Thu), metaredux: clj-refactor.el 4.0 (Thu), metaredux: CIDER 2.0: Sky is the Limit (Wed).
FreeBSD governance and cleanup: FBSD Foundation: Welcomes Dave Cottlehuber (today), FBSD Foundation: Understanding the Board's Role (today), Slashdot: FreeBSD 16 Retires the Last of Its GPL Code (Wed).

Scan (15 min)

Friday signal (short)
- Kimi K3 2.8T-A50B — Opus 4.8-class at Sonnet 5 pricing, Latent Space, 07-17
- The human-in-the-loop is tired, Pydantic (via HN), 07-17
- Spot birds not golf, Simon Willison, 07-17
- Runtime logs now show cache reasons, Vercel, 07-17
- FreeBSD Foundation welcomes Dave Cottlehuber, FBSD Foundation, 07-17
- Stepping Through Macros in CIDER, Planet Clojure, 07-17
- claude-code v2.1.212, claude-code-releases, 07-17
HN Friday tabs
- The human-in-the-loop is tired, HN, 07-17
- Pebble Mega Update — July 2026, HN, 07-17
- EEG shows brain can simultaneously encode two speech streams, HN, 07-17
- GrapheneOS recommended for domestic abuse victims, HN, 07-17
- How Has Roman Concrete Lasted for Millennia? 1,900-Year-Old Latrine Offers Clues, HN, 07-17
- M 3.9 Experimental Explosion – 147 Km ENE of Ponce Inlet, Florida, HN, 07-17
- UIUC AI Teaching Assistant, HN, 07-17
Slashdot Friday
- Astronomers Find an Atmosphere On a Nearby Earth-Like Planet, Slashdot, 07-17
- Truth Social To Sell Wall Street Firms the 'Fastest' Access To Trump's Posts, Slashdot, 07-17
MIT Technology Review Friday
- The risk of weather data sabotage is rising, MIT Tech Review, 07-17
- There's a lot of hype around perimenopause. Don't buy it., MIT Tech Review, 07-17
Corp-eng Friday
- Runtime logs now show cache reasons, Vercel, 07-17
- v2.1.212, claude-code-releases, 07-17
- Microsoft just open-sourced the app that put Comic Sans on the map, The New Stack, 07-17
Hackaday Friday
- Wireless LCD Streaming for the ANENG AN870 Multimeter, Hackaday, 07-17
- A Sloshing-Mercury-Powered Neon Light, Hackaday, 07-17
- White Rails are the Infrastructure Hack We Didn't Know We Needed, Hackaday, 07-17
Aviation / cert (Leeham)
- Bjorn's Corner: Aircraft structures Part 10 — Composite certification, Leeham News, 07-17
Mathematics (Terence Tao)
- Two more apps: visualizing the zeta process and the motions of the heavens, Terence Tao, 07-17
Pinboard (jwalsh) Friday
- Atlassian CLI, Pinboard, 07-17
- Teamwork Graph CLI, Pinboard, 07-17
- FreeBSD Laptop Support Continues Improving With WiFi, GPU & Audio Driver Work, Pinboard, 07-17
- EXPRESS 2026 — SPLASH/ISSTA 2026, Pinboard, 07-17
- Events & Conferences — IEEE Reliability Society, Pinboard, 07-17

Tail

Two Apache-2.0 frontier open-weight launches in a single week (Inkling Thu, Kimi K3 Fri) is the week's shape — the American Apache open model frame Latent Space used for Inkling on Thursday now competes with a bigger Chinese-provenance model on Friday. Vercel AI Gateway had both same-day; Databricks had Inkling same-day; nobody has published a Kimi K3 hosted-inference partnership yet as of the Friday morning cut.
arxiv-cs-ai: yesterday's giant drop (Oracle Agent Memory, Experience Memory Graph, Compaction as Epistemic Failure, CAVA, AgentCompass, Interventional Grounding Audits, Harness Handbook, Self-Improvements Survey) is still the framing arXiv content — no new drop yet on Friday morning (expected midday).
bitsavers (6 feeds): connected, 0 items on Friday (sparse archive, expected).

Feed silences (diagnostic)

Anthropic first-party, Cloudflare, GitHub Blog, Interconnects, Kenneth Payne, Martin Fowler, Microsoft Research, Charity Majors, Alignment Forum, Quanta Magazine, Grafana Labs, 404 Media, Pluralistic, Andrej Bauer, Databricks, Hugging Face Blog, OpenAI, Nature Machine Intelligence, Hillel Wayne, Schneier on Security, Apple ML Research, Google Research, EFF Deeplinks, wal.sh site, TechCrunch, Latent Space (second post): no primary posts yet on Friday morning (expected — most publish Friday afternoon US/EU or defer to Monday).
Netflix Tech Blog, James Bornholt: persistent DNS/TLS errors continue (unchanged from Sun/Mon/Tue/Wed/Thu).

Build provenance

build: 2026-07-17 | crawler-sha: 5fe7ab8 (Walsh-Research/1.2, compliance v1.3) | feeds: 65 active (78 configured, incl. 12 corp-eng, 6 bitsavers, 5 generated) | items-considered: 4187 (14d, incl. 2424 arXiv) | warehouse: 25977 items | published: 7 | note: Friday morning. Kimi K3 (2.8T-A50B) is the week's second open-weight frontier launch — Latent Space frames it as "the largest open model ever released, Opus 4.8-class at Sonnet 5 pricing." Two Apache-licensed frontier models shipped in one week (Kimi K3 Fri, Thinking Machines Inkling Thu) with same-day Vercel AI Gateway availability for both. Willison keeps the pelican benchmark alive as the sanity check. Yesterday's arXiv agent-memory + governance cluster (Oracle Agent Memory, Experience Memory Graph, Compaction as Epistemic Failure, CAVA, AgentCompass) still frames the read. Pydantic ships "The human-in-the-loop is tired" — the practitioner counterpart to CAVA's runtime attestation proposal. FreeBSD Foundation adds Dave Cottlehuber to the board (Clojure-adjacent). CIDER lands macro stepping. Truth Social to sell Wall Street "fastest" access to Trump posts. MIT Tech Review flags weather-data sabotage. Astronomers claim atmosphere on nearby Earth-like planet.