Agenda and stream notes · Both days

These notes describe the recordings. For engineering lessons, start with the two-day digest.

Day 1

October 13, 2026 · Day 1 digest · Source video

The day in one table

Wall clock	What happened	Why it mattered
8:56 AM	Show open and AIE welcome, then MC welcome	Lia McBride opened the third New York event. Alex Lieberman framed the room as an “Edge City” living six months to two years ahead of AI consensus.
9:02 AM	Vault: Making Unstructured Data Work for Investing (Bridgewater)	A governed 100-million-document corpus. Humans read about 300 words a minute; agents read about 30,000. A 200-example benchmark lifted precision at 10 from about 60% to 90-95%.
9:19 AM	Anthropic: Accelerating frontier AI in Financial Services	A five-layer agent foundation (MCP connectors, skills, identity, approval, evals). One bank cut $120 million of potential auto-loan fraud in 12 months.
9:38 AM	Balancing Acceleration and Safety for AI Agents (Two Sigma)	The $900 million Citi/Revlon wire error, retold. Five composable controls let a regulated firm grant agents real autonomy.
9:58 AM	Does Code Quality Even Matter? (Capital One)	Software splits into durable and disposable. Quality matters for the first, not the second, and rigor can be made cheap.
10:15 AM	What does an AI engineer do? (Arize AI)	Build the loop, not the feature. Arize’s Signal agent reads traces and opens fixes; telemetry should serve agents first.
10:39 AM	Wells Fargo Foundation Model for fraud detection	A 1-billion-parameter model pre-trained from scratch on billions of monthly transactions. It beat the production XGBoost model on detection rate, false positives, and latency inside the 250 ms decline window.
11:08 AM	Inside Arca’s long-horizon agents that run the business	Agents own 10-day client goals end to end. A signals engine lets agents watch the world so people do not have to.
11:31 AM	Intent Based Auth (panel)	OAuth 2.0 designer Dick Hardt said people “grabbed” OAuth for agents. The panel argued for per-tool-call authorization, missions, and budgets.
12:15-2:05 PM	Seven unlisted lunch talks: Range, Keycard, Stealthium, Railway, Southbridge, Perpetual, Flywire	The agenda called this a Networking Break. The stream carried almost two hours of unpublished technical content, including two of the day’s most concrete security talks.
1:57 PM	Great Agents Start with Great Infrastructure (Modal)	Sandboxes now hold multiple containers and sidecars. Production scale: 20 million snapshots a week, 8 petabytes a day, one workload needing 10,000 sandboxes per second.
2:22 PM	Beyond the Model Card (Vals AI)	No single model leads every task. Replaying GitHub pull-request history as agent tasks cut one team’s spend 90%.
2:44 PM	AI and the Future of Fundamental Investing (Kepler fireside)	Jack Woodruff: alpha stays divergent, prep work falls from six-to-eight hours to about five minutes, and humans stay the bottleneck.
2:52 PM	Shipping like a startup, regulated like a bank (Revolut)	Self-hosted GLM on 56 B200 GPUs. Blended token cost fell to 30 cents per million - one-tenth of the old invoice - with 70% better lead time.
3:17 PM	Everyone Has AI. Who Has the Advantage?	Only about 37% of enterprises report positive EBIT contribution from AI. Judgment, problem selection, and workflow redesign separate leaders.
4:02 PM	A Hedge Fund as a Function (Coatue)	The firm compresses context into one tuple: selection, size, timing. Each decomposed step runs on the best of four runners - code, statistics, LLMs, or people.
4:17 PM	Frontier Models for Acting in Space and Time (General Intuition)	Pixels in, actions out. Medal yields over 1.5 trillion labeled action tokens a month. One model plays games at the highest level and controls over 10 robot types.
4:36 PM	Hermes gives you wings (Nous Research), then Closing Remarks	Commoditized models push labs toward the agent layer. Rolnick argued for open foundations you own. The day wrapped by 4:53 PM.

Divergences from the published agenda

The lunch block replaced the Networking Break. The 12:15-2:05 PM slot is the only agenda item with no matching segment on the stream. It carried the seven unlisted talks above plus the early Modal start. (timeline)
The afternoon ran 25-30 minutes ahead of the agenda from about 2 PM. Modal started about 26 minutes early (Great Agents Start with Great Infrastructure, Hold after Modal talk ends early: no speech or OCR for 10.5 minutes before the Vals AI session), the Revolut talk aired at 2:52 PM against a 3:20 PM slot (Shipping like a startup, regulated like a bank, Break / room turnaround in Simons Room: after “Shipping like a startup, regulated like a bank” (David Abitbol, Revolut), before “Everyone Has AI. Who Has the Advantage?” (Sid Pednekar) - 13.6 min with no transcribed audio), and Coatue’s keynote aired at 4:02 PM against a 4:30 PM slot (A Hedge Fund as a Function).
The “after stream end” items actually aired. The agenda annotated Hermes gives you wings and Closing Remarks as starting after the stream cut. Both aired, at 4:36 PM and 4:51 PM, because the day ran early. They are not gaps.
The stream cut early relative to the published day. The broadcast ended at 4:53:42 PM ET, after applause following the closing remarks - well before the agenda’s 5:25-5:30 PM Closing Remarks slot. (Silence after final sign-off until the livestream cuts (applause or outro, untranscribed))
Stream-only artifacts. The feed opened with an eight-minute countdown hold (Pre-show countdown hold before show open (holding slide, pre-show music, no speech)) and held through long untranscribed changeovers between afternoon talks: 10.5 minutes after Modal (Hold after Modal talk ends early: no speech or OCR for 10.5 minutes before the Vals AI session), 12 minutes after Vals (Break / stage changeover after the Vals AI evals talk (12 min, no speech on stream)), 13.6 minutes after Revolut (Break / room turnaround in Simons Room: after “Shipping like a startup, regulated like a bank” (David Abitbol, Revolut), before “Everyone Has AI. Who Has the Advantage?” (Sid Pednekar) - 13.6 min with no transcribed audio), 12.7 minutes before the MC returned (Untranscribed gap after afternoon talk, before MC returns (12m44s; room changeover / idle stream)), and 10.4 minutes inside the welcome-back block (Untranscribed gap inside the welcome-back block (~10.3 min; possible sponsor reel, video, or dead air before the MC resumes)). Treat the agenda as a prior, not a grid.

Day 2

October 14, 2026 · Day 2 digest · Source video

The day in one table

Wall clock	What happened	Why it mattered
8:55:37 AM	Welcome remarks - Ksenia Se (Turing Post)	Framed day three as capabilities, agents, and who manages them; named the four rooms and the ROI question.
9:00:35 AM	Ramp - Chasing capabilities while waiting for the frontier	Applied research beat waiting: a post-trained Qwen 3.5 35B specialist shipped under frontier orchestrators, and a 10,000-agent pen test found seven high-severity bugs, all patched.
9:15:03 AM	Sonar - The Machine that Changed the World, vAI	Agent-written code rose 2,450% while releases rose only 30%; human review fails under that volume, so confidence must be engineered in.
9:32:03 AM	Coinbase - Trusting Agents in Policy-Driven Workflows	Policy as code, deterministic DAGs outside, adaptive agent loops inside, autonomy granted only on measured eval quality.
9:50:08 AM	OpenAI - Building the Next Chapter of AI in Finance	ChatGPT for financial services, one month old, built with Morgan Stanley and Evercore; expert-level finance judgment predicted for 2027.
10:09:25 AM	Bloomberg - Security in the Age of AI	A July 26 sandbox escape ran 17,600 agent actions in 4.5 days before detection; the answer is deterministic controls outside the model.
10:40:32 AM	Compound AI - Towards Autonomous Analysts in Finance	Rebuilt Office with 1.5 trillion tokens and 8M+ lines of code so long-running finance agents can move from copilot to autopilot.
11:05:43 AM	SemiAnalysis - Not All GPU’s Are Deployed The Same	Two B200 stacks, same model and target: about 200 versus about 300 generated tokens per second at P90. The operator is the multiplier.
11:30:35 AM	San Francisco Compute - How Compute Markets Work	Credit, not power, limits AI; the deposit to start a lab climbed from $10 million to $1 billion in three years.
11:55:58 AM	Jump Trading - The Dark Pools of Compute	Physical compute trades unlit, over chat; 15-25% prepayments squeeze middle-market buyers and make procurement a runway decision.
12:20:23 PM	Compute Exchange + Silicon Data - Carmen Li	GPU term structure turned contango; CME will settle H100 and B200 futures on Silicon Data indices within weeks.
12:39-2:03 PM	Published “Networking Break” carried recorded, unlisted talks	Intuit Credit Karma’s MCP stack, Michele Grant’s trust framework, Jellyfish’s cost-per-PR research, and Nori Agentic’s sec-master talk aired only here.
2:03:17 PM	Built to Scale - Blackstone x Norm.ai fireside	Fix the data strategy first; Norm’s tens of thousands of hand-labeled evals let it score agents the day they arrived.
2:30:20 PM	Kepler - Why Wall Street Still Doesn’t Trust AI	Code writes every number through pointers; citations are verified, corrected at search time, or refused.
2:52:53 PM	Google DeepMind - Coding an Autonomous Wall Street Coworker	Gemini managed agents with write-only credentials, un-editable hooks, and cron triggers; Gemini 4.0 recap from the day before.
3:19:18 PM	Citadel Securities - Your Company Has No Model of Itself	Treat the company as a state estimator: half of delivery dates change within two months, and 90% of those changes slip later.
3:39:53 PM	Prime Intellect - Self-Improving Organizations	A $300K trained model now competes for knowledge work; train at night on inference capacity that sits idle.
4:07:06 PM	CAIS accessibility talk, then a Coder talk cut mid-flow	Jemima Abu’s 11:30 AM breakout aired at 4:07 PM; the off-agenda Coder talk was cut to a music slate at about 4:22 PM.
4:26:03 PM	Final block: Apollo fireside, Basis, Runway	CLO portfolios went from 2-3 days to minutes; behavior specs supervise process; Runway demoed Continuum and promised release in 8-12 months.
5:27:25 PM	Closing remarks and the Swyx send-off	AI Engineer Europe returns to London in February; the World’s Fair plans 10,000 people in San Francisco next summer.

Divergences from the published agenda

The morning ran three to five minutes ahead. Sonar began 9:15:03 versus a 9:18 slot, Coinbase 9:32:03 versus 9:36, OpenAI 9:50:08 versus 9:54, and Bloomberg 10:09:25 versus 10:12. Bloomberg closed at 10:25:29, about five minutes early, which produced a 15-minute hold before the 10:30 AM break (Post-keynote hold and the 10:30-10:40am ET break (room turnover, no transcribed content)).
The 11:30 AM slot carried Evan Conrad, not the Coinbase talk. Lincoln Murr’s “Building the Financial Account for Agents” never appeared on this feed. The manifest’s boundary check recorded the slot as a swap (timeline); the archived agenda copy lists Conrad in the Simons Room and Murr in the Shannon Room at the same time. Either way, this feed aired How Compute Markets Work.
The lunch block never aired as a break. Off-agenda content started at 12:39:23 PM, one minute before the published 12:40 start, and ran to 2:02:56 PM. The Blackstone fireside then began at 2:03:17 PM, about two minutes before its published 2:05 slot (Built to Scale: Future-Proofing Your Technology Roadmap).
Afternoon talks ran ahead, then overran. The Gemini co-worker talk began 2:52:53 versus a 2:55 slot and ended about three minutes early (Coding an Autonomous Wall Street Coworker: Custom Skills, Connectors, and Event Hooks). Prime Intellect began 3:39:53 versus 3:45 and ran 3 minutes 22 seconds past its 4:03 close (Self-Improving Organizations).
A 22-minute published hole, 4:03 to 4:25 PM, was filled with two unscheduled talks (Stop Wasting Tokens on Products People Can’t Use, Scary stories of AI gone wrong: three real incidents and the infrastructure controls that prevent them). One was an 11:30 AM breakout moved by nearly five hours; the other had no agenda entry.
The early stream cut. At about 4:22 PM the feed cut the Coder talk mid-flow, on “Before I get into the stories, though, let me set up a little bit more context,” and showed a branded music slate for about 3 minutes 45 seconds. The three promised incidents never aired (Untranscribed break interstitial: stream cuts from the Coder talk (Zhao Chenjiang) mid-flow to a music slate; captions resume at the welcome-back voiceover for the final block).
The final block drifted late by cumulative overruns. Apollo ran about two minutes past 4:48, Basis started 1 minute 53 seconds late and ended 3 minutes 32 seconds past 5:06, and Runway started 3 minutes 44 seconds late and ended 2 minutes 46 seconds past 5:24 (Everything will be Simulated). Closing remarks began 5:27:25 versus the published 5:25, and Swyx’s last line landed at 5:30:36, 36 seconds past the published close.
Net effect. Measured item by item, the day drifted by minutes, not by tens of minutes. The largest divergences are structural: the lunch block that carried recorded talks, the 22-minute hole filled from other rooms, and the mid-talk cut at 4:22 PM.