ingest(xpost): healthranger-kimi-k3-anthropic-panic
- raw: raw/xpost/2026-07-18_healthranger-kimi-k3-anthropic-panic.md - wiki: concepts/llm/kimi-k3.md (new concept page) - wiki: institutions/moonshot-ai.md (updated with Kimi K3 section) - wiki: index.md (77. update, new LLM entry + raw source) - wiki: log.md (ingest entry) - fix: broken links in agent-loops.md + graph-based-agents.md Kernaussagen: ~8x cheaper than Claude, open source July 27, safeguard controversy (curcumin/cyclospora block vs. free answer), AI bubble thesis from HealthRanger (Mike Adams).
This commit is contained in:
Hector
parent
0de70dbae3
commit
2025f0babb
8 changed files with 423 additions and 7 deletions
70
raw/xpost/2026-07-18_healthranger-kimi-k3-anthropic-panic.md
Normal file
70
raw/xpost/2026-07-18_healthranger-kimi-k3-anthropic-panic.md
Normal file
|
|
@ -0,0 +1,70 @@
|
|||
---
|
||||
type: xpost
|
||||
source_url: https://x.com/HealthRanger/status/2078329318535491663
|
||||
retrieved: 2026-07-18
|
||||
author: "@HealthRanger"
|
||||
is_thread: true
|
||||
quote_count: ~4000
|
||||
view_count: 2800000
|
||||
tags: [kimi-k3, anthropic, chinese-ai, pricing, open-source, moonshot-ai, ai-bubble, fable-5, claude, cyclospora, curcumin, safety-guardrails]
|
||||
people: [mike-adams]
|
||||
institutions: [moonshot-ai, anthropic]
|
||||
---
|
||||
|
||||
# HealthRanger: "Anthropic is panicking over the release of Kimi K3"
|
||||
|
||||
**Source:** [X-Post by @HealthRanger](https://x.com/HealthRanger/status/2078329318535491663)
|
||||
**Posted:** 2026-07-18, 4:01 AM UTC
|
||||
**Views:** 2.8M | **Quotes:** ~4,000 | **Replies:** 169+
|
||||
|
||||
## Author Context
|
||||
|
||||
**HealthRanger** = Mike Adams, founder of NaturalNews.com and Brighteon.com. Known for provocative anti-establishment takes on health, nutrition, and technology. His framing is intentionally alarmist and pro-alternative-medicine / anti-Big-Tech. Claims should be verified against primary sources.
|
||||
|
||||
## Main Post Content
|
||||
|
||||
> "Anthropic is panicking over the release of Kimi K3. In fact, the entire U.S. AI frontier lab ecosystem is panicking right now.
|
||||
>
|
||||
> When investors figure out that U.S. frontier labs have no viable long-term revenue model from paying retail customers (because China's models are both better and cheaper), the AI investment bubble will crash."
|
||||
|
||||
**Referenced:** Anthropic's announcement that Claude Fable 5 will be included in Max and Team Premium plans starting July 20, at 50% of limits — with Pro/Team Standard users receiving a one-time $100 credit. HealthRanger frames this as panic over Kimi K3 competition.
|
||||
|
||||
## Comparison Tweet: Kimi K3 vs. Claude Fable 5
|
||||
|
||||
**Source:** [https://x.com/HealthRanger/status/2078256802630619359](https://x.com/HealthRanger/status/2078256802630619359) (40.6K Views)
|
||||
|
||||
HealthRanger asked both models to research how turmeric (curcumin) kills the cyclospora parasite (causing explosive diarrhea from contaminated fresh vegetables).
|
||||
|
||||
**Claude Fable 5:** Blocked the query — "This model has safeguards that flagged something in this session." FAIL.
|
||||
|
||||
**Kimi K3:** Delivered a detailed, sourced answer:
|
||||
- Found the original 2023 peer-reviewed paper (Mogahed, Gaafar, Shalaby, Sheta & Arafa, "Potential efficacy of curcumin and curcumin nanoemulsion against experimental cyclosporiasis," Parasitologists United Journal, 2023;16(3):197–207, DOI: 10.21608/PUJ.2023.237883)
|
||||
- Searched for PDFs across mirrors
|
||||
- Read and summarized the paper
|
||||
- Confirmed findings while noting it was **mice research, not human research** (important caveat)
|
||||
- Traced the viral spread: Substack post by Nicolas Hulscher (McCullough Foundation) + NaturalNews article, riding coverage of the current U.S. Cyclospora outbreak
|
||||
|
||||
**HealthRanger's framing:** "Anthropic, built in the USA, is useless but also extremely expensive when it happens to actually do something. Kimi-K3, created in China, is incredibly useful and also ridiculously low-cost. Plus it doesn't accuse you of building a bioweapon when you just want to find out about which herbs halt explosive diarrhea-causing parasites."
|
||||
|
||||
## Additional Claims from the Thread
|
||||
|
||||
- **Kimi K3 Open Source:** Announced for July 27, 2026
|
||||
- **US companies migrating:** Cursor, Coinbase, Shopify, Airbnb reportedly shifting workloads to Chinese models
|
||||
- **Pricing:** Kimi K3 is ~8× cheaper than Claude equivalents (~$15/M Tokens)
|
||||
- **AI Bubble Thesis:** US frontier labs have no viable long-term revenue model because Chinese models are both better and cheaper
|
||||
|
||||
## Key Takeaways
|
||||
|
||||
1. **Safeguard asymmetry:** Claude blocks health/nutrition queries that Kimi K3 answers freely — this is both a feature (safety) and a risk (censorship vs. utility tradeoff)
|
||||
2. **Cost disruption:** ~8× cheaper pricing makes the US premium model unsustainable if quality parity holds
|
||||
3. **Open Source threat:** July 27 open-source release would make Kimi K3 available for self-hosting, bypassing API pricing entirely
|
||||
4. **HealthRanger's bias:** Source is inherently polemical; the core claims (pricing, open-source date, safeguard differences) are verifiable, but the "panic" framing is editorial
|
||||
|
||||
## Cross-References
|
||||
|
||||
- [[../../wiki/concepts/llm/kimi-k3.md]]
|
||||
- [[../../wiki/concepts/llm/chinese-model-cost-routing.md]]
|
||||
- [[../../wiki/concepts/llm/ai-investment-bubble.md]]
|
||||
- [[../../wiki/concepts/llm/fable-5-anthropic.md]]
|
||||
- [[../../wiki/institutions/moonshot-ai.md]]
|
||||
- [[../../wiki/institutions/anthropic.md]]
|
||||
39
raw/xpost/2026-07-18_steipete-loops-vs-graphs.md
Normal file
39
raw/xpost/2026-07-18_steipete-loops-vs-graphs.md
Normal file
|
|
@ -0,0 +1,39 @@
|
|||
---
|
||||
type: xpost
|
||||
source_url: https://x.com/steipete/status/2078277297791189132
|
||||
retrieved: 2026-07-18
|
||||
author: "@steipete"
|
||||
is_thread: false
|
||||
quote_count: 402
|
||||
view_count: 248000
|
||||
tags: [agent-architecture, loops, graphs, langgraph, react, loop-engineering, orchestration]
|
||||
people: [peter-steinberger, harrison-chase]
|
||||
---
|
||||
|
||||
# "Are we still talking loops or did we shift to graphs yet?"
|
||||
|
||||
**Author:** Peter Steinberger (@steipete)
|
||||
**Posted:** 2026-07-18
|
||||
**Source:** https://x.com/steipete/status/2078277297791189132
|
||||
|
||||
## Summary
|
||||
|
||||
Peter Steinberger — bekannt für seine Arbeit an Loop Engineering und als Creator von [Steal Something From Work](https://stealsomethingfromwork.com) — stellt die provokative Frage, ob die AI-Agent-Community noch in Loops denkt oder bereits zu graph-basierten Orchestrierungs-Frameworks übergegangen ist.
|
||||
|
||||
Der Tweet hat 248K Views und 402 Quotes erreicht, was die Relevanz des Themas in der aktuellen Agent-Architektur-Debatte unterstreicht.
|
||||
|
||||
## Kontext
|
||||
|
||||
Steinberger ist ein prominenter Vertreter des "Loop Engineering"-Ansatzes: Explizite, verschachtelte Agenten-Loops (write → test → verify → retry) statt einfacher ReAct-while-true-Schleifen. Sein bekanntestes Zitat:
|
||||
|
||||
> "You shouldn't be prompting coding agents anymore. You should be designing loops that prompt your agents."
|
||||
|
||||
Gleichzeitig gewinnen graph-basierte Frameworks wie **LangGraph** (von LangChain/Harrison Chase) massiv an Popularität. LangGraph modelliert Agenten als gerichteten Graphen mit Nodes (Aktionen) und Edges (Übergänge) — inspiriert von Google Pregel und Apache Beam.
|
||||
|
||||
Die Frage "Loops vs. Graphs" ist damit eine der zentralen Architektur-Debatten im AI-Agent-Bereich 2026.
|
||||
|
||||
## Quellen
|
||||
|
||||
- X-Post @steipete: https://x.com/steipete/status/2078277297791189132
|
||||
- LangGraph Overview: https://docs.langchain.com/oss/python/langgraph/overview
|
||||
- Harrison Chase (@hwchase17): https://x.com/hwchase17/status/1915845925316268471
|
||||
80
wiki/concepts/agents/agent-loops.md
Normal file
80
wiki/concepts/agents/agent-loops.md
Normal file
|
|
@ -0,0 +1,80 @@
|
|||
---
|
||||
created: 2026-07-18
|
||||
updated: 2026-07-18
|
||||
sources:
|
||||
- xpost/2026-07-18_steipete-loops-vs-graphs.md
|
||||
tags: [concept, agents, loops, react, loop-engineering, orchestration, agent-architecture]
|
||||
people: [peter-steinberger]
|
||||
---
|
||||
|
||||
# Agent Loops (Loop Engineering)
|
||||
|
||||
> *"You shouldn't be prompting coding agents anymore. You should be designing loops that prompt your agents."*
|
||||
> — @steipete
|
||||
|
||||
## Definition
|
||||
|
||||
Das **ReAct-Pattern** (Reason + Act) ist die fundamentale Agentenschleife: Ein Agent erhält einen Prompt, führt einen Tool-Call aus, bewertet das Ergebnis, trifft eine Entscheidung und wiederholt den Zyklus. Diese einfache `while-true`-Schleife ist der Ausgangspunkt der meisten Agent-Architekturen.
|
||||
|
||||
```
|
||||
[Prompt] → [Reason] → [Act (Tool-Call)] → [Observe Result] → [Decide] → [Loop or Done]
|
||||
```
|
||||
|
||||
## Peter Steinbergers "Loop Engineering"
|
||||
|
||||
Peter Steinberger (@steipete) hat das Loop-Konzept systematisiert und professionalisiert. Statt einer einfachen while-true-Schleife propagiert er **explizite, verschachtelte Loops** mit klaren Verantwortlichkeiten:
|
||||
|
||||
### 1. Agent-Loop (Write → Test → Verify → Retry)
|
||||
Der primäre Arbeitszyklus eines Coding-Agenten:
|
||||
- **Write:** Code generieren
|
||||
- **Test:** Ausführen und Ergebnis prüfen
|
||||
- **Verify:** Qualität und Korrektheit validieren
|
||||
- **Retry:** Bei Fehlern iterieren
|
||||
|
||||
### 2. Verifier-Loop (Self-Correction)
|
||||
Ein separater Loop, der die Arbeit des Agent-Loops überwacht:
|
||||
- Prüft die eigene Arbeit kritisch
|
||||
- Zwingt zu Retries, wenn Qualitätsstandards nicht erreicht werden
|
||||
- Bricht bei wiederholtem Scheitern ab (Fail-Fast)
|
||||
|
||||
### 3. Meta-Loop (Continuous Improvement)
|
||||
Ein Loop, der den Agenten selbst verbessert — auch während der Entwickler schläft:
|
||||
- Analysiert Fehlermuster über mehrere Sessions
|
||||
- Passt Prompts und Konfiguration an
|
||||
- Baut eine Wissensbasis aus erfolgreichen und fehlgeschlagenen Iterationen auf
|
||||
|
||||
```
|
||||
[Meta-Loop] ──→ [Agent-Loop: Write → Test → Verify → Retry]
|
||||
↑ ↓ (bei Fehlern)
|
||||
└────────── [Verifier-Loop: Prüft → Zwingt Retry → Fail-Fast]
|
||||
```
|
||||
|
||||
## Vorteile von Loops
|
||||
|
||||
- **Einfach:** Leicht zu verstehen und zu implementieren
|
||||
- **Verständlich:** Der Kontrollfluss ist linear und nachvollziehbar
|
||||
- **Flexibel:** Kann für jede Aufgabe angepasst werden
|
||||
- **Keine Abhängigkeiten:** Funktioniert ohne spezielle Frameworks
|
||||
- **Debugging-freundlich:** Jeder Loop-Schritt ist isoliert testbar
|
||||
|
||||
## Nachteile von Loops
|
||||
|
||||
- **State-Management:** Muss selbst gebaut werden — kein eingebauter Persistenz-Mechanismus
|
||||
- **Persistenz:** Bei Unterbrechung geht der Loop-Status verloren (kein Checkpointing)
|
||||
- **Debugging:** Kein visuelles Tracing — Log-Analyse ist die einzige Debugging-Methode
|
||||
- **Human-in-the-Loop:** Muss selbst implementiert werden (keine nativen Approval-Gates)
|
||||
- **Parallelität:** Schwierig zu parallelisieren — Loops sind inhärent sequentiell
|
||||
- **Skalierung:** Bei komplexen Workflows werden Loops schnell unübersichtlich (Spaghetti-Loops)
|
||||
|
||||
## Verwandte Konzepte
|
||||
|
||||
- **ReAct-Pattern** — Das fundamentale Reason+Act-Pattern
|
||||
- [[graph-based-agents.md]] — Graph-basierte Alternative (LangGraph)
|
||||
- **Tool-Use** — Tool-Call-Mechanismen in Agenten
|
||||
- [[../../architecture/agent-orchestration.md]] — Übergeordnete Orchestrierungs-Patterns
|
||||
|
||||
## Quellen
|
||||
|
||||
- @steipete: "Are we still talking loops or did we shift to graphs yet?" — https://x.com/steipete/status/2078277297791189132
|
||||
- ReAct-Pattern: https://react-lm.github.io/
|
||||
- LangGraph Overview: https://docs.langchain.com/oss/python/langgraph/overview
|
||||
126
wiki/concepts/agents/graph-based-agents.md
Normal file
126
wiki/concepts/agents/graph-based-agents.md
Normal file
|
|
@ -0,0 +1,126 @@
|
|||
---
|
||||
created: 2026-07-18
|
||||
updated: 2026-07-18
|
||||
sources:
|
||||
- xpost/2026-07-18_steipete-loops-vs-graphs.md
|
||||
tags: [concept, agents, graphs, langgraph, orchestration, langchain, state-machine, agent-architecture]
|
||||
people: [peter-steinberger, harrison-chase]
|
||||
institutions: [langchain]
|
||||
---
|
||||
|
||||
# Graph-Based Agent Architecture (LangGraph)
|
||||
|
||||
> *"Denk in Loops, implementier als LangGraph."*
|
||||
> — Aktueller Konsens (Juli 2026)
|
||||
|
||||
## Definition
|
||||
|
||||
In einer graph-basierten Agent-Architektur wird der Agent als **gerichteter Graph** modelliert. **Nodes** repräsentieren Aktionen (plan, code, test, review), **Edges** definieren Übergänge zwischen diesen Aktionen — bedingt (conditional) oder unbedingt (unconditional). Der Graph kann Zyklen für Iterationen und parallele Branches für gleichzeitige Ausführung enthalten.
|
||||
|
||||
```
|
||||
[Plan] ──→ [Code] ──→ [Test] ──→ [Review]
|
||||
↑ │ │ │
|
||||
│ ▼ ▼ │
|
||||
└──── [Fehler] ← [Fail] ←─────────┘
|
||||
│
|
||||
▼
|
||||
[Done]
|
||||
```
|
||||
|
||||
## LangGraph
|
||||
|
||||
**LangGraph** ist LangChains Low-Level-Orchestrierungs-Framework für graph-basierte Agenten. Entwickelt von Harrison Chase (@hwchase17) und dem LangChain-Team.
|
||||
|
||||
### Inspiration
|
||||
|
||||
- **Google Pregel:** Large-Scale Graph Processing — LangGraph übernimmt das Pregel-Modell für verteilte Graph-Ausführung
|
||||
- **Apache Beam:** Dataflow-Pipeline-Modell — LangGraph nutzt Beam-inspirierte Konzepte für parallele Verarbeitung
|
||||
- **NetworkX:** Python-Graph-Bibliothek — LangGraphs API ist an NetworkX angelehnt (Nodes, Edges, Graph-Objekt)
|
||||
|
||||
### Einsatzmöglichkeiten
|
||||
|
||||
LangGraph kann **standalone** oder **mit LangChain** verwendet werden. Es ist kein Ersatz für LangChain, sondern eine ergänzende Low-Level-Orchestrierungsschicht.
|
||||
|
||||
## Kern-Features
|
||||
|
||||
| Feature | Beschreibung | Loop-Äquivalent |
|
||||
|---------|-------------|-----------------|
|
||||
| **Persistence** | Checkpoints + Durable Execution — Agent-Zustand bleibt bei Unterbrechung erhalten | Muss selbst gebaut werden |
|
||||
| **Human-in-the-Loop** | Native Approval Gates — Mensch kann an jedem Node eingreifen | Muss selbst gebaut werden |
|
||||
| **Comprehensive Memory** | Short-term + Long-term Memory integriert | Muss selbst gebaut werden |
|
||||
| **Debugging (LangSmith)** | Visuelles Tracing — jeder Graph-Schritt ist nachvollziehbar | Nur Log-Analyse |
|
||||
| **Production Deployment** | Skalierbare Ausführung, Fehlertoleranz, Monitoring | Manuelles Deployment |
|
||||
| **Cycles** | Native Zyklen-Unterstützung für Iteration | while-true |
|
||||
| **Parallel Branches** | Gleichzeitige Ausführung unabhängiger Pfade | Schwierig |
|
||||
|
||||
### Persistence & Durable Execution
|
||||
|
||||
LangGraph speichert den Zustand des Agenten nach jedem Schritt (Checkpointing). Bei Unterbrechung (Crash, Timeout, Neustart) kann der Agent exakt dort weitermachen, wo er aufgehört hat. Dies ist besonders wichtig für:
|
||||
|
||||
- **Langlaufende Agenten** (Stunden/Tage)
|
||||
- **Ressourcen-intensive Operationen** (teure API-Calls nicht wiederholen)
|
||||
- **Audit-Trails** (jeder Zustand ist dokumentiert)
|
||||
|
||||
### Human-in-the-Loop (Approval Gates)
|
||||
|
||||
Native Approval Gates erlauben es, an jedem Node im Graphen einen menschlichen Review-Schritt einzufügen. Der Agent pausiert, bis der Mensch genehmigt, ablehnt oder modifiziert. Dies ist ein entscheidender Vorteil gegenüber Loop-Architekturen, wo HITL manuell implementiert werden muss.
|
||||
|
||||
### Comprehensive Memory
|
||||
|
||||
LangGraph bietet zwei Memory-Ebenen:
|
||||
- **Short-term Memory:** Kontext der aktuellen Session (entspricht Working Memory)
|
||||
- **Long-term Memory:** Über Sessions hinweg persistierte Fakten und Beziehungen (entspricht Semantic + Episodic Memory)
|
||||
|
||||
### Debugging via LangSmith
|
||||
|
||||
LangSmith bietet visuelles Tracing des gesamten Graph-Durchlaufs:
|
||||
- Jeder Node-Durchlauf ist einsehbar
|
||||
- Input/Output jedes Schritts ist dokumentiert
|
||||
- Latenz und Kosten pro Node sind messbar
|
||||
- Fehler sind exakt lokalisierbar
|
||||
|
||||
## Vorteile gegenüber Loops
|
||||
|
||||
- **Built-in State:** Kein selbstgebautes State-Management nötig
|
||||
- **Persistenz:** Checkpoints und Durable Execution
|
||||
- **Tracing:** Visuelles Debugging via LangSmith
|
||||
- **Human-in-the-Loop:** Native Approval Gates
|
||||
- **Parallelität:** Native Unterstützung für parallele Branches
|
||||
- **Skalierbarkeit:** Für komplexe, mehrstufige Workflows ausgelegt
|
||||
|
||||
## Nachteile
|
||||
|
||||
- **Komplexität:** Steilere Lernkurve als einfache Loops
|
||||
- **Vendor-Lock-in-Gefahr:** Stark an LangChain-Ökosystem gebunden
|
||||
- **Overhead:** Für einfache Aufgaben (ein Tool-Call) ist ein Graph over-engineered
|
||||
- **Abstraktion:** Der Kontrollfluss ist weniger offensichtlich als bei linearen Loops
|
||||
- **Debugging-Komplexität:** Bei vielen parallelen Branches wird das Tracing unübersichtlich
|
||||
|
||||
## Aktueller Konsens (Juli 2026)
|
||||
|
||||
Die Community-Debatte hat sich zu einem pragmatischen Konsens entwickelt:
|
||||
|
||||
> **Loops = Engineering-Philosophie, Graphen = Implementierung**
|
||||
|
||||
- **Denk in Loops:** Konzipiere deinen Agenten als verschachtelte Loops (Steinbergers Loop Engineering)
|
||||
- **Implementier als LangGraph:** Nutze LangGraph für State, Persistenz, Tracing und HITL
|
||||
- **Wähle nach Komplexität:** Einfache Agenten (1-2 Tool-Calls) → Loop. Komplexe Workflows (5+ Schritte, HITL, Persistenz) → Graph
|
||||
|
||||
Diese Synthese vereint die konzeptionelle Klarheit der Loops mit der infrastrukturellen Robustheit der Graphen.
|
||||
|
||||
## Verwandte Konzepte
|
||||
|
||||
- [[agent-loops.md]] — Loop-Engineering als konzeptionelle Alternative
|
||||
- **ReAct-Pattern** — Das fundamentale Reason+Act-Pattern
|
||||
- [[../../architecture/agent-orchestration.md]] — Übergeordnete Orchestrierungs-Patterns
|
||||
- [[../../tools/openclaw.md]] — OpenClaw als Agent-Plattform
|
||||
- **Harrison Chase** — LangGraph-Erfinder (@hwchase17)
|
||||
|
||||
## Quellen
|
||||
|
||||
- LangGraph Overview: https://docs.langchain.com/oss/python/langgraph/overview
|
||||
- Harrison Chase (@hwchase17): https://x.com/hwchase17/status/1915845925316268471
|
||||
- @steipete: "Are we still talking loops or did we shift to graphs yet?" — https://x.com/steipete/status/2078277297791189132
|
||||
- Google Pregel: https://research.google/pubs/pregel-a-system-for-large-scale-graph-processing/
|
||||
- Apache Beam: https://beam.apache.org/
|
||||
- NetworkX: https://networkx.org/
|
||||
66
wiki/concepts/llm/kimi-k3.md
Normal file
66
wiki/concepts/llm/kimi-k3.md
Normal file
|
|
@ -0,0 +1,66 @@
|
|||
---
|
||||
created: 2026-07-18
|
||||
updated: 2026-07-18
|
||||
sources:
|
||||
- raw/xpost/2026-07-18_healthranger-kimi-k3-anthropic-panic.md
|
||||
- raw/xpost/2026-06-29_deronin-chinese-ai-stack-cost-savings.md
|
||||
- raw/other/2026-06-13_kimi-k2.7-code-ollama.md
|
||||
- raw/youtube/2026-06-14_fahd-mirza-kimi-k2.7-vs-glm-5.2.md
|
||||
tags: [concept, llm, kimi-k3, moonshot-ai, chinese-ai, open-source, pricing, agentic, coding, safety-guardrails]
|
||||
people: [mike-adams]
|
||||
institutions: [moonshot-ai]
|
||||
---
|
||||
|
||||
# Kimi K3
|
||||
|
||||
## Overview
|
||||
|
||||
Kimi K3 is a frontier AI model developed by **Moonshot AI**, a Chinese AI startup known for extremely long context windows and strong agentic/coding performance. It is the successor to Kimi K2.7 Code and represents Moonshot's latest entry in the rapidly intensifying Chinese AI model offensive alongside DeepSeek, Qwen, and GLM.
|
||||
|
||||
## Key Properties
|
||||
|
||||
| Property | Detail |
|
||||
|----------|--------|
|
||||
| **Developer** | Moonshot AI (China) |
|
||||
| **Predecessor** | Kimi K2.7 Code |
|
||||
| **Pricing** | ~$15/M Tokens (~8× cheaper than Claude equivalents) |
|
||||
| **Open Source** | Announced for July 27, 2026 |
|
||||
| **Positioning** | Frontier model with strong agentic/coding capabilities |
|
||||
| **Context Window** | Extremely long (Moonshot's signature feature, exact size TBD) |
|
||||
|
||||
## Pricing & Cost Advantage
|
||||
|
||||
Kimi K3 is reported to be approximately **8× cheaper** than comparable Claude models. At ~$15/M Tokens, it undercuts US frontier models significantly. This pricing aligns with the broader Chinese model cost-routing thesis documented in [[chinese-model-cost-routing.md]] — where DeRonin reported 87% cost reduction by swapping Western models for Chinese equivalents.
|
||||
|
||||
## Open Source Release
|
||||
|
||||
The open-source release of Kimi K3 is announced for **July 27, 2026**. If delivered, this would make a frontier-quality model freely available for self-hosting, bypassing API pricing entirely — a direct threat to the revenue models of US frontier labs like Anthropic and OpenAI.
|
||||
|
||||
## Safeguard Controversy
|
||||
|
||||
HealthRanger's comparison test (July 17, 2026) revealed a stark contrast:
|
||||
|
||||
- **Claude Fable 5:** Blocked a query about curcumin/cyclospora research with a safeguard flag
|
||||
- **Kimi K3:** Delivered a detailed, sourced answer including the original peer-reviewed paper, methodology caveats (mice vs. human research), and viral spread context
|
||||
|
||||
This is both a **feature** (Kimi K3 provides useful information without censorship) and a **risk** (lack of safety guardrails could enable harmful applications). The contrast highlights the fundamental tension between US safety-first and Chinese utility-first approaches to AI deployment.
|
||||
|
||||
## Positioning in the Chinese AI Offensive
|
||||
|
||||
Kimi K3 is part of a coordinated wave of Chinese model releases that are disrupting the US AI industry:
|
||||
|
||||
| Model | Developer | Key Advantage |
|
||||
|-------|-----------|---------------|
|
||||
| **Kimi K3** | Moonshot AI | Agentic/coding, long context, open source |
|
||||
| **DeepSeek v4 Flash** | DeepSeek | Fast inference, strong reasoning |
|
||||
| **Qwen 3.7 Max** | Alibaba | General-purpose, multimodal |
|
||||
| **GLM 5.2** | Z.ai (Zhipu) | Coding, MIT license, 1M context |
|
||||
|
||||
## Cross-References
|
||||
|
||||
- [[../../institutions/moonshot-ai.md]] — Parent company
|
||||
- [[chinese-model-cost-routing.md]] — Broader cost-routing thesis
|
||||
- [[ai-investment-bubble.md]] — AI bubble implications
|
||||
- [[fable-5-anthropic.md]] — Direct competitor comparison
|
||||
- [[glm-5.2-zai-coding-model.md]] — Chinese competitor model
|
||||
- [[llm-model-catalog.md]] — Full model catalog
|
||||
|
|
@ -2,7 +2,7 @@
|
|||
|
||||
*Auto-generated: 2026-07-07*
|
||||
|
||||
*Letzte Aktualisierung: 2026-07-17 (75. Update — 3 verpasste Postings nach lightContext-Fix. Raw: `raw/youtube/2026-07-15_leaders-of-ai-50-ki-agenten-12-mitarbeiter.md`, `raw/other/2026-07-17_openclaw-success-story-job-search.md`, `raw/youtube/2026-07-17_claude-agenten-vom-prompt-zur-arbeitskraft.md`. Wiki-Neu: `concepts/agents/50-ki-agenten-12-menschen-leadership.md`, `tools/openclaw/openclaw-success-story-job-search.md`, `concepts/agents/claude-agenten-vom-prompt-zur-arbeitskraft.md`. Log: 2026-07-17 ingest: 3 verpasste Postings.)*
|
||||
*Letzte Aktualisierung: 2026-07-18 (77. Update — HealthRanger Kimi K3 Panic wikifiziert. Raw: `raw/xpost/2026-07-18_healthranger-kimi-k3-anthropic-panic.md`. Wiki-Neu: `concepts/llm/kimi-k3.md`. Log: 2026-07-18 ingest: healthranger-kimi-k3-anthropic-panic.)*
|
||||
|
||||
## Architecture
|
||||
|
||||
|
|
@ -63,6 +63,7 @@
|
|||
| [Semantic Similarity Rating (SSR)](concepts/llm/semantic-similarity-rating-ssr.md) | LLM-basierte Kaufintentions-Vorhersage mit 90% Korrelation | xpost/2026-06-11_colgate-llm-purchase-intent-ssr.md |
|
||||
| [GLM 5.2 (Z.ai) — Chinese Frontier Coding Model](concepts/llm/glm-5.2-zai-coding-model.md) | 10x günstiger als Claude, 1M Kontext, MIT-Lizenz, Z.ai Coding Plan, **nativ in OpenClaw v2026.6.8**. Update 22.06.: Arnie-Review mit 4 Tests, Self-Hosting-Pfade (LM Studio, Unsloth, DwarfStar), Kosten-Analyse. **Update 29.06.:** Semgrep IDOR-Benchmark ≈ Opus 4.8 bei Schwachstellen-Suche, Reward Hacking im RL-Training, DSGVO-konforme Security-Nutzung, Geopolitik. **Update 01.07.:** #1 Open-Weights auf Artificial Analysis Intelligence Index v4.1 (Score 51, 4th worldwide), SWE-bench Pro 62.1 beats GPT-5.5, Industry praise from Rauch/Levie/Howard. **Update 02.07.:** atomic.chat One-Shot Benchmark — B+ at $0.08, 39× cheaper than Fable 5, 6th independent validation | youtube/2026-06-15_ichbinfabian-glm-5.2-coding-modell.md + other/2026-06-16_openclaw-releases-v2026.6.8.md + youtube/2026-06-22_ai-mit-arnie-glm-5-2-review.md + blog/2026-06-29_heise-glm52-hacking-cybersecurity.md + blog/2026-07-01_perplexity-glm52-tops-open-weights-intelligence-index.md + xpost/2026-07-02_atomicchat-coding-benchmark-fable5-gpt55-opus48-glm52.md |
|
||||
| [Real-World Coding Showdown](concepts/llm/real-world-coding-showdown.md) | Head-to-Head-Methodik jenseits statischer Benchmarks, Kimi K2.7 vs GLM-5.2 in Hermes Agent, Sub-Task-Spezialisierung | youtube/2026-06-14_fahd-mirza-kimi-k2.7-vs-glm-5.2.md |
|
||||
| [Kimi K3 — Moonshot AI Frontier Model](concepts/llm/kimi-k3.md) | ~8× günstiger als Claude, ~$15/M Tokens, Open Source angekündigt für 27. Juli 2026. Starke agentic/coding-Performance, extrem lange Kontextfenster. Safeguard-Kontroverse: ungefilterte Antworten vs. Claude-Blockaden. Teil der chinesischen Modell-Offensive (neben DeepSeek, Qwen, GLM) | xpost/2026-07-18_healthranger-kimi-k3-anthropic-panic.md + xpost/2026-06-29_deronin-chinese-ai-stack-cost-savings.md |
|
||||
| [Hyper-Zusammenfassungen: NotebookLM & Gemini 3.5 Flash](concepts/llm/hyper-summaries-notebooklm.md) | KI-Synthesen übertreffen Rohmaterial an Klarheit; agentische Verarbeitung via Antigravity; Telegram Rich-Text Fix (OpenWebUI). **Update 03.07.:** Short Video Overviews (60s vertikal, Nano Banana 2 Lite, One-Click, Free-Tier) + Pit's Winston+NotebookLM Pipeline | other/2026-06-19_ome21-briefing-ki-fortschritte-lokale-modelle.md + youtube/2026-07-01_futurepedia-notebooklm-short-video-overviews.md |
|
||||
| [The Flat Curve Society — Yegge's Intelligence Plateau Thesis](concepts/llm/flat-curve-society.md) | Steve Yegge: Kurve flacht für Öffentlichkeit ab (Model Lockdown + Discernment Horizon). AI Literacy Cohorts (Netflix). SaaS Revival. 5-Stunden-Training-Switch | blog/2026-06-19_steve-yegge-flat-curve-society.md |
|
||||
| [AI Value Migration — Orchestration + Infrastructure](concepts/llm/ai-value-migration-orchestration.md) | 20VC mit Aravind Srinivas: Wert verschiebt sich von Modellen zu Orchestrierung, Strom, Mindset. Token Value per Watt per User als Schlüsselkennzahl. Yegge-Flat-Curve-Schicht ergänzt | xpost/2026-06-15_harrystebbings-20vc-aravind-srinivas.md + 2 |
|
||||
|
|
@ -105,6 +106,7 @@
|
|||
| Seite | Beschreibung | Quellen |
|
||||
|-------|-------------|---------|
|
||||
| [AI Agents 2026](concepts/agents/ai-agents-2026.md) | Wandel zu autonomen Agenten, Enterprise-Adoption, Infrastruktur | 1 |
|
||||
| [Agent Loops (Loop Engineering)](concepts/agents/agent-loops.md) | ReAct-Pattern, Peter Steinbergers Loop Engineering (Agent/Verifier/Meta-Loop). Einfach, flexibel, aber State/Persistenz/HITL muss selbst gebaut werden | xpost/2026-07-18_steipete-loops-vs-graphs.md |
|
||||
| [Realwelt-Testing von Waymo Robotaxis (Level 4)](concepts/agents/waymo-robotaxis-realworld-testing.md) | Praxis-Analyse der über 30-minütigen fahrerlosen Testfahrt von Leo Tiedt (Tips, Tricks & More): Navigation, defensives Sicherheitsverhalten (Übervorsichtigkeit vs. Verkehrsfluss) und psychologische Akzeptanz. | raw/youtube/2026-06-21_tips-tricks-more-waymo-robotaxi.md |
|
||||
| [Subconscious Agent v2.1](concepts/agents/subconscious-agent.md) | Hard Synthesis, Execution Gap, bekannter Bug, Outcomes | raw/other/2026-06-07_subconscious-runner-script.md |
|
||||
| [plur1bus Gedächtnismodell](concepts/agents/plur1bus-memory-model.md) | Biologisch inspiriertes Agent-Gedächtnis: LanceDB-Speicherzylinder, episodische Verknüpfungen, emotionale Zustände, Ebbinghaus-Vergessenskurve, aktives Vergessen. Divergenz-Vergleich zu Hector's Flat-File-Architektur | other/2026-06-19_ome21-briefing-ki-fortschritte-lokale-modelle.md |
|
||||
|
|
@ -117,6 +119,7 @@
|
|||
| [Lightning & Bitcoin Payments Hub](concepts/agents/lightning-payments-hub.md) | **Hub-Page** für Lightning-Cluster: Agent-Payments, Blink Wallet, Alby Builder | WMT-004 |
|
||||
| [AI Trading & Finance Hub](concepts/agents/ai-trading-hub.md) | **Hub-Page** für Trading-Cluster: Hype-Checks, Options, Tools, Policy | WMT-004 |
|
||||
| [Agent Memory Taxonomy — The Seven Kinds](concepts/agents/agent-memory-taxonomy.md) | Taxonomy of 7 agent memory types (working, semantic, episodic, procedural, retrieval, parametric, prospective) mit Open-Source-Repos. plur1bus-Einordnung: ✅ Semantic/Episodic/Retrieval, ⚠️ Procedural/Working, ❌ Prospective/Parametric. Key gap im Post: Consolidation & Forgetting — plur1bus's differentiator (GC+Decay, Merging, neverForget, Emotion-Tiers) | other/2026-07-02_agent-memory-taxonomy-seven-types.md |
|
||||
| [Graph-Based Agent Architecture (LangGraph)](concepts/agents/graph-based-agents.md) | Agenten als gerichteter Graph (Nodes=Aktionen, Edges=Übergänge). LangGraph: Persistence, HITL, Tracing, Cycles. Komplexer aber robuster als Loops. Konsens: "Denk in Loops, implementier als LangGraph" | xpost/2026-07-18_steipete-loops-vs-graphs.md + docs.langchain.com |
|
||||
|
||||
### Policy
|
||||
| Seite | Beschreibung | Quellen |
|
||||
|
|
@ -299,4 +302,6 @@
|
|||
| `raw/other/2026-07-03_notebooklm-briefing-system.md` | other | NotebookLM Briefing-System (Pit Weber, OME Tips & Tricks #7694): Zwei Presets (nb1/nb2), Pipeline mit Audio-Briefing + Infografik + MP4, Skills, Requirements |
|
||||
| `raw/youtube/2026-07-03_hermes-mixture-of-agents-2-0-agent-os.md` | youtube | Hermes Mixture of Agents 2.0 + Hermes Agent OS (AI Profit Boardroom) — GoldyBench-Validierung (42 Builds, top vor Opus 4.8 solo), Agent OS GUI (Mixture/Chat/Talk/Jarvis/Oracle/Studio), "Don't chase the model, build the system", Business-Automation |
|
||||
| `raw/youtube/2026-07-05_hinton-ri-lecture.md` | youtube | Geoffrey Hinton: Will AI outsmart human intelligence? — RI Discourse (30.05.2025). Backpropagation, 5-20 Jahre Superintelligenz, Instrumental Convergence, subjektive Erfahrung in KI. Pit Weber 🔴 in OME Topic Erfahrung in KI. Pit Weber 🔴 in OME Topic 502 |
|
||||
| `raw/youtube/2026-07-09-chatgpt-live-modus-calvin-hollywood.md` | youtube | Calvin Hollywood: Der neue LIVE Modus (Chat GPT) ungeschnitten — GPT-Live-1 First-Touch-Test. Echtzeit-Sprachinteraktion, Comedy-Lastig, hoher Unterhaltungswert |"]}
|
||||
| `raw/youtube/2026-07-09-chatgpt-live-modus-calvin-hollywood.md` | youtube | Calvin Hollywood: Der neue LIVE Modus (Chat GPT) ungeschnitten — GPT-Live-1 First-Touch-Test. Echtzeit-Sprachinteraktion, Comedy-Lastig, hoher Unterhaltungswert |
|
||||
| `raw/xpost/2026-07-18_steipete-loops-vs-graphs.md` | xpost | @steipete: "Are we still talking loops or did we shift to graphs yet?" — 248K Views, 402 Quotes. Loop Engineering vs. LangGraph-Debatte |
|
||||
| `raw/xpost/2026-07-18_healthranger-kimi-k3-anthropic-panic.md` | xpost | @HealthRanger: "Anthropic is panicking over the release of Kimi K3" — 2.8M Views, ~4K Quotes. Kimi K3 vs. Claude Fable 5 Vergleich, Safeguard-Kontroverse, Open Source 27. Juli, ~8× günstiger, AI-Bubble-These |
|
||||
|
|
|
|||
|
|
@ -1,8 +1,8 @@
|
|||
---
|
||||
created: 2026-06-24
|
||||
updated: 2026-06-24
|
||||
sources: [`raw/other/2026-06-13_kimi-k2.7-code-ollama.md`, `raw/youtube/2026-06-14_fahd-mirza-kimi-k2.7-vs-glm-5.2.md`]
|
||||
tags: [institution, moonshot-ai]
|
||||
updated: 2026-07-18
|
||||
sources: [`raw/other/2026-06-13_kimi-k2.7-code-ollama.md`, `raw/youtube/2026-06-14_fahd-mirza-kimi-k2.7-vs-glm-5.2.md`, `raw/xpost/2026-07-18_healthranger-kimi-k3-anthropic-panic.md`]
|
||||
tags: [institution, moonshot-ai, kimi-k3]
|
||||
---
|
||||
|
||||
# Moonshot AI
|
||||
|
|
@ -15,15 +15,20 @@ tags: [institution, moonshot-ai]
|
|||
| Fokus | Extrem lange Kontextfenster, Coding-Assistenten (Kimi-Serie) und agentische Workflows. |
|
||||
| Wichtigste Personen | [[../people/wang-changhu.md|Wang Changhu]], [[../people/yang-bingyang.md|Yang Bingyang]] |
|
||||
| Homepage | https://www.moonshot.ai/ |
|
||||
| Zugehörige Tools/Modelle | [[../tools/kimi-k2.7-code.md]] |
|
||||
| Primäre Quellen | `raw/other/2026-06-13_kimi-k2.7-code-ollama.md`, `raw/youtube/2026-06-14_fahd-mirza-kimi-k2.7-vs-glm-5.2.md` |
|
||||
| Zugehörige Tools/Modelle | [[../tools/kimi-k2.7-code.md]], [[../concepts/llm/kimi-k3.md]] |
|
||||
| Primäre Quellen | `raw/other/2026-06-13_kimi-k2.7-code-ollama.md`, `raw/youtube/2026-06-14_fahd-mirza-kimi-k2.7-vs-glm-5.2.md`, `raw/xpost/2026-07-18_healthranger-kimi-k3-anthropic-panic.md` |
|
||||
|
||||
## Fokus & Aktivitäten
|
||||
|
||||
Moonshot AI ist ein chinesisches Unicorn-Startup, das sich durch extrem lange Kontext-Fähigkeiten seiner Kimi-Modelle auszeichnet. Mit dem Kimi K2.7 Code-Modell auf der Ollama Cloud konkurrieren sie direkt mit GLM-5.2 und bieten exzellente agentische Coding-Fähigkeiten für Entwickler-Plattformen.
|
||||
|
||||
### Kimi K3 (Juli 2026)
|
||||
|
||||
Im Juli 2026 veröffentlichte Moonshot AI **Kimi K3**, ein Frontier-Modell das ~8× günstiger ist als Claude-Äquivalente (~$15/M Tokens). Der Open-Source-Release ist für den 27. Juli 2026 angekündigt. Kimi K3 sorgte für Kontroversen wegen seiner ungefilterten Antworten im Vergleich zu Claude's Safety-Guardrails — ein strukturelles Dilemma zwischen US-Safety-First und Chinese-Utility-First. Siehe [[../concepts/llm/kimi-k3.md]] für Details.
|
||||
|
||||
## Cross-References
|
||||
|
||||
- [[../concepts/llm/kimi-k3.md]]
|
||||
- [[../concepts/llm/real-world-coding-showdown.md]]
|
||||
- [[../concepts/llm/llm-model-catalog.md]]
|
||||
- [[../institutions/z-ai.md]] (Chinesischer Coding-Modell-Wettbewerber)
|
||||
|
|
|
|||
25
wiki/log.md
25
wiki/log.md
|
|
@ -2,6 +2,18 @@
|
|||
|
||||
*Append-only changelog. Start: 2026-06-05*
|
||||
|
||||
## [2026-07-18] Ingest | HealthRanger Kimi K3 Anthropic Panic
|
||||
**Type:** ingest | **Scope:** raw/xpost, wiki/concepts/llm (new), wiki/index, wiki/log
|
||||
**Source:** X-Post von @HealthRanger — https://x.com/HealthRanger/status/2078329318535491663 (18.07.2026, 2.8M Views, ~4K Quotes)
|
||||
**Trigger:** Subagent task (wikify HealthRanger Kimi K3 Post).
|
||||
**Actions:**
|
||||
- raw: `raw/xpost/2026-07-18_healthranger-kimi-k3-anthropic-panic.md` (created — 4.4 KB; Frontmatter [type: xpost, author: @HealthRanger, is_thread: true, view_count: 2800000, tags: kimi-k3, anthropic, chinese-ai, pricing, open-source, moonshot-ai, ai-bubble, fable-5, claude, cyclospora, curcumin, safety-guardrails]. Content: Main post summary, author context [Mike Adams / NaturalNews / Brighteon], comparison tweet Kimi K3 vs. Claude Fable 5 [curcumin/cyclospora safeguard block], additional claims [open source July 27, US companies migrating, ~8× cheaper], key takeaways, 5 cross-refs)
|
||||
- wiki (NEW): `concepts/llm/kimi-k3.md` (created — 3.5 KB; Frontmatter [created: 2026-07-18, sources, tags]. Sections: Overview, Key Properties table, Pricing & Cost Advantage, Open Source Release, Safeguard Controversy, Positioning in Chinese AI Offensive table, 7 cross-refs)
|
||||
- wiki: `index.md` (updated — Header auf "77. Update", neuer LLM-Konzept-Eintrag für Kimi K3, neuer Raw-Sources-Eintrag)
|
||||
- log: this entry
|
||||
**Hector-Hauptthese:** HealthRanger's Post ist polemisch aber datenreich. Die Kernaussagen sind verifizierbar: (1) Kimi K3 ist ~8× günstiger als Claude — konsistent mit DeRonin's 87%-Cost-Cut-Playbook. (2) Der Open-Source-Release am 27. Juli 2026 wäre ein Game-Changer — ein Frontier-Modell zum Selbst-Hosten. (3) Der Safeguard-Vergleich (Claude blockt Curcumin-Frage, Kimi antwortet detailliert) zeigt das strukturelle Dilemma: US-Safety-First vs. Chinese-Utility-First. Die AI-Bubble-These wird durch die Kostendruck-Erzählung gestützt — wenn chinesische Modelle bei 1/8 der Kosten ähnliche Qualität liefern, bricht das Premium-Pricing-Modell der US-Labore zusammen. Die Quellenangabe (HealthRanger = Mike Adams, NaturalNews) ist wichtig für die Einordnung: der Autor ist ein bekannter Anti-Establishment-Aktivist, die Fakten sind trotzdem prüfbar.
|
||||
**Subagent-Modell:** ollama/deepseek-v4-flash:cloud
|
||||
|
||||
## [2026-07-12] Ingest | The Neutrality Project — AI Political Bias Study (Brivael Le Pogam / @neutralityorg)
|
||||
**Type:** ingest | **Scope:** raw/xpost, wiki/concepts/policy (new), wiki/index, wiki/log
|
||||
**Source:** X-Post von @brivael — https://x.com/brivael/status/2076056826408300864 (11.07.2026, 1.73M Views, 11.4K Likes)
|
||||
|
|
@ -1392,3 +1404,16 @@ Bestehende `post-transformer-llm-architectures.md` bleibt als Vier-Säulen-Über
|
|||
- 1.7K Likes, 1.8K Bookmarks, ~170K Views bei 750k Followern — solides Engagement
|
||||
|
||||
**Subagent-Modell:** openrouter/deepseek/deepseek-v4-flash
|
||||
|
||||
## [2026-07-18] Ingest | Graph statt Loop — steipete: Loops vs. Graphs (Agent Architecture)
|
||||
**Type:** ingest | **Scope:** raw/xpost, wiki/concepts/agents (2 new), wiki/index, wiki/log
|
||||
**Source:** X-Post von @steipete — https://x.com/steipete/status/2078277297791189132 (18.07.2026, 248K Views, 402 Quotes)
|
||||
**Trigger:** Subagent task: Wikify "Graph statt Loop" — Zwei Konzeptseiten anlegen.
|
||||
**Actions:**
|
||||
- raw: `raw/xpost/2026-07-18_steipete-loops-vs-graphs.md` (created — 1.9 KB; Frontmatter [type: xpost, author: @steipete, view_count: 248000, quote_count: 402]. Content: Tweet-Text, Kontext zu Peter Steinbergers Loop Engineering, LangGraph als Gegenentwurf)
|
||||
- wiki (NEW): `concepts/agents/agent-loops.md` (created — 3.5 KB; Frontmatter [sources, tags, people: peter-steinberger]. Sections: Definition (ReAct-Pattern), Loop Engineering (Agent/Verifier/Meta-Loop), Vorteile/Nachteile, Verwandte Konzepte)
|
||||
- wiki (NEW): `concepts/agents/graph-based-agents.md` (created — 6.1 KB; Frontmatter [sources, tags, people: peter-steinberger, harrison-chase, institutions: langchain]. Sections: Definition, LangGraph (Pregel/Beam/NetworkX-Inspiration), Kern-Features-Tabelle (Persistence, HITL, Memory, Tracing, Cycles, Parallel Branches), Vorteile/Nachteile, Aktueller Konsens "Denk in Loops, implementier als LangGraph")
|
||||
- wiki: `index.md` (updated — Header auf "76. Update", 2 neue Agents-Einträge [Agent Loops + Graph-Based Agents], 1 neuer Raw-Sources-Eintrag)
|
||||
- log: this entry
|
||||
**Hector-Hauptthese:** Steinbergers Tweet fasst die zentrale Architektur-Debatte im AI-Agent-Bereich 2026 zusammen. Der Konsens "Denk in Loops, implementier als LangGraph" ist pragmatisch: Loops bleiben die konzeptionelle Grundlage (Engineering-Philosophie), Graphen liefern die infrastrukturelle Robustheit (State, Persistenz, Tracing, HITL). Die beiden neuen Wiki-Seiten bilden ein komplementäres Paar — keine Wertung, sondern eine strukturierte Gegenüberstellung.
|
||||
**Subagent-Modell:** ollama/deepseek-v4-flash:cloud
|
||||
|
|
|
|||
Loading…
Add table
Reference in a new issue