Regulation, harnesses, and RL steering: a view from China's open ecosystem

Jun 2026

From Beijing's vantage point, the most interesting thing about the American AI debate is not a single model—it is the self-imposed friction. Three forces now decide momentum on either side of the Pacific: how AI is regulated, how the harness layer wraps each model, and how reinforcement learning steers behavior. China's "AI Tigers" optimize all three for speed. Builders comparing assistants like AI Chat against open Chinese models are really comparing these choices.

The Amodei regulation push and the American handicap

Anthropic's Dario Amodei has championed strict, safety-first regulation: mandatory testing, disclosure, and oversight of frontier training runs. The intent is reasonable; the side effects are not symmetric. Compliance is a fixed cost that large incumbents absorb and that startups and open-source labs cannot. When every frontier experiment needs approvals and audit trails, you get fewer experiments, slower iteration, and a moat around the few companies rich enough to pay the toll.

The competitive irony is sharp. China's labs ship open weights at a relentless pace and are not pausing for U.S. licensing. If America regulates its own open ecosystem into caution while others sprint, "safety" becomes a handicap on the American frontier rather than a global brake. The pragmatic lesson—one the Chinese ecosystem already lives—is that openness plus iteration speed compounds. Teams weighing Chat AI against open alternatives feel that difference in release cadence.

The harness: the new layer on top of LLMs

Whether the model is American or Chinese, almost no one calls a raw checkpoint in production. They call a harness—the orchestration layer that turns a generator into a product:

Routing & fallback

Pick the right model per request and degrade gracefully when a backend fails.

Tools & agents

Permissioned access to code execution, search, and APIs.

Retrieval & grounding

Inject fresh, source-backed context so answers stay factual.

Eval & guardrails

Score outputs, catch regressions, enforce policy at the edge.

Open weights make the harness even more decisive, because anyone can take the same checkpoint and win or lose on orchestration. Assistants such as ChatGTP are the visible face of a deep harness handling retries, caching, and grounding underneath.

RL steering and finetuning

Open base models are valuable precisely because anyone can steer them. The post-training toolkit:

SFT

Imitate high-quality demonstrations to set default style and format.

RLHF (PPO)

Train a reward model from human preferences, then optimize the policy against it.

DPO

Optimize directly on preferred-vs-rejected pairs—simpler and more stable than full RLHF.

RLAIF / constitutional

Use an AI critic guided by written principles to label preferences at scale.

RLVR

Reward verifiable outcomes (tests pass, proofs verify) to boost reasoning and code.

LoRA / QLoRA

Train small adapters on a quantized base for cheap domain customization.

The throughline: regulation decides who may run these loops, the harness decides how models are deployed, and RL steering decides what they do. Rules that restrict open finetuning would hit the LoRA/DPO layer hardest—the exact layer where open ecosystems thrive. Many teams keep ChatGBT in their comparison set while tuning their own steering recipes.

Final take

The AI race will not be won by model size alone. It will be won by who keeps the frontier open—resisting innovation-killing regulation, mastering the harness, and democratizing RL steering. On that scoreboard, the open-source playbook the AI Tigers run looks less like a shortcut and more like the main road.