Yang Zhilin Predicts Long-Context AI Will Replace Fine-Tuning

Yang Zhilin’s most useful claim for marketers is also the one that should make marketing-ops teams sit up straight: “Fine-tuning may not exist in the long term.” In the ChinaTalk interview, the Moonshot AI founder argues that personalization will come less from training a model on a user’s data and more from putting the user’s full interaction history directly into the model’s context, producing what he calls “unreplicable, direct dialogue.”[1]

That is not a generic AI optimism point. It is a specific bet about where personalization work moves. If Yang is right, the operating center of gravity shifts from “How do we fine-tune a model on customer records?” to “Can we assemble, permission, clean, and pass the right customer history into context at the moment of use?” For anyone reading Yang Zhilin’s Moonshot AI founder interviews through a marketing AI lens, that is the practical question.

Two caveats matter before turning the quote into a roadmap. First, this is Yang’s forecast, not a settled market conclusion. Second, most of the interview material circulating in English is translated or summarized from Chinese-language conversations, so the safest reading is to stay close to the exact claim: long-context systems may reduce the need for fine-tuning as the default path to personalization. That is already a big enough idea without inflating it.

One more housekeeping point prevents a common search-result trap: Yang Zhilin’s Moonshot AI is the company behind Kimi. It is not the marketing conversion-rate-optimization company using the moonshot-ai.com domain. The distinction matters because this article is about Yang’s technical argument for long-context AI, not a vendor pitch for a marketing platform.

The Claim Is About Context, Not Just Bigger Models

Yang’s prediction depends on a particular view of context windows. In a LinkedIn interview, he describes a “Moore’s Law for context length,” where context windows grow exponentially, and he suggests that this could make current fine-tuning approaches obsolete within “2-3 years.”[2] That timeline should not be treated as a procurement calendar. It is a founder’s technical forecast. But it does create a useful planning horizon: if your personalization stack assumes that every durable customer preference must be baked into model weights, that assumption now deserves pressure-testing.

The old personalization pattern is familiar. A brand collects behavioral data, campaign responses, CRM fields, purchase records, support tickets, loyalty status, product preferences, and content engagement. Then a team tries to turn that material into segments, scores, features, embeddings, model-training sets, or rules. The customer’s history is compressed into something the system can act on.

Long-context changes the compression problem. Instead of reducing the customer to a thin profile before the model sees them, the system can expose a much richer history at inference time: prior conversations, abandoned carts, support escalations, preferences the user actually stated, recent browsing behavior, campaign fatigue, and the last offer they ignored. The model does not need to permanently learn that individual inside its weights if it can read the relevant history when making the next decision.

That is the marketing translation of Yang’s argument. The value does not come from saying “fine-tuning is dead” as if every current workflow should be thrown out. It comes from recognizing that the bottleneck may move. Today, many teams spend serious time debating model adaptation. In a long-context environment, the harder work may be identity resolution, consent capture, event taxonomy, retrieval, deduplication, freshness, and deciding which history is safe and useful to place in front of the model.

A 10,000-Word Instruction Is a Marketing Data Strategy

In a LessWrong-curated interview, Yang uses the idea of a “10,000-word instruction” to explain what becomes possible when context is abundant. With enough context, the user can describe the ideal experience in detail instead of training a model toward that experience indirectly.[3] For marketers, the phrase is less interesting as an instruction-writing trick than as a data architecture hint.

Imagine a customer experience system that does not begin with a static customer label. It begins with a living brief: what the customer bought, what they returned, what they asked support, what content they read, what they said they care about, what they opted out of, what tone they respond to, what they explicitly corrected, and which offers have already been exhausted. That is not a 10,000-word instruction written by a copywriter. It is a customer history assembled into a usable operating context.

This is where the forecast becomes operational. A marketing team that believes Yang’s direction does not need to stop all fine-tuning work. It does need to stop treating context as an afterthought. The durable asset is not only the trained model. It is the structured, permissioned, retrievable history that lets a general model behave as if it understands the customer in the moment.

That changes what “AI readiness” means. A cleaned CRM field is useful, but it is often too thin. A campaign event table is useful, but it may not explain why the customer acted. A support transcript is useful, but only if it can be safely retrieved and filtered. The long-context version of personalization rewards teams that can preserve the texture of interaction without drowning the model in irrelevant, stale, or non-consented material.

The Trust Threshold Is Not a Soft Issue

Yang’s personalization vision has a condition attached. He has argued that users need to become “friends” with AI before they are willing to feed it personal data, and coverage of his comments frames this trust threshold as a major obstacle to the broader vision.[4] That is the part marketers should take most seriously, because it keeps the long-context argument from collapsing into “collect more data and win.”

Marketing already has a rough equivalent in zero-party data: the information a customer intentionally shares because they expect a better experience in return. The analogy is imperfect. A preference-center answer is not the same as giving an AI access to a long trail of personal interactions. But the commercial logic is similar. The more sensitive the context, the more the user needs to understand what is being used, why it is being used, and how the benefit comes back to them.

A full interaction-history strategy therefore has two jobs. The first is technical: assemble context the model can use. The second is relational: earn permission to use it. If the second job fails, the first one becomes a liability. A model that can read every interaction is powerful only when the customer believes the company has the right to connect those dots.

This is the friction that makes sweeping “fine-tuning is over” takes unhelpful. Fine-tuning can happen inside a controlled enterprise boundary. Long-context personalization often asks for more visible, more granular, and more situational use of customer history. That raises the standard for consent design, preference management, retention rules, data minimization, and internal access controls. The marketing team does not get to treat trust as a brand sentiment metric while operations treats context as a plumbing problem. They are the same implementation surface.

Emergent Use Cases Are a Signal, Not Proof of Market Adoption

The LinkedIn interview includes a useful example: Kimi users independently discovered resume screening as a use case, which Yang presents as evidence that long-context systems can support cross-scenario inference and user-led applications.[2] For a marketer, the important point is not resume screening itself. It is the way a capability designed around large context can surface uses the builder did not prescribe in advance.

That maps cleanly to the messy middle of marketing systems. A team may begin by using long context for email personalization, then discover that the same interaction history improves sales-assist briefs, suppresses badly timed offers, rewrites onboarding sequences, or prevents a paid media audience from being built around customers who just complained to support. These are not separate AI strategies. They are different views of the same customer memory.

Still, a discovered use case is not adoption evidence. It does not tell us how often enterprises will deploy the pattern, how much lift it will produce, or whether the economics beat existing segmentation, retrieval, and fine-tuning approaches. It only supports the narrower conclusion: when context windows become large enough, users may find valuable workflows that were hard to specify in advance.

What Marketers Should Change Now

The right response is not to cancel every fine-tuning project. Some tuned systems may still be cheaper, safer, more predictable, or already good enough. The better response is to make sure the next year of marketing infrastructure work does not assume that model weights are the only place customer intelligence can live.

• Audit which personalization decisions currently depend on compressed customer representations: segments, scores, tags, journey stages, lead grades, and propensity models.
• Identify where richer interaction history would change the decision: support context before promotion, explicit preference before inferred interest, recent behavior before lifetime average.
• Design a context layer that can retrieve only the relevant history for a specific task instead of dumping every available record into a model request.
• Tie each new context source to a permission basis, user-facing benefit, retention rule, and fallback experience when permission is absent.
• Run limited experiments against existing segmentation or fine-tuning workflows, measuring whether long-context decisions improve customer experience and operational reliability.

The experiment should be narrow enough to govern. A lifecycle team could test whether adding recent support history reduces inappropriate upsell messages. A content team could test whether stated preferences outperform inferred topic affinity. A paid media team could test whether context-aware suppression prevents spend against customers in a negative service state. These tests do not require a grand replacement theory. They require one decision point where richer context should produce a visibly better action.

The hardest part will be handoff. Marketing does not own every useful signal. Support owns tickets. Product owns usage. Sales owns conversation notes. Data teams own identity stitching and governance. Legal and privacy teams own acceptable use. If long-context personalization becomes real, the winning capability is not a message template. It is the organization’s ability to make cross-functional customer memory available in a controlled, explainable way.

A Disciplined Reading of Yang’s Forecast

Yang’s argument deserves attention because it points to a different personalization architecture: feed the model the customer’s full relevant history as context, rather than assuming the durable advantage comes from fine-tuning on customer data. His “2-3 years” forecast raises the urgency, but it does not remove the need for evidence inside each marketing environment.[2]

The practical bet is cautious but clear. Start designing for full interaction-history context now. Do it where user permission, data quality, and decision accountability are strong enough to make the test honest. Keep fine-tuning where it is proven, economical, and governed. The replacement claim may or may not arrive on Yang’s timeline; the preparation work is useful either way.

References

1. Moonshot AI’s AGI Vision, ChinaTalk
2. Interview with Yang Zhilin: Moonshot AI, March Toward Endless, LinkedIn
3. Interviews with Moonshot AI’s CEO Yang Zhilin, LessWrong
4. Moonshot AI’s Founder: His Pursuit, aiproem Substack