Field research · transcript + metadata corpus

What a typical AI Engineer talk looks like

Descriptive analysis of 566 sessions (459 conference talks) across five AI Engineer events, 2025–2026. Built entirely from public YouTube captions and metadata — no video analysis. Everything below is measured from the corpus.

459 conference talks median 18.7 min 5 events 2025–2026
18.7 min
median talk length
16.3–20.6
middle-50% range (min)
82%
open with a self-intro
91%
link a CTA in the description

1Duration — the format is an ~18–20 min slot

Conference talks (4–45 min, n=459). Distribution is tight and consistent across every event — long-form sessions (107) are workshops/keynotes/streams, excluded here.

4
6
8
10
12
14
16
18
20
22
24
26
28
30
32
34
36
38
40
42
44
talk length (minutes) · 2-min bins
EventSessionsTalksMedian min
AIE World's Fair 2026383420.1
AIE Europe (Paris) 202623617919.1
AIE World's Fair 202524521118.2
AI Engineer Summit NY 2025181817.9
AIE CODE 2025 Labs291718.1
Distribution (all talks)
Median18.7 min
25th–75th percentile16.3–20.6 min
90th percentile24.7 min
Mean19.1 min

→ Build for ~18 minutes of content plus a short Q&A. Going past ~25 min is unusual for a talk slot.

2How talks open

Share of talks matching each pattern in the first ~120 spoken words.

Self-introductionself-intro (I'm / my name)
82%
Company / team mention
53%
Problem / hookquestion, 'imagine', 'raise your hand'
43%
Explicit agenda'today I'll…', 'this talk'
39%
Greeting / thanks
35%

3What the body contains

Share of talks whose transcript shows each signal. The corpus is demo- and evidence-driven, not slideware.

Code / GitHub / open-source
64%
Benchmark / eval numbers
53%
Live demo
50%
'Lessons learned' framing
24%
Hiring / recruiting
5%
Structure — 45% of videos are chaptered, with a median of 12 segments. Most common section titles: Introduction Conclusion Key Takeaways Best Practices Architecture Q&A

4Themes — practitioner & agent-centric

Most frequent talk-title keywords across all 566 sessions (subscript = count).

agents112building68agent53engineering30mcp30code29coding24evals23llms21context18enterprise17agentic16llm16scale16lessons15voice15workshop15models14production13apps13need12software12rag12full12frontier11knowledge11open11model10

5Who speaks — companies

Title convention is Talk Title — Speaker, Company. Top companies by talk count skew to frontier labs + AI infra/devtools.

Google DeepMind27Microsoft13Anthropic9Braintrust8OpenAI7Google7WorkOS6GitHub6AWS6Cloudflare5Neo4j5Cursor5Arize4OpenClaw4ElevenLabs4Elastic4Prime Intellect3Bright Data3Snorkel3RunPod3Turbopuffer3Cline3Stripe3Modal3

6Talk structures — one prototypical arc, seven variations

Derived from 199 talks with human-authored chapter markers. First, the canonical arc — the order beats tend to appear in (label shows how often that beat shows up; beats clustered mid-talk are spaced out here for legibility):

Intro / who am I33%
Problem / why it matters43%
Approach / how it works37%
Results / evidence27%
Demo / example24%
Close / takeaways57%
Q&A14%
start → finish · % = share of talks containing that beat

Most talks are a variation on that arc. The seven recurring shapes, by prevalence:

30%

Survey / State-of-X

You have broad perspective on a fast-moving area (a 'state of', a category map).
  1. Frame the landscape
  2. Tour the key developments / options
  3. Where it's heading
  4. Takeaways
Recursive Model ImprovementLee Robinson · Cursor Everything we knew about software has changedTheo Browne · @t3dotgg 2026 State of AI EngineeringBarr Yaron · Amplify Partners
17%

Build-Along (problem → build → demo)

You shipped a concrete system and can show it running end-to-end.
  1. Intro + the problem
  2. Design the approach
  3. Build it up in stages
  4. Live demo
  5. Close
Notion's Token TownSarah Sachs · Notion Through the AI FogManoj Nair · Snyk In the Land of AI Agents, the Verifiers Are KingTariq Shaukat · Sonar
17%

Lessons / Field Notes

You ran something in production and earned hard-won opinions. Very well received.
  1. Set the context (what we built)
  2. Result up front
  3. Lesson 1 · Lesson 2 · Lesson 3…
  4. One thing to remember
How I deleted 95% of my agent skills and got better resultsNick Nisi · WorkOS RAG Agents in Prod: 10 Lessons We LearnedDouwe Kiela · Contextual AI Your Attention Is the BottleneckZack Proser · WorkOS
14%

Problem → Solution (deep-dive)

A single technical idea deserves a rigorous, linear walkthrough.
  1. The problem, in depth
  2. Why existing answers fall short
  3. Our approach
  4. Why it works / results
HTML Is All Agents NeedJames Russo · HeyGen Perception AgentsAntje Barth · Amazon AGI Lab Software engineering is not about writing codeBenoit Schillings · Google DeepMind
10%

Thesis / Opinion (argument-led)

You want to change how the room thinks — a POV talk, often titled as a quote or question.
  1. Provocative claim
  2. Evidence / reasoning
  3. Address the counter-argument
  4. Land the point
Should AI Engineers Still Read Code in 2026?Alex Volkov · ThursdAI Beyond Components: Designing Generative UIRuben Casas · Postman Don't Build Agents, Build Skills Instead
8%

Demo-Driven / Product

The product speaks for itself; you spend most of the slot in the tool.
  1. Quick why
  2. Mostly live demo
  3. Q&A
What if the network was the sandbox?Remy Guercio · Tailscale How Google DeepMind Runs Agents at ScaleKP Sawhney & Ian Ballantyne · Google DeepMind Lobster Trap: OpenClaw in ContainersSally Ann O'Malley · Red Hat
3%

System + Results

An infra/scale story where the metrics are the headline.
  1. The system
  2. The numbers
  3. How we got there
  4. Close
Scaling GitHub for your AgentsSam Morrow · GitHub Optimizing inference for voice modelsPhilip Kiely · Baseten Building Generative Image & Video models at ScaleSander Dieleman · Google DeepMind

7Who's on stage — presenter pedigree

The room is builders, not academics: two-thirds are from startups/scaleups, and the seniority curve is a barbell of founders and hands-on ICs.

Company type
Startup / scaleup 67%Frontier lab / Big Tech 17%Independent / creator 1%Unclassified 16%
Seniority (highest signal wins)
Founder / C-level 19%IC — engineer / DevRel / PM 21%Staff+ / Principal / Lead 10%Researcher / Scientist / PhD 8%VP / Head / Director 6%Role not stated 36%
Role signals in bios (non-exclusive)
Plain engineer (SWE/ML/AI)
43%
Researcher / Scientist / PhD
21%
Founder / Co-founder / CEO
16%
Distinguished / Principal / Staff / Fellow
8%
VP / Head of / Director
7%
OSS creator / maintainer / author
7%
CTO / Chief (technical)
6%
DevRel / Advocate
6%
Product / Design
6%
Credential markers
names a well-known project (DSPy, LangChain, PyTorch, Next.js…)8%'previously at' an ex-employer15%big-tech alumni (ex-Google/Meta/OpenAI…)3%prior startup / founder exit9%book / O'Reilly author4%

8Presenter archetypes — and how each earns authority

Five recurring speaker profiles. Most talks map to one; knowing which you are tells you how to open and what to lean on.

~19% (founder/C-level)

The Builder-Founder

Who: Founder or CTO of an AI startup — usually the one who wrote the code.
How they establish authority: Names their company + framework, 'we built', ships a live demo of their own product. Credibility = the thing exists and works.
Sam BhagwatFounder & CEO, Mastra Bereket HabtemeskelCEO, Better Auth Douwe Kielacreator of RAG / Contextual AI
~17% at big labs

The Frontier-Lab Insider

Who: Researcher, MTS, or VP at OpenAI / Anthropic / Google DeepMind / Microsoft / NVIDIA.
How they establish authority: Speaks from inside the model factory; 'state of', capability previews, research-grounded. Authority = the org's name + first-hand access.
Benoit SchillingsVP of Research, Google DeepMind Sander DielemanGoogle DeepMind Antje BarthMTS, Amazon AGI Lab
~21% ICs (largest role bucket)

The Practitioner-IC

Who: Staff/senior engineer or ML/AI engineer at a scaleup shipping AI in production.
How they establish authority: 'Here's what we learned running this in prod' — lessons/field-notes, real eval numbers, gotchas. Credibility = scars, not titles.
Nick NisiWorkOS Zack ProserWorkOS Sarah SachsNotion
~7% cite a project

The OSS Maintainer

Who: Creator/core-contributor of a tool the audience already uses.
How they establish authority: Deep on one library; the project IS the credential. Often previews a next version.
Maxime RivestCore Contributor, DSPy Isaac MillerLead Maintainer, DSPy Sam MorrowGitHub
~6% DevRel + indies

The Creator / Voice-of-the-Community

Who: Developer advocate, well-known online voice, or independent (@handle instead of a corp).
How they establish authority: Opinion/thesis talks, provocative titles, POV on where the field is going. Authority = audience trust + reach.
Theo Browne@t3dotgg Alex VolkovThursdAI Addy OsmaniGoogle (creator/author)

9Surprising / unexpected from the transcripts

Non-obvious things the corpus revealed — the cultural signals that don't show up in a CFP form.

58%
actively de-hype

Talks use 'it's not magic', 'demystify', 'nothing special', 'just a…' — the culture rewards candor over spectacle.

62% / 52%
bring hard numbers

62% cite token/latency/cost figures; 52% quote a % improvement. Vague claims stand out (badly).

21%
open on a filler word

1 in 5 transcripts literally start with 'So…', 'Okay', 'Alright' — the vibe is conversational, not oratorical.

10%
swear a little

Casual profanity ('badass', 'shit', 'damn') appears in ~10% of talks. The register is engineer-casual, not corporate.

10%
acknowledge the room

'I'm the last speaker before lunch / between you and drinks' — situational, human openings are common and land well.

16%
poll the audience

'Raise your hand if…' / 'how many of you…' — a light interaction hook used in ~1 in 6 talks.

176 wpm
talk FAST

Median pace is ~176 words/min (typical presentation is 130-150). ~18 min at this clip packs a lot — dense, not padded.

5 talks
title = a quote

Some of the best-received titles are a bare provocative sentence in quotes, e.g. “Software Fundamentals Matter More Than Ever”.

Title conventions
median 7 words'Hook: Subtitle' colon 32%ends in '?' 3%

10Featured talks — what got the most reach

Highest-viewed talks in the corpus. Note how they cluster right around the 16–22 min mark. Click to open on YouTube.

11How tied are talks to the speaker's product?

The single clearest cultural rule, measured: own-company name mentions per 1,000 transcript words (n=403 talks with a parseable company). This is what the CFP's “no vendor-only talks” policy looks like in practice.

Never names own company
47%
Light (<1 / 1k words)
23%
Moderate (1–3 / 1k)
20%
Heavy (>3 / 1k)
10%

Median talk names its own company 1× total (0.27/1k words). Only 5% put their product in the title; 80% reference neutral/3rd-party tools (GPT, Claude, LangChain…).

The product-heavy minority (>3 mentions/1k — the exception, ~10%)
Scaling AI Agents Without Breaking RelTemporal · 16.6/1kStop Using RAG as MemoryZep · 11.0/1kUnder 5 minutes to a deployed LLM endpRunPod · 8.4/1kWindsurf everywhere, doing everythingWindsurf · 8.3/1k

→ Anchor the talk on a transferable lesson; let the product be the vehicle, mentioned once or twice, not the subject. For the FinServ mainstage specifically, the CFP asks vendors to put a customer/champion on stage instead.

12Calls-to-action — soft, not salesy

In-description links, and the phrases speakers use to close. Overt recruiting is rare — CTAs are "try the thing / find me".

Description has ≥1 link
91%
Personal / company site
52%
LinkedIn
45%
X / Twitter
36%
GitHub
18%
Closing phrases (talks using each, in final ~20%)
“thank you”333“qr code”39“try it”35“check out”33“reach out”31“find me”21“join us”16“come find”14“sign up”14“try out”13“our booth”13“come talk”11“we're hiring”8“scan the”6

The template a submission should fit

  • Length: design ~18 min of content (16–21 min is the sweet spot) + brief Q&A.
  • Arc: hook / problem → who you are + why credible → the meat (live demo, code, real eval numbers) → key takeaways.
  • Evidence over claims: ~64% show code/open-source, ~53% cite benchmarks, ~50% run a live demo. Bring receipts.
  • Angle: practitioner and production-focused (agents, evals, MCP, context, scale) — "lessons learned" framing plays well.
  • Pick a structure: the safest defaults are Lessons/Field-Notes (if you shipped it) or Build-Along (if you can demo it). Match the arc: intro early, close + takeaways in the last ~20%.
  • Know your archetype: founder → show the product works; IC → show the scars; researcher → show the frontier. The room skews startup builders, so hard-won production lessons over theory.
  • Tone: de-hype it (58% do), bring hard numbers (62% do), talk like an engineer not a keynote — a filler-word open is totally normal.
  • Title: Punchy Hook: Subtitle — Your Name, Company (~7 words); a provocative quote or question also works.
  • CTA: a single link / QR to "try it" or "find me" — skip the hard sell and the hiring pitch.