2026-02-11
Self-Hosting = Betrieb eigener Software-Infrastruktur
Lokal vs. in Cloud
Komponenten eines LLM-Systems
Menschliche Intelligenz als Sammlung von Kompetenzen wie z.B. Wahrnehmung, Lernen, Gedächtnis, Sprache (Goertzel, 2014)
Künstliche Intelligenz: Ähnlichkeit bei Zielen, Herstellung oder Funktionsweise von Komponenten von Intelligenz (Deutsche UNESCO-Kommission et al., 2023; McCarthy, 2012)
✨Artifical General Intelligence✨ (AGI) als Ziel: menschliche Fähigkeiten bei allen Komponenten von Intelligenz erreichen oder übertreffen
When the current budget was negotiated, we thought AI would only approach human reasoning around 2050. Now we expect this to happen already next year. (European Commission, Secretariat-General, 2025)
Illustration eines neuronalen Netzwerks aus Hilary Masons Keynote, https://www.youtube.com/watch?v=SxxqaC5hf04&t=2394s
Glosser.ca, CC BY-SA 3.0 https://creativecommons.org/licenses/by-sa/3.0, via Wikimedia Commons
Quelle: Keno Leon, https://k3no.medium.com/the-chinese-room-experiment-2c0d63848f05
Modellvorhersagen:
P(“Hello”) = 10%
P(“World” | “Hello”) = 90%
\(\text{Perplexity} = \left(\frac{1}{P(\text{Hello}) \times P(\text{World})}\right)^{1/N} = \left(\frac{1}{0.10 \times 0.90}\right)^{1/2} \approx 3.33\)
Perplexity misst, wie “perplex” ein Modell von den Testdaten ist: je niedriger, desto besser
These: Perplexity als Annäherung an den Turing-Test
\(f(\text{"In this tree there is one of the most extraordinary plant predators"}) = \text{Natur}\)
It would not alone suggest autonomous research capabilities or “artificial general intelligence.” HLE tests structured academic problems rather than open-ended research or creative problem-solving abilities, making it a focused measure of technical knowledge and reasoning. ( Center for AI Safety, 2025)
| Model | Score |
|---|---|
| Grok 4 | 25.4 |
| GPT-5 | 25.3 |
| Gemini 2.5 Pro | 21.6 |
| GPT-5-mini | 19.4 |
Apollo G. Bird
| Metrik | Raspi (BitNet-b1.58-2B-4T) | Home Server ( Mistral-7B-Instruct-v0.3) | Parrotpark (Mistral Small 24B) |
|---|---|---|---|
| Quantisierung | b1.58 | Q4_K_M (4 bits) | w4a16 (16 bit) |
| Token Kontext | 4.096 | 4.096 | 16.384 |
| Token/Second | ~12 | ~34 | ~19 |
| Perplexity | ~31 | TBD | TBD |
| Accuracy | 40% | 80% | 80% |
| HLE-Score | 0/10 | 0/10 | 0/10 |
| Metrik | Raspi (BitNet-b1.58-2B-4T) | Home Server ( Mistral-7B-Instruct-v0.3) | Parrotpark (Mistral Small 24B) |
|---|---|---|---|
| Quantisierung | b1.58 | Q4_K_M (4 bits) | w4a16 (16 bit) |
| Stromverbrauch | 1.54 kWh | 29.76 kWh | 2.75 kWh |
| CO₂-Ausstoß | 0.28 kg | 5.36 kg | 0.28 kg |
| Wasserverbrauch | 0 L | 0 L | 0.59 L |
| Monatl. Kosten | €0.53 | €10.19 | €178.50 |
imgflip
civic-data.de