The five profiles
DIGEST ships with five reader profiles: Student, Researcher, Industry Pro, Curious Adult, and Quick Scan. The summarizer agent takes the same arXiv abstract and a profile-specific prompt and emits a tuned summary.
The profiles aren't a personalisation gimmick. They're a deliberate constraint on the summarizer to optimise for a specific reading mode. Same paper, different "what should I take away from this in 90 seconds".
What changes between profiles
Take Mixtral 8x7B as the working example. Same abstract goes through five prompts. Here's how the summaries differ, side by side.
Student
Mixtral uses a Mixture of Experts setup — 8 small networks per layer, a router picks 2 per token. So the model is big in parameters but cheap to run.
Defines "Mixture of Experts" inline. Explains the "router picks 2 per token" mechanism in plain language. Avoids citing prior work the student probably hasn't read.
Researcher
Mixtral extends Mistral 7B's architecture with top-2 routing across 8 expert FFNs. Slightly higher per-token compute than Switch Transformer's top-1, traded for routing stability.
Assumes Mistral 7B + Switch Transformer are known. Calls out the methodological delta (top-2 vs top-1) and the trade-off (stability vs compute). No definitions; you have the vocabulary.
Industry Pro
Mixtral 8x7B ships as open weights. 47B total params (94 GB FP16), 13B active per token, ~6× faster than Llama 2 70B. Catch: full VRAM required even though only 2/8 experts fire.
Production numbers up front. Capacity vs compute decoupled. The catch ("full VRAM") is the deployment surprise — most production readers would otherwise budget for the active param count.
Curious Adult
Big language models are usually one giant network. Mixtral is different: 8 specialist networks plus a switchboard that picks 2 per word. The result is as good as the biggest models but ~6× faster.
Analogy ("switchboard") plus the headline finding ("as good as the biggest models"). Zero unexplained jargon.
Quick Scan
Mixtral 8x7B: open-weights Sparse MoE. 47B params, 13B active. Matches GPT-3.5 at ~6× speed. Needs full VRAM.
Bullet-compact. Headline + nuance + catch in three sentences. Mobile-readable.
Why these five
We considered more profiles. "Quantitative trader" and "policy advisor" came up. We dropped them because the gain from a 6th or 7th profile didn't justify the additional summarizer complexity at v1.
The five we picked cover the cases that meaningfully change the summary shape:
- Student vs Researcher = vocabulary level
- Industry Pro vs Researcher = "what does it cost in production" vs "what does the method actually do"
- Curious Adult = analogy + zero jargon
- Quick Scan = the length / structure constraint
If your reading mode falls between two of these (e.g. PhD candidate switching between Researcher mode and Industry Pro mode depending on the paper), the right move on a Pro account is to pick the profile that matches your reading mode today and switch when it changes.
What's coming
Profile customisation (free-text "for an X reader" prompts) is on the roadmap post-launch. For now, the five fixed profiles are the contract: testable, comparable across users, and small enough to keep summary quality consistent.
If you have a sixth profile you'd genuinely use, email hello@digest.ltd. Concrete examples move the needle.