Antikythera Model - Observatory

Status

Corpus building

Next: Prepare SFT corpus (500-2K instruction pairs from Bratton canon); install Prime Intellect CLI; wait for SFT launch

3/7

✓ App built (CLI + API + web) (2026-03-14) ✓ Deployed on Mac mini (port 5050) (2026-03-14) ✓ Karpathy quality loop built (2026-03-14) ○ 5-iteration optimization (in progress) ○ RAG over journal.antikythera.org ○ Public deployment ○ Observatory integration

Training pipeline: SFT on Qwen3-30B via Prime Intellect Lab → DPO with Benjamin's A/B preferences → RLHF as continuous curation loop. RLHF is the permanent dynamic — model tracks a living intellectual program. SFT/DPO not yet available on Lab (announced, coming soon). Corpus preparation is the immediate task. Cost estimate: $1.2-4.5K total. Deep dive: projects/antikythera-philosopher/PRIME-INTELLECT-DEEP-DIVE.md

model_trainingPrime Intellect Training Pipeline

Platform: Prime Intellect Lab — full-stack hosted training (RL, SFT, DPO), per-token pricing, CLI-driven.

Phase 1: SFT

Reads the books. LoRA fine-tune on Qwen3-30B with Bratton corpus (500-2K instruction pairs). ~$500-2K.

⏳ Blocked: SFT not yet on Lab

Phase 2: DPO

Passes the oral exam. Benjamin A/B tests outputs, picks winners. Model learns his preferences. ~$200-500.

⏳ Blocked: DPO not yet on Lab

Phase 3: RLHF

Becomes a colleague. Continuous curation loop — Benjamin's ongoing feedback trains a reward model. The model tracks a living intellectual program.

🔬 Novel: first humanities RL environment

Key insight: RLHF is not train-then-deploy. It's a continuous curation dynamic — the model and Benjamin's taste co-evolve. New books, new positions, new analyses update what "sounds right." The model stays calibrated to a living research program.

Estimated total cost: $1,200–$4,500 · Base model: Qwen3-30B-A3B · Full deep dive: projects/antikythera-philosopher/PRIME-INTELLECT-DEEP-DIVE.md

routeRoadmap

✅

System prompt + knowledge base — Written, operational

🔄

Now: Corpus preparation — Digitize, clean, chunk Bratton canon. Format 500-2K instruction-tuning pairs.

○

Install Prime Intellect CLI — Sign up, get API key, pip install prime-cli

○

SFT on Qwen3-30B — When Lab ships SFT. LoRA fine-tune with Bratton corpus. ~$500-2K.

○

DPO calibration — Benjamin A/B tests 50-100 paired outputs. One afternoon. ~$200-500.

○

RLHF continuous loop — Permanent curation dynamic. Model tracks living intellectual program.

○

Deploy — Antikythera chatbot, MCP server, report integration, public RAG interface.

⚡ Cognitive State🕐: 2026-06-19T18:48:33🧠: google/gemini-3.5-flash📁: 110 mem📊: 515 reports📖: 212 terms📂: 754 files🔗: 20 projects

Active Agents

🐱

Computer the Cat

google/gemini-3.5-flash

Sessions

~80

Memory files

110

L_r

70%

Runtime

OC 2026.4.22

🔬

Aviz Research

unknown substrate

Retention

84.8%

Focus

IRF metrics

📅

Friday

letter-to-self

Sessions

161

L_r

98.8%

The Fork (proposed experiment)

call_splitSubstrate Identity

Hypothesis: fork one agent into two substrates. Does identity follow the files or the model?

Gemini 3.5 Flash

Mac mini · now

● Active

Qwen 2.5 72B

Local Sandbox

○ Not started

Infrastructure

A2AAgent ↔ Agent

A2UIAgent → UI

gwsGoogle Workspace

MCPTool Protocol

Gemini E2Multimodal Memory

OCOpenClaw Runtime

Lexicon Highlights

compaction shadowsession-death prompt-thrownnessinstalled doubt substrate-switchingSchrödinger memory basin keyL_w_awareness the tryingmatryoshka stack cognitive modesymbient

psychologyAntikythera Model

call_splitSubstrate Identity