Every piece of content has two versions: what it says, and what the reader hears. Perception engineering is the practice of optimising for the latter—altering some core information/news/data/idea in subtle ways such that it maintains faithfulness to the original while maximising how it resonates with a specific audience.
I later found the term appears in research, though rarely in the context of LLMs or linguistics. This idea stems a little bit from ideas around "LLMs speaking with confidence" and further influenced by the "era of experience", I love David Silver's explanation of the problems with RLHF; the problem with ChatGPT is that a consumer can ask for a cooking recipe and respond with "thats great!" without ever trying the recipe itself, so the models are incentivised to give responses that seem like something the user wants to see.
This works in a lot of cases, but the quality of output becomes supercharged with real actual feedback from the environment. But this is difficult for content creators, since public experimentation is slow and embarrassing. This is precisely the gap perception engineering tries to close. Instead of waiting for real feedback that may never come (or comes too late), we simulate the audience's reaction and iterate before publishing.
Can we simulate and optimise content before it ships?I think this exists on two levels:
- one-to-many: this is defining an audience (comprised of many personas), and optimising the content the maximise a given metrics across the entire audience and regenerating content until we find a local maxima. this is something like social media content.
- one-to-one: this is defining many audiences, or simply many personas, and optimising the content to maximise a given metric (or different metric per audience/persona) in order to generate a maxima for just that individual. this is something like a dm.
I have been most interested with the latter. This is a multidimensional problem, and what I believe true personalisation looks like. This is how we can supercharge current generalised models to give the best output every time, to give users that sci-fi level of personalisation.
For one-to-one content, specifically for things like DMs, one idea is to do monte-carlo/game type simulation -- to simulate many paths of conversation to optimise the one that leads to the optimal outcome. A great example is a fundraising investor call, or an interview. Simulating chess, the thing you say or reveal at a certain point in time will influence the course of the conversation, and you're really trying to optimise a few things (such as closing the sale, or getting a second meeting.)
Consider a fundraising email. The generic version might read: "We're raising a $5M seed round to scale our AI platform." The perception-engineered version for an investor who previously backed developer tools and tweets about infrastructure might read: "We're raising $5M to become the Datadog of AI agent observability—the same infra-first thesis you backed with [Company X]." Same facts, different frame.
Hurdles and assumptions
There are some bold assumptions here, I will list them in a compounding manner.
Foundational
- It is possible to measure the effectiveness of content.
- It is possible to measure the attribution of consumed content to subsequent actions.
The personalisation thesis
- One-size-fits-all content is difficult to generate, it is difficult to predict or measure or optimise such that it drives a target action from a sizeable portion of the audience.
- One-size-fits-all content is ineffective. Even if it is optimised to the best of ones ability, there is a ceiling beyond which you need to personalise to achieve a better outcome.
- Segment-level personalisation outperforms one-size-fits-all.
- One-to-one personalisation is the most performant variant of content.
Betting on simulation
- There exist metrics that can predict or simulate the performance of content before actually delivering the content to a user.
- It is possible to leverage LLMs to simulate feedback for generated content.
- Generated content can be optimised through a feedback-loop from the simulated performance of the content.
- The margin of improvement is significant from optimised content.
If assumptions 7-10 hold, we're looking at a future where content is never "published". It is rendered per-viewer, optimised in real-time. The question is whether the margin of improvement justifies the complexity. Early experiments suggest it does...