34 Research Pitches, Talks, and Seminars
This chapter is part of a book in active development and has not yet been through the author’s review. Content may change as the review advances.
The research decision. Build a spoken spine that survives your venue’s actual length. A thirty-second pitch, a three-minute walk, and a twenty-minute seminar are not one artifact at different speeds; each makes a different promise to the audience, and the claim must survive both compression and expansion without growing stronger in the shorter version.
34.1 Why this decision matters
The decision on the table: what gets cut when you say your claim out loud, and what never gets cut.
“I stop at maybe forty posters in an afternoon. What I remember is not the one with the most data. It is the one whose presenter told me, in one clean breath, exactly what they found and exactly where it stops. The ones who oversold lost me on the second sentence.” — a skeptical peer working the room at a research conference
A finished poster is silent. The moment a visitor stops, your voice takes over, and speaking is faster than reading, so you compress on the fly. That is where claims quietly grow. Under a little pressure, “was associated with” slides into “causes,” and “on data it never saw” drops out. The peer above is not testing your data. She is testing whether the sentence in your mouth still matches the sentence on your board. This chapter gives you a pitch you can say cold, under pressure, that never widens the claim your evidence earned.
34.2 The concept
A pitch is a short, prepared spoken version of your project, built to fit a fixed slice of time (Bourne 2007). Example: the thirty seconds you give a visitor who pauses, and the guided tour you give the one who stays, are two pitches of different lengths.
Every pitch shares one skeleton, the pitch architecture: the fixed set of beats each version moves through, which is hook, claim, evidence, boundary, invitation (Alley 2013).
- Hook: one sentence a stranger outside your field cares about. Example: “a help desk drowns in tickets every morning and cannot tell which will drag on.”
- Claim: your headline, said plainly and inside its limits.
- Evidence: the finding that earns the claim, one spoken sentence at a time.
- Boundary sentence: one sentence naming what the claim does not show. This is the beat pressure will try to drop, so you rehearse it until it is reflex.
- Invitation: a reason to look closer.
The three layers are that one skeleton at three lengths: a 30-second hook (hook, claim, boundary, invitation, with no evidence yet), a 90-second walk (the hook plus two or three evidence beats), and a full pitch of three to five minutes. Deciding what goes in each length is the choice this chapter equips.
One rule governs all three. Compression is a shorter version of a claim that keeps its boundary. Example: “on data the model had never seen, it flagged four of every five slow tickets” compresses to “on new tickets, it caught most of the slow ones.” Inflation is a version that quietly makes the claim bigger. Example: the same finding becomes “my model predicts which tickets will be slow.” One lost a detail; the other lost the truth. The failure to hunt is the excited-mouth upgrade: the moment your spoken verb outruns your written one and the boundary moves while nobody watches the words. Example: you wrote “is associated with” and your mouth says “causes.”
34.3 A worked example
You built a small machine-learning model for a campus help desk. A held-out set is a slice of your data the model never learned from, kept aside so its score is honest. Example: you trained on 4,000 past tickets and tested on 1,000 the model had never seen. On that held-out set, when the model flagged a ticket “slow,” it was right about 78% of the time. That right-when-it-flags rate has a name: precision is the share of a model’s flags that turn out correct. Your poster headline reads: “On held-out tickets, the model flagged slow ones with 78% precision.”
Now you build the pitch on the architecture.
- Hook: “A help desk drowns in tickets every morning and cannot tell which ones will drag on for days.”
- Claim, compressed for speech: “On tickets it had never seen, my model flagged the slow ones, and when it raised a flag it was right about three times in four.”
- Boundary sentence, the beat you rehearse to reflex: “That number is precision on a held-out slice, not overall accuracy, and the model still misses some slow tickets.”
- Invitation: “The surprising part is which tickets fooled it. Want the walk?”
Watch the two ways this goes wrong when you speak it. Drop the words “held-out” and you have claimed the model works on any ticket, which your test never showed. Swap “precision” for “accuracy” and you have quoted a number your poster does not report. Both are the excited-mouth upgrade, and both feel natural because the shorter sentence sounds cleaner. The fix is mechanical: mark the boundary sentence and say it at every length, even the 30-second one.
Building a spoken version around one claim and its boundary, rather than compressing the whole paper, is standard advice for research talks (Bourne 2007).
The block below scores the model on the held-out tickets and prints both numbers the pitch depends on: the one in the claim, and the one the boundary sentence protects you from being asked about.
import numpy as np, pandas as pd
SEED = 464
rng = np.random.default_rng(SEED)
# 1,000 held-out tickets the model never trained on.
held_out = 1000
slow = rng.random(held_out) < 0.22 # 22% really do drag on
flagged = np.where(slow, rng.random(held_out) < 0.62, # caught
rng.random(held_out) < 0.037) # false alarms
precision = slow[flagged].mean()
recall = flagged[slow].mean()
print(f"tickets flagged 'slow' : {flagged.sum()}")
print(f"precision (right when it flags) : {precision*100:.0f}%")
print(f"recall (share of slow ones caught): {recall*100:.0f}%")
print(f"\nthe headline number is the first one. the second is what the")
print("boundary sentence is about: the model still misses some slow tickets")34.4 An AI failure case
You paste your 90-second walk and ask a tool to compress it to a 30-second hook. Back comes a crisp, confident line: “My model predicts slow help-desk tickets 78% of the time.” It reads beautifully, and it is wrong twice. It dropped “held-out,” so it now claims the model works on any ticket, and it turned precision into a general “78% of the time,” a number your poster never reports. The sentence is shorter and smoother than yours, which is exactly the trap.
You catch it by putting the compressed line next to your poster’s boundary sentence and reading them word for word. “Predicts … 78% of the time” does not match “precision on a held-out slice.” Then you say the line to a friend, who asks, “so it’s right 78% of the time on any ticket?” That question is the widened claim coming back at you. You keep your own boundary sentence and drop the tool’s version.
If your talk carries slides, the slides serve the spoken spine and never replace it. Build the talk first, add a slide only where a listener needs to see what they cannot hold in memory, and give every quantitative visual the figure discipline from the poster chapter: honest axes, values on the marks, and uncertainty shown where the design licenses it, as an interval or an honest caveat. A seminar is the same spine grown to the venue’s length, not a different animal: problem, gap, and question, then method, evidence, boundary and limitations, and the next step, each section expanded until it fills its share of the time, with the boundary still spoken aloud.
34.5 It is your turn
You are working inside Studio 10: Prepare to publish or present. This lesson adapts your work to the talk format. Work the format your venue actually asks for; the studio page routes you.
Your draft is finished and criticized; this step is the version you say out loud, at two lengths, without the claim growing on the way.
The hands-on half of this section lives in the chapter’s companion notebook: open it in Colab with the badge at the top, and work the steps there.
Commit your own version first, then delegate. Each prompt is a checkable job, not a request for a verdict.
Expect two or three rounds each. The first compression is usually too smooth, so you point at the qualifier it dropped and ask again. Agentic tools will iterate toward whatever sounds best on its own terms, which is exactly the pressure that widens claims, so your boundary sentence has to be something you hold rather than something you delegate.
Three calls stay yours. You decide which claim your evidence actually supports, where the claim must stop (the exact boundary sentence), and whether a compressed sentence still says what the poster says. A tool can trim words and flag jargon. It cannot decide the shorter sentence is still true, and it will not be standing at the board when the stranger asks. You say the sentence. You own it.
Write your boundary sentence first, on its own line, and mark it. It is the beat that gets dropped under pressure, so it goes in before anything else.
Build the 30-second version: hook, claim, boundary, invitation, with no evidence beats yet.
Locate the jargon and translate it.
Here is one sentence from my poster: "[paste it]". My headline claim and its boundary are: "[paste them]". List every word a curious stranger with no training in my field could not decode. Then give two plain rewrites a first-year in any major would follow, each keeping my claim exactly as narrow as it is now.After running, verify: read each rewrite next to your boundary and reject any that dropped a qualifier like “held-out” or turned “flagged” into “predicts.” Counters silent scope change (a “plain” rewrite that quietly widens the claim while it simplifies).
A second angle, optional:
List the words that inflate my claim.
Here is my 30-second hook: "[paste it]". Here is my poster's boundary sentence: "[paste it]". Play a hurried visitor. List every word in the hook that says something bigger than the boundary allows. Rewrite nothing; just name the words.After running, verify: if it flags nothing and calls the hook “strong,” push back and demand the single worst offender. Counters sycophantic agreement (praise reviews your ego, not your words).
Build the 3-minute version: the same skeleton with two or three evidence beats and the same boundary sentence, word for word.
Expand to your venue’s actual length: grow the 3-minute spine section by section (problem and gap, question, method, evidence, boundary and limitations, next step) until it fills the time, and write the transformation memo line for what each expansion added.
Say the versions out loud to someone outside your field and have them repeat your claim back to you. If what comes back is bigger than what you wrote, the words they added are your revision list.
Red-team the pitch as a hurried stranger.
Play a skeptical conference visitor with thirty seconds of patience, from outside my field. Here is my full pitch: "[paste it]". Name the exact sentence where you lose interest or stop believing me, and the single place I promise more than I show.After running, verify: check its answer against a real practice run with a friend, and ask it “what is the one thing you left out?” Counters illusion of completeness (a polished “it’s clear” that skips the moment a stranger actually drifts).
Log it in your AI Research Ledger, and verify at least one compressed sentence with a named method from the Verification Guide, such as peer reasoning. An AI reviewer may run the check with you; the decision to accept or reject stays yours.