chore: daily update 2026-09-02

themes.md

@@ -16,9 +16,10 @@ Index kept narrow by design: only themes that map to Robin's actual work (life-s
 | `class-organoid-3d-bioprinting` | Recurring organoid, 3D bioprinting, and 3D-cell-culture automation coverage; this window adds a 2026-07-28 HN drug-discovery essay (owlposting "Why haven't organoids solved all of drug discovery?", 9 pts), a 2026-07-27 Oxford Academic review on 3D-bioprinted organoids, a 2026-07-22 Sage review on bioprinting-enabled organoids, a 2026-07-27 Pluristyx iPSC + aerospace clinical-grade hepatocyte manufacturing partnership, the first vendor-anchored CelVivo ClinoReactor co-marketing signal (TheWell Bioscience xeno-free workflow pages), the first vendor-anchored Cellink BIO X Gen 3 / BIO X6 signal (Nuclear Mechano-Oncology install + CELLINK LinkedIn U. Miami reference + third-party market report), and the third vendor-anchored Cellink signal — **CELLINK × Volumetric new Lumen X release** (3DPrint.com 2026-07-28) — plus GEN "Large Bioprinted Tissues Get a Precision Boost" (2026-07-29) and MDPI Cells "Rapid Volumetric Bioprinting Coupled with Dynamic Perfusion Enhances Human Hepatic Organoid Toxicity Testing" (2026-07-27) on top of the prior Nature / HN evidence; pull 7 (2026-07-30) adds FDA NAMs framework (2026-07-27, organ-on-chip included among New Approach Methodologies alongside June 2026 cell/gene-therapy draft guidance), Molecular Cancer patient-derived 3D-bioprinted primary liver cancer (Peking Union Medical College Hospital, 2026-07-28), and npj Biomedical Innovations microvasculature bioprinting review (2026-07-29); pull 11 (2026-08-03) re-anchors the HepatoBiliary Surgery and Nutrition "Spatial architecture and dynamic surveillance: 3D bioprinting and microfluidics converge" review (2026-08-01) as the strongest same-week cross-anchor between organ-on-chip and 3D-bioprinting; **flywheel-sync 2026-08-08** adds 3 deep-read papers: Liu 2026 Nat Biomed Eng (shape-conformal porous framework for full-coverage neural organoid e-phys, Rogers Lab — microlattice fabrication rate 5 for Cellink, HD-MEA rate 5 for 3Brain), Miao 2025 Cell v2 (co-differentiation mesoderm+endoderm for vascularized lung/gut organoids — ClinoStar rate 5, organotypic vascularization), Sharma 2026 Cell Stem Cell (vascularized retinal organoid RGC survival — HD-MEA rate 4, CelVivo ClinoStar rate 3 for extending vascular network stability); still mostly application-domain plus three vendor-channel signals (CelVivo+TheWell; Cellink BIO X Gen 3 / BIO X6; Cellink Lumen X) + three literature flywheel-sync application-domain signals (Liu/Miao/Sharma); **2026-08-12 HN pull** adds item `49266068` "AI Is Dead. Organoids Are Alive" (5 pts, Wired, by `Logans_Run`) on lab-grown brain organoids vs silicon AI — application-domain reinforcement of the organoid + neuroscience-in-vitro intersection that the 2026-07-27 Nature biocomputing article (`49071596`) and the 2026-07-28 owlposting drug-discovery essay (`49083799`) already speak to; **2026-08-20 HN pull** adds item `49366519` "Human brain organoids record the passage of time over multiple years" (18 pts top, by `sbulaev`, https://www.nature.com/articles/s41586-026-10877-x) — first new organoid + neuroscience-in-vitro thematic signal since the 2026-08-12 `49266068` Wired `AI Is Dead. Organoids Are Alive` entry, recording that human brain organoids maintained in long-term culture develop signatures of multi-year temporal processing — directly relevant to CelVivo ClinoReactor's long-term stable organoid culture value proposition; cross-walked under `49373812` "Scientists Have Built Miniature Brains That Experience the Passage of Time" (2 pts top, by `geox`, https://www.sciencealert.com/unprecedented-scientists-have-built-miniature-brains-that-experience-the-passage-of-time — Science Alert user-press write-up of the same Nature s41586-026-10877-x paper); **second neural-organoid HN signal in 9 pulls** (after the 2026-07-27 Nature biocomputing article `49071596` and the 2026-08-12 Wired `49266068`), qualifying for Last-seen refresh by the recurrence criterion alone; **2026-08-23/24 git-repo-1 LDR burst (FIRST burst)** adds 5 root-level single-shot entries that **strengthen the class-organoid-3d-bioprinting evidence footprint**: (a) **CelVivo App Note 2026 shear stress ClinoStar vs orbital shaker** — CelVivo-direct competitive-killer document (HepG2/C3A 15d + embedded Caco-2 + brain organoid); (b) **Aiello & Bagni 2025 biorxiv Rotating Chamber brain organoid CFD** — first CFD-quantified RC shear-stress study (< 20 mPa, CelVivo-adjacent); (c) **Ajongbolo 2025 Cells brain organoid + assembloid review** — canonical brain-organoid review (CelVivo ClinoStar-adjacent + 3Brain HD-MEA-adjacent), first comprehensive brain-organoid review since the 2026-07-29 Organoids 2026 review; (d) **Kim 2025 Nat Commun EGaIn liquid-metal 3D MEA** — 3rd next-gen HD-MEA approach alongside Liu 2026 + 3Brain Accura-3D (3Brain-adjacent competitor); (e) **Tian 2025 Nat Commun hiPSC epicardium mTOR maturation** — cardiac-epicardium application-domain expansion (no vendor caLine truncated
 | `china-cde-cell-gene-therapy-rules-2026` | China CDE (under NMPA) published technical guidelines defining scope & classification of cell and gene therapy drugs; standardizes CMC planning for Chinese pharma/CDMO/biotech customers | 2026-07-24 | 2026-07-24 | source-backed | web-search | 1 | watchlist | [Tavily: cisema.com 2026-07-24](/sources/web-search.md) |
 | `china-order-no-818-clinical-translation` | China's Order No. 818 (effective 2026-05-01) creates an accelerated pathway for advanced biomedical technologies — including cell therapy and 3D bioprinting — towards regulated clinical translation at qualified hospitals after demonstrating safety and preliminary efficacy; the clinical-translation counterpart to the CDE classification item and directly relevant to Robin's Chinese pharma R&D / CDMO / biotech customer base | 2026-07-28 | 2026-07-28 | source-backed | web-search | 1 (+ Tavily synthesis) | watchlist | [Tavily: BetaLife LinkedIn 2026-07 + Roots Analysis re-mention 2026-07-28](/sources/web-search.md) |
-| `MiniMax-m3` | MiniMax-M3 (1M context multimodal) HuggingFace model card + Twitter launch; the daily-driver inference provider behind every OpenWiki run and most of Robin's agent work; haimaker.ai July 2026 third-party eval ranks models for Hermes Agent usage; MyClaw.ai 2026-07-29 confirms M3 + M2.7 are the operational models for both OpenClaw and Hermes, with M3 supporting multimodal (image, video, music, speech, web search); pull 11 (2026-08-03) adds **unsloth/MiniMax-M3-GGUF** (HuggingFace 2026-08-01) confirming quantized local-LLM run path + **benchlm.ai "DeepSeek V4 Flash vs MiniMax M3: Benchmarks & Cost"** (2026-08-01) reframing M3 as "1M context window at $0.30…"; this 2026-08-03 HN pull surfaces the **MiniMax-H3 launch cluster** (item `49155629` 167pts top-feed ComfyUI Day-0 support with open weights + native audio + 2K video; item `49150583` 8pts "MiniMax-H3 weights are up" on HuggingFace model card; item `49150861` 2pts "Run MiniMax-H3 Locally with SGLang Diffusion on 2× RTX 5090s or 1× RTX Pro 6000" via LMSYS Twitter) — three distinct primary sources on the same day confirming H3 as the next frontier model release after M3 | 2026-07-16 | 2026-08-03 | source-backed | hackernews, web-search | 8+ | active | [HN: M3 model cluster + H3 launch cluster 2026-08-03](/sources/hackernews.md); [Tavily: Hermes Agent best-models eval](/sources/web-search.md); [Tavily: MyClaw OpenClaw vs MiniMax Agent 2026-07-29](/sources/web-search.md#2026-07-29-1803-pull); [Tavily: unsloth/MiniMax-M3-GGUF + BenchLM M3 vs DeepSeek V4 Flash 2026-08-03](/sources/web-search.md#2026-08-03-1808-pull) |
+| `MiniMax-m3` | MiniMax-M3 (1M context multimodal) HuggingFace model card + Twitter launch; the daily-driver inference provider behind every OpenWiki run and most of Robin's agent work; haimaker.ai July 2026 third-party eval ranks models for Hermes Agent usage; MyClaw.ai 2026-07-29 confirms M3 + M2.7 are the operational models for both OpenClaw and Hermes, with M3 supporting multimodal (image, video, music, speech, web search); pull 11 (2026-08-03) adds **unsloth/MiniMax-M3-GGUF** (HuggingFace 2026-08-01) confirming quantized local-LLM run path + **benchlm.ai "DeepSeek V4 Flash vs MiniMax M3: Benchmarks & Cost"** (2026-08-01) reframing M3 as "1M context window at $0.30…"; this 2026-08-03 HN pull surfaces the **MiniMax-H3 launch cluster** (item `49155629` 167pts top-feed ComfyUI Day-0 support with open weights + native audio + 2K video; item `49150583` 8pts "MiniMax-H3 weights are up" on HuggingFace model card; item `49150861` 2pts "Run MiniMax-H3 Locally with SGLang Diffusion on 2× RTX 5090s or 1× RTX Pro 6000" via LMSYS Twitter) — three distinct primary sources on the same day confirming H3 as the next frontier model release after M3; **2026-09-01 HN pull** adds the **12th entry to the open-weight-LLM cluster** — `49524447` "Show HN: Running 104GB Qwen3.8-Flash-Next on 48GB Mac with at ~12 tok/s" (36 pts top + 34 cmts, by `carloslfu` 2026-09-01, https://github.com/carloslfu/slotstream — Show HN local-LLM deployment of Qwen3.8-Flash-Next on a 48GB Mac via the slotstream project); the open-weight-LLM cluster now totals **~3,354+ pts across 12 items in 10 pulls over 13 days** — the strongest sustained engagement cluster of the entire 2026-08-07 streak — joining the 2026-08-21 DeepSeek V4 Flash vision 379 pts + 2026-08-19 Unsloth Dynamic 3.0 GGUFs 314 pts + 2026-08-23 GLM-5.3 vs Anthropic/OpenAI 141 pts + 2026-08-23 Qwen 3.8 27B reverse-engineering 234 pts + 2026-08-25 Qwen 3.8-Flash-Next "releasing tomorrow" 244 pts + 2026-08-26 GLM-5.3-Flash 530 pts + 2026-08-26 Qwen3.8-Flash-Next release 441 pts + 2026-08-26 Z.ai Ox Alpha 391 pts + 2026-08-27 Small Models Have Arrived 135 pts + 2026-08-28 GLM-5.3 is now open-weight 268 pts + 2026-08-30 Tencent Hy4 preview 364 pts + **2026-09-01 Qwen3.8-Flash-Next on 48GB Mac 36 pts**; the cluster corroborates the `MiniMax-m3` row's open-weight comparison anchor + the unsloth/MiniMax-M3-GGUF local-LLM run path; Last seen stays at 2026-08-03 because the 2026-09-01 pull does not surface any MiniMax-M3 or MiniMax-H3 release | 2026-07-16 | 2026-08-03 | source-backed | hackernews, web-search | 9+ | active | [HN: M3 model cluster + H3 launch cluster 2026-08-03](/sources/hackernews.md); [HN: Qwen3.8-Flash-Next on 48GB Mac 2026-09-01 — 12th open-weight-LLM cluster entry, ~3,354+ pts across 12 items in 10 pulls over 13 days](/sources/hackernews.md#latest-update--2026-09-01-1800-utc); [Tavily: Hermes Agent best-models eval](/sources/web-search.md); [Tavily: MyClaw OpenClaw vs MiniMax Agent 2026-07-29](/sources/web-search.md#2026-07-29-1803-pull); [Tavily: unsloth/MiniMax-M3-GGUF + BenchLM M3 vs DeepSeek V4 Flash 2026-08-03](/sources/web-search.md#2026-08-03-1808-pull) |
 | `hermes-agent-v0-19-quicksilver` | Hermes Agent v0.19 "Quicksilver" release (2026-07-25): first-token delay 4.3s → 0.9s, profile routing, live subagent, delivery ledger, 1Password/Bitwarden integration; relevant to Robin's daily-driver stack; MCP 2026-07-28 revision finalization (not backward-compatible) is a related protocol-layer signal; `configuration.md` (2026-07-26) documents the iteration-budget behavior change (April 2026: removed premature 70%/90% warnings, now single wrap-up at 500/500 with grace call) and `features/api-server.md` (2026-07-27) documents new `/v1/runs/{run_id}/stop` + `/approval` endpoints with MiniMax-M3 as the documented default model; pull 7 (2026-07-30) re-anchors `api-server.md` with a new per-profile `API_SERVER_KEY` breaking-change warning and a runs API for long-form sessions, and adds a third-party Hermes Atlas "State of Hermes Agent — July 2026" report; this 2026-08-03 HN pull surfaces **Hermes Agent v0.20.0** (item `49159200`, NousResearch GitHub releases tag `v2026.8.3`, 1pt on HN day-of — the next major version after v0.19 Quicksilver); pull 11 (2026-08-03) adds a **Hermes Agent 0.19.1 GitHub bug report** (NousResearch issues/77000, 2026-08-02, score 0.790, "gpt-5.6-luna + opencode-go") confirming patch-version activity on the v0.19.x track between the v0.19 Quicksilver release and the v0.20.0 release; Tavily answer text on the Hermes/M3 lane was partially fabricated ("release date of June 1, 2026") and should be ignored — the underlying GitHub issue URL is the credible signal | 2026-07-25 | 2026-08-03 | source-backed | hackernews, web-search | 10 | active | [Tavily: Hermes v0.19 Quicksilver + MCP revision 2026-07-27](/sources/web-search.md); [Tavily: hermes-agent configuration.md + api-server.md + MyClaw 2026-07-29](/sources/web-search.md#2026-07-29-1803-pull); [Tavily: api-server.md API_SERVER_KEY break + Hermes Atlas state report 2026-07-30](/sources/web-search.md#2026-07-30-1804-pull); [Tavily: Hermes Agent 0.19.1 GitHub bug report 2026-08-03](/sources/web-search.md#2026-08-03-1808-pull); [HN: Hermes Agent v0.20.0 release 2026-08-03](/sources/hackernews.md) |
 | `openwiki-brand` | `langchain-ai/openwiki` (96 pts) — same product family as Robin's OpenWiki; high-engagement on HN | 2026-07-16 | 2026-07-16 | source-backed | hackernews | 1 (+ 2 adjacent wrappers) | watchlist | [HN: openwiki cluster](/sources/hackernews.md#openwiki-brand-signal) |
+<!-- openwiki: broken internal link [/sources/hackernews.md#latest-update--2026-08-31-1800-utc] heading anchor "latest-update--2026-08-31-1800-utc" does not exist in "/sources/hackernews.md". Fix the href or restore the target, then delete this comment. -->
 | `agent-memory-architecture` | Shared agent-memory architecture using Karpathy's LLM Wiki + Google's Open Knowledge Format (OKF) + MCP — let "every Local AI share ONE permanent memory"; this 2026-07-29 Tavily pull adds four more corroborated sources (Hackernoon Moorcheh.ai "Agent Memory Has a Lock-In Problem. Open Formats Are How We Fix It." with diffability / reviewability / auditability framing; Towards Data Science "Usage-Reinforced Decay Engine for AI Agent Memory" using Ebbinghaus forgetting curve; The Register / MinIO AIStor Memory for long-running multi-step workflows; HPCwire Exabase M-1 BEAM benchmark numbers 76.9% / 75.0% / 68.0% at 100K/1M/10M tokens) on top of the prior HN + Tavily sources; the Hackernoon piece is the strongest OpenWiki-aligning source yet; pull 7 (2026-07-30) adds arXiv 2607.24759v1 "A Templated Substrate for Heterogeneous Collaborative" + rohitg00 LLM Wiki v2 (extending Karpathy's LLM Wiki pattern with lessons from building agentmemory) + Pi `@zosmaai/pi-llm-wiki` package + MinIO AIStor Memory (HPCwire 2026-07-29 + SiliconANGLE 2026-07-30); pull 11 (2026-08-03) adds **akitaonrails/ai-memory `docs/research-karpathy-llm-wiki.md`** (GitHub 2026-07-28, score 0.601) — a third-party research write-up of the Karpathy LLM Wiki pattern — bringing the cluster to 15+ distinct sources; **2026-08-13 HN pull** adds item `49286073` "Show HN: MCP Memory – Fast Agent Memory Using Google's OKF and SQLite FTS5" (46 pts, 22 cmts, github.com/fellowgeek/mcp-memory) — **the first HN item since the 2026-07-16 `langchain-ai/openwiki` brand signal to explicitly name Google's OKF frontmatter** (the same schema OpenWiki uses at `/index.md` frontmatter `okf_version: "0.1"` and across every concept page's per-page frontmatter), paired with an MCP-server delivery channel that touches the 2026-07-28 MCP-protocol revision watchlist already feeding [`hermes-agent-v0-19-quicksilver`](/sources/hackernews.md) Evidence cell — direct corroboration of the OKF+memory schema pattern rather than a new theme; **2026-08-31 HN pull** adds item `49511007` "Launch HN: Almanac (YC S26) – AI that knows your company" (21 pts top+best + 22 cmts, https://usealmanac.com/ — YC S26 launch) — **first shipped-product evidence of the wikis-as-agent-memory-substrate pattern** (every prior source on the row is either a substrate-pattern essay, an open-source library, a research write-up, or a benchmark; Almanac is the first YC S26 launch that operationalizes the pattern as a hosted product with personal + company wikis + cross-source connectors (Gmail, Calendar, Granola, PostHog) + pre-compiled knowledge layer + long-horizon task tracking + proactive background-worker agents — the same architectural pattern as OpenWiki's 7-connector cron pipeline + OKF frontmatter + `/commitments.md` long-horizon tracking + `~/.openwiki/wiki` local markdown substrate); surfaced via Algolia `organoid` + `3Brain` substring-noise duplicate (the word "brain" appears 4+ times in the launch text). Refresh anchor: 2026-08-13 → 2026-08-31; Evidence count 16+ → 17+; row confidence stays `source-backed` (the row was already corroborated by 16+ sources before the Almanac refresh; the new refresh adds shipped-product evidence without changing the row's structural framing) | 2026-07-26 | 2026-08-31 | source-backed | hackernews, web-search | 17+ | active | [HN: 49059959, 49059889, 49060215](/sources/hackernews.md); [HN: MCP Memory using Google's OKF 2026-08-13](/sources/hackernews.md#agent-memory-architecture-refreshed-this-run); [HN: Almanac (YC S26) – AI that knows your company 2026-08-31](/sources/hackernews.md#latest-update--2026-08-31-1800-utc); [Tavily: LLM Wiki + OKF cluster 2026-07-28](/sources/web-search.md); [Tavily: Moorcheh.ai + TDS + MinIO + Exabase cluster 2026-07-29](/sources/web-search.md#2026-07-29-1803-pull); [Tavily: arXiv 2607.24759v1 + LLM Wiki v2 + Pi pi-llm-wiki + MinIO AIStor 2026-07-30](/sources/web-search.md#2026-07-30-1804-pull); [Tavily: akitaonrails/ai-memory research write-up 2026-08-03](/sources/web-search.md#2026-08-03-1808-pull) |
 
 ## Notes