Anyone can have a post-graduate assistant on any subject.

I have a friend who works for Roche. He's pharma-literate. I'm not. We have an evening conversation coming up about cholesterol and statins. I wanted to walk in calibrated rather than tribal — neither deferring to authority nor parroting the contrarian YouTube position. So I commissioned a 267-page footnoted research synthesis. It took about 45 minutes of agent runtime. A simulated five-reviewer hospital panel scored it 74 out of 100 and found no invented citations in their sample. The whole thing is live, with figures, debate playbook, and external review attached.

→ Read the synthesis at articulate-ai.work/statins

The claim, said plainly

For most of human history, deep specialist knowledge has been gated by three things: time (a graduate degree is years), money (Harvard is expensive), and access (you need to know which professor to email). The Harvard Business Review framing of this is "generative AI transforms knowledge work" — fine as far as it goes, but it understates the part that matters: the gate is gone, not just lowered. Most people who want to understand a complex domain — to make a sound decision about their own health, their own retirement, their own legal exposure, their own technical stack — have had to choose between three bad options. One: trust the headlines. Headlines are wrong about contested science about half the time. Two: trust the authority you happen to know. Your GP is excellent at general medicine and may have read the last cardiology paper six years ago. Three: read a single popular book. Books are out of date the moment they ship, and the contrarian books outsell the careful ones.

None of those is what a postgraduate student gets. A postgrad gets a long, careful, supervised read of the actual literature — not the headlines, not the popular book, not the off-the-cuff specialist opinion. They get the evidence hierarchy, the named participants in each dispute, the trials that matter, the trials that don't, the steelmanned best case for every side, and a calibrated landing position that survives being argued against.

That is now a 45-minute task on a desktop AI workflow. Not a "summary." Not a "chat with the literature." An actual synthesis, the kind you'd hand to a supervisor and get back with marginal corrections rather than a request to do it again.

What the demonstration actually is

A friend asked me how to debate someone informed at Roche on cholesterol and statins. I treated it as a test case. The brief was specific: PhD-grade synthesis, ≥100 pages, full footnotes, steelman both sides, land calibrated. The output is at articulate-ai.work/statins. It includes:

• A full literature arc from Ancel Keys (1953) through the 2024–2026 frontier (STAREE, PREVENTABLE, the Lp(a) RNA therapeutics).
• The biology — mevalonate pathway, lipoprotein particles, atherogenesis, the failed HDL hypothesis — with six original SVG figures.
• The disputes — the consensus camp (Baigent, Sabatine, Ridker, Nissen) and the heterodox camp (Ravnskov, Kendrick, DuBroff, Malhotra) — by name, with their strongest arguments before the rebuttals.
• The harms ledger — SAMSON, the diabetes signal, cognition, rhabdomyolysis, asymmetric AE reporting.
• An adjudication chapter with NNT/NNH by patient stratum and a one-sentence calibrated position.
• A debate playbook — eight conversation-ready cards for the actual exchange.
• An external review panel of five simulated reviewers (lipidologist, cardiology trialist, EBM methodologist, pharmacology lecturer, clinical pharmacist) who scored it 74/100 and listed its specific weaknesses, unedited.

How it was built

Five research agents in parallel, each writing one part of the document from a brief of about 1,000 words. A grounding pass beforehand pulled the 2024–2026 trial findings via web search so the recent material had verified citations to cite from. A synthesis layer bound the parts into one coherent stance. A separate independent agent ran the review panel. Total wall-clock time, about 45 minutes. Total agent token consumption, on the order of half a million tokens. SuperSebastian owns this surface — research synthesis as a routine output, not a project — and the synthesis was the demonstration that shipped during my first week running AI as colleagues. The orchestration pattern sits on top of Anthropic's published research programme.

The crucial architectural choice was decomposing by epistemic boundary rather than chapter boundary. The biology of the mevalonate pathway is a different sort of work from the historiography of the trial era, which is different again from mapping the dispute between named participants. Each agent worked in its strongest mode rather than being asked to switch modes between biology, history, and dispute mapping.

What it isn't

It is not medical advice. It is not infallible. The review panel found real weaknesses — uneven steelmanning in three places, an over-precise NNT/NNH table, an asymmetric "what would change my mind" section, a handful of pharmacology gaps. None of those weaknesses invalidates the substantive analysis, but readers should keep them in mind. The full critique is published alongside the synthesis. Showing the work transparently is better than presenting only the polished surface.

Most importantly, it is not a substitute for a clinician who knows your specific numbers. What it is is the substrate for a real conversation — with the clinician, with the friend at Roche, with yourself.

Why this matters beyond statins

The same orchestration produces the same calibre of work on any subject where the literature is publicly accessible. Pension policy. Property law in your jurisdiction. The methane-leak literature. The actual evidence base for a supplement your in-laws keep recommending. A contested historical question. A technical stack you've never touched but need to inherit. A regulatory regime you're suddenly subject to.

The bottleneck is no longer access to the literature. The bottleneck is the framing of the question — what counts as evidence, who the named participants in the dispute are, what would change your mind. That's the work the person commissioning the synthesis still has to do. The synthesis itself is no longer scarce.

This is the part of the current AI moment that I think operators underweight. Everyone is busy looking at productivity gains inside their existing workflows. The bigger move is that decisions you previously couldn't make — because the knowledge gap was structural — are now decisions you can make from a position of evidence. The information asymmetry between an informed lay reader and a domain specialist hasn't disappeared, but it has narrowed sharply, and on questions of have I read the right things it has nearly closed. The Hype Radar is the filter that separates which capabilities like this are worth installing now from which ones to skip.

What this means for Articulate AI work

The same engine that produces the statins synthesis produces a competitive brief, a regulatory landscape, a pricing teardown, an attribution audit, a contract-review pack. The synthesis-as-a-service surface is one of the deliverables embedded in the Marketing Engine Pilot. The cost to a client is roughly an afternoon of analyst time billed at a senior rate, not a graduate student year billed at full overhead.