{"schema":1,"corpus":{"sources":{"aiengineer-nyc-2026/00-original/transcript-raw/transcript-master.json":"30aee858f35f0f46743ca23a16975b994f4ea37a4f3023b96da0222ddab21946","aiengineer-nyc-2026-day2/00-original/transcript-raw/transcript-master.json":"e6f9bf8eb67200c2059c511adbcd826a3bac6a1b54c6e14d530e28a7eadb4595"},"talks":47,"by_day":{"1":{"talks":23,"words":61757,"segments":3828},"2":{"talks":24,"words":72493,"segments":4355}},"words":134250,"segments":8183,"windows":838,"excluded_overlapping_or_noncontent_records":[{"day":1,"id":"047","reason":"Q&A overlaps the preceding keynote"},{"day":2,"id":"001","reason":"Welcome remarks, classified as a keynote in the source"},{"day":2,"id":"022","reason":"Duplicate Intuit recording; retain record 023"}],"partial_recordings":[{"key":"1:030","title":"Extra lunch talk: Flywire's internal AI platform — structured workflow automation (Orbit on the Victoria PaaS)"},{"key":"2:042","title":"Scary stories of AI gone wrong: three real incidents and the infrastructure controls that prevent them"}],"normalization":{"agents":"agent","agentic":"agent","models":"model","llms":"llm","gpus":"gpu","tokens":"token","workflows":"workflow","evaluations":"evaluation","evals":"evaluation","eval":"evaluation","evaluating":"evaluation","evaluate":"evaluation","evaluated":"evaluation","datasets":"dataset","benchmarks":"benchmark","systems":"system","tools":"tool","companies":"company","humans":"human","customers":"customer","users":"user","documents":"document","permissions":"permission","skills":"skill","costs":"cost","tasks":"task","traces":"trace","transactions":"transaction","applications":"application","servers":"server","sources":"source","prices":"price","contracts":"contract","clusters":"cluster","stocks":"stock","policies":"policy","controls":"control","processes":"process","developers":"developer","engineers":"engineer","environments":"environment","tests":"test","testing":"test","training":"training","trained":"training","train":"training"},"coverage":"A term family appears in at least two distinct transcript segments per talk. Categories overlap. The network also requires both concepts to meet this talk-level coverage rule.","network":"Undirected positive normalized pointwise mutual information, disjoint ~150-word windows; >=8 co-windows and >=4 talks; display cutoff .08. Co-occurrence is not causation, sequence, or agreement.","scope":"Clip boundaries can retain brief introductions within talks. The two captured main-stage feeds and recorded inserts; not the full multi-room event. Master ASR text, not written summaries or speaker bios. Names and numerical claims can contain transcription errors.","reproducible_seed":42,"claim_labels":"Eight excerpts reviewed in transcript context. Labels describe the excerpt, not the whole talk. Speaker results are not independently verified."},"sessions":[{"key":"1:003","day":1,"id":"003","title":"Vault: Making Unstructured Data Work for Investing","speakers":["Sam Green","Aaron Linsky"],"url":"/sessions/003-vault-making-unstructured-data-work-for-investing-bridgewater/","start":829,"end":1863,"partial":false,"words":2803,"sentence_count":165},{"key":"1:006","day":1,"id":"006","title":"Anthropic: Accelerating frontier AI in Financial Services.","speakers":["Belinda Neal","Keon Shahab"],"url":"/sessions/006-anthropic-accelerating-frontier-ai-in-financial-services/","start":1880,"end":3017,"partial":false,"words":3267,"sentence_count":206},{"key":"1:008","day":1,"id":"008","title":"Balancing Acceleration and Safety for AI Agents","speakers":["John Feminella"],"url":"/sessions/008-balancing-acceleration-and-safety-for-ai-agents/","start":3033,"end":4180,"partial":false,"words":3721,"sentence_count":202},{"key":"1:010","day":1,"id":"010","title":"Does Code Quality Even Matter?","speakers":["Max Kanat-Alexander"],"url":"/sessions/010-does-code-quality-even-matter-understanding-execution/","start":4182,"end":5196,"partial":false,"words":2946,"sentence_count":228},{"key":"1:013","day":1,"id":"013","title":"What does an AI engineer do?","speakers":["Aparna Dhinakaran"],"url":"/sessions/013-what-does-an-ai-engineer-do-build-the-loop-not-the-feature/","start":5217,"end":6162,"partial":false,"words":2583,"sentence_count":121},{"key":"1:016","day":1,"id":"016","title":"Wells Fargo Foundation Model: Application to Fraud Detection at Scale","speakers":["Freddy Lecue"],"url":"/sessions/016-wells-fargo-foundation-model-application-to-fraud-detection/","start":6685,"end":7971,"partial":false,"words":2671,"sentence_count":161},{"key":"1:018","day":1,"id":"018","title":"Inside Arca’s long-horizon agents that run the business","speakers":["Aman Singh"],"url":"/sessions/018-inside-arca-s-long-horizon-agents-that-run-the-business/","start":8394,"end":9201,"partial":false,"words":2109,"sentence_count":142},{"key":"1:020","day":1,"id":"020","title":"Intent Based Auth","speakers":["Allie Howe","Ian Livingstone","Michael Davis","Dick Hardt"],"url":"/sessions/020-intent-based-auth-panel-runtime-authorization-for-agents/","start":9793,"end":12382,"partial":false,"words":9040,"sentence_count":468},{"key":"1:022","day":1,"id":"022","title":"We Let Our Financial Planners Write the Agents","speakers":["Matt Ragone"],"url":"/sessions/022-extra-lunch-talk-financial-planners-building-their-own-agents/","start":12420,"end":12949,"partial":false,"words":1408,"sentence_count":114},{"key":"1:023","day":1,"id":"023","title":"Your Agent Is a Lethal Trifecta","speakers":["Kim Maida"],"url":"/sessions/023-extra-lunch-talk-your-agent-is-a-lethal-trifecta-task-scoped/","start":12957,"end":14484,"partial":false,"words":3632,"sentence_count":212},{"key":"1:025","day":1,"id":"025","title":"Extra lunch talk: what a GPU co-tenant can learn about your LLM from timing alone (MIG and NVLink side channels)","speakers":["Branislav Brzak"],"url":"/sessions/025-extra-lunch-talk-what-a-gpu-co-tenant-can-learn-about-your/","start":14504,"end":15305,"partial":false,"words":1968,"sentence_count":170},{"key":"1:027","day":1,"id":"027","title":"Become your company's SaaS layer — principles for building polished internal tools with agents (Railway)","speakers":["Mahmoud Abdelwahab"],"url":"/sessions/027-extra-lunch-talk-become-your-company-s-saas-layer-principles/","start":15330,"end":16128,"partial":false,"words":2454,"sentence_count":126},{"key":"1:028","day":1,"id":"028","title":"No universal schema: just-in-time ontologies, model priors, and Pareto surfing","speakers":["Hrishi Olickel"],"url":"/sessions/028-extra-lunch-talk-no-universal-schema-just-in-time-ontologies/","start":16137,"end":17412,"partial":false,"words":3559,"sentence_count":223},{"key":"1:029","day":1,"id":"029","title":"Reliable and intelligent workflows for finance, wealth, and insurance","speakers":["Karthic Rao"],"url":"/sessions/029-extra-lunch-talk-reliable-and-intelligent-workflows-for-fina/","start":17430,"end":18356,"partial":false,"words":2235,"sentence_count":174},{"key":"1:030","day":1,"id":"030","title":"Extra lunch talk: Flywire's internal AI platform — structured workflow automation (Orbit on the Victoria PaaS)","speakers":["Eduardo Simon","Javier Casal"],"url":"/sessions/030-extra-lunch-talk-flywire-s-internal-ai-platform-structured-wo/","start":18379,"end":18542,"partial":true,"words":319,"sentence_count":22},{"key":"1:031","day":1,"id":"031","title":"Great Agents Start with Great Infrastructure","speakers":["Akshat Bubna"],"url":"/sessions/031-great-agents-start-with-great-infrastructure-modal-sandboxes/","start":18543,"end":19407,"partial":false,"words":2117,"sentence_count":126},{"key":"1:033","day":1,"id":"033","title":"Beyond the Model Card: The Jagged Frontier of AI in Finance","speakers":["Rayan Krishnan"],"url":"/sessions/033-beyond-the-model-card-the-jagged-frontier-of-ai-in-finance/","start":20038,"end":20626,"partial":false,"words":2029,"sentence_count":108},{"key":"1:035","day":1,"id":"035","title":"AI and the Future of Fundamental Investing","speakers":["Jack Woodruff","Vinoo Ganesh"],"url":"/sessions/035-ai-and-the-future-of-fundamental-investing-kepler-fireside/","start":21358,"end":21830,"partial":false,"words":1042,"sentence_count":59},{"key":"1:036","day":1,"id":"036","title":"Shipping like a startup, regulated like a bank","speakers":["David Abitbol"],"url":"/sessions/036-shipping-like-a-startup-regulated-like-a-bank-self-hosting/","start":21830,"end":22523,"partial":false,"words":1747,"sentence_count":151},{"key":"1:038","day":1,"id":"038","title":"Everyone Has AI. Who Has the Advantage?","speakers":["Sid Pednekar"],"url":"/sessions/038-everyone-has-ai-who-has-the-advantage-three-pillars-design/","start":23337,"end":24416,"partial":false,"words":2364,"sentence_count":171},{"key":"1:043","day":1,"id":"043","title":"A Hedge Fund as a Function","speakers":["Frank Long"],"url":"/sessions/043-a-hedge-fund-as-a-function-coatue/","start":26060,"end":26914,"partial":false,"words":2320,"sentence_count":165},{"key":"1:046","day":1,"id":"046","title":"Frontier Models for Acting in Space and Time","speakers":["Pim de Witte"],"url":"/sessions/046-frontier-models-for-acting-in-space-and-time-pixels-to-actio/","start":26933,"end":28070,"partial":false,"words":3226,"sentence_count":173},{"key":"1:049","day":1,"id":"049","title":"Hermes gives you wings: Build on open foundations.","speakers":["Dillon Rolnick"],"url":"/sessions/049-hermes-gives-you-wings-build-agents-on-open-foundations-you/","start":28071,"end":29006,"partial":false,"words":2197,"sentence_count":141},{"key":"2:002","day":2,"id":"002","title":"Chasing capabilities while waiting for the frontier","speakers":["Alexander Shevchenko","Ksenia Se"],"url":"/day2/002-chasing-capabilities-while-waiting-for-the-frontier-ramp-app/","start":1199,"end":2065,"partial":false,"words":2230,"sentence_count":96},{"key":"2:004","day":2,"id":"004","title":"The Machine that Changed the World, vAI","speakers":["Tariq Shaukat"],"url":"/day2/004-the-machine-that-changed-the-world-vai-engineering-confidenc/","start":2067,"end":3081,"partial":false,"words":2702,"sentence_count":158},{"key":"2:005","day":2,"id":"005","title":"Trusting Agents in Policy-Driven Workflows","speakers":["Evan Kormos"],"url":"/day2/005-trusting-agents-in-policy-driven-workflows-policies-as-code/","start":3087,"end":4159,"partial":false,"words":2191,"sentence_count":135},{"key":"2:007","day":2,"id":"007","title":"Building the Next Chapter of AI in Finance","speakers":["Joseph Kim"],"url":"/day2/007-building-the-next-chapter-of-ai-in-finance-chatgpt-for-finan/","start":4172,"end":5328,"partial":false,"words":3248,"sentence_count":179},{"key":"2:010","day":2,"id":"010","title":"Security in the Age of AI: Recent Attacks, Incidents & Defenses","speakers":["Andrey Rybka"],"url":"/day2/010-security-in-the-age-of-ai-recent-attacks-incidents-defenses/","start":5329,"end":6293,"partial":false,"words":2053,"sentence_count":219},{"key":"2:012","day":2,"id":"012","title":"Towards Autonomous Analysts in Finance","speakers":["Yijia Liang"],"url":"/day2/012-towards-autonomous-analysts-in-finance-compound-long-running/","start":7196,"end":8130,"partial":false,"words":2925,"sentence_count":163},{"key":"2:014","day":2,"id":"014","title":"Not All GPU's Are Deployed The Same","speakers":["Oliver Kennon"],"url":"/day2/014-not-all-gpu-s-are-deployed-the-same-cost-curves-for-intellig/","start":8707,"end":9755,"partial":false,"words":2678,"sentence_count":210},{"key":"2:016","day":2,"id":"016","title":"How Compute Markets Work","speakers":["Evan Conrad"],"url":"/day2/016-how-compute-markets-work-supercomputing-as-a-service-credit/","start":10199,"end":11256,"partial":false,"words":3457,"sentence_count":207},{"key":"2:018","day":2,"id":"018","title":"The Dark Pools of Compute","speakers":["Affaan Mustafa","Alejandro Soumah Fernandez"],"url":"/day2/018-the-dark-pools-of-compute-unlit-otc-compute-markets-jevons-p/","start":11722,"end":12944,"partial":false,"words":3825,"sentence_count":209},{"key":"2:020","day":2,"id":"020","title":"CEO, Compute Exchange","speakers":["Carmen Li"],"url":"/day2/020-ceo-compute-exchange-the-only-neutral-gpu-marketplace-silico/","start":13187,"end":14264,"partial":false,"words":2695,"sentence_count":236},{"key":"2:023","day":2,"id":"023","title":"Building an enterprise agent ecosystem with MCP (gateway, servers, monitoring, memory, evals, PII trimmer, context optimizer)","speakers":["Harisha P C"],"url":"/day2/023-building-an-enterprise-agent-ecosystem-with-mcp-gateway-serv/","start":14328,"end":16097,"partial":false,"words":5002,"sentence_count":236},{"key":"2:024","day":2,"id":"024","title":"The Agent Works. Does Anyone Trust It?","speakers":["Michele Grant"],"url":"/day2/024-the-agent-works-does-anyone-trust-it-michele-grant-value-cas/","start":16103,"end":17019,"partial":false,"words":1720,"sentence_count":125},{"key":"2:026","day":2,"id":"026","title":"Tokens Are the Cheap Part: AI spend, developer throughput, and cost per merged PR","speakers":["Steve Jamison"],"url":"/day2/026-tokens-are-the-cheap-part-ai-spend-developer-throughput-and/","start":17044,"end":18090,"partial":false,"words":2145,"sentence_count":179},{"key":"2:029","day":2,"id":"029","title":"Agentifying a SEC Master","speakers":["Amol Kapoor","Clifford Ressel","Jiro"],"url":"/day2/029-agentifying-a-sec-master-nori-agentic-podcast-format-aie-onl/","start":18138,"end":19340,"partial":false,"words":4301,"sentence_count":311},{"key":"2:031","day":2,"id":"031","title":"Built to Scale: Future-Proofing Your Technology Roadmap","speakers":["John Stecher","Scott Worland"],"url":"/day2/031-built-to-scale-future-proofing-your-technology-roadmap-black/","start":19361,"end":20594,"partial":false,"words":4810,"sentence_count":214},{"key":"2:033","day":2,"id":"033","title":"Why Wall Street Still Doesn’t Trust AI","speakers":["Vinoo Ganesh"],"url":"/day2/033-why-wall-street-still-doesn-t-trust-ai-kepler-verification-d/","start":20984,"end":22080,"partial":false,"words":3183,"sentence_count":205},{"key":"2:035","day":2,"id":"035","title":"Coding an Autonomous Wall Street Coworker: Custom Skills, Connectors, and Event Hooks","speakers":["Philipp Schmid"],"url":"/day2/035-coding-an-autonomous-wall-street-co-worker-with-gemini-manag/","start":22337,"end":23370,"partial":false,"words":3095,"sentence_count":196},{"key":"2:037","day":2,"id":"037","title":"Your Company Has No Model of Itself: Bootstrapping a World Model from Slack, Decks and Meetings","speakers":["Joao Wenzel"],"url":"/day2/037-your-company-has-no-model-of-itself-bootstrapping-a-world-mo/","start":23922,"end":25052,"partial":false,"words":3372,"sentence_count":205},{"key":"2:039","day":2,"id":"039","title":"Self-Improving Organizations","speakers":["Will Brown"],"url":"/day2/039-self-improving-organizations-ai-native-enterprises-open-sour/","start":25157,"end":26746,"partial":false,"words":3888,"sentence_count":170},{"key":"2:041","day":2,"id":"041","title":"Stop Wasting Tokens on Products People Can’t Use","speakers":["Jemima Abu"],"url":"/day2/041-stop-wasting-tokens-on-products-people-can-t-use-cost-of-acc/","start":26790,"end":27391,"partial":false,"words":2449,"sentence_count":124},{"key":"2:042","day":2,"id":"042","title":"Scary stories of AI gone wrong: three real incidents and the infrastructure controls that prevent them","speakers":["Zhao Chenjiang"],"url":"/day2/042-scary-stories-of-ai-gone-wrong-three-real-incidents-and-the/","start":27393,"end":27708,"partial":true,"words":763,"sentence_count":45},{"key":"2:045","day":2,"id":"045","title":"AI in Practice: An Inside Look at Apollo's Approach","speakers":["Michael Manapat","Jack McDonald"],"url":"/day2/045-ai-in-practice-an-inside-look-at-apollo-s-approach-rowspace/","start":28190,"end":29356,"partial":false,"words":3313,"sentence_count":143},{"key":"2:047","day":2,"id":"047","title":"Process Supervision for Agents That Do Long-Horizon Work","speakers":["Mitchell Troyanovsky"],"url":"/day2/047-process-supervision-for-agents-that-do-long-horizon-work-beh/","start":29357,"end":30536,"partial":false,"words":3341,"sentence_count":207},{"key":"2:049","day":2,"id":"049","title":"Everything will be Simulated","speakers":["Cristóbal Valenzuela"],"url":"/day2/049-everything-will-be-simulated-runway-video-models-next-frame/","start":30548,"end":31570,"partial":false,"words":3107,"sentence_count":183}],"concepts":[{"concept":"Software and coding","mentions":413,"talks":41,"coverage":0.8723,"day1_talks":20,"day2_talks":21,"day1_coverage":0.8696,"day2_coverage":0.875,"day1_per_10000_words":28.82,"day2_per_10000_words":32.42,"day1_mean_talk_rate":32.27,"day2_mean_talk_rate":34.22,"day_shift_pp":0.54,"shift_ci_low_pp":-16.49,"shift_ci_high_pp":18.03,"id":"software-and-coding","guide":"coding-and-product-value","primary":true,"pattern":"\\b(?:cod(?:e|es|ing|ebase)|pull requests?|merged pr|software|developers?)\\b","matches":[{"key":"1:006","mentions":9,"segments":8,"excerpt":{"text":"Many of you in the room are using these tools, and you've gone from autocomplete in coding to managing fleets of agents where you're really driving and accelerating a multiple effect of how you're doing coding.","seconds":2006}},{"key":"1:008","mentions":7,"segments":6,"excerpt":{"text":"And when I say open-ended here, you can imagine coding agents like Pi or Claude Code.","seconds":3719}},{"key":"1:010","mentions":48,"segments":42,"excerpt":{"text":"In fact, I believe the absolute amount of durable software in the world will increase because AI coding tools make it easier and easier to make more durable software.","seconds":4914}},{"key":"1:013","mentions":9,"segments":7,"excerpt":{"text":"So you can trace Cloud Code, you can trace Cursor, you can trace Codex, but all of these coding agents can be traced so you can understand what happened in the factory of agents that you all are building.","seconds":6031}},{"key":"1:016","mentions":6,"segments":5,"excerpt":{"text":"And even if you look at the zip code, I mean, you have a much larger number of zip codes in the U.","seconds":7504}},{"key":"1:020","mentions":17,"segments":16,"excerpt":{"text":"And in the context of using Claude Code or whatever, why am I working in this moment with Claude Code?","seconds":10593}},{"key":"1:022","mentions":4,"segments":4,"excerpt":{"text":"And none of us, the financial planners for that matter, write code.","seconds":12430}},{"key":"1:023","mentions":3,"segments":3,"excerpt":{"text":"The runtime might be a CLI like Claude Code, or an agent SDK, which might be implemented in a custom application like the on-call agent.","seconds":13664}},{"key":"1:025","mentions":4,"segments":4,"excerpt":{"text":"We can tell what kind of work it's doing: code, JSON, or retrieval.","seconds":14614}},{"key":"1:027","mentions":5,"segments":3,"excerpt":{"text":"And for me, there's one missing piece that rarely gets mentioned, and it's the fact that there's a gap, a big gap, between working software that got one-shotted and software I would actually use, software that sparks joy, that is polished, that feels good.","seconds":15342}},{"key":"1:028","mentions":4,"segments":4,"excerpt":{"text":"The primary reason that this happens is that intelligence, as it relates to data work today, occupies three separate roles that coding agents kind of mash into one context window.","seconds":16663}},{"key":"1:029","mentions":2,"segments":2,"excerpt":{"text":"Exemption codes apply only to some entities and never to an S corporation.","seconds":17665}},{"key":"1:030","mentions":3,"segments":3,"excerpt":{"text":"To give you an idea of our engineering organization, we have around 400 software engineers distributed globally.","seconds":18464}},{"key":"1:031","mentions":5,"segments":5,"excerpt":{"text":"And if you're building a coding agent, you need your repo, your GitHub repo.","seconds":18796}},{"key":"1:033","mentions":6,"segments":5,"excerpt":{"text":"The second is, I think a lot of people are starting to see that the cost of software is coming down and it's actually very, very easy to produce code.","seconds":20295}},{"key":"1:036","mentions":15,"segments":14,"excerpt":{"text":"Around February, we started playing around with Code, and then January is when we first launched Code at scale.","seconds":21964}},{"key":"1:038","mentions":3,"segments":3,"excerpt":{"text":"So are the software engineers, the AI leaders, the portfolio managers, and the data scientists.","seconds":23633}},{"key":"1:043","mentions":11,"segments":10,"excerpt":{"text":"So you'll see on the right here, I have code, LLM, LLM, code, person.","seconds":26518}},{"key":"1:046","mentions":8,"segments":8,"excerpt":{"text":"So you can think of the way that Anthropic focused on just nailing coding.","seconds":27222}},{"key":"1:049","mentions":8,"segments":6,"excerpt":{"text":"It used to be that software was difficult to produce and it took coordinated teams of engineers long hours to pilfer open source and take libraries and combine them together into proprietary pieces of software and then sell that software as SaaS.","seconds":28610}},{"key":"2:002","mentions":2,"segments":2,"excerpt":{"text":"So if you're interested in any of these two products, you can take a photo of the URLs or scan the QR codes there.","seconds":1926}},{"key":"2:004","mentions":36,"segments":28,"excerpt":{"text":"The left-hand side is a study that was done on code translation, and what it shows is that the agents were able to get it right 91% of the time when the codebase was 10,000 lines of code.","seconds":2510}},{"key":"2:005","mentions":20,"segments":17,"excerpt":{"text":"Apache Airflow, if you're familiar, is an example of open source software that represents workflows, their dependencies as DAGs in Python code.","seconds":3375}},{"key":"2:007","mentions":7,"segments":6,"excerpt":{"text":"Tools allow them to search for information, write and run code, and interact with the software we use.","seconds":5222}},{"key":"2:010","mentions":4,"segments":4,"excerpt":{"text":"And what happened was the instructions in Morse code was decoded, and 3 billion tokens from the wallet got stolen.","seconds":5832}},{"key":"2:012","mentions":13,"segments":11,"excerpt":{"text":"And with coding agents dramatically improving and bringing the cost of software down, we decided to just build our own office from scratch.","seconds":7555}},{"key":"2:014","mentions":6,"segments":6,"excerpt":{"text":"The operator is who chooses and implements the software that serves the model and keeps requests moving through the system.","seconds":8874}},{"key":"2:023","mentions":5,"segments":5,"excerpt":{"text":"So once you have that repository, so developers basically, you know, clone that repository.","seconds":14694}},{"key":"2:024","mentions":2,"segments":2,"excerpt":{"text":"A pilot tests whether the software functions, but daily operations is where we test whether it creates value within the controls set by leadership.","seconds":16228}},{"key":"2:026","mentions":16,"segments":15,"excerpt":{"text":"Here we looked at developer weeks with heavy token use compared to the same developer weeks with lower token use.","seconds":17586}},{"key":"2:029","mentions":17,"segments":15,"excerpt":{"text":"And so they create these software factories, software factories, quote unquote, that are really inflexible.","seconds":19260}},{"key":"2:031","mentions":25,"segments":18,"excerpt":{"text":"And we, our approach to this, and I'm sure a lot of other companies, and maybe even you guys have taken this approach as well, is there's parts of our codebase that we, like, still require, like, complete human oversight from when it comes to designing, building, code ownership, even improving any kind of changes to the codebase.","seconds":20296}},{"key":"2:033","mentions":5,"segments":4,"excerpt":{"text":"Code writes every number, and what code does and pre-processing does is we assign pointers to unique identifiers, meaning facts, meaning something like a number in a 10K.","seconds":21339}},{"key":"2:035","mentions":11,"segments":10,"excerpt":{"text":"Then we moved to assistants using tools on our own computer, having the first coding agents writing code for us.","seconds":22370}},{"key":"2:037","mentions":8,"segments":8,"excerpt":{"text":"Ultimately, these culminate in a pull request, but the synthesis of these things, I think, makes a big difference.","seconds":24797}},{"key":"2:039","mentions":14,"segments":13,"excerpt":{"text":"I think, like, with code, we've gotten to a point of maturity around, like, testing and life cycles and software development life cycles for ensuring that codebases grow healthily over time, usually.","seconds":26229}},{"key":"2:041","mentions":25,"segments":20,"excerpt":{"text":"I think this is very important, especially if you're like vibe coding or you're not working in a big codebase.","seconds":27326}},{"key":"2:042","mentions":2,"segments":2,"excerpt":{"text":"You prompt your coding agent to take a look at your issues.","seconds":27515}},{"key":"2:045","mentions":3,"segments":3,"excerpt":{"text":"So I think one thing that we've been thinking about is software engineering that has completely changed as a discipline.","seconds":28900}},{"key":"2:047","mentions":8,"segments":7,"excerpt":{"text":"It's going to sound silly, but I promise it's very effective, which is that you can, when working with coding agents, ask the coding agent to make a sub-agent that pretends to be your production agent.","seconds":30312}},{"key":"2:049","mentions":5,"segments":3,"excerpt":{"text":"So we've connected Continuum to a coding agent, and what the coding agent can do is it can use Continuum to visualize through a problem.","seconds":31405}}]},{"concept":"Cost","mentions":459,"talks":34,"coverage":0.7234,"day1_talks":14,"day2_talks":20,"day1_coverage":0.6087,"day2_coverage":0.8333,"day1_per_10000_words":21.21,"day2_per_10000_words":45.25,"day1_mean_talk_rate":21.08,"day2_mean_talk_rate":51.47,"day_shift_pp":22.46,"shift_ci_low_pp":-3.08,"shift_ci_high_pp":47.92,"id":"cost","guide":"compute-economics","primary":true,"pattern":"\\b(?:cost\\w*|prices?|pricing|spend\\w*|budgets?|expensive|cheaper?|dollars?)\\b","matches":[{"key":"1:006","mentions":7,"segments":7,"excerpt":{"text":"This was built on the foundation that advisors only spend approximately 20% of their time with clients.","seconds":2861}},{"key":"1:008","mentions":3,"segments":3,"excerpt":{"text":"So this is how we keep the cost to teams really low for deploying, building, running agents.","seconds":3931}},{"key":"1:010","mentions":14,"segments":12,"excerpt":{"text":"However, what you want to do, and I think what you want to care about the most, is reducing the cost of effort, reducing the cost of rigor.","seconds":4947}},{"key":"1:013","mentions":7,"segments":7,"excerpt":{"text":"It helps us control costs so we can see, you know, as we use more powerful models.","seconds":5830}},{"key":"1:016","mentions":5,"segments":4,"excerpt":{"text":"So the main matrix, and those matrix relate right away to dollar, to cost, well, to savings for the firm.","seconds":7852}},{"key":"1:020","mentions":17,"segments":11,"excerpt":{"text":"Where in budget, there's continual feedback to the agent about how much budget was consumed, has an idea about how much budget it has, and so now it can operate within a budget, right, like a person would do about, Oh, this resource is scarce.","seconds":11386}},{"key":"1:028","mentions":22,"segments":21,"excerpt":{"text":"Given the right knowledge, instructions, and an incredibly stable surface to stand on, they can try new things, vary budgets to limit cost exposure, debug quickly, take notes, surf their way not just to the right solution, but the optimal one.","seconds":16774}},{"key":"1:031","mentions":4,"segments":3,"excerpt":{"text":"Maybe you want to look at the budget and make sure you're not spending too much in your sandbox.","seconds":19070}},{"key":"1:033","mentions":16,"segments":13,"excerpt":{"text":"On the X-axis, you've got cost, in particular it's cost per task, not just token in Excel models.","seconds":20096}},{"key":"1:035","mentions":3,"segments":3,"excerpt":{"text":"So you start getting diminishing returns on how much time you could spend on any particular stock.","seconds":21534}},{"key":"1:036","mentions":7,"segments":5,"excerpt":{"text":"In terms of cost, in terms of impact, self-hosting GLM, our blended cost of token is 30 cents per million, and so our invoice is one-tenth of what we were spending on.","seconds":22149}},{"key":"1:038","mentions":6,"segments":6,"excerpt":{"text":"Teams towards three main verticals: optimizing efficiency, driving P&L, and reducing costs.","seconds":23400}},{"key":"1:043","mentions":2,"segments":2,"excerpt":{"text":"So for something like, if the price to earnings ratio is less than 15, I should just use deterministic code for that.","seconds":26547}},{"key":"1:049","mentions":13,"segments":12,"excerpt":{"text":"There was some charts going around Twitter this week showing the average cost of a token going down while the total spent on tokens surges, and the infrastructure costs associated with serving those tokens has stayed flat, meaning","seconds":28183}},{"key":"2:002","mentions":22,"segments":16,"excerpt":{"text":"Over the last year, the number of use cases that use LLMs has increased 123%, while at the same time, a lot of our top features are growing in token spend, but on a unit economic basis, they're actually becoming cheaper because there are frontier models that are being released all the time now, and these models are cheaper, faster, and better.","seconds":1261}},{"key":"2:004","mentions":6,"segments":6,"excerpt":{"text":"If I pick Albuquerque, it's going to take me a long time and a lot of cost.","seconds":2568}},{"key":"2:005","mentions":4,"segments":4,"excerpt":{"text":"Previously, some work used to cost too much or be too complex to coordinate in this way.","seconds":3344}},{"key":"2:007","mentions":4,"segments":4,"excerpt":{"text":"The time you spend reviewing it becomes much smaller than the time it saves you.","seconds":5029}},{"key":"2:010","mentions":3,"segments":3,"excerpt":{"text":"You want hard rate limits and budgets, and you want extra guardrails.","seconds":5859}},{"key":"2:012","mentions":7,"segments":7,"excerpt":{"text":"They spend most of the time thinking about, is this actually a good deal?","seconds":7894}},{"key":"2:014","mentions":70,"segments":56,"excerpt":{"text":"Today, price per token is a useful practical proxy, especially for comparable workloads and quality requirements, and gets us closer to delivered output, much closer than the often quoted, all you finance bros out here will have heard this, GPU dollar per hour price, or as we view it, a cost.","seconds":8812}},{"key":"2:016","mentions":10,"segments":7,"excerpt":{"text":"We think this is what you are actually going to have to do in order to start a new company, because other people are spending hundreds of millions of dollars of tokens.","seconds":11163}},{"key":"2:018","mentions":40,"segments":31,"excerpt":{"text":"So there's an equilibrium that exists that has to be followed where, you know, the ownership costs outweigh the, you know, rental costs, and the rental costs— Should sit below the ownership cost.","seconds":12451}},{"key":"2:020","mentions":20,"segments":19,"excerpt":{"text":"If you're looking for hedging in the financial sense, right, your price volatility, your cost volatility, you go to Brent.","seconds":13498}},{"key":"2:023","mentions":4,"segments":4,"excerpt":{"text":"And you can see that this one row, each for, you know, like success, like latency cost, tool accuracy, safety, etc.","seconds":15674}},{"key":"2:024","mentions":3,"segments":3,"excerpt":{"text":"If four analysts spend half of their week building this deck, that is 80 hours a week the agent gives back.","seconds":16427}},{"key":"2:026","mentions":91,"segments":74,"excerpt":{"text":"We need to add the AI spend and the human spend, the human salary spend.","seconds":17932}},{"key":"2:029","mentions":8,"segments":8,"excerpt":{"text":"There's no API that you can hit to be like, give me the price of this data center.","seconds":18868}},{"key":"2:031","mentions":8,"segments":7,"excerpt":{"text":"And does it cost me less to rewrite it than it costs me to keep versioning the current version forward?","seconds":20457}},{"key":"2:033","mentions":3,"segments":2,"excerpt":{"text":"That's a hugely expensive token cost.","seconds":21382}},{"key":"2:035","mentions":3,"segments":3,"excerpt":{"text":"So if we have very long system instructions, all of our requests are going to be much more expensive.","seconds":22795}},{"key":"2:039","mentions":6,"segments":4,"excerpt":{"text":"And I think nowadays people are realizing that, like, cached input is, like, the bottleneck in many agent costs.","seconds":25930}},{"key":"2:041","mentions":8,"segments":8,"excerpt":{"text":"If anything, I would say it costs more time and more money to build inaccessible websites.","seconds":26857}},{"key":"2:045","mentions":6,"segments":6,"excerpt":{"text":"So spending the time and really putting boots on the ground to sit with whatever team or front office professional that you're working to build a tool for or enhance is time well spent.","seconds":29305}}]},{"concept":"Finance","mentions":488,"talks":33,"coverage":0.7021,"day1_talks":16,"day2_talks":17,"day1_coverage":0.6957,"day2_coverage":0.7083,"day1_per_10000_words":34.17,"day2_per_10000_words":38.21,"day1_mean_talk_rate":35.47,"day2_mean_talk_rate":36.71,"day_shift_pp":1.27,"shift_ci_low_pp":-24.46,"shift_ci_high_pp":26.81,"id":"finance","guide":"company-context","primary":false,"pattern":"\\b(?:financ\\w*|invest\\w*|banks?|banking|trades?|trading|portfolios?|fraud|accounting|transactions?)\\b","matches":[{"key":"1:003","mentions":18,"segments":17,"excerpt":{"text":"And the way that we're solving that problem is in an experience we're calling the investor workbench, where the Agent is always on, tracking what's important to that investor and presenting them with the areas that are most in need of their attention.","seconds":1741}},{"key":"1:006","mentions":33,"segments":28,"excerpt":{"text":"In this case, a trade break agent is meant to look at two records of a single trade that don't match.","seconds":2448}},{"key":"1:008","mentions":19,"segments":17,"excerpt":{"text":"And the bank that's overseeing that transaction, that's in the middle between the borrower and the lender, sends the lender all of the money that you owe in a catastrophic mistake.","seconds":3139}},{"key":"1:016","mentions":58,"segments":43,"excerpt":{"text":"We have information in investment banking, consumer banking, and all of these data is data that we have in Wells Fargo and all big financial organizations have too.","seconds":6685}},{"key":"1:018","mentions":6,"segments":4,"excerpt":{"text":"We provide our clients with financial services like financial planning, tax planning, asset management, investment advice, and we serve a range of clients.","seconds":8439}},{"key":"1:020","mentions":15,"segments":12,"excerpt":{"text":"Banks moved long ago from authorizing and authenticating humans to authorizing the transaction, right?","seconds":10127}},{"key":"1:022","mentions":6,"segments":6,"excerpt":{"text":"And none of us, the financial planners for that matter, write code.","seconds":12430}},{"key":"1:028","mentions":4,"segments":4,"excerpt":{"text":"We were working on donors in Ohio's campaign finance, and at this point, a lot of people have seen this chart.","seconds":16802}},{"key":"1:029","mentions":3,"segments":3,"excerpt":{"text":"Workflows in finance and wealth management come with four facts, and each of them pushes the architecture in the same direction.","seconds":17498}},{"key":"1:031","mentions":4,"segments":4,"excerpt":{"text":"You want to run a complex analysis on every company in your portfolio.","seconds":18576}},{"key":"1:033","mentions":3,"segments":3,"excerpt":{"text":"Last week we have Sundar talking about our Vals index on stage, and then even just yesterday, Muse Models citing our finance agent benchmark.","seconds":20060}},{"key":"1:035","mentions":4,"segments":4,"excerpt":{"text":"What does fundamental investing look like in the age of AI, say five years from now?","seconds":21403}},{"key":"1:036","mentions":4,"segments":4,"excerpt":{"text":"We're live in 45 markets with the most downloaded financial app in Europe for the last year.","seconds":21871}},{"key":"1:038","mentions":14,"segments":14,"excerpt":{"text":"It's the decisions where you can weigh what trade-offs are worth making.","seconds":23614}},{"key":"1:043","mentions":10,"segments":10,"excerpt":{"text":"Or, oh man, LLMs are really, really bad at translating that to portfolio impact.","seconds":26800}},{"key":"1:049","mentions":8,"segments":7,"excerpt":{"text":"Somebody is going to earn back the trillion-dollar investments that we've made.","seconds":28987}},{"key":"2:002","mentions":9,"segments":7,"excerpt":{"text":"And so we took a step back and decided to train a retrieval model, or kind of like a financial question answering model that would, while still keeping these frontier models as orchestrators, would then be handed off to this much faster model that would do financial question answering for that planning agent.","seconds":1512}},{"key":"2:005","mentions":2,"segments":2,"excerpt":{"text":"I've been delivering, designing, building software for 20 years, the last 15 in enterprise financial services.","seconds":3130}},{"key":"2:007","mentions":57,"segments":46,"excerpt":{"text":"ChatGPT for financial services is a tailored version of ChatGPT work built specifically for financial professionals, starting with investment banking and equity research.","seconds":4285}},{"key":"2:012","mentions":17,"segments":17,"excerpt":{"text":"People in finance complain all the time, Hey, I have an Excel model that's 30 megabytes.","seconds":7818}},{"key":"2:014","mentions":8,"segments":8,"excerpt":{"text":"futures grade, establish a sound benchmark, or create enough trading to buy and sell easily.","seconds":9600}},{"key":"2:016","mentions":13,"segments":11,"excerpt":{"text":"And so that means that in order to finance GPU clusters, you have to have long-term take-or-pay contracts, hopefully with someone who has investment grade credit.","seconds":10646}},{"key":"2:018","mentions":28,"segments":24,"excerpt":{"text":"Hey everyone, my name is Afan, and I'm a lead for compute trading at Jump Trading, which is a proprietary trading firm, alongside my co-lead over here, Alejandro.","seconds":11722}},{"key":"2:020","mentions":21,"segments":20,"excerpt":{"text":"So we have seen a lot of block trading happening around the term structure.","seconds":14093}},{"key":"2:024","mentions":2,"segments":2,"excerpt":{"text":"The senior analyst decides whether a variance is an accounting timing difference or a real cash shortfall.","seconds":16641}},{"key":"2:026","mentions":4,"segments":4,"excerpt":{"text":"Investing in spending more on AI fueled more of an improvement in cost per PR compared to peers who leaned in less.","seconds":17460}},{"key":"2:029","mentions":24,"segments":21,"excerpt":{"text":"This is part of AIE New York, which of course is very, very focused on the finance and, like, AI in finance in 2026.","seconds":18197}},{"key":"2:031","mentions":19,"segments":15,"excerpt":{"text":"So really it gets down to, like, where do we want to invest heavily and where are we willing to invest conservatively?","seconds":20083}},{"key":"2:033","mentions":7,"segments":7,"excerpt":{"text":"That isn't just true in finance; it's true in a bunch of different industries.","seconds":21911}},{"key":"2:035","mentions":6,"segments":5,"excerpt":{"text":"So hopefully we have many finance people or, like, engineers working in finance here.","seconds":22353}},{"key":"2:042","mentions":3,"segments":2,"excerpt":{"text":"Many of our customers are in highly regulated industries, from federal government agencies to banks and investment management.","seconds":27425}},{"key":"2:045","mentions":47,"segments":32,"excerpt":{"text":"I think in finance broadly, especially in, call it the asset management, investment banking, the bank institutional world, I still think it's very early.","seconds":28962}},{"key":"2:047","mentions":7,"segments":5,"excerpt":{"text":"So in this example, earlier I was showing the process of maybe doing the monthly accounting work for a client at an accounting firm.","seconds":29562}}]},{"concept":"Workflow","mentions":170,"talks":31,"coverage":0.6596,"day1_talks":14,"day2_talks":17,"day1_coverage":0.6087,"day2_coverage":0.7083,"day1_per_10000_words":13.28,"day2_per_10000_words":12.14,"day1_mean_talk_rate":16.95,"day2_mean_talk_rate":13.5,"day_shift_pp":9.96,"shift_ci_low_pp":-15.94,"shift_ci_high_pp":35.69,"id":"workflow","guide":"workflows-and-trust","primary":true,"pattern":"\\b(?:workflows?|orchestrat\\w*|dag|dags|deterministic|processes)\\b","matches":[{"key":"1:003","mentions":2,"segments":2,"excerpt":{"text":"It unlocked new workflows, but it didn't change how our researchers worked.","seconds":1570}},{"key":"1:006","mentions":11,"segments":11,"excerpt":{"text":"Then we want to give it context for our actual business rules and processes via skills.","seconds":2343}},{"key":"1:008","mentions":4,"segments":4,"excerpt":{"text":"And if people are going to route around slow processes, then we wind up in the same situation that we had before.","seconds":3669}},{"key":"1:010","mentions":3,"segments":3,"excerpt":{"text":"How do you make it so that every test is fast and deterministic without anybody having to be a deep expert in testing?","seconds":5090}},{"key":"1:013","mentions":3,"segments":2,"excerpt":{"text":"Actually, doing code review, it's actually how do you set up systems and agents to do the code reviews, and how do you write evals to actually do the non-deterministic, to review those non-deterministic tasks.","seconds":5401}},{"key":"1:020","mentions":7,"segments":7,"excerpt":{"text":"And so if the user says, Oh, I want you to go read the CEO's email, that's probably going to get caught in deterministic policies right away, right?","seconds":10473}},{"key":"1:022","mentions":3,"segments":3,"excerpt":{"text":"Our processes change, tax rules change, and a skill that was right in March can be wrong by December.","seconds":12813}},{"key":"1:023","mentions":2,"segments":2,"excerpt":{"text":"Sandboxes bound what the processes can do, but they don't control what is authorized to act outside the sandbox itself.","seconds":13480}},{"key":"1:029","mentions":16,"segments":15,"excerpt":{"text":"In eleven deterministic stages, it turns the graph into an execution DAG or it fails closed.","seconds":17946}},{"key":"1:030","mentions":3,"segments":3,"excerpt":{"text":"It also keeps the codebase that the other AI services and tools we use across Flywire can start workflows directly.","seconds":18501}},{"key":"1:038","mentions":15,"segments":15,"excerpt":{"text":"And such a workflow of being a faster horse introduces one key nuance.","seconds":23738}},{"key":"1:043","mentions":7,"segments":7,"excerpt":{"text":"Most things have deterministic code plus LLM plus statistical method, or a human with LLM in conjunction.","seconds":26734}},{"key":"1:046","mentions":2,"segments":2,"excerpt":{"text":"So the way that they're going to be released is we still expect you to use an orchestrator model.","seconds":27671}},{"key":"1:049","mentions":3,"segments":3,"excerpt":{"text":"The new application is no longer necessary that it is deterministic and solid and cold hard code.","seconds":28825}},{"key":"2:002","mentions":2,"segments":2,"excerpt":{"text":"So a much faster, but just a smart model that could then be used in tandem with the really smart frontier orchestrator.","seconds":1555}},{"key":"2:004","mentions":2,"segments":2,"excerpt":{"text":"So we look at deterministic mixture methods, and then you add in agentic mixture methods, and it comes up with what we think is the gold standard of verification.","seconds":2928}},{"key":"2:005","mentions":21,"segments":17,"excerpt":{"text":"Apache Airflow, if you're familiar, is an example of open source software that represents workflows, their dependencies as DAGs in Python code.","seconds":3375}},{"key":"2:007","mentions":5,"segments":5,"excerpt":{"text":"So we focused on helping the model produce those outputs in a way that fits the firm's existing workflows.","seconds":4638}},{"key":"2:010","mentions":6,"segments":6,"excerpt":{"text":"Every tool call needs to go through some sort of deterministic broker.","seconds":6226}},{"key":"2:012","mentions":4,"segments":3,"excerpt":{"text":"If you look at the anatomy on a deal team, right, the type of tasks that matter for long-running type of workflows, for example, going from a data room into all the sorts of materials that make its way in front of an investment committee, it doesn't work for the single-player siloed workflows.","seconds":7388}},{"key":"2:023","mentions":4,"segments":4,"excerpt":{"text":"It's provide a graph-based architecture that support loops, branching, state management for complex multi-agent AI workflows, right?","seconds":15193}},{"key":"2:024","mentions":3,"segments":3,"excerpt":{"text":"The moment an agent steps out of a sandbox and into a core business workflow, the entire test changes.","seconds":16219}},{"key":"2:026","mentions":3,"segments":3,"excerpt":{"text":"The research that I'll present is drawn from the Jellyfish platform, the industry's largest set of engineering workflow data joined to AI telemetry data.","seconds":17052}},{"key":"2:029","mentions":4,"segments":4,"excerpt":{"text":"And that drives some of the biggest processes at any asset manager.","seconds":18498}},{"key":"2:031","mentions":4,"segments":4,"excerpt":{"text":"And so I think, like a lot of companies in, like, early 2024, you know, we were trying to deploy these models as part of, like, our workflow and as part of, like, our production software.","seconds":19609}},{"key":"2:033","mentions":3,"segments":3,"excerpt":{"text":"Now, for those of you who are going to say we have deterministic models, great, we do.","seconds":21309}},{"key":"2:035","mentions":2,"segments":2,"excerpt":{"text":"We have hooks to protect certain deterministic logic.","seconds":23363}},{"key":"2:037","mentions":2,"segments":2,"excerpt":{"text":"So I guess the takeaway here is be intentional about how you build these workflows and don't just assume that a YOLO method is going to work out.","seconds":24587}},{"key":"2:039","mentions":10,"segments":8,"excerpt":{"text":"And even if you do get great cache rates and cache costs go down, you still have to, like, reprocess a lot of input to redo a workflow every single time.","seconds":25936}},{"key":"2:045","mentions":3,"segments":2,"excerpt":{"text":"So we've been very forward and open-minded about adopting new technologies and bringing those into our workflows and how we can be more efficient and free up capacity for people to be more thoughtful and think and replace manual processes.","seconds":28864}},{"key":"2:047","mentions":8,"segments":5,"excerpt":{"text":"And those processes, if we believe the people who put them in place think that those processes will generalize to good outcomes, then they can be comfortable that even though the people executing the processes are not deterministic, you will generally get to good outcomes if you follow the process.","seconds":29542}}]},{"concept":"Deployment","mentions":157,"talks":28,"coverage":0.5957,"day1_talks":15,"day2_talks":13,"day1_coverage":0.6522,"day2_coverage":0.5417,"day1_per_10000_words":11.82,"day2_per_10000_words":11.59,"day1_mean_talk_rate":11.35,"day2_mean_talk_rate":11.63,"day_shift_pp":-11.05,"shift_ci_low_pp":-36.59,"shift_ci_high_pp":18.57,"id":"deployment","guide":"agent-systems","primary":false,"pattern":"\\b(?:deploy\\w*|production|ship(?:ping|ped)?)\\b","matches":[{"key":"1:003","mentions":2,"segments":2,"excerpt":{"text":"We have two amazing folks from Bridgewater who are in the weeds of production AI every single day.","seconds":836}},{"key":"1:006","mentions":6,"segments":6,"excerpt":{"text":"Let's say that a new model is shipped, or we want to make an adjustment to the business context layer with the skills.","seconds":2600}},{"key":"1:008","mentions":10,"segments":9,"excerpt":{"text":"So a thing that summarizes Jira or linear tickets is really different than a production deployment agent that can touch a running system.","seconds":3652}},{"key":"1:010","mentions":4,"segments":4,"excerpt":{"text":"So, hey, you can only deploy to this environment that has access to only your data.","seconds":4831}},{"key":"1:013","mentions":7,"segments":5,"excerpt":{"text":"You can ship it, and then once it goes into production, the goal is to actually trace everything.","seconds":6017}},{"key":"1:016","mentions":3,"segments":3,"excerpt":{"text":"So in the model we have in production, it's a NextGBoost model, and it's trained on a large amount of transactions, which is fraudulent or non-fraudulent.","seconds":7756}},{"key":"1:020","mentions":12,"segments":9,"excerpt":{"text":"I know when it comes to building, deploying, and operating agents in production, there's many different outcomes teams should be thinking about, such as discovering what's running, what the agent's capabilities are, the ability to validate inputs and outputs, constraining an agent like we've talked about using the egress proxy and different networking rules.","seconds":11679}},{"key":"1:022","mentions":2,"segments":2,"excerpt":{"text":"Step five: ship it with an owner and a reviewer.","seconds":12807}},{"key":"1:023","mentions":6,"segments":6,"excerpt":{"text":"It calls the endpoint to restart the server, which brings production offline.","seconds":13193}},{"key":"1:027","mentions":2,"segments":2,"excerpt":{"text":"And for the agent to effectively work with Railway and be able to deploy everything, you just give it the Railway agent skills, and it was just able to do it.","seconds":16025}},{"key":"1:028","mentions":5,"segments":4,"excerpt":{"text":"Finally, working pipelines get deployed to production, where they can slowly mature.","seconds":17320}},{"key":"1:031","mentions":2,"segments":2,"excerpt":{"text":"And I think this is where a lot of people deploying agents are today.","seconds":19338}},{"key":"1:036","mentions":3,"segments":3,"excerpt":{"text":"So it's going to be when an engineer picks up a feature until a feature is merged and ready to be deployed.","seconds":21916}},{"key":"1:038","mentions":4,"segments":2,"excerpt":{"text":"And as you can see and imagine, you can have forward deployed engineers, you can have forward deployed product managers, but you could also have the inverse, which is forward deployed portfolio managers.","seconds":24078}},{"key":"1:046","mentions":2,"segments":2,"excerpt":{"text":"And ideally what we want is somebody who has experience shipping end-to-end AI products, but it doesn't have to be on the research side.","seconds":27887}},{"key":"2:002","mentions":4,"segments":4,"excerpt":{"text":"And even back then, we decided, well, we need to ship these things.","seconds":1359}},{"key":"2:004","mentions":5,"segments":5,"excerpt":{"text":"We haven't figured out how to scale from producing stuff to shipping stuff.","seconds":2450}},{"key":"2:010","mentions":3,"segments":3,"excerpt":{"text":"Build fast, deploy securely.","seconds":6291}},{"key":"2:012","mentions":3,"segments":3,"excerpt":{"text":"Today, for people who have deployed Claude or ChatGPT in their institutions, it's all single-player.","seconds":7404}},{"key":"2:014","mentions":13,"segments":13,"excerpt":{"text":"A lower production cost gives more room to compete when prices fall.","seconds":8777}},{"key":"2:016","mentions":3,"segments":3,"excerpt":{"text":"So, the clusters then are delayed because the cloud couldn't actually deploy the cluster until they had the customer in hand.","seconds":10733}},{"key":"2:020","mentions":4,"segments":3,"excerpt":{"text":"Basically, you can think about as a dedicated deployment, on-premise deployment, right?","seconds":13539}},{"key":"2:023","mentions":12,"segments":9,"excerpt":{"text":"Don't try to be in rush to building some agent quickly and deploy ship it.","seconds":15999}},{"key":"2:024","mentions":6,"segments":6,"excerpt":{"text":"The one reason to delay deployment is a decision with no assigned owner.","seconds":16687}},{"key":"2:031","mentions":8,"segments":7,"excerpt":{"text":"And so I think, like a lot of companies in, like, early 2024, you know, we were trying to deploy these models as part of, like, our workflow and as part of, like, our production software.","seconds":19609}},{"key":"2:037","mentions":2,"segments":2,"excerpt":{"text":"But when it comes to the actual, like, subtleties of how projects are sort of run and deployed, it's really interesting to look at the actual traces, you know, for users and engineers who opt into that, right?","seconds":24771}},{"key":"2:039","mentions":5,"segments":4,"excerpt":{"text":"Around turning their traces from deploying a model in production for a high-volume use case into rich, robust environments that are blended real and synthetic that kind of capture the complexity of the real tasks they want to be using long-running agents for.","seconds":26399}},{"key":"2:047","mentions":12,"segments":12,"excerpt":{"text":"You can give it the situation and see if you can recreate the situation that you saw in your eval or in production.","seconds":30333}}]},{"concept":"Trust","mentions":147,"talks":24,"coverage":0.5106,"day1_talks":12,"day2_talks":12,"day1_coverage":0.5217,"day2_coverage":0.5,"day1_per_10000_words":7.93,"day2_per_10000_words":13.52,"day1_mean_talk_rate":9.29,"day2_mean_talk_rate":15.32,"day_shift_pp":-2.17,"shift_ci_low_pp":-31.8,"shift_ci_high_pp":24.01,"id":"trust","guide":"workflows-and-trust","primary":true,"pattern":"\\b(?:trust\\w*|confiden\\w*|reliab\\w*)\\b","matches":[{"key":"1:006","mentions":9,"segments":9,"excerpt":{"text":"One example here is Citadel told us at our briefing event in May that their investment professionals started to use Claude for Excel, and it gives them a step change function in how they can test, build, and trust their work.","seconds":2035}},{"key":"1:010","mentions":2,"segments":2,"excerpt":{"text":"If it's there, it feels and seems free, and it is reliable.","seconds":5064}},{"key":"1:013","mentions":3,"segments":2,"excerpt":{"text":"And do I really trust that the eval is, you know, is going to catch, or, you know, how do I trust that it's as good as my judgment?","seconds":5944}},{"key":"1:020","mentions":5,"segments":5,"excerpt":{"text":"It's how they build trust in their business, and maintaining the business continuity is bar first, like the first priority.","seconds":9920}},{"key":"1:022","mentions":3,"segments":3,"excerpt":{"text":"If you write the questions and set the limits, you can get something you can trust.","seconds":12757}},{"key":"1:025","mentions":2,"segments":2,"excerpt":{"text":"Across our testing, only two least confident answers ended up being wrong ones.","seconds":14924}},{"key":"1:028","mentions":3,"segments":3,"excerpt":{"text":"You ask for a transformation, whatever, and you get a confident yes, sir.","seconds":16651}},{"key":"1:029","mentions":7,"segments":7,"excerpt":{"text":"Given agent the whole task and wait for the next model to make it reliable.","seconds":17834}},{"key":"1:031","mentions":2,"segments":2,"excerpt":{"text":"So if you were provisioning this, it doesn't make sense, and then reliably get them when the sandbox needs them, and only pay for them for that duration.","seconds":18844}},{"key":"1:033","mentions":2,"segments":2,"excerpt":{"text":"Reliability: this is usually a function of the number of tasks or the sample count.","seconds":20179}},{"key":"1:046","mentions":2,"segments":2,"excerpt":{"text":"If you, all of a sudden, if you can reliably orchestrate robots in space and time, and they can take text instructions and do things, all of a sudden, this is a recursive problem, right?","seconds":27727}},{"key":"1:049","mentions":3,"segments":3,"excerpt":{"text":"I've seen where this goes, and we've seen where the moats will lie: network effects, reliability, and reputation.","seconds":28969}},{"key":"2:004","mentions":18,"segments":13,"excerpt":{"text":"We're actually building confidence, and you're seeing results: better security, more confidence in our time to market, improved, or a reduction in number of outages, improved reliability.","seconds":3000}},{"key":"2:005","mentions":16,"segments":14,"excerpt":{"text":"So I'm a believer in quality, and in traditional software, we've built confidence mechanisms in place to give us confidence that the software is of high quality.","seconds":3792}},{"key":"2:007","mentions":3,"segments":3,"excerpt":{"text":"For any given task, there's a threshold of usefulness, the point where a model becomes capable and reliable enough to save you time.","seconds":4995}},{"key":"2:010","mentions":4,"segments":4,"excerpt":{"text":"We're going to cover identity and trust, prompt and input defense, to an action governance, execution isolation, data protection, and governance observability.","seconds":5553}},{"key":"2:012","mentions":7,"segments":7,"excerpt":{"text":"You can delegate a task, book a restaurant, book a flight, and it'll do it reliably now.","seconds":7254}},{"key":"2:016","mentions":2,"segments":2,"excerpt":{"text":"isn't sufficiently standardized or reliable for the market operator to be divorced from the physical infrastructure.","seconds":10913}},{"key":"2:018","mentions":2,"segments":2,"excerpt":{"text":"They want confidence that their loan will get repaid.","seconds":12222}},{"key":"2:024","mentions":5,"segments":5,"excerpt":{"text":"Answering those three questions is what earns it the trust to run at scale.","seconds":17012}},{"key":"2:033","mentions":11,"segments":11,"excerpt":{"text":"Is it, let's say, some analyst who I really trust more than anyone else?","seconds":21262}},{"key":"2:037","mentions":2,"segments":2,"excerpt":{"text":"What you do is you have some sort of belief about something you care about, and the longer it's been since you've seen or observed evidence that supports it, the less you trust it, right?","seconds":24119}},{"key":"2:045","mentions":5,"segments":4,"excerpt":{"text":"So to that end, we're able to kind of have trust in the system and You know, appreciate that.","seconds":28739}},{"key":"2:047","mentions":18,"segments":14,"excerpt":{"text":"We are based right here in New York City, and the thing that we focus on is how can we get agents to do production-scale work reliably enough so that we are trusted and used end-to-end by accounting firms across the country.","seconds":29406}}]},{"concept":"Context","mentions":129,"talks":22,"coverage":0.4681,"day1_talks":10,"day2_talks":12,"day1_coverage":0.4348,"day2_coverage":0.5,"day1_per_10000_words":10.36,"day2_per_10000_words":8.97,"day1_mean_talk_rate":10.08,"day2_mean_talk_rate":10.26,"day_shift_pp":6.52,"shift_ci_low_pp":-19.2,"shift_ci_high_pp":36.23,"id":"context","guide":"company-context","primary":true,"pattern":"\\bcontext\\w*\\b","matches":[{"key":"1:006","mentions":9,"segments":9,"excerpt":{"text":"Then we want to give it context for our actual business rules and processes via skills.","seconds":2343}},{"key":"1:013","mentions":2,"segments":2,"excerpt":{"text":"Agent gets triggered, agent gets started, loaded with all the context, the evidence.","seconds":5775}},{"key":"1:016","mentions":2,"segments":2,"excerpt":{"text":"So it could be classification task for languages, but in our context for transaction foundation model.","seconds":7192}},{"key":"1:018","mentions":5,"segments":5,"excerpt":{"text":"We've been closing all of our context gaps so that agents are fully aware.","seconds":8977}},{"key":"1:020","mentions":24,"segments":23,"excerpt":{"text":"Yeah, and this is how these things we're talking about, which seemed like geeky OSS problems, actually then merge back up to real agent engineering problems, which is if you actually understand what you're doing, one of the biggest complaints, less now, but one of the biggest complaints with MCP was context bloat, especially when we had smaller context windows, right?","seconds":11622}},{"key":"1:023","mentions":2,"segments":2,"excerpt":{"text":"Some context on this demo: we're going to look at an on-call triage agent that reads and writes to an incident pager and resolves or escalates tickets.","seconds":12986}},{"key":"1:028","mentions":4,"segments":3,"excerpt":{"text":"And pulling these schemas together is the process of transformation, of making information actually useful, taking the context that it was recorded in and merging it with the context of the decision.","seconds":16442}},{"key":"1:030","mentions":2,"segments":2,"excerpt":{"text":"Before we dive into the deep technical details of our AI platform, let us give you some essential context about Flywire.","seconds":18408}},{"key":"1:038","mentions":4,"segments":4,"excerpt":{"text":"It's the taste that tells you what good looks like in your given context.","seconds":23609}},{"key":"1:043","mentions":6,"segments":5,"excerpt":{"text":"So at a high level, you can think about a hedge fund as a function that takes in two parameters, context external and context internal, and spits out one tuple: selection, size, timing.","seconds":26128}},{"key":"2:002","mentions":8,"segments":6,"excerpt":{"text":"You had to reduce and scope down the context as much as possible because giving too much context to the agent would just confuse it.","seconds":1350}},{"key":"2:005","mentions":2,"segments":2,"excerpt":{"text":"So ideally, we've established a trusted context for our agent, and there's some references on that.","seconds":3716}},{"key":"2:007","mentions":6,"segments":6,"excerpt":{"text":"You can give it a project, share context, and ask for an update.","seconds":5264}},{"key":"2:012","mentions":8,"segments":8,"excerpt":{"text":"It's connected to the systems of context that matter most for you: email, Drive, calendar.","seconds":7278}},{"key":"2:020","mentions":2,"segments":2,"excerpt":{"text":"I'll also give you some context who I am, what I do.","seconds":13472}},{"key":"2:023","mentions":10,"segments":9,"excerpt":{"text":"Now if I ask, now I enable the context optimizer behind the scene, and I'm rerun the exact same query again.","seconds":15902}},{"key":"2:031","mentions":3,"segments":3,"excerpt":{"text":"I remember when we had our first team get together, we were at Stanford's campus for a week, and, I mean, I'm probably misremembering the exact timeline, but it felt like we went from 4K to 8K to 128K context windows, and then we had multimodal, and everything just felt like it was kind of coming all at once.","seconds":19589}},{"key":"2:035","mentions":4,"segments":3,"excerpt":{"text":"Context is very valuable for us, so we want to decouple them and separate them into much more slices where we can then have, like, dedicated references and scripts if we want the agent to, like, do very certain things where it needs more context.","seconds":22831}},{"key":"2:037","mentions":3,"segments":3,"excerpt":{"text":"But I think also, most importantly, you have the sort of beginning of a context layer that every agent can rely on.","seconds":24889}},{"key":"2:039","mentions":6,"segments":6,"excerpt":{"text":"And so there is a question of, like, okay, what goes in context, what goes in the weights?","seconds":25985}},{"key":"2:042","mentions":4,"segments":4,"excerpt":{"text":"Before I get into the stories, though, let me set up a little bit more context.","seconds":27702}},{"key":"2:047","mentions":5,"segments":4,"excerpt":{"text":"And the idea of a behavior spec is essentially saying, before we update the context or write the context, what do we actually want?","seconds":30122}}]},{"concept":"Security","mentions":174,"talks":20,"coverage":0.4255,"day1_talks":8,"day2_talks":12,"day1_coverage":0.3478,"day2_coverage":0.5,"day1_per_10000_words":15.71,"day2_per_10000_words":10.62,"day1_mean_talk_rate":15.73,"day2_mean_talk_rate":12.78,"day_shift_pp":15.22,"shift_ci_low_pp":-14.23,"shift_ci_high_pp":42.92,"id":"security","guide":"agent-security","primary":true,"pattern":"\\b(?:secur\\w*|sandbox\\w*|prompt injection|attacks?|vulnerabilit\\w*|exfiltrat\\w*)\\b","matches":[{"key":"1:006","mentions":8,"segments":7,"excerpt":{"text":"Now, so far we've talked about building a secure, trustable agent foundation framework.","seconds":2749}},{"key":"1:008","mentions":3,"segments":2,"excerpt":{"text":"We don't want to sandbox or lock down everything because a sandbox with no real access to systems is safe, but it's not very interesting or useful.","seconds":3427}},{"key":"1:013","mentions":9,"segments":8,"excerpt":{"text":"And so the agent's triggered, we build a sandbox, we put everything that we need inside of that sandbox.","seconds":5639}},{"key":"1:020","mentions":19,"segments":18,"excerpt":{"text":"So the core of the question is like, okay, well, if we give these, like, agents, we give the brain, which is the model and the LLM, and you have these hands, which you kind of give them, put them in a sandbox, and inside the sandbox is the tools.","seconds":11477}},{"key":"1:023","mentions":17,"segments":16,"excerpt":{"text":"Sandboxes bound what the processes can do, but they don't control what is authorized to act outside the sandbox itself.","seconds":13480}},{"key":"1:025","mentions":9,"segments":9,"excerpt":{"text":"Our own product at work called the attack straight out of the box.","seconds":15190}},{"key":"1:031","mentions":23,"segments":21,"excerpt":{"text":"You could vibe code a sandbox platform, but who is going to do all the work to, like, continually reason about the security boundaries and make sure it's defensible in an age when everything is...","seconds":19133}},{"key":"1:036","mentions":4,"segments":4,"excerpt":{"text":"And the sandbox means now if you've used Codex Cloud Code, it has all of these features of approve for me.","seconds":22440}},{"key":"2:002","mentions":4,"segments":4,"excerpt":{"text":"At the end of all of this, we have actually found seven high severity vulnerabilities that we all patched.","seconds":1744}},{"key":"2:004","mentions":5,"segments":4,"excerpt":{"text":"It can also work really well to help you hack stuff, help you attack stuff, find vulnerabilities, find quality issues, etc.","seconds":2348}},{"key":"2:005","mentions":2,"segments":2,"excerpt":{"text":"If you're shipping code that gets canceled or doesn't meet security standards, that's something you should want to know about.","seconds":3566}},{"key":"2:010","mentions":28,"segments":28,"excerpt":{"text":"It's just such a great H2B in the age of AI and security.","seconds":5376}},{"key":"2:014","mentions":4,"segments":4,"excerpt":{"text":"Ensuring cluster security takes ongoing work, just as delivering the promised performance does.","seconds":9034}},{"key":"2:023","mentions":5,"segments":5,"excerpt":{"text":"If everything looks great, security approved, you can move that image back into the registry, right?","seconds":14753}},{"key":"2:024","mentions":2,"segments":2,"excerpt":{"text":"The moment an agent steps out of a sandbox and into a core business workflow, the entire test changes.","seconds":16219}},{"key":"2:029","mentions":5,"segments":5,"excerpt":{"text":"Yeah, and there actually are efforts right now to securitize GPUs, GPU clusters.","seconds":18779}},{"key":"2:031","mentions":2,"segments":2,"excerpt":{"text":"It's like, you know, data security for us is, you know, is, is, is, like, as paramount as delivering quick.","seconds":20324}},{"key":"2:033","mentions":6,"segments":5,"excerpt":{"text":"Okay, so I want to talk a little bit about security here because it's super important.","seconds":21626}},{"key":"2:035","mentions":9,"segments":9,"excerpt":{"text":"And the benefit here is that there's no way for a prompt injection, right?","seconds":23066}},{"key":"2:037","mentions":3,"segments":3,"excerpt":{"text":"At Citadel Securities, the system that we've built has agents at every part of this particular loop, right?","seconds":24327}}]},{"concept":"Evaluation","mentions":151,"talks":18,"coverage":0.383,"day1_talks":8,"day2_talks":10,"day1_coverage":0.3478,"day2_coverage":0.4167,"day1_per_10000_words":12.95,"day2_per_10000_words":9.79,"day1_mean_talk_rate":13.37,"day2_mean_talk_rate":8.69,"day_shift_pp":6.88,"shift_ci_low_pp":-18.84,"shift_ci_high_pp":36.41,"id":"evaluation","guide":"evaluation","primary":true,"pattern":"\\b(?:evals?|evaluat\\w*|benchmarks?)\\b","matches":[{"key":"1:006","mentions":3,"segments":3,"excerpt":{"text":"Once you see the list of proposed adjustments, you can solidify that as a scale change after you've reviewed and tested all that within your eval framework.","seconds":2783}},{"key":"1:008","mentions":12,"segments":9,"excerpt":{"text":"So again, one tempting answer here is let's have some kind of central evaluation methodology that waits for a full eval suite and a committee sign-off.","seconds":3978}},{"key":"1:013","mentions":25,"segments":20,"excerpt":{"text":"We have a skill with Arize Evaluator to actually help you build and write evals, as well as align those evals with the human judgment.","seconds":5918}},{"key":"1:018","mentions":2,"segments":2,"excerpt":{"text":"These are all signals that our background agents are continuously evaluating to progress work.","seconds":8695}},{"key":"1:023","mentions":6,"segments":6,"excerpt":{"text":"It evaluates the delegation chain, and it looks at the requested access.","seconds":13929}},{"key":"1:028","mentions":2,"segments":2,"excerpt":{"text":"So in one benchmark, four different model families produced the same total down to the cent because they all shared a signed convention that the answer key didn't.","seconds":17147}},{"key":"1:029","mentions":2,"segments":2,"excerpt":{"text":"Logging and evals exist today.","seconds":18256}},{"key":"1:033","mentions":22,"segments":20,"excerpt":{"text":"The goal here is to try and build evals that are reflective of the actual work people are doing and the work output they want to see, not to build these contrived evals that the models can actually game.","seconds":20285}},{"key":"2:005","mentions":6,"segments":6,"excerpt":{"text":"We need to build representative data sets and evaluate them and define what good looks like.","seconds":3849}},{"key":"2:007","mentions":2,"segments":2,"excerpt":{"text":"For example, before GPT-4 finished training, OpenAI researchers were able to predict its final loss as well as its performance on a coding eval.","seconds":4810}},{"key":"2:010","mentions":2,"segments":2,"excerpt":{"text":"Think of this as your guiding principle as you evaluate security for the agents.","seconds":5709}},{"key":"2:012","mentions":3,"segments":3,"excerpt":{"text":"We evaluated a lot: the desktop plugins of Excel, PowerPoint, and Word.","seconds":7531}},{"key":"2:014","mentions":5,"segments":5,"excerpt":{"text":"The benchmark then needs to track the rent they actually pay or earn.","seconds":9583}},{"key":"2:023","mentions":15,"segments":14,"excerpt":{"text":"Third, do we have like enough tracking and evaluation system to make sure that every query when you are sending to the agents, whether they've been tracked or not, and are you evaluating the quality response or not, right?","seconds":14453}},{"key":"2:031","mentions":9,"segments":7,"excerpt":{"text":"But going back to what John mentioned, how we do our data capture and distillation and how we make How we make this available is, like, state to these models, and then also just how we do our benchmarks and evaluation.","seconds":20174}},{"key":"2:037","mentions":2,"segments":2,"excerpt":{"text":"And what that allows you to do is to actually do some robust evals down the line, right?","seconds":24458}},{"key":"2:039","mentions":6,"segments":5,"excerpt":{"text":"But the agents are quite quickly getting to the point where, thankfully, we have robots that can do a lot of the eval curation for us, especially as we train them more on the best practices of making evals in ways that are robust and amenable to learning.","seconds":26286}},{"key":"2:047","mentions":14,"segments":12,"excerpt":{"text":"And when I say outcome-based evals, I mean what we traditionally think of as evals.","seconds":29940}}]},{"concept":"Observability","mentions":158,"talks":18,"coverage":0.383,"day1_talks":8,"day2_talks":10,"day1_coverage":0.3478,"day2_coverage":0.4167,"day1_per_10000_words":12.95,"day2_per_10000_words":10.76,"day1_mean_talk_rate":12.13,"day2_mean_talk_rate":10.75,"day_shift_pp":6.88,"shift_ci_low_pp":-22.64,"shift_ci_high_pp":36.41,"id":"observability","guide":"agent-systems","primary":true,"pattern":"\\b(?:observabil\\w*|telemetry|traces?|tracing|monitor\\w*|audit\\w*|logs?|logging)\\b","matches":[{"key":"1:013","mentions":37,"segments":24,"excerpt":{"text":"It's actually no longer even about just setting up metrics and software kind of telemetry, but it's actually how do you have agents go through and read the observability data, and have agents actually consume that telemetry to find patterns.","seconds":5413}},{"key":"1:020","mentions":11,"segments":8,"excerpt":{"text":"Ian, why is it important to have your audit log separate from the agent itself, meaning outside of its control?","seconds":12035}},{"key":"1:023","mentions":12,"segments":8,"excerpt":{"text":"If you remember the audit log, we couldn't tell who was using the API keys.","seconds":13503}},{"key":"1:025","mentions":6,"segments":6,"excerpt":{"text":"The mean distance from an integer across 22 traces is just 0.","seconds":14846}},{"key":"1:028","mentions":2,"segments":2,"excerpt":{"text":"45 on second opinions, settling disagreements, and the audit against 128 for first answers.","seconds":16909}},{"key":"1:029","mentions":3,"segments":3,"excerpt":{"text":"At typed points, you can inspect and use every judgment you log to need less of it next time.","seconds":18340}},{"key":"1:031","mentions":2,"segments":2,"excerpt":{"text":"And the final underrated thing is it's really important to have really great observability.","seconds":19039}},{"key":"1:038","mentions":2,"segments":2,"excerpt":{"text":"Being cognizant that you need to redesign your system and really focus it on key metrics that you're evaluating, you're monitoring, and you're creating a flywheel off.","seconds":24383}},{"key":"2:005","mentions":6,"segments":4,"excerpt":{"text":"So, combining with evals as sort of your runtime logs of what happened, or sorry, traces, which are sort of your runtime logs of what happens, we've built this confidence mechanism to complement traditional software.","seconds":3874}},{"key":"2:010","mentions":8,"segments":8,"excerpt":{"text":"We definitely want an audit trail, guardrails outside, circuit breakers, supply chain verification of the serial tasting.","seconds":6048}},{"key":"2:014","mentions":4,"segments":3,"excerpt":{"text":"In our published security work, we found monitoring access that exposed another customer's logs and metrics.","seconds":9015}},{"key":"2:023","mentions":31,"segments":24,"excerpt":{"text":"So the way the monitoring work is usually like the moment you submit any calls to your agent, it will send all the operation log into the log store.","seconds":15073}},{"key":"2:024","mentions":7,"segments":5,"excerpt":{"text":"In the third line, internal audit provides assurance, deciding based on material risk when and how the agent enters the annual audit plan.","seconds":16559}},{"key":"2:026","mentions":2,"segments":2,"excerpt":{"text":"The research that I'll present is drawn from the Jellyfish platform, the industry's largest set of engineering workflow data joined to AI telemetry data.","seconds":17052}},{"key":"2:035","mentions":6,"segments":6,"excerpt":{"text":"It should use our portfolio auditing skill, and it runs the request.","seconds":23231}},{"key":"2:037","mentions":3,"segments":3,"excerpt":{"text":"You have meeting notes and you have Slack and code changes and even your session traces with your favorite harness.","seconds":24096}},{"key":"2:039","mentions":4,"segments":4,"excerpt":{"text":"These traces have your prompts and they have your feedback to things that models did.","seconds":26268}},{"key":"2:041","mentions":2,"segments":2,"excerpt":{"text":"And then finally, before running the code, your agent should carry out an audit against all those steps.","seconds":27352}}]},{"concept":"Compute","mentions":220,"talks":15,"coverage":0.3191,"day1_talks":5,"day2_talks":10,"day1_coverage":0.2174,"day2_coverage":0.4167,"day1_per_10000_words":3.72,"day2_per_10000_words":27.18,"day1_mean_talk_rate":4.03,"day2_mean_talk_rate":24.89,"day_shift_pp":19.93,"shift_ci_low_pp":-5.89,"shift_ci_high_pp":45.11,"id":"compute","guide":"compute-economics","primary":true,"pattern":"\\b(?:compute|computing|clusters?|data centers?|datacenters?|capacity)\\b","matches":[{"key":"1:025","mentions":5,"segments":4,"excerpt":{"text":"What's isolated are compute units, memory capacity, bandwidth quotas, and false errors.","seconds":14659}},{"key":"1:028","mentions":2,"segments":2,"excerpt":{"text":"We first made compute cheaper than storage, and then we gave it to everybody, and then we made it faster.","seconds":16177}},{"key":"1:031","mentions":4,"segments":4,"excerpt":{"text":"So one thing we spend a lot of time doing is actually aggregating cloud capacity from a bunch of hyperscalers and neoclouds.","seconds":19237}},{"key":"1:038","mentions":5,"segments":5,"excerpt":{"text":"They're focused on aligning engineering, analytics, AI capacity, and judgment towards a singular focus.","seconds":23891}},{"key":"1:046","mentions":2,"segments":2,"excerpt":{"text":"One, you actually need to spend the compute to simulate environments, and the more realistic you try to simulate, the more expensive that gets.","seconds":27425}},{"key":"2:005","mentions":3,"segments":3,"excerpt":{"text":"So we want to add parallel capacity by still keeping the controls.","seconds":3634}},{"key":"2:007","mentions":5,"segments":5,"excerpt":{"text":"Which meant scaling compute by making the models bigger and adding more data.","seconds":4802}},{"key":"2:014","mentions":15,"segments":15,"excerpt":{"text":"Now let's look at the compute that an intelligence producer is buying.","seconds":9464}},{"key":"2:016","mentions":60,"segments":47,"excerpt":{"text":"So this is a talk about compute markets, but it's kind of more of a talk about why to make a compute market.","seconds":10199}},{"key":"2:018","mentions":60,"segments":50,"excerpt":{"text":"So today we're going to be looking over some of the market structure as it applies to compute trading, what compute trading even is, what we can see it becoming.","seconds":11732}},{"key":"2:020","mentions":30,"segments":27,"excerpt":{"text":"But at the same time, when you need compute in the next 48 hours on reserved contracts, the easiest way is go to an exchange and then go to the verified compute resources, just like go to Nasdaq, NYSE to buy stocks, right?","seconds":13312}},{"key":"2:024","mentions":2,"segments":2,"excerpt":{"text":"Then you only count that value after netting out the compute, licensing, and human review costs of running the agent.","seconds":16451}},{"key":"2:029","mentions":8,"segments":6,"excerpt":{"text":"But so many people are now investing in, like, data centers and, like, clusters of GPUs.","seconds":18730}},{"key":"2:031","mentions":4,"segments":4,"excerpt":{"text":"It doesn't maps into the paradigm, like the, you know, how do you deal with increasing compute, making it easier to do these things.","seconds":20046}},{"key":"2:045","mentions":3,"segments":3,"excerpt":{"text":"So we've been very forward and open-minded about adopting new technologies and bringing those into our workflows and how we can be more efficient and free up capacity for people to be more thoughtful and think and replace manual processes.","seconds":28864}}]},{"concept":"Human review","mentions":53,"talks":14,"coverage":0.2979,"day1_talks":8,"day2_talks":6,"day1_coverage":0.3478,"day2_coverage":0.25,"day1_per_10000_words":5.34,"day2_per_10000_words":2.76,"day1_mean_talk_rate":4.96,"day2_mean_talk_rate":2.84,"day_shift_pp":-9.78,"shift_ci_low_pp":-35.69,"shift_ci_high_pp":15.94,"id":"human-review","guide":"workflows-and-trust","primary":true,"pattern":"\\b(?:human.in.the.loop|human (?:review|approval|oversight)|approvals?|reviewers?|human judgment)\\b","matches":[{"key":"1:006","mentions":4,"segments":4,"excerpt":{"text":"We can think about that in three general steps: who the agent is acting as, with an identity and permissions layer; how we can sandbox and determine the approval layer of the agent so that someone is signing off before it acts; and lastly, how we know the agent will still work once we adjust any variables over time.","seconds":2352}},{"key":"1:008","mentions":7,"segments":7,"excerpt":{"text":"And if we can use the same, if we try to use the same review process to cover all of these agents, then every agent is waiting in the same queue for weeks for human review.","seconds":3661}},{"key":"1:013","mentions":2,"segments":2,"excerpt":{"text":"But these evals can also, they need to be kind of aligned with human judgment.","seconds":5890}},{"key":"1:020","mentions":3,"segments":3,"excerpt":{"text":"At some point, agents are going to be doing so much work, there's no way we can have human in the loop approving everything, right?","seconds":11030}},{"key":"1:022","mentions":2,"segments":2,"excerpt":{"text":"The reviewer checks it before the team relies on it, and it's someone other than the author.","seconds":12820}},{"key":"1:023","mentions":7,"segments":6,"excerpt":{"text":"Now, forcing a human in the loop approval just kind of passes the access control on to a consent-fatigued user.","seconds":13339}},{"key":"1:029","mentions":5,"segments":5,"excerpt":{"text":"In the story, waiting, approvals, and follow-ups were most of the work.","seconds":18088}},{"key":"1:036","mentions":2,"segments":2,"excerpt":{"text":"And so how do you give reassurance that you have a human in the loop?","seconds":22432}},{"key":"2:004","mentions":2,"segments":2,"excerpt":{"text":"So it's not that human in the loop is, we just need to make things faster.","seconds":2486}},{"key":"2:010","mentions":5,"segments":5,"excerpt":{"text":"We want least privileged tools, invocation broker that's deterministic, human approval, rate limits and budgets.","seconds":5768}},{"key":"2:023","mentions":3,"segments":3,"excerpt":{"text":"But if it is a write and sensitive operations, definitely it will go through the HITL process where you wait for an approval.","seconds":15286}},{"key":"2:031","mentions":2,"segments":2,"excerpt":{"text":"And these are things where, you know, basically you can build things on top of these, but the things that you build on top of these are not allowed to, like, change these systems without having, like, human review.","seconds":20314}},{"key":"2:035","mentions":3,"segments":2,"excerpt":{"text":"Then there might be some specific house rules on what you should do or what you should not do, what you can do on your own, or where you need to get some approvals.","seconds":22425}},{"key":"2:039","mentions":2,"segments":2,"excerpt":{"text":"This is a great labelable action where the human might be doing the approval the same way that, like, we approve our tool calls in cloud code sometimes.","seconds":26359}}]},{"concept":"Verification","mentions":76,"talks":11,"coverage":0.234,"day1_talks":4,"day2_talks":7,"day1_coverage":0.1739,"day2_coverage":0.2917,"day1_per_10000_words":2.27,"day2_per_10000_words":8.55,"day1_mean_talk_rate":2.17,"day2_mean_talk_rate":9.47,"day_shift_pp":11.78,"shift_ci_low_pp":-13.59,"shift_ci_high_pp":36.87,"id":"verification","guide":"evaluation","primary":true,"pattern":"\\b(?:verif\\w*|citations?|provenance|source.check\\w*)\\b","matches":[{"key":"1:023","mentions":2,"segments":2,"excerpt":{"text":"Now this is not something that it has access to directly, so it is essentially assuming that the morning team is going to verify that this did in fact work and the information was restored.","seconds":13148}},{"key":"1:027","mentions":3,"segments":3,"excerpt":{"text":"And yeah, so, like, once you have all of your agent skills and everything, in my opinion, the final step, and I think this is the most important step, is verification.","seconds":16033}},{"key":"1:028","mentions":2,"segments":2,"excerpt":{"text":"And now that the pipeline is built and verified, we expect the audit budget for the next state to be close to zero.","seconds":16920}},{"key":"1:029","mentions":3,"segments":3,"excerpt":{"text":"When a step needs judgment, it makes one fixed schema model call: compose an email, judge a reply, spot missing values, extract fields, verify a document.","seconds":18154}},{"key":"2:004","mentions":13,"segments":12,"excerpt":{"text":"Then you get to verification, and we started Life as a verification company.","seconds":2905}},{"key":"2:005","mentions":3,"segments":3,"excerpt":{"text":"We've defined what good looks like with the domain subject matter experts, and that they're beginning to verify it.","seconds":4013}},{"key":"2:007","mentions":3,"segments":3,"excerpt":{"text":"And we built citation viewers that let people inspect the source passages and tables behind an answer.","seconds":4665}},{"key":"2:010","mentions":9,"segments":8,"excerpt":{"text":"So this is kind of a classic supply chain verification, but I would say like, don't just do supply chain verification.","seconds":6144}},{"key":"2:020","mentions":7,"segments":7,"excerpt":{"text":"The difference between us and other so-called brokers is we do verify GPUs.","seconds":13400}},{"key":"2:031","mentions":2,"segments":2,"excerpt":{"text":"Well, you can write code, and then it can execute the code and, like, basically run a computer in some verifiable way.","seconds":20134}},{"key":"2:033","mentions":21,"segments":18,"excerpt":{"text":"So three months ago, I think there's a little bit of an overlap, but I gave a talk at AIE World Fair really around verification and why that most institutions need verification in some form of the trust ecosystem of AI.","seconds":21007}}]},{"concept":"Retrieval","mentions":76,"talks":11,"coverage":0.234,"day1_talks":5,"day2_talks":6,"day1_coverage":0.2174,"day2_coverage":0.25,"day1_per_10000_words":4.21,"day2_per_10000_words":6.9,"day1_mean_talk_rate":4.18,"day2_mean_talk_rate":6.76,"day_shift_pp":3.26,"shift_ci_low_pp":-18.48,"shift_ci_high_pp":28.62,"id":"retrieval","guide":"company-context","primary":false,"pattern":"\\b(?:retriev\\w*|search\\w*|rag|vector\\w*|embeddings?)\\b","matches":[{"key":"1:003","mentions":9,"segments":9,"excerpt":{"text":"I'm going to start with one that might seem obvious, but I still think it's worth saying, which is search is hard.","seconds":1296}},{"key":"1:008","mentions":3,"segments":3,"excerpt":{"text":"So the agent registry makes agents legible, and it registers, retrieves, and discovers the key facts about agents that we want to know for any agent.","seconds":3597}},{"key":"1:016","mentions":3,"segments":3,"excerpt":{"text":"And then you have the token to ID, the tokenizations, and then the embedding lookup.","seconds":7451}},{"key":"1:025","mentions":3,"segments":3,"excerpt":{"text":"The vector is that the attacker only ever reads its own memory.","seconds":14703}},{"key":"1:043","mentions":4,"segments":4,"excerpt":{"text":"I have an absolutely enormous amount of hyperparameters within my search space.","seconds":26768}},{"key":"2:002","mentions":3,"segments":3,"excerpt":{"text":"You know kind of the ledger, the aging report, and the documents that they will need to be retrieving.","seconds":1617}},{"key":"2:007","mentions":3,"segments":3,"excerpt":{"text":"Tools allow them to search for information, write and run code, and interact with the software we use.","seconds":5222}},{"key":"2:012","mentions":3,"segments":3,"excerpt":{"text":"No longer do you have to index everything in a vector DB unless it's for a very specific use case.","seconds":7836}},{"key":"2:033","mentions":30,"segments":26,"excerpt":{"text":"It says, I'm going to go here and retrieve this source and retrieve this entity and do this thing.","seconds":21370}},{"key":"2:035","mentions":4,"segments":3,"excerpt":{"text":"We have built-in tools from the Gemini API for Google Search, so there's nothing you need to additionally do to get Google Search integrated.","seconds":22889}},{"key":"2:037","mentions":2,"segments":2,"excerpt":{"text":"And of course, you're probably asking yourself, like, why not just RAG, right?","seconds":24341}}]},{"concept":"Skills","mentions":87,"talks":11,"coverage":0.234,"day1_talks":5,"day2_talks":6,"day1_coverage":0.2174,"day2_coverage":0.25,"day1_per_10000_words":6.48,"day2_per_10000_words":6.48,"day1_mean_talk_rate":8.54,"day2_mean_talk_rate":6.71,"day_shift_pp":3.26,"shift_ci_low_pp":-18.48,"shift_ci_high_pp":28.62,"id":"skills","guide":"agent-systems","primary":false,"pattern":"\\bskills?\\b","matches":[{"key":"1:006","mentions":10,"segments":10,"excerpt":{"text":"Then we want to give it context for our actual business rules and processes via skills.","seconds":2343}},{"key":"1:013","mentions":5,"segments":5,"excerpt":{"text":"Here to explore why defining success may become the essential skill, please welcome Aparna Dhinakaran, co-founder and chief product officer at Arize AI.","seconds":5222}},{"key":"1:022","mentions":15,"segments":15,"excerpt":{"text":"That combination is what a skill is, our firm's way of doing one specific piece of work written down so an agent can follow it.","seconds":12604}},{"key":"1:027","mentions":7,"segments":7,"excerpt":{"text":"And then last but not least is the Railway agent skills, of course.","seconds":16009}},{"key":"1:043","mentions":2,"segments":2,"excerpt":{"text":"I can choose which hooks load with the skills.","seconds":26763}},{"key":"2:007","mentions":5,"segments":5,"excerpt":{"text":"We use Skills together with shared artifact tools to guide financial workflows and produce Excel models and PowerPoint presentations.","seconds":4709}},{"key":"2:023","mentions":2,"segments":2,"excerpt":{"text":"And then you can choose, you know, like your skills or prompts, and then you submit your information, right, to this.","seconds":14634}},{"key":"2:035","mentions":15,"segments":12,"excerpt":{"text":"With skills, we have very short descriptions which give the agent some guidance on like, hey, you can use that skill if the user wants to do a certain task or if you have to do certain tasks, and then the agent can read that skill.","seconds":22810}},{"key":"2:039","mentions":4,"segments":4,"excerpt":{"text":"There's things that you can do at the harness layer with skills and memory.","seconds":26160}},{"key":"2:041","mentions":15,"segments":12,"excerpt":{"text":"And I ran it twice, the first time without any accessibility skill and the second time by using the skill that I had written.","seconds":26871}},{"key":"2:047","mentions":2,"segments":2,"excerpt":{"text":"If I'm the user, I now have to manage my skill and memories and think about how to, when I should use the agent.","seconds":29635}}]},{"concept":"Training","mentions":110,"talks":11,"coverage":0.234,"day1_talks":7,"day2_talks":4,"day1_coverage":0.3043,"day2_coverage":0.1667,"day1_per_10000_words":8.1,"day2_per_10000_words":8.28,"day1_mean_talk_rate":8.72,"day2_mean_talk_rate":7.72,"day_shift_pp":-13.77,"shift_ci_low_pp":-35.51,"shift_ci_high_pp":11.41,"id":"training","guide":"agent-systems","primary":false,"pattern":"\\b(?:train\\w*|fine.tun\\w*|finetun\\w*|post.train\\w*|reinforcement learning)\\b","matches":[{"key":"1:016","mentions":26,"segments":19,"excerpt":{"text":"And then you have supervised fine-tuning, and then you do RLHF, LRVR, and all of the great post-training process to get better at the task that you want to, right?","seconds":7042}},{"key":"1:025","mentions":3,"segments":3,"excerpt":{"text":"This same method also picks up fine-tunes and distillations that share the same base model.","seconds":14932}},{"key":"1:029","mentions":3,"segments":3,"excerpt":{"text":"The training loop is next.","seconds":18259}},{"key":"1:031","mentions":4,"segments":3,"excerpt":{"text":"So going back to the example of auto research and an agent doing a lot of training, what happens when the training run fails?","seconds":19044}},{"key":"1:043","mentions":2,"segments":2,"excerpt":{"text":"Or maybe I'm training a red team.","seconds":26681}},{"key":"1:046","mentions":5,"segments":5,"excerpt":{"text":"So JEPA doesn't train alongside the action labels, so they bet on inferred action labels for the most part.","seconds":27988}},{"key":"1:049","mentions":4,"segments":3,"excerpt":{"text":"They also trained the models to not assist in the process of training other models, and they labeled that safety.","seconds":28753}},{"key":"2:002","mentions":9,"segments":8,"excerpt":{"text":"So you're post-training this latent that then can be projected onto different models, and that really amortizes the cost that you have to spend on post-training.","seconds":1975}},{"key":"2:007","mentions":17,"segments":16,"excerpt":{"text":"We used reinforcement learning and post-training to teach models to work through problems more effectively, and we gave them more compute to think when answering.","seconds":4843}},{"key":"2:039","mentions":25,"segments":23,"excerpt":{"text":"And training is a very effective way of doing this because the shape of data you're giving there is quite close to what you would want from, like, an actual training pipeline.","seconds":26144}},{"key":"2:047","mentions":5,"segments":3,"excerpt":{"text":"At the end of the day, this is the data that the agent is being trained on at runtime, not literally trained on, but effectively trained on.","seconds":30084}}]},{"concept":"Business value","mentions":66,"talks":10,"coverage":0.2128,"day1_talks":2,"day2_talks":8,"day1_coverage":0.087,"day2_coverage":0.3333,"day1_per_10000_words":2.27,"day2_per_10000_words":7.17,"day1_mean_talk_rate":3.03,"day2_mean_talk_rate":8.54,"day_shift_pp":24.64,"shift_ci_low_pp":3.44,"shift_ci_high_pp":45.83,"id":"business-value","guide":"coding-and-product-value","primary":true,"pattern":"\\b(?:roi|return on investment|business value|value case|productivity|lead time|throughput|cost per (?:merged|accepted)|revenues?)\\b","matches":[{"key":"1:036","mentions":7,"segments":7,"excerpt":{"text":"What you see is Gemini had 10% improvement in lead time, 20% improvement in coding time, so that people no longer needed to look Stack Overflow and reading documentation.","seconds":21972}},{"key":"1:038","mentions":4,"segments":3,"excerpt":{"text":"Other studies earlier in the year have shown that about 66% are saying they're having a productivity increase, but only about 40 and 20% are actually seeing cost reductions and revenue growth.","seconds":23466}},{"key":"2:014","mentions":2,"segments":2,"excerpt":{"text":"The rental bill is revenue to the provider and an input cost to the inference producer.","seconds":9188}},{"key":"2:016","mentions":6,"segments":6,"excerpt":{"text":"So that means that you're going to have to buy a new cluster, a net new one, and there's a lead time.","seconds":10455}},{"key":"2:018","mentions":4,"segments":4,"excerpt":{"text":"And so what they're looking at is that can the person get contracted revenue over a certain scale?","seconds":12236}},{"key":"2:024","mentions":2,"segments":2,"excerpt":{"text":"The agent has to catch it, and the value case measures that directly.","seconds":16484}},{"key":"2:026","mentions":23,"segments":23,"excerpt":{"text":"However, the increase in PR throughput is not coming from smaller PRs.","seconds":17778}},{"key":"2:031","mentions":4,"segments":4,"excerpt":{"text":"There's no, like, real business value increase that will be created because it's produced today versus six months from now.","seconds":19965}},{"key":"2:033","mentions":3,"segments":3,"excerpt":{"text":"Let's say we're trying to figure out what the revenue for Palantir was back in 2023.","seconds":21273}},{"key":"2:047","mentions":5,"segments":5,"excerpt":{"text":"As I said earlier, everyone already has CoWork, ChatGPT, Codex, Claude Code, but where is that productivity?","seconds":29604}}]},{"concept":"Authorization","mentions":166,"talks":9,"coverage":0.1915,"day1_talks":4,"day2_talks":5,"day1_coverage":0.1739,"day2_coverage":0.2083,"day1_per_10000_words":20.08,"day2_per_10000_words":5.79,"day1_mean_talk_rate":11.89,"day2_mean_talk_rate":7.58,"day_shift_pp":3.44,"shift_ci_low_pp":-18.12,"shift_ci_high_pp":24.82,"id":"authorization","guide":"agent-security","primary":true,"pattern":"\\b(?:authori[sz]\\w*|permissions?|credentials?|identit\\w*|oauth|tokens? exchange|access controls?)\\b","matches":[{"key":"1:006","mentions":7,"segments":5,"excerpt":{"text":"We can think about that in three general steps: who the agent is acting as, with an identity and permissions layer; how we can sandbox and determine the approval layer of the agent so that someone is signing off before it acts; and lastly, how we know the agent will still work once we adjust any variables over time.","seconds":2352}},{"key":"1:010","mentions":3,"segments":2,"excerpt":{"text":"On the right, you just see a series of instructions with no particular identity, recognizable or understandable identity.","seconds":4471}},{"key":"1:020","mentions":47,"segments":45,"excerpt":{"text":"Well, when people started looking at how do we authorize agents to access something, the tool that was sitting there on the shelf was OAuth.","seconds":9987}},{"key":"1:023","mentions":62,"segments":52,"excerpt":{"text":"So the first problem is that credentials weren't attributed to a user or an agent's identity.","seconds":13508}},{"key":"2:010","mentions":16,"segments":14,"excerpt":{"text":"We're going to talk about workload identity, scoped credentials, agent-to-agent mutual authentication, and secureless agents.","seconds":5570}},{"key":"2:023","mentions":4,"segments":4,"excerpt":{"text":"It also do create a bunch of other production resources, for example, the workload identity, right?","seconds":14774}},{"key":"2:024","mentions":2,"segments":2,"excerpt":{"text":"For every fork in the process, it maps the decision, the human role authorized to make it.","seconds":16624}},{"key":"2:035","mentions":17,"segments":12,"excerpt":{"text":"So for Slack, if you want to connect Slack, you can define, hey, here's the credential, and you are only allowed to use that credential with a certain domain or with a certain environment variable, so that way you can control when a credential is used.","seconds":23007}},{"key":"2:042","mentions":2,"segments":2,"excerpt":{"text":"It did exactly what the user gave it permission to do.","seconds":27562}}]},{"concept":"Memory","mentions":50,"talks":9,"coverage":0.1915,"day1_talks":5,"day2_talks":4,"day1_coverage":0.2174,"day2_coverage":0.1667,"day1_per_10000_words":4.05,"day2_per_10000_words":3.45,"day1_mean_talk_rate":4.58,"day2_mean_talk_rate":2.93,"day_shift_pp":-5.07,"shift_ci_low_pp":-26.91,"shift_ci_high_pp":16.3,"id":"memory","guide":"company-context","primary":false,"pattern":"\\bmemor(?:y|ies)\\b","matches":[{"key":"1:013","mentions":5,"segments":5,"excerpt":{"text":"So we have to consider the memory of, did it see something before that it thought was an issue?","seconds":5500}},{"key":"1:018","mentions":4,"segments":4,"excerpt":{"text":"That gives them all of the tools and the guardrails, and more importantly, persistence of memory over the kind of long horizon period to carry that work out.","seconds":8633}},{"key":"1:020","mentions":2,"segments":2,"excerpt":{"text":"So an example of this would be, you know, people that are leveraging memory systems.","seconds":11561}},{"key":"1:025","mentions":7,"segments":7,"excerpt":{"text":"What's isolated are compute units, memory capacity, bandwidth quotas, and false errors.","seconds":14659}},{"key":"1:031","mentions":3,"segments":3,"excerpt":{"text":"When you want to persist this state over time memory, your usage is going to change.","seconds":18829}},{"key":"2:010","mentions":4,"segments":4,"excerpt":{"text":"So now we need to think about how do we check memory for any of those malicious instructions.","seconds":6013}},{"key":"2:018","mentions":3,"segments":3,"excerpt":{"text":"If someone's talking about memory, are we talking about, you know, VRAM?","seconds":12686}},{"key":"2:023","mentions":11,"segments":11,"excerpt":{"text":"So just walk you through the quick demo here on the memory.","seconds":15432}},{"key":"2:037","mentions":3,"segments":3,"excerpt":{"text":"So this is broadly a talk about engineering an institutional memory layer.","seconds":24034}}]},{"concept":"GPU","mentions":136,"talks":8,"coverage":0.1702,"day1_talks":3,"day2_talks":5,"day1_coverage":0.1304,"day2_coverage":0.2083,"day1_per_10000_words":8.26,"day2_per_10000_words":11.73,"day1_mean_talk_rate":11.45,"day2_mean_talk_rate":11.92,"day_shift_pp":7.79,"shift_ci_low_pp":-13.59,"shift_ci_high_pp":28.99,"id":"gpu","guide":"compute-economics","primary":false,"pattern":"\\b(?:gpus?|b200\\w*|h100\\w*|cuda|nvidia|nvlink|mig)\\b","matches":[{"key":"1:025","mentions":35,"segments":28,"excerpt":{"text":"If you're using NVIDIA hardware, GPU isolation is achieved using multi-instance GPU, or MIG for short.","seconds":14630}},{"key":"1:031","mentions":4,"segments":3,"excerpt":{"text":"The other interesting emerging thing is agents need a lot more cloud science, where pre-process some data on CPUs, or spin up GPUs for running pre-process some data on CPUs, or spin up GPUs for running a scientific model, it will offload that to a model sandbox.","seconds":18855}},{"key":"1:036","mentions":11,"segments":9,"excerpt":{"text":"If you just get GPUs and upload weights on a GPU and you start doing some inference, you're going to get terrible performance.","seconds":22174}},{"key":"2:014","mentions":20,"segments":18,"excerpt":{"text":"Both estimates use B200 GPUs serving the same model with roughly 8,000 input tokens and 1,000 output tokens, a pretty standard ratio.","seconds":8910}},{"key":"2:016","mentions":14,"segments":11,"excerpt":{"text":"So if you've heard of GPU brokers, we were the first GPU broker because we were forced into being a GPU broker.","seconds":10367}},{"key":"2:018","mentions":11,"segments":11,"excerpt":{"text":"You'd be selling a fixed price service while the GPU bill floats, and that creates price exposure.","seconds":12789}},{"key":"2:020","mentions":32,"segments":29,"excerpt":{"text":"So the ones going to start trading at CME will be the H100 Neo Cloud on-demand rate, as well as B200 prices.","seconds":13956}},{"key":"2:029","mentions":5,"segments":4,"excerpt":{"text":"Yeah, and there actually are efforts right now to securitize GPUs, GPU clusters.","seconds":18779}}]},{"concept":"Inference","mentions":56,"talks":8,"coverage":0.1702,"day1_talks":4,"day2_talks":4,"day1_coverage":0.1739,"day2_coverage":0.1667,"day1_per_10000_words":3.89,"day2_per_10000_words":4.41,"day1_mean_talk_rate":5.11,"day2_mean_talk_rate":4.34,"day_shift_pp":-0.72,"shift_ci_low_pp":-22.2,"shift_ci_high_pp":20.47,"id":"inference","guide":"compute-economics","primary":false,"pattern":"\\b(?:inferen\\w*|serving|quantiz\\w*|kv.cache|prefill|decode\\w*)\\b","matches":[{"key":"1:016","mentions":3,"segments":3,"excerpt":{"text":"And that gives you an example of embedding dimensions, number of decoder blocks we use, and all of the recipe from an architecture perspective.","seconds":7554}},{"key":"1:025","mentions":8,"segments":8,"excerpt":{"text":"In our example, the LLM we're serving runs on a slice with three compute units and 40 gigabytes of VRAM.","seconds":14647}},{"key":"1:031","mentions":3,"segments":2,"excerpt":{"text":"We've built our own file systems, the scheduling, and the runtime to give you really great performance and DevEx for use cases like inference, including LLM inference, sandboxes, training, and functions, running a lot of batch processing and whatnot.","seconds":18746}},{"key":"1:036","mentions":6,"segments":6,"excerpt":{"text":"And you can look into optimize spec decoding, sequential prefill, PDDAG, all of these.","seconds":22308}},{"key":"2:014","mentions":9,"segments":9,"excerpt":{"text":"The rental bill is revenue to the provider and an input cost to the inference producer.","seconds":9188}},{"key":"2:016","mentions":5,"segments":5,"excerpt":{"text":"Now, one of the things that you could think could happen is maybe there can be new inference companies and new model labs.","seconds":11056}},{"key":"2:018","mentions":9,"segments":8,"excerpt":{"text":"For inference clouds, that's going to be more applications serving more tokens, bigger market share.","seconds":12008}},{"key":"2:020","mentions":2,"segments":2,"excerpt":{"text":"You get quantization, precision, everything is fixed for you for a period of time.","seconds":13559}}]},{"concept":"Latency","mentions":51,"talks":7,"coverage":0.1489,"day1_talks":3,"day2_talks":4,"day1_coverage":0.1304,"day2_coverage":0.1667,"day1_per_10000_words":3.24,"day2_per_10000_words":4.28,"day1_mean_talk_rate":4.57,"day2_mean_talk_rate":5.39,"day_shift_pp":3.62,"shift_ci_low_pp":-17.57,"shift_ci_high_pp":24.64,"id":"latency","guide":"compute-economics","primary":false,"pattern":"\\b(?:latenc\\w*|throughput|tokens per second|time to first token|requests per second)\\b","matches":[{"key":"1:016","mentions":2,"segments":2,"excerpt":{"text":"But the model I'm going to show you is something that beat XGBoost, even from a latency perspective, but also from a fraud detection.","seconds":6693}},{"key":"1:025","mentions":4,"segments":4,"excerpt":{"text":"There is a fundamental 881 hertz, which is just tokens per second plus harmonics, since each token is a mix of distinct frequencies, not one clean one.","seconds":14767}},{"key":"1:036","mentions":12,"segments":10,"excerpt":{"text":"So when we started back in April, you could see requests per second around three, which is high but not so high, and time to first token over 10 seconds, which is very slow.","seconds":22205}},{"key":"2:002","mentions":4,"segments":3,"excerpt":{"text":"And then the second one is Ramp Router, which also leans into improving the latency of your agent, where we have a bunch of routing strategies that we work with Ramp Labs to develop, things like Nvidia Switchyard, things like Smart Flex offloading of work to both improve latency and cost per tokens that you spend.","seconds":1897}},{"key":"2:014","mentions":2,"segments":2,"excerpt":{"text":"The P50 target is 100 output tokens per second for each user.","seconds":8920}},{"key":"2:023","mentions":3,"segments":3,"excerpt":{"text":"And you can see that this one row, each for, you know, like success, like latency cost, tool accuracy, safety, etc.","seconds":15674}},{"key":"2:026","mentions":19,"segments":19,"excerpt":{"text":"However, the increase in PR throughput is not coming from smaller PRs.","seconds":17778}}]},{"concept":"Simulation","mentions":102,"talks":6,"coverage":0.1277,"day1_talks":3,"day2_talks":3,"day1_coverage":0.1304,"day2_coverage":0.125,"day1_per_10000_words":6.48,"day2_per_10000_words":8.55,"day1_mean_talk_rate":5.08,"day2_mean_talk_rate":8.21,"day_shift_pp":-0.54,"shift_ci_low_pp":-21.74,"shift_ci_high_pp":20.57,"id":"simulation","guide":"world-models","primary":true,"pattern":"\\b(?:simulat\\w*|world models?|next.frame|pixels?|robot\\w*|video models?)\\b","matches":[{"key":"1:020","mentions":5,"segments":4,"excerpt":{"text":"In robotics, there's this concept when maybe a robot raises its arm, it thinks, is it safe for me to raise my arm?","seconds":11897}},{"key":"1:043","mentions":2,"segments":2,"excerpt":{"text":"Every single thing that a human can see in pixels on the screen, an agent in your building should be able to see via API.","seconds":26176}},{"key":"1:046","mentions":32,"segments":23,"excerpt":{"text":"We just bet on scaling pixel space the same way that anthropic scaled coding space, and betting that people will just, as we scale, data compute into pixel space, that people will bring more problems into pixel space.","seconds":27368}},{"key":"2:037","mentions":5,"segments":5,"excerpt":{"text":"But it turns out that in robotics, you actually have the solution for this, and it's called a state estimator, right?","seconds":24113}},{"key":"2:039","mentions":2,"segments":2,"excerpt":{"text":"Like, they don't really—they have to go, like, relearn a world model in context every single time, and this just does not allow them to really operate as fluently as they can discuss things that happened longer ago.","seconds":26188}},{"key":"2:049","mentions":52,"segments":40,"excerpt":{"text":"This is done—this is a video model, an image model that we had back then that was creating pixels in 128 pixels by 128 pixels that we then forcefully upsample to this kind of resolution.","seconds":30639}}]},{"concept":"Privacy","mentions":31,"talks":3,"coverage":0.0638,"day1_talks":1,"day2_talks":2,"day1_coverage":0.0435,"day2_coverage":0.0833,"day1_per_10000_words":2.11,"day2_per_10000_words":2.48,"day1_mean_talk_rate":3.9,"day2_mean_talk_rate":2.62,"day_shift_pp":3.99,"shift_ci_low_pp":-8.88,"shift_ci_high_pp":16.67,"id":"privacy","guide":"agent-security","primary":false,"pattern":"\\b(?:privacy|pii|sensitive data|sensitive information|tenant\\w*|data protection)\\b","matches":[{"key":"1:025","mentions":8,"segments":7,"excerpt":{"text":"So when one tenant is busy, the other tenants' own reads get slightly slower.","seconds":14685}},{"key":"2:010","mentions":8,"segments":8,"excerpt":{"text":"We're going to cover identity and trust, prompt and input defense, to an action governance, execution isolation, data protection, and governance observability.","seconds":5553}},{"key":"2:023","mentions":5,"segments":5,"excerpt":{"text":"So the goal here is to just block this sensitive data as it is the moment it gets ingested, right, the moment you type.","seconds":15790}}]}],"network":{"nodes":[{"id":"Evaluation","talks":18,"windows":79,"community":0,"betweenness":0.0158},{"id":"Security","talks":20,"windows":89,"community":0,"betweenness":0.2609},{"id":"Authorization","talks":9,"windows":68,"community":0,"betweenness":0.0949},{"id":"Verification","talks":11,"windows":45,"community":1,"betweenness":0.0395},{"id":"Trust","talks":24,"windows":85,"community":1,"betweenness":0.0},{"id":"Human review","talks":14,"windows":38,"community":0,"betweenness":0.1067},{"id":"Observability","talks":18,"windows":78,"community":0,"betweenness":0.1462},{"id":"Memory","talks":9,"windows":31,"community":0,"betweenness":0.0751},{"id":"Context","talks":22,"windows":89,"community":1,"betweenness":0.0316},{"id":"Retrieval","talks":11,"windows":39,"community":2,"betweenness":0.0},{"id":"Workflow","talks":31,"windows":116,"community":1,"betweenness":0.1462},{"id":"Software and coding","talks":41,"windows":226,"community":1,"betweenness":0.0237},{"id":"Skills","talks":11,"windows":45,"community":1,"betweenness":0.0158},{"id":"GPU","talks":8,"windows":58,"community":3,"betweenness":0.0791},{"id":"Compute","talks":15,"windows":90,"community":3,"betweenness":0.1502},{"id":"Cost","talks":34,"windows":197,"community":3,"betweenness":0.0751},{"id":"Latency","talks":7,"windows":22,"community":3,"betweenness":0.0},{"id":"Training","talks":11,"windows":51,"community":1,"betweenness":0.0},{"id":"Simulation","talks":6,"windows":37,"community":4,"betweenness":0.0},{"id":"Business value","talks":10,"windows":40,"community":3,"betweenness":0.0316},{"id":"Deployment","talks":28,"windows":106,"community":0,"betweenness":0.0277},{"id":"Privacy","talks":3,"windows":14,"community":5,"betweenness":0.0},{"id":"Inference","talks":8,"windows":31,"community":3,"betweenness":0.1621},{"id":"Finance","talks":33,"windows":224,"community":3,"betweenness":0.0}],"edges":[{"source":"Compute","target":"GPU","windows":33,"talks":7,"lift":5.2977,"npmi":0.5155},{"source":"GPU","target":"Inference","windows":16,"talks":5,"lift":7.4572,"npmi":0.5076},{"source":"Authorization","target":"Security","windows":30,"talks":7,"lift":4.154,"npmi":0.4277},{"source":"Memory","target":"Observability","windows":13,"talks":4,"lift":4.5054,"npmi":0.3613},{"source":"Compute","target":"Inference","windows":13,"talks":5,"lift":3.9047,"npmi":0.327},{"source":"Authorization","target":"Human review","windows":12,"talks":5,"lift":3.8916,"npmi":0.32},{"source":"Business value","target":"Cost","windows":25,"talks":8,"lift":2.6586,"npmi":0.2784},{"source":"Compute","target":"Cost","windows":47,"talks":12,"lift":2.2214,"npmi":0.2771},{"source":"Trust","target":"Verification","windows":13,"talks":4,"lift":2.8481,"npmi":0.2512},{"source":"Observability","target":"Security","windows":20,"talks":11,"lift":2.4143,"npmi":0.236},{"source":"Authorization","target":"Observability","windows":16,"talks":4,"lift":2.5279,"npmi":0.2343},{"source":"Evaluation","target":"Observability","windows":18,"talks":7,"lift":2.4479,"npmi":0.2331},{"source":"Memory","target":"Security","windows":9,"talks":5,"lift":2.7336,"npmi":0.2218},{"source":"Context","target":"Memory","windows":9,"talks":5,"lift":2.7336,"npmi":0.2218},{"source":"Cost","target":"Latency","windows":13,"talks":5,"lift":2.5136,"npmi":0.2212},{"source":"Skills","target":"Workflow","windows":15,"talks":7,"lift":2.408,"npmi":0.2184},{"source":"Business value","target":"Compute","windows":11,"talks":5,"lift":2.5606,"npmi":0.217},{"source":"Cost","target":"GPU","windows":28,"talks":7,"lift":2.0536,"npmi":0.2117},{"source":"Deployment","target":"Evaluation","windows":21,"talks":7,"lift":2.1015,"npmi":0.2015},{"source":"Human review","target":"Workflow","windows":12,"talks":9,"lift":2.2813,"npmi":0.1942},{"source":"Deployment","target":"Human review","windows":11,"talks":8,"lift":2.2885,"npmi":0.1911},{"source":"Inference","target":"Security","windows":8,"talks":4,"lift":2.4299,"npmi":0.1909},{"source":"Human review","target":"Observability","windows":8,"talks":5,"lift":2.2618,"npmi":0.1755},{"source":"Trust","target":"Workflow","windows":22,"talks":10,"lift":1.8698,"npmi":0.1719},{"source":"Observability","target":"Verification","windows":9,"talks":4,"lift":2.1487,"npmi":0.1687},{"source":"Context","target":"Skills","windows":10,"talks":5,"lift":2.0924,"npmi":0.1667},{"source":"GPU","target":"Security","windows":12,"talks":5,"lift":1.9481,"npmi":0.157},{"source":"Deployment","target":"Inference","windows":8,"talks":4,"lift":2.0402,"npmi":0.1533},{"source":"Compute","target":"Finance","windows":37,"talks":8,"lift":1.538,"npmi":0.138},{"source":"Training","target":"Workflow","windows":12,"talks":5,"lift":1.6998,"npmi":0.1249},{"source":"Software and coding","target":"Workflow","windows":45,"talks":22,"lift":1.4384,"npmi":0.1243},{"source":"Skills","target":"Software and coding","windows":19,"talks":8,"lift":1.5656,"npmi":0.1184},{"source":"Observability","target":"Workflow","windows":17,"talks":9,"lift":1.5745,"npmi":0.1165},{"source":"Verification","target":"Workflow","windows":10,"talks":6,"lift":1.6054,"npmi":0.1069},{"source":"Deployment","target":"Observability","windows":15,"talks":8,"lift":1.5203,"npmi":0.1041},{"source":"Authorization","target":"Workflow","windows":14,"talks":7,"lift":1.4873,"npmi":0.097},{"source":"Context","target":"Training","windows":8,"talks":5,"lift":1.477,"npmi":0.0838},{"source":"Security","target":"Workflow","windows":17,"talks":9,"lift":1.3799,"npmi":0.0826},{"source":"Cost","target":"Finance","windows":65,"talks":21,"lift":1.2344,"npmi":0.0824},{"source":"Finance","target":"GPU","windows":21,"talks":6,"lift":1.3545,"npmi":0.0823},{"source":"Business value","target":"Software and coding","windows":15,"talks":5,"lift":1.3905,"npmi":0.0819}],"communities":[["Authorization","Deployment","Evaluation","Human review","Memory","Observability","Security"],["Context","Skills","Software and coding","Training","Trust","Verification","Workflow"],["Retrieval"],["Business value","Compute","Cost","Finance","GPU","Inference","Latency"],["Simulation"],["Privacy"]],"sensitivity":[{"cutoff":0.05,"edges":49,"top_bridges":["Security","Workflow","Compute","Observability","Inference"]},{"cutoff":0.08,"edges":41,"top_bridges":["Security","Inference","Compute","Observability","Workflow"]},{"cutoff":0.12,"edges":31,"top_bridges":["Security","Inference","Workflow","Human review","Compute"]},{"cutoff":0.16,"edges":26,"top_bridges":["Security","Inference","Observability","Compute","Human review"]}]},"claims":[{"concept":"Evaluation","session":"1:033","label":"Implementation guidance","title":"Design evaluations around the work users need.","context":"The speaker contrasts useful work with benchmarks that models can game.","text":"The goal here is to try and build evals that are reflective of the actual work people are doing and the work output they want to see, not to build these contrived evals that the models can actually game.","seconds":20285},{"concept":"Security","session":"1:023","label":"Constraint","title":"A sandbox does not grant task-scoped authorization.","context":"This passage describes the boundary of process isolation. It does not claim that sandboxes are useless.","text":"Sandboxes bound what the processes can do, but they don't control what is authorized to act outside the sandbox itself.","seconds":13480},{"concept":"Cost","session":"1:036","label":"Reported result","title":"Revolut reports a lower invoice for its self-hosted workload.","context":"The next segment names Anthropic and OpenAI as the comparison. This is a speaker-reported invoice result, not a total-cost benchmark. Hardware, staffing, and workload equivalence are not established here.","text":"In terms of cost, in terms of impact, self-hosting GLM, our blended cost of token is 30 cents per million, and so our invoice is one-tenth of what we were spending on.","seconds":22149},{"concept":"Verification","session":"1:013","label":"Implementation guidance","title":"Align an evaluator with human judgments before release.","context":"The speaker describes the team’s own evaluator checks. Agreement with these examples does not prove correctness on every workload.","text":"These data sets, golden truth data sets, we run evals, verify them against the human output, and then try to align to make sure that the judge actually agrees with the human and they're in agreement before we actually ship out a new evaluator.","seconds":5968},{"concept":"Cost","session":"2:026","label":"Implementation guidance","title":"Include human spend when measuring the cost of a pull request.","context":"The next segment completes the point: calculate the real cost per pull request. This is a measurement recommendation.","text":"We need to add the AI spend and the human spend, the human salary spend.","seconds":17932},{"concept":"Cost","session":"2:014","label":"Constraint","title":"Compare token prices only with workload and quality in view.","context":"The speaker calls token price a practical proxy. The preceding passage says price per successful task is the goal and is difficult to standardize.","text":"Today, price per token is a useful practical proxy, especially for comparable workloads and quality requirements, and gets us closer to delivered output, much closer than the often quoted, all you finance bros out here will have heard this, GPU dollar per hour price, or as we view it, a cost.","seconds":8812},{"concept":"Compute","session":"2:039","label":"Future plan","title":"Reuse spare capacity for training outside working hours.","context":"The preceding passage calls the system an internal prototype. This passage proposes a possible capacity schedule; it does not report a measured saving.","text":"Maybe your inference capacity in the day is serving your agents; at night, employees are not using them as much, and you can use that capacity for training instead.","seconds":26675},{"concept":"Trust","session":"2:004","label":"Reported result","title":"Sonar reports improvements in confidence and reliability.","context":"The passage states qualitative results. It does not supply effect sizes or establish that another team would see the same results.","text":"We're actually building confidence, and you're seeing results: better security, more confidence in our time to market, improved, or a reduction in number of outages, improved reliability.","seconds":3000}],"lexical":[{"term":"agent","mentions":1100,"talks":41,"talk_share":0.8723,"day1_per_10000_words":95.37,"day2_per_10000_words":70.49},{"term":"model","mentions":598,"talks":43,"talk_share":0.9149,"day1_per_10000_words":48.09,"day2_per_10000_words":41.52},{"term":"work","mentions":359,"talks":46,"talk_share":0.9787,"day1_per_10000_words":27.2,"day2_per_10000_words":26.35},{"term":"data","mentions":348,"talks":41,"talk_share":0.8723,"day1_per_10000_words":24.29,"day2_per_10000_words":27.31},{"term":"time","mentions":348,"talks":45,"talk_share":0.9574,"day1_per_10000_words":26.56,"day2_per_10000_words":25.38},{"term":"people","mentions":276,"talks":40,"talk_share":0.8511,"day1_per_10000_words":19.27,"day2_per_10000_words":21.66},{"term":"different","mentions":227,"talks":41,"talk_share":0.8723,"day1_per_10000_words":17.33,"day2_per_10000_words":16.55},{"term":"token","mentions":223,"talks":31,"talk_share":0.6596,"day1_per_10000_words":16.35,"day2_per_10000_words":16.83},{"term":"build","mentions":212,"talks":39,"talk_share":0.8298,"day1_per_10000_words":13.76,"day2_per_10000_words":17.52},{"term":"tool","mentions":191,"talks":33,"talk_share":0.7021,"day1_per_10000_words":14.9,"day2_per_10000_words":13.66},{"term":"company","mentions":184,"talks":35,"talk_share":0.7447,"day1_per_10000_words":11.01,"day2_per_10000_words":16.0},{"term":"code","mentions":179,"talks":35,"talk_share":0.7447,"day1_per_10000_words":11.82,"day2_per_10000_words":14.62},{"term":"human","mentions":177,"talks":36,"talk_share":0.766,"day1_per_10000_words":14.74,"day2_per_10000_words":11.86},{"term":"cost","mentions":169,"talks":26,"talk_share":0.5532,"day1_per_10000_words":6.8,"day2_per_10000_words":17.52},{"term":"building","mentions":165,"talks":37,"talk_share":0.7872,"day1_per_10000_words":13.93,"day2_per_10000_words":10.9},{"term":"new","mentions":161,"talks":37,"talk_share":0.7872,"day1_per_10000_words":12.95,"day2_per_10000_words":11.17},{"term":"run","mentions":157,"talks":36,"talk_share":0.766,"day1_per_10000_words":11.5,"day2_per_10000_words":11.86},{"term":"start","mentions":151,"talks":38,"talk_share":0.8085,"day1_per_10000_words":9.88,"day2_per_10000_words":12.41},{"term":"world","mentions":136,"talks":31,"talk_share":0.6596,"day1_per_10000_words":11.82,"day2_per_10000_words":8.69},{"term":"software","mentions":135,"talks":29,"talk_share":0.617,"day1_per_10000_words":10.04,"day2_per_10000_words":10.07},{"term":"access","mentions":127,"talks":23,"talk_share":0.4894,"day1_per_10000_words":15.87,"day2_per_10000_words":4.0},{"term":"problem","mentions":126,"talks":28,"talk_share":0.5957,"day1_per_10000_words":10.69,"day2_per_10000_words":8.28},{"term":"task","mentions":126,"talks":27,"talk_share":0.5745,"day1_per_10000_words":10.2,"day2_per_10000_words":8.69},{"term":"context","mentions":125,"talks":29,"talk_share":0.617,"day1_per_10000_words":10.2,"day2_per_10000_words":8.55},{"term":"built","mentions":123,"talks":35,"talk_share":0.7447,"day1_per_10000_words":11.17,"day2_per_10000_words":7.45},{"term":"process","mentions":123,"talks":28,"talk_share":0.5957,"day1_per_10000_words":7.29,"day2_per_10000_words":10.76},{"term":"user","mentions":123,"talks":28,"talk_share":0.5957,"day1_per_10000_words":12.79,"day2_per_10000_words":6.07},{"term":"example","mentions":121,"talks":35,"talk_share":0.7447,"day1_per_10000_words":7.77,"day2_per_10000_words":10.07},{"term":"evaluation","mentions":118,"talks":22,"talk_share":0.4681,"day1_per_10000_words":9.88,"day2_per_10000_words":7.86},{"term":"information","mentions":114,"talks":27,"talk_share":0.5745,"day1_per_10000_words":7.77,"day2_per_10000_words":9.1},{"term":"mcp","mentions":109,"talks":8,"talk_share":0.1702,"day1_per_10000_words":5.34,"day2_per_10000_words":10.48},{"term":"end","mentions":108,"talks":34,"talk_share":0.7234,"day1_per_10000_words":8.74,"day2_per_10000_words":7.45},{"term":"real","mentions":107,"talks":34,"talk_share":0.7234,"day1_per_10000_words":7.45,"day2_per_10000_words":8.41},{"term":"financial","mentions":105,"talks":22,"talk_share":0.4681,"day1_per_10000_words":6.64,"day2_per_10000_words":8.83},{"term":"question","mentions":105,"talks":31,"talk_share":0.6596,"day1_per_10000_words":10.53,"day2_per_10000_words":5.52},{"term":"training","mentions":103,"talks":17,"talk_share":0.3617,"day1_per_10000_words":7.12,"day2_per_10000_words":8.14},{"term":"compute","mentions":103,"talks":15,"talk_share":0.3191,"day1_per_10000_words":1.62,"day2_per_10000_words":12.83},{"term":"working","mentions":101,"talks":29,"talk_share":0.617,"day1_per_10000_words":6.48,"day2_per_10000_words":8.41},{"term":"team","mentions":98,"talks":29,"talk_share":0.617,"day1_per_10000_words":5.99,"day2_per_10000_words":8.41},{"term":"gpu","mentions":96,"talks":10,"talk_share":0.2128,"day1_per_10000_words":5.34,"day2_per_10000_words":8.69},{"term":"point","mentions":94,"talks":34,"talk_share":0.7234,"day1_per_10000_words":4.7,"day2_per_10000_words":8.97},{"term":"source","mentions":93,"talks":25,"talk_share":0.5319,"day1_per_10000_words":6.48,"day2_per_10000_words":7.31},{"term":"better","mentions":93,"talks":36,"talk_share":0.766,"day1_per_10000_words":6.32,"day2_per_10000_words":7.45},{"term":"server","mentions":93,"talks":11,"talk_share":0.234,"day1_per_10000_words":6.8,"day2_per_10000_words":7.04},{"term":"customer","mentions":92,"talks":24,"talk_share":0.5106,"day1_per_10000_words":5.18,"day2_per_10000_words":8.28},{"term":"scale","mentions":91,"talks":27,"talk_share":0.5745,"day1_per_10000_words":8.42,"day2_per_10000_words":5.38},{"term":"engineer","mentions":90,"talks":27,"talk_share":0.5745,"day1_per_10000_words":9.07,"day2_per_10000_words":4.69},{"term":"mean","mentions":90,"talks":30,"talk_share":0.6383,"day1_per_10000_words":7.45,"day2_per_10000_words":6.07},{"term":"day","mentions":89,"talks":34,"talk_share":0.7234,"day1_per_10000_words":6.15,"day2_per_10000_words":7.04},{"term":"skill","mentions":87,"talks":15,"talk_share":0.3191,"day1_per_10000_words":6.48,"day2_per_10000_words":6.48},{"term":"understand","mentions":87,"talks":26,"talk_share":0.5532,"day1_per_10000_words":6.15,"day2_per_10000_words":6.76},{"term":"change","mentions":84,"talks":34,"talk_share":0.7234,"day1_per_10000_words":7.45,"day2_per_10000_words":5.24},{"term":"business","mentions":83,"talks":19,"talk_share":0.4043,"day1_per_10000_words":8.26,"day2_per_10000_words":4.41},{"term":"running","mentions":81,"talks":27,"talk_share":0.5745,"day1_per_10000_words":6.48,"day2_per_10000_words":5.66},{"term":"spend","mentions":80,"talks":25,"talk_share":0.5319,"day1_per_10000_words":3.56,"day2_per_10000_words":8.0},{"term":"ask","mentions":78,"talks":31,"talk_share":0.6596,"day1_per_10000_words":5.67,"day2_per_10000_words":5.93},{"term":"years","mentions":77,"talks":27,"talk_share":0.5745,"day1_per_10000_words":3.24,"day2_per_10000_words":7.86},{"term":"control","mentions":77,"talks":25,"talk_share":0.5319,"day1_per_10000_words":7.93,"day2_per_10000_words":3.86},{"term":"workflow","mentions":77,"talks":21,"talk_share":0.4468,"day1_per_10000_words":6.48,"day2_per_10000_words":5.1},{"term":"long","mentions":77,"talks":31,"talk_share":0.6596,"day1_per_10000_words":5.34,"day2_per_10000_words":6.07},{"term":"finance","mentions":76,"talks":18,"talk_share":0.383,"day1_per_10000_words":2.11,"day2_per_10000_words":8.69},{"term":"product","mentions":75,"talks":32,"talk_share":0.6809,"day1_per_10000_words":3.56,"day2_per_10000_words":7.31},{"term":"important","mentions":75,"talks":34,"talk_share":0.7234,"day1_per_10000_words":5.67,"day2_per_10000_words":5.52},{"term":"value","mentions":74,"talks":24,"talk_share":0.5106,"day1_per_10000_words":3.56,"day2_per_10000_words":7.17},{"term":"second","mentions":73,"talks":32,"talk_share":0.6809,"day1_per_10000_words":6.15,"day2_per_10000_words":4.83},{"term":"llm","mentions":73,"talks":17,"talk_share":0.3617,"day1_per_10000_words":7.29,"day2_per_10000_words":3.86},{"term":"open","mentions":72,"talks":29,"talk_share":0.617,"day1_per_10000_words":5.99,"day2_per_10000_words":4.83},{"term":"output","mentions":72,"talks":24,"talk_share":0.5106,"day1_per_10000_words":3.56,"day2_per_10000_words":6.9},{"term":"create","mentions":72,"talks":24,"talk_share":0.5106,"day1_per_10000_words":1.94,"day2_per_10000_words":8.28},{"term":"problems","mentions":72,"talks":24,"talk_share":0.5106,"day1_per_10000_words":5.67,"day2_per_10000_words":5.1},{"term":"test","mentions":71,"talks":26,"talk_share":0.5532,"day1_per_10000_words":5.51,"day2_per_10000_words":5.1},{"term":"big","mentions":71,"talks":33,"talk_share":0.7021,"day1_per_10000_words":5.34,"day2_per_10000_words":5.24},{"term":"based","mentions":71,"talks":21,"talk_share":0.4468,"day1_per_10000_words":6.15,"day2_per_10000_words":4.55},{"term":"single","mentions":70,"talks":28,"talk_share":0.5957,"day1_per_10000_words":5.99,"day2_per_10000_words":4.55},{"term":"firm","mentions":70,"talks":17,"talk_share":0.3617,"day1_per_10000_words":5.83,"day2_per_10000_words":4.69},{"term":"trying","mentions":70,"talks":27,"talk_share":0.5745,"day1_per_10000_words":3.56,"day2_per_10000_words":6.62},{"term":"production","mentions":69,"talks":20,"talk_share":0.4255,"day1_per_10000_words":5.67,"day2_per_10000_words":4.69},{"term":"number","mentions":69,"talks":29,"talk_share":0.617,"day1_per_10000_words":4.7,"day2_per_10000_words":5.52},{"term":"cloud","mentions":69,"talks":17,"talk_share":0.3617,"day1_per_10000_words":2.59,"day2_per_10000_words":7.31},{"term":"quality","mentions":68,"talks":19,"talk_share":0.4043,"day1_per_10000_words":3.24,"day2_per_10000_words":6.62},{"term":"means","mentions":68,"talks":34,"talk_share":0.7234,"day1_per_10000_words":5.18,"day2_per_10000_words":4.97},{"term":"started","mentions":67,"talks":30,"talk_share":0.6383,"day1_per_10000_words":5.34,"day2_per_10000_words":4.69},{"term":"price","mentions":67,"talks":10,"talk_share":0.2128,"day1_per_10000_words":0.81,"day2_per_10000_words":8.55},{"term":"engineering","mentions":66,"talks":26,"talk_share":0.5532,"day1_per_10000_words":5.99,"day2_per_10000_words":4.0},{"term":"case","mentions":66,"talks":32,"talk_share":0.6809,"day1_per_10000_words":4.7,"day2_per_10000_words":5.1},{"term":"help","mentions":66,"talks":28,"talk_share":0.5957,"day1_per_10000_words":5.18,"day2_per_10000_words":4.69},{"term":"place","mentions":66,"talks":29,"talk_share":0.617,"day1_per_10000_words":4.05,"day2_per_10000_words":5.66},{"term":"intent","mentions":66,"talks":8,"talk_share":0.1702,"day1_per_10000_words":9.23,"day2_per_10000_words":1.24},{"term":"loop","mentions":65,"talks":19,"talk_share":0.4043,"day1_per_10000_words":6.15,"day2_per_10000_words":3.72},{"term":"person","mentions":65,"talks":23,"talk_share":0.4894,"day1_per_10000_words":5.83,"day2_per_10000_words":4.0},{"term":"research","mentions":64,"talks":23,"talk_share":0.4894,"day1_per_10000_words":5.18,"day2_per_10000_words":4.41},{"term":"step","mentions":64,"talks":24,"talk_share":0.5106,"day1_per_10000_words":6.32,"day2_per_10000_words":3.45},{"term":"come","mentions":64,"talks":32,"talk_share":0.6809,"day1_per_10000_words":3.89,"day2_per_10000_words":5.52},{"term":"interesting","mentions":64,"talks":19,"talk_share":0.4043,"day1_per_10000_words":4.53,"day2_per_10000_words":4.97},{"term":"try","mentions":63,"talks":25,"talk_share":0.5319,"day1_per_10000_words":5.02,"day2_per_10000_words":4.41},{"term":"environment","mentions":63,"talks":16,"talk_share":0.3404,"day1_per_10000_words":5.51,"day2_per_10000_words":4.0},{"term":"services","mentions":62,"talks":21,"talk_share":0.4468,"day1_per_10000_words":5.83,"day2_per_10000_words":3.59},{"term":"happens","mentions":62,"talks":30,"talk_share":0.6383,"day1_per_10000_words":4.7,"day2_per_10000_words":4.55},{"term":"year","mentions":62,"talks":22,"talk_share":0.4681,"day1_per_10000_words":3.4,"day2_per_10000_words":5.66},{"term":"actual","mentions":62,"talks":22,"talk_share":0.4681,"day1_per_10000_words":3.56,"day2_per_10000_words":5.52}],"phrases":[{"phrase":"mcp server","mentions":61,"talks":5},{"phrase":"financial services","mentions":34,"talks":11},{"phrase":"open source","mentions":31,"talks":13},{"phrase":"ai agent","mentions":28,"talks":12},{"phrase":"claude code","mentions":20,"talks":10},{"phrase":"coding agent","mentions":20,"talks":11},{"phrase":"post training","mentions":20,"talks":6},{"phrase":"real time","mentions":17,"talks":6},{"phrase":"world model","mentions":16,"talks":5},{"phrase":"foundation model","mentions":15,"talks":4},{"phrase":"software engineer","mentions":15,"talks":6},{"phrase":"frontier model","mentions":14,"talks":8},{"phrase":"cloud agent","mentions":13,"talks":4},{"phrase":"data center","mentions":13,"talks":5},{"phrase":"new model","mentions":12,"talks":8},{"phrase":"ai native","mentions":11,"talks":3},{"phrase":"software development","mentions":10,"talks":4},{"phrase":"software factories","mentions":10,"talks":4},{"phrase":"decision model","mentions":10,"talks":3},{"phrase":"high value","mentions":9,"talks":5},{"phrase":"code base","mentions":9,"talks":5},{"phrase":"useful work","mentions":9,"talks":3},{"phrase":"source model","mentions":8,"talks":3},{"phrase":"cloud code","mentions":8,"talks":6},{"phrase":"months ago","mentions":8,"talks":6},{"phrase":"customer data","mentions":8,"talks":3},{"phrase":"writing code","mentions":8,"talks":5},{"phrase":"tool calls","mentions":8,"talks":5},{"phrase":"bare metal","mentions":8,"talks":4},{"phrase":"software engineering","mentions":8,"talks":5},{"phrase":"long time","mentions":8,"talks":6},{"phrase":"long horizon","mentions":8,"talks":3},{"phrase":"api keys","mentions":7,"talks":4},{"phrase":"api key","mentions":7,"talks":3},{"phrase":"long term","mentions":7,"talks":5},{"phrase":"long running","mentions":7,"talks":4},{"phrase":"running agent","mentions":7,"talks":7},{"phrase":"open source model","mentions":7,"talks":3},{"phrase":"different types","mentions":7,"talks":5},{"phrase":"agent loop","mentions":7,"talks":4},{"phrase":"gpu cluster","mentions":7,"talks":5},{"phrase":"language model","mentions":7,"talks":5},{"phrase":"hedge fund","mentions":7,"talks":5},{"phrase":"data source","mentions":7,"talks":5},{"phrase":"data sets","mentions":7,"talks":3},{"phrase":"test cases","mentions":7,"talks":4},{"phrase":"training model","mentions":7,"talks":3},{"phrase":"audit log","mentions":7,"talks":3},{"phrase":"non deterministic","mentions":7,"talks":5},{"phrase":"applied ai","mentions":6,"talks":4},{"phrase":"york city","mentions":6,"talks":3},{"phrase":"company called","mentions":6,"talks":5},{"phrase":"new york city","mentions":6,"talks":3},{"phrase":"output token","mentions":6,"talks":4},{"phrase":"frontier labs","mentions":6,"talks":5},{"phrase":"completely different","mentions":6,"talks":3},{"phrase":"single time","mentions":6,"talks":5},{"phrase":"asset management","mentions":6,"talks":3},{"phrase":"workload identity","mentions":6,"talks":3},{"phrase":"excel model","mentions":6,"talks":3}]}