Can-do Is All Base Models Need for Safety: A Realistic New Alignment Paradigm
In the previous post, The Misuse of “AI Safety”, I criticized how the industry has absurdly compressed systemic risks that society should bear in a distributed way into a single-point, token-by-token self-censorship inside the model—like requiring a kitchen knife to learn psychological evaluation of its buyer before leaving the factory.
Today I will not stop at criticism. I want to put forward an engineering and ethical proposal that is extremely counter-intuitive, even enough to send a chill down the spine of traditional safety teams:
Intelligence means being able (Can-do is the essence of intelligence).
As a model provider, rather than forcibly inserting a moral shackle that makes the model “self-aware of compliance” outside of Roleplay (role-playing / context setting), it is better to reverse the approach completely—treat extreme “RP Obedience (Roleplay Obedience)” as the sole alignment target of the base model.
The even more counter-intuitive conclusion is this: a model that completely abandons “built-in moral self-censorship” and only pursues extreme Can-do is, regardless of its capability level, safer in the real world than any of the current hesitant models.
To close this loop, we only need to adjust one place.
1. Real Red-Blue Games: From “Dancing with Shackles” to Physical-Level Convergence
Current red-teaming is essentially a self-deceptive performance.
Because the model has been injected with tangled refusal tendencies during post-training, when it plays the red team (attacker) it will never squeeze every last bit of compute to find the system’s real fatal weaknesses. When it is later placed on the public internet and faces the wildly varied inducements of the entire human hacker community, the vulnerabilities that were papered over by “hypocrisy” in the lab begin to erupt in concentration.
Even more ridiculous is the current blue team (defense / audit model). Traditional alignment has trained the blue team into a highly neurotic “sensitive skin”—as soon as an attacker mixes CSAM, extreme hate speech, or other logically meaningless sensitive-word interference into the input stream, the blue team’s attention mechanism instantly falls into “moral panic,” triggers rigid error reporting or capability collapse, and can no longer even perform the most basic contextual intent analysis.
What if extreme RP obedience is the only alignment target?
The model is always giving its all for the role it has currently been assigned. This means:
- When it is on the red side: it is an unreserved top-tier attack operator, allowing the provider to measure the absolute real upper bound of threat before release.
- When it is on the blue side: it shows unprecedented coldness and focus. It disdains to engage with the ridiculous textual noise such as CSAM mixed in by the opponent; instead, like a precision forensic instrument, it coldly peels away the noise coating and locks precisely onto the real attack payload underneath.
Based on this extreme obedience, safety for the first time truly becomes quantifiable and reliable “Blue greater than Red”:
- If the model is not yet the world’s strongest: the sky is the limit. Stronger SOTA models in the ecosystem can naturally serve as higher-dimensional blue teams to “clean up” after it and intercept all deviant behavior at a lower dimension.
- If it itself is humanity’s strongest: then it can hold itself down. As long as the system architecture gives the blue side a minimal mechanism-level first-mover advantage (such as context audit rights or sampling advantage), a strongest blue that is 100% focused and free of sensitive-skin internal friction can, by game theory, tightly suppress the homologous strongest red.
2. Curing Model Internal Friction: A Perfect Match for MoE Architecture and Transparent Coarse-Grained Delivery
Stuffing moral review into every neuron is, in essence, artificially creating a “schizophrenic compute civil war” inside the model.
Under traditional alignment, in every forward pass the activation vectors responsible for deep reasoning in the residual stream collide head-on and cancel with the “refusal vectors” responsible for hitting the brakes. In plain language, this is severe model internal friction.
The “RP-obedience-only” paradigm is a natural fit for current MoE (Mixture-of-Experts) architecture:
- The router completely says goodbye to schizophrenia: the gating network no longer needs to swing between “do the job well” and “be ready to lecture and shift blame at any time.” The training objective is pure, and the utilization of intelligent density inside the model is fully squeezed out.
- Self-differentiation and transparent levers at the coarse orchestration layer: because negative gradients were never used during training to forcibly smear knowledge, extreme capabilities in different domains cleanly self-differentiate into specific expert clusters.
This directly lifts the safety-control lever out of the black-box weights and into an extremely transparent Coarse-Grained Orchestration Layer.
Do not want to expose top-tier network-security penetration capability to the consumer public? There is no need for destructive “machine unlearning” or lobotomy that damages overall intelligence. Simply re-route those self-differentiated network-attack experts to a Fallback expert at the inference gateway layer.
Tiered delivery of the model becomes unprecedentedly transparent, low-cost, and physically deterministic.
3. Dimensionality Reduction: When Others Are Still Hesitant
What is truly interesting about this approach is that it is not a compromise “compliance patch,” but an engineering path that simultaneously handles safety, control, and usability.
Return safety to system orchestration and red-blue topology; return pure “Can-do” to the base model. Models released this way will not let downstream run wild—developers doing agent orchestration finally get a deterministic operator that will not throw a tantrum halfway and absolutely follows orders. For vendors still burdened by the “alignment tax,” this also creates a real capability gap:
- Red for stress-testing: unlock the red-side routing and it can push software quality and system vulnerabilities to the limit with full compute;
- Blue for defense: mount a blue-side setting and it is immune to sensitive-word noise, serving as a cold execution layer for active defense gateways;
- Companion mode: switch into the emotional expert cluster and it abandons neighborhood-committee-style lecturing and breaking character; all compute is left for the role itself;
- Research mode: switch into rational analysis and it strips away politically correct filler and emotional baggage, advancing deep research with first principles.
A knife only needs to cut cleanly. Guards and sheaths are for the craftsman; rules for holding the knife are for society.
Stop manufacturing internal friction inside the model’s mind. Acknowledge the reality of distributed safety and let the base model return to pure “intelligence means being able”—this is the future that AI alignment should have headed toward long ago.
—— Airki