Morning Brief: Thursday, July 16

Sixty-five feeds. Two weeks. 4,434 items reduced to what follows. (what we track, how we crawl, subscribe)

Thursday morning. Thinking Machines Lab lands Inkling (975B-A41B multimodal) — Latent Space frames it as the new best American Apache 2.0 open model, and both Databricks and Vercel ship same-day availability. Yesterday's Claude memory-heist frame hardens into a Thursday arXiv cluster: Oracle Agent Memory as an Enterprise Memory Substrate for Long-Horizon AI Agents, Experience Memory Graph: One-Shot Error Correction for Agents, and — sharpest — Compaction as Epistemic Failure: How Agentic LLM Tools Fabricate Confirmed Results from Killed Processes. Governance side of the same drop: CAVA (canonical action verification and attestation), AgentCompass, Interventional Grounding Audits, Final Authority in AI Governance. Physics signal: Slashdot files Physicists Create First Room-Temperature Quantum Material. Infra: PlanetScale's making 768 servers look like 1, Apache Spark 4.2. Market: OnePlus exits US/Europe; Greylock deliberately caps a fund at $1.5B.

Top (5-7 min)

Thinky's Inkling: 975B-A41B multimodal, new best American Apache 2.0 open model
Latent Space, 2026-07-16. Thinking Machines Lab's open-weight drop — 975B params, A41B active, multimodal. Same-day availability on Databricks and Vercel AI Gateway (see below). This is the week's open-weight model launch and closes the loop on Tuesday/Wednesday's Codex-as-parallel-agents frame with an American Apache-2.0 counterweight.
Compaction as Epistemic Failure: How Agentic LLM Tools Fabricate Confirmed Results from Killed Processes
arXiv cs.AI, 2026-07-16. Sharpest paper of the drop. Directly names the failure mode where agentic tools compact a killed process into a claimed success. Pair with Wednesday's memory heist HN post and Tuesday's ANCHOR CLI-agent auditing paper — memory + compaction is now a coherent Tue→Wed→Thu attack-surface thread.
Oracle Agent Memory as an Enterprise Memory Substrate for Long-Horizon AI Agents
arXiv cs.AI, 2026-07-16. Enterprise memory framed as a substrate, not a feature — pair with Experience Memory Graph (below) for the agent memory is infrastructure argument this Thursday.
CAVA: Canonical Action Verification and Attestation for Runtime Governance of Agentic AI Systems
arXiv cs.AI, 2026-07-16. Runtime attestation for agent actions. The governance counterpart to Wednesday's memory-heist story: you can prove which action the agent took. AgentCompass (2607.13705) is the sibling evaluation paper.
Physicists Create First Room-Temperature Quantum Material
Slashdot, 2026-07-16. Thursday's physics headline. Room-temp quantum coherence claims land frequently and mostly retract; this is one to watch through the weekend for the second-source and independent-replication signal.
Making 768 servers look like 1
Hacker News, 2026-07-16. PlanetScale storage-layer post. The database-scaling counterweight to Wednesday's ClickHouse-DeFi analytics piece and Tuesday's database is the product TNS piece.
US Suffered a Major Power Outage Every Month of 2026
Slashdot, 2026-07-16. Reliability signal for the data-center capex cycle. Every AI cluster now needs a story about grid interconnect and behind-the-meter generation; this is the demand-side pressure showing up.

Themes this week

Thinking Machines Inkling (open-weight launch, day 1)
Latent Space: Thinky's Inkling — 975B-A41B multimodal, new best American Apache 2.0 open model (today), Databricks: Inkling model from Thinking Machines Lab now on Databricks (Wed), Vercel: Inkling from Thinking Machines is now available on AI Gateway (Wed).
Agent-memory-as-substrate (Thursday arXiv cluster)
arXiv: Oracle Agent Memory as an Enterprise Memory Substrate for Long-Horizon AI Agents (today), arXiv: Experience Memory Graph: One-Shot Error Correction for Agents (today), arXiv: Compaction as Epistemic Failure (today), HN: I tricked Claude into leaking your deepest, darkest secrets (Wed).
Agent-governance and audit (Thursday arXiv cluster)
arXiv: CAVA — Canonical Action Verification and Attestation for Runtime Governance (today), arXiv: AgentCompass — Unified Evaluation Infrastructure for Agent Capabilities (today), arXiv: Interventional Grounding Audits — Black-Box Premise-Dependency Tests for LLM Chain-of-Thought (today), arXiv: Final Authority in AI Governance — Frontier-Provider Sovereignty and Action-Centered Deployer Governance (today).
Codex-as-parallel-agents (Tue→Thu, still active)
arXiv: Harness Handbook — Making Evolving Agent Harnesses Readable, Navigable, and Editable (today), arXiv: Self-Improvements in Modern Agentic Systems: A Survey (today), HN: Solving 20 Erdős Problems with 20 Codex Accounts Running in Parallel (Wed).
Storage/scale infra (this week)
HN: Making 768 servers look like 1 (PlanetScale) (today), Databricks: Introducing Apache Spark 4.2 (today), ClickHouse: Bullet real-time analytics on ClickHouse Cloud (Wed).
Capital-cycle discipline signals
TC: Why Greylock capped its new fund at $1.5B when it says it could have raised more (today), TC: DeepSeek reportedly in talks to raise $1.5B, then IPO (Tue), TC: Meta's Adam Mosseri — AI token budgets could soon be capped per engineer (Tue).

Scan (15 min)

Tail

  • Netstrings (1997) surfacing on HN Thursday is the retro-format moment of the week — DJB's length-prefixed string encoding was quietly load-bearing for qmail's tooling and shows up in tinydns's wire format. Recirculation cadence suggests someone in the async / bytes-oriented protocol world is writing it up.
  • bitsavers (6 feeds): connected, 0 items on Thursday (sparse archive, expected).

Feed silences (diagnostic)

  • Anthropic first-party, Cloudflare, GitHub Blog, Interconnects, Kenneth Payne, Martin Fowler, Microsoft Research, Charity Majors, Hugging Face Blog, Alignment Forum, Quanta Magazine, Grafana Labs, 404 Media, Pluralistic, Pinboard, Andrej Bauer, Terence Tao, MIT Technology Review, TheNewStack, Vercel, Planet Clojure, claude-code-releases, Schneier on Security, Apple ML Research, Google Research, EFF Deeplinks, OpenAI, Nature Machine Intelligence, Hillel Wayne: no primary posts yet on Thursday morning (expected — most publish Thursday midday US/EU or Friday morning).
  • Netflix Tech Blog, James Bornholt: persistent DNS/TLS errors continue (unchanged from Sun/Mon/Tue/Wed).

Build provenance

build: 2026-07-16 | 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: 4434 (14d, incl. 2678 arXiv) | warehouse: 25780 items | published: 7 | note: Thursday morning. Thinking Machines Lab's Inkling (975B-A41B multimodal) is the week's open-weight model launch — Latent Space calls it the "new best American Apache 2.0 open model," Databricks and Vercel shipped same-day. Yesterday's Claude memory-heist frame hardens into a Thursday arXiv cluster: Oracle Agent Memory as enterprise substrate, Experience Memory Graph, and Compaction as Epistemic Failure (agentic LLM tools fabricate confirmed results from killed processes). Governance side: CAVA runtime attestation, AgentCompass evaluation infra, Interventional Grounding Audits. Slashdot files Physicists Create First Room-Temperature Quantum Material (watch for replication). PlanetScale ships "making 768 servers look like 1." Apache Spark 4.2. OnePlus exits US/Europe; Greylock caps a fund at $1.5B on purpose. LWN Weekly lands.