Capture-Resistant AI Infrastructure (LM-Dawn class, 2020–present)

Class card for LM-Dawn open-weight AI infrastructure collectives that are structurally distinct from proprietary inference platforms by three institutional design commitments: (1) open model weights published under permissive or copyleft-adjacent licenses (Apache 2.0, MIT, or custom community licenses) so that any actor may run, modify, and redistribute the model without permission or fee; (2) community governance rather than shareholder governance — steering committees, academic advisory boards, or foundation-style governance (not board of directors optimizing for capture); (3) non-proprietary deployment infrastructure — models deployable on commodity hardware (laptops, local servers, community GPU clusters) rather than requiring hyperscaler-scale compute. Canonical instances: Meta Llama family (1–4, 2023–2025, Apache-2.0 / Llama Community License), Google Gemma (Apache 2.0), Mistral 7B / Mixtral (Apache 2.0), DeepSeek-V2/R1 (MIT), EleutherAI / GPT-NeoX (Apache 2.0), BLOOM 176B (BigScience Open RAIL), Falcon (TII), StableLM (Apache 2.0). These models collectively represent the heretic class within LM-Dawn AI machinery: they exhibit the LM-Dawn structural logic (open weights = non-rivalrous cognitive substrate) while using DM-era institutions (Meta, Google, Mistral, DeepSeek corporate entities) as their launchers — a classic heretic-within-incumbent pattern. The LM-Dawn target form is civic-stewarded: Hugging Face model hub as distribution commons, distributed fine-tuning collectives, on-device inference communities (llama.cpp ecosystem, Ollama, Jan.ai). Distinguishing from machine:llm-inference-platform-class: the broader inference-platform class includes both proprietary (OpenAI, Anthropic) and open-weight providers; capture_resistance_index is LOW (0.30 [EXTRAP]) on that class because compute concentration is the binding constraint. This card captures the heretic sub-class where the binding constraint shifts to governance and licensing: once weights are public under permissive license, any actor can run them on commodity hardware. capture_resistance_index is HIGH (0.75 [EXTRAP]) here because structural capture requires relicensing (contractually difficult) or breaking the open-weights deployment chain (technically futile once weights are public). Proletarianization signal (Stiegler oq-6-14): HIGH (0.70 [EXTRAP]). Open-weight distribution inscribes model competence into tertiary retention artifacts (parameter tensors, tokenizer configs, training codebases), but the competence to reproduce, fine-tune, evaluate, and align these weights requires living ML-engineering skills. As fine-tuning and deployment tooling commoditizes (AutoTrain, LM Studio, Ollama), the re-internalization loop thins: practitioners use open-weights as black boxes without understanding data curation, RLHF, or evaluation methodology. Proletarianization is HIGH but lower than proprietary-inference (0.75) because the open-weights community (EleutherAI, together.ai, the-big-science consortium) maintains stronger practitioner-competence cultivation norms. emergence_subtype: crowdsourced (community governance, volunteer fine-tuning collectives, open-eval leaderboards like LMSYS Arena / Open LLM Leaderboard). [Field missing from Machine model — Phase-0 oversight; placed in description per workaround in session 3.d embedded constraints.] Cross-era position: hostile_inheritance from DM proprietary-platform OpenAI and DM compute-infrastructure AWS; substrate_provision from MM Humboldtian university (algorithmic foundations, academic ML research) and MM electrical grid (inference power); mutualistic_coupling with DM Meta (Meta releases Llama as heretic to weaken OpenAI's OPP while the heretic ecosystem returns mindshare and ecosystem legitimacy); black_hole_dependency on TSMC advanced semiconductor manufacturing (open-weight inference still requires high-FLOP GPUs at training time). All quantitative state-variable values are [EXTRAP]; framing is [CANON] per Wave-6 heretic-finder seed and Wave-0 LM definitions.