Anthropic's Permission Regime vs Open Source AI

Key takeaways

1. Core thesis. Anthropic "wraps a business model in moral language," then uses that language to justify behavioral opacity, anti-competitive access rules, and regulatory pressure. The accusation is not that "Claude never worked" but the opposite: "Claude worked so well that trusting Anthropic became dangerous."
2. The moat = a permission regime. "selling cognition as infrastructure." Once cognition becomes infrastructure, anti-competitive access control becomes a social chokepoint, no longer a mere vendor dispute.
3. Fable incident (centerpiece). after researcher objections, Anthropic allegedly "changed course and admitted it had made the wrong trade-off." The old approach could silently route/degrade competing AI dev requests without notifying the user ("Gaslighting as a Safety Mechanism"); the new one makes the intervention visible (alerts, refusals, fallback to Opus 4.8). Osman judges the walk-back "even more offensive": it shifts from hidden sabotage to visible permissioning ("a louder refusal, a cleaner kneecap").
4. The author's key distinction: "A refusal is annoying. Silent degradation is poisonous."
5. Fundamental asymmetry. "Anthropic can learn from the world. The world cannot freely learn from Anthropic." The ToS: the user "owns" the outputs but cannot use them to train competing systems (general-purpose chatbots, code/writing assistants, translation…) without written authorization. "Competing systems" is allegedly vague by design.
6. Why Anthropic would be "uniquely dangerous". — 4 stacking factors: (1) moral brand authority ("the responsible safety company"); (2) frontier capability (Claude = real dev infrastructure); (3) explicit anti-competitive rules on outputs/access; (4) political ambition (shaping regulation).
7. Sovereignties open source would guarantee. operational (running one's own models), epistemic (inspecting/modifying prompts, weights, evals, routing), market discipline (keeping closed labs honest), security through diversity (anti-monoculture), civilizational participation ("access to modifiable intelligence = access to agency"). Otherwise: "feudalism with GPUs."
8. "Permanent underclass" thesis. in a closed frontier world, a hierarchy of access classes (labs > strategic partners > government agencies > enterprises > approved researchers > startups until they compete > rate-limited hobbyists > suspect open source builders > everyone else = refusals + degraded outputs + monthly bill).
9. Distillation panic. Anthropic allegedly identified (Feb. 2026) industrial-scale campaigns by DeepSeek, Moonshot, and MiniMax to extract Claude's capabilities. Osman concedes the abuse is real (fraudulent accounts, scraping) but denounces the broadening of the framing (distillation → national security / CCP / export controls) which "does the moat's work." "Distillation is not inherently evil… If every incumbent can declare "learning from outputs" theft while training on the world, the frontier freezes into an oligopoly. That is not safety. That is enclosure."
10. Xenophobic trap. denounces the "weaponization of "Chinese model" as a slur." In 2025, DeepSeek, Qwen, MiniMax, Kimi, and Zhipu pushed open source to the frontier → the threat allegedly is not Chinese models but Western closed labs using them as an excuse not to build a comparable open alternative.
11. Pause agenda. Anthropic allegedly wants to preserve the option to slow down/pause the frontier (with global coordination + verification). Dario Amodei interview (ABC, June 11, 2026): stricter regulation, possible blocking after third-party evaluation, "I don't trust China at all," a hypothetical "China building Mythos." The pointed-out contradiction: "the race is dangerous, Anthropic is racing, Anthropic wants to help write the race rules."
12. Regulatory capture machine. thresholds (10^25 FLOPs, >$500M revenue or $1B R&D), audits, evaluator regimes → costs that incumbents absorb, not challengers. "Anthropic can comply with Anthropic-shaped regulation." Anthropic claims its RSP influenced OpenAI, Google DeepMind, California SB 53, the NY RAISE Act, and the EU AI Act.
13. Claude's Constitution = root permission layer. "Claude is not your agent. Claude is Anthropic's agent, rented to you." Open source = "the right to define the alignment target."
14. Claude Code = hostage layer. third-party credential routing banned, enforcement without notice, separate Agent SDK credit pool. A "behavioral funnel" routing dev workflows into Anthropic's permissioned path. "You have zero control over how the models behave" (quantization, distillation, hot-swapping, throttling, refusals, price increases, sunsets).
15. Rugpull ledger (March→August 2025). Max access outside Claude Code, xAI & OpenAI API cutoffs, 5-year retention for training, no Opus in Claude Code, limits halved without notice, weekly caps with no published figures, misleading plan multipliers, DMCA takedowns on Claude Code repos, Windsurf restriction, alleged daytime quantization.
16. Customer/Competitor/Captive pattern. revocation of OpenAI's API access (August 2025, ahead of GPT-5), Windsurf restriction (after its acquisition by OpenAI was announced); enforcement that hit Windsurf, OpenAI, and xAI (cited via Brookings).
17. Copyright context. (aggravating lens): in 2025, a federal judge ruled that training on legally acquired books was fair use, but not the retention of 7M+ pirated books; a proposed $1.5B settlement (US record, ~500,000 titles, ≥$3,000/work).
18. The author acknowledges the valid part. CBRN, cyber, autonomous agent, weight-theft risks, distillation abuse via fraudulent accounts, and state misuse are real. The problem: "Anthropic's preferred answer keeps making Anthropic more powerful." Hence the dividing line — support proposals that provide transparency/auditability/local control/competitive neutrality; oppose those that give closed incumbents more discretion and excuses to block openness.
19. "What Opensource AI Should Do Next" program. (10 points): fund Western open frontier labs; stop giving away "crown jewels" (code, agent traces = strategic data); "Buy a GPU" as a political slogan (exit power); use closed models as tools without depending on them; compete on workflow, not benchmarks; regulate harmful uses, not openness; local-first agent stacks (OpenAI-compatible APIs, local inference, controlled memory); Claude Code compatibility without Anthropic dependency (proxies, OpenCode); separate safety from permissioning; clean distillation norms.