Day 1 Afternoon Keynote | Glean:GO 2026  🔗 **Source:** [Day 1 Afternoon Keynote | Glean:GO 2026](https://www.youtube.com/watch?v=GuwoULakxdY) > [!note]+ Document Notes > tags: [enterprise-ai, ai-agents, ai-governance, token-economics, knowledge-management, developer-productivity] > author: Glean > source: Glean (YouTube) > reference: https://www.youtube.com/watch?v=GuwoULakxdY > date created: 2026-09-21 00:00 > > > [!info] Generated analysis — not the author's words. Sections below are machine-written > > from the captured source. Only text inside `> [!quote]` blocks is verbatim from the author. > > Do not quote or attribute any other line to the named author. > > # 🏷️ Title: Day 1 Afternoon Keynote | Glean:GO 2026 > <!-- deep-read:sections start --> ## 🧭 Table of contents - [📝 Summary](#-summary) - [📖 Main Content](#-main-content) - [🗒️ Notes](#️-notes) - [📚 References](#-references) - [🗣️ Paraphrase](#️-paraphrase) - [📰 Newsletter](#-newsletter) - [✅ Action Items](#-action-items) - [🧪 Hypotheses](#-hypotheses) - [📊 Methodology](#-methodology) - [❓ Questions](#-questions) - [🔮 Future Research](#-future-research) - [💡 Implications](#-implications) - [🧠 Learning Styles Assessment](#-learning-styles-assessment) - [📊 Accessibility Assessment](#-accessibility-assessment) - [🌍 DEIB Assessment](#-deib-assessment) - [✅ Logical Validity](#-logical-validity) - [🛡️ DARVO Analysis](#️-darvo-analysis) - [🎯 Conclusion](#-conclusion) - [🔄 Refactor Appendix](#-refactor-appendix--day-1-afternoon-keynote--gleango-2026) ### 🔗 References - [Day 1 Afternoon Keynote | Glean:GO 2026](https://www.youtube.com/watch?v=GuwoULakxdY) — the source recording, 57:36, uploaded 2026-09-21. - [Glean](https://www.glean.com/) — vendor hosting the event. - Patrick Wendell's post on how Databricks optimized coding-agent costs, referenced on X at [00:26:41]. No URL given on stage; search X for the Databricks coding-cost note. - OpenRouter token-consumption study cited at [00:39:21]. No URL given on stage. ## 📝 Summary - **What happened.** Three back-to-back conversations closed day one of Glean:GO 2026 in San Francisco: a customer fireside with Ericsson's chief data officer and eBay's CIO, a founder conversation between Glean co-founder Vish and Databricks co-founder Patrick, and a four-person panel moderated by Sharon Goldman with Cisco, Deloitte, and Glean's Arvind. The session closed with day-two logistics and the inaugural Glean:GO AI Innovator Awards. - **Why it matters for a Glean customer.** This session ships zero product announcements. Everything Glean-specific is deployment evidence from three named accounts — eBay (three years in, search first, now "no data source internally... that's not connected"), Ericsson (POC fall 2024, rollout early 2025, 50–70% of employees active), and Databricks (10,000 employees, multiple uses per day). If you're cataloguing what Glean *shipped*, this is not the session; if you're building an internal case for expansion, the adoption and governance numbers here are the reusable part. - **Key stakeholders.** Joel (chief data officer, Ericsson), John Felmmeer (CIO, eBay), Patrick (co-founder, Databricks), Vish (co-founder, Glean), Sharon Goldman (moderator), Arvind (Glean), Ashish (Deloitte), and the Cisco speaker rendered by the captions as "G2." - **Risks and tradeoffs.** Every cost figure on stage is qualitative — no dollar amounts, no per-seat comparisons, no latency or benchmark numbers. The most quantified claims (60% of compute on inference, 14x agent token growth, 5x agent-to-human token ratio, FY27 liftoff, Cisco's 24-month no-legacy-code prediction) come from one speaker and are sourced to either an uncited OpenRouter study or nothing at all. - **Bottom line.** Treat this as field evidence and market framing, not a roadmap. The durable takeaways are operational: eBay's per-user AI spend dashboard, Ericsson's shift from pre-screening checklists to runtime control environments, Databricks' pattern of using a second agent to review the first agent's code, and Deloitte's per-persona "workbench." ## 📖 Main content ### The session is customer evidence, not product news No Glean feature was announced, dated, or given a GA/beta/preview status in these 58 minutes. Every Glean capability discussed — enterprise search, chat, connectors, the knowledge graph, agent building, and an agent publication and visibility workflow — is described in the past tense by customers running it in production. Availability status for all of them: **unstated in this session**, though production usage at eBay, Ericsson, Databricks, and Deloitte implies GA. The one forward-looking Glean item is a cross-reference, not an announcement: eBay's CIO says at [00:17:08] that he heard "earlier in the keynotes about agents with their own identity," and Arvind says at [00:52:24] that proactive AI got time "today." Both point back to the morning keynote. Status of agent identity and proactive AI from *this* session: **unstated**. ### Three deployment stories, three different starting points eBay and Ericsson both entered through enterprise search, and both describe the same second act: search becomes the default entry point, then the connector surface expands, then the knowledge graph gets good enough that removing it becomes politically impossible. eBay started with roughly half a dozen connectors and now claims near-total internal coverage. Ericsson's differentiator in the buying decision was agents — Joel says Glean was "early on agents" before launch, and enabling employees to build their own was part of why Ericsson bought. Databricks arrived by accident. Patrick, who owns internal security and application deployment, first met Glean as a security review he nearly blocked, then got converted by an employee who had already deployed it without asking. ### Cost control is the shared operational problem Three distinct cost mechanisms surface: 1. **Visibility first, controls second** (eBay). A dashboard shipped in the last couple of months shows every user their own AI spend across tools. The stated purpose is behavior change, not blocking. 2. **Power-curve triage** (Ericsson). Token consumption follows a power curve, so a small number of individuals account for most spend, and coaching those individuals is enough control for now. 3. **Population-scale unbounded risk** (Databricks). Consumption-based coding tools multiplied by more than 2,000 engineers is the exposure Patrick spends his effort on. None of the three gives a dollar figure, a per-seat comparison, or a budget. Contested claim worth flagging: eBay's argument at [00:13:56] that usage-based pricing makes tool sprawl cost-neutral assumes constant per-task token cost across tools, which is not true if two tools route to different models at different prices. ### Where quality bottlenecks move Patrick's argument is the most carefully reasoned segment. Coding models are good enough that writing an individual module is no longer the bottleneck, so the bottleneck relocates — to code review, and past that to architecture and taste. His explanation for why models still fail at architecture is a claim about latent context: an experienced engineer brings intuition from a decade of prior jobs that is not representable in the requirements document, and tools that add context help but are bounded by what can be written down. He names the feedback-loop asymmetry as the mechanism: small decisions get fast feedback, architectural decisions get feedback over years. The Cisco speaker makes the same bottleneck-relocation argument at [00:49:11] independently, which is the strongest convergence in the session. ### The market-forces framing The Cisco speaker lays out four forces: inference is the new workload, agents are the new workforce, tokenomics, and control. The supporting numbers are the densest in the session and also the least sourced — see Logical Validity below. Terms worth defining for a reader who wasn't there: - **AI mesh** — Ericsson's internal name for the connective and governance layer between AI systems. Not a Glean product. - **Workbench** — Deloitte's per-persona bundle of tools and agents. Not a Glean product. - **AI front doors** — one of four themes named by the moderator at [00:05:35] as running through the whole day; the single-entry-point problem eBay describes at [00:14:37]. - **Tokenomics** — the Cisco speaker's term for the cost-per-token curve and its downstream effects. ## 🗒️ Notes #### ⏱️ [00:00:02](https://www.youtube.com/watch?v=GuwoULakxdY&t=2) Host opens day one closing keynote - Three segments announced: Joel (Ericsson) and John (eBay); Glean co-founder Vish with Databricks co-founder Patrick; a panel with "G2 from Cisco, Ashish from Deote and Arind." - Caption quality note: the transcript renders Deloitte as "Deote"/"Deoid," Arvind as "Arind"/"Arvin," and appears to garble the Cisco speaker's name to "G2." Treat all name spellings in this file as ASR-derived unless confirmed elsewhere. #### ⏱️ [00:00:54](https://www.youtube.com/watch?v=GuwoULakxdY&t=54) Speaker roles - Joel: chief data officer at Ericsson, also running the program to scale enterprise AI from micro-productivity to workflows to enterprise-level outcomes, plus European AI governance. - John Felmmeer: CIO at eBay, responsible for all internally used AI tooling. #### ⏱️ [00:02:29](https://www.youtube.com/watch?v=GuwoULakxdY&t=149) eBay's three-year Glean journey - Started "just about three years ago" as a search experience; chat was just starting out. - eBay described as a 30-year-old company with information spread across systems and two prior failed internal knowledge-management attempts. #### ⏱️ [00:03:46](https://www.youtube.com/watch?v=GuwoULakxdY&t=226) Ericsson's timeline and adoption rate - Engagement started roughly two years before the event; POC in fall 2024; rollout beginning of 2025. - Entry use case was enterprise search — Joel says the internal search bar "was terrible." - Agents were decisive in the purchase: Glean was early on agents before launch, and Ericsson wanted employees building their own. > [!quote] [00:04:24] I think we are now somewhere between like 50 to 70% of our um employees actively using. #### ⏱️ [00:04:59](https://www.youtube.com/watch?v=GuwoULakxdY&t=299) Data priority over business priority - Ericsson is 150 years old this year; eBay 30. - Deploying search surfaced concrete data defects: internal policies not tagged to a legal entity, oversharing, unknown locations of personal data. - Joel's framing: fixing these builds the trust needed for more sophisticated AI, especially over unstructured data. #### ⏱️ [00:05:35](https://www.youtube.com/watch?v=GuwoULakxdY&t=335) The four day-one themes - Moderator names them: enterprise context, token efficiency, the role of governance, and AI front doors. This is the explicit overlap seam with the morning keynote. #### ⏱️ [00:06:17](https://www.youtube.com/watch?v=GuwoULakxdY&t=377) Connectors as the buying reason - eBay started with about half a dozen connectors; John says there is now effectively no internal data source not connected through Glean. - Stated mechanism for context quality: the more Glean is used, the better the answers, via the knowledge graph built over two and a half years. > [!quote] [00:07:30] if I tried to take glean away from from my employee base I you know I would have a revolution on my hands #### ⏱️ [00:08:10](https://www.youtube.com/watch?v=GuwoULakxdY&t=490) Ericsson on token tradeoffs - Joel reframes it away from cost-versus-quality: baseline is (1) visibility on token spend and (2) ability to control it. - Second axis is sovereignty and vendor lock-in as much as cost; flexibility in the stack is a safeguard. > [!quote] [00:08:50] so far the token consumption is very much following a power curve as well #### ⏱️ [00:09:27](https://www.youtube.com/watch?v=GuwoULakxdY&t=567) eBay's per-user AI spend dashboard - Shipped in the last couple of months. Shows all users what they are spending on AI across tools. - Cross-reference: John says he covered this in an earlier session the same day — so the dashboard is duplicate material for anyone processing that session. > [!quote] [00:09:27] people don't really understand the cost of AI. - Stated user misconception: people assume AI is an all-you-can-eat enterprise tool like Google Docs or Microsoft Word rather than usage-based. - Reported effect at [00:10:42]: immediate behavior change, users surprised by their own spend. #### ⏱️ [00:11:19](https://www.youtube.com/watch?v=GuwoULakxdY&t=679) Ericsson's governance generations - Second or third revision of governance since LLMs arrived. - Generation one: trimming duplicative steps from procurement and compliance to buy faster. - Current shift: pre-screening checklists (paper or PowerPoint, for data transfer, privacy review, IT security review) can't keep up. > [!quote] [00:11:57] We are going to have to rely on control environments, policies, and then active monitoring and managing the deviations. #### ⏱️ [00:12:35](https://www.youtube.com/watch?v=GuwoULakxdY&t=755) AI mesh and a governance agent - Ericsson is building an internal layer it calls an AI mesh, with oversight at the most senior levels. - Built an agent for the responsible AI process: upload an IT architecture, and it populates the questionnaires and returns a position. - Joel's conclusion: autonomy has to be managed at runtime. #### ⏱️ [00:13:15](https://www.youtube.com/watch?v=GuwoULakxdY&t=795) eBay on AI sprawl - Per-seat licensing made sprawl expensive; usage-based pricing changed the calculus. > [!quote] [00:13:56] so when we move to a pure usage model, which we're largely at right now, uh, then I don't worry as much about sprawl there. - Tools named as competing front doors at eBay: ChatGPT, Glean, Claude, Gemini. - Remaining unsolved problem: internal customer experience. Users don't know where to go to ask, or which tool to build an agent in. #### ⏱️ [00:15:50](https://www.youtube.com/watch?v=GuwoULakxdY&t=950) Ericsson on grassroots sprawl and agent publication - Glean helped manage grassroots agent building; Ericsson and Glean worked out a publication process for promoting agents and controlling who sees what. Availability status of that workflow: unstated. - Joel's forward concern is decision loops: where agent traces and tool calls accrue, who gets to learn across workflows, and staying in control of that. #### ⏱️ [00:17:08](https://www.youtube.com/watch?v=GuwoULakxdY&t=1028) eBay's two-year wish — agents in the org chart - John wants agents with their own identity working side by side with employees, referencing the morning keynote. - His model of future work: employees, contractors, outside consultants, and agents, with the mix varying by department. Orchestration between them is the thing still to be built. #### ⏱️ [00:19:02](https://www.youtube.com/watch?v=GuwoULakxdY&t=1142) Ericsson's two wishes - Wish one, the boring one: outcomes at the organizational or domain level — his example is responding one week faster half the time in a process. - Wish two: work changing fundamentally. His worked example is asking Glean for a project status while the team was on vacation, then asking whether the plan was executable; the answer was no, citing multiple documents with different deadlines and no single source of truth. - Extension he wants: daily detection of stalled decisions and his own ambiguity, not just red/yellow/green status. - Closing analogy at [00:20:12]: he wants the first-Waymo-ride feeling at work; the captions render Waymo as "Whimo," and he notes it spoke to him in Swedish. #### ⏱️ [00:22:11](https://www.youtube.com/watch?v=GuwoULakxdY&t=1331) Databricks nearly blocked the Glean purchase - Patrick owns internal security and application deployment at Databricks. He received a security approval request for a vendor he'd never heard of. - His objection: the product indexes every important application's data. - An unnamed employee had already deployed it without approval, demoed it on their laptop, and answered a live question about why a project was running behind. Patrick declines to name the person so Glean's sales team can't recruit them. #### ⏱️ [00:24:16](https://www.youtube.com/watch?v=GuwoULakxdY&t=1456) Databricks scale - 10,000 employees; average usage described as multiple times a day. - Vish's framing: the security questions Databricks asked as an early customer improved the product. #### ⏱️ [00:24:52](https://www.youtube.com/watch?v=GuwoULakxdY&t=1492) AI in the Databricks engineering org - Glean was one of Databricks' first AI tools and remains pervasive for knowledge workers. - Biggest change in the last year: coding tools, driven by better coding models and a proliferation of harnesses. > [!quote] [00:25:30] almost every line of code today at data bricks is authored initially by some AI - Qualifier in the same breath: human reviewed and edited, but the first stab is usually AI. #### ⏱️ [00:26:06](https://www.youtube.com/watch?v=GuwoULakxdY&t=1566) The unbounded-cost problem > [!quote] [00:26:06] we have more than 2,000 engineers at data bricks. - Mechanism: consumption-based pricing means a single aggressive user can accrue almost unbounded cost; multiply by the engineering population. - Bottleneck observation: writing an individual piece of code is no longer the bottleneck, so bottlenecks appear elsewhere in the software supply chain. - Vish points the audience at Patrick's post on X about how Databricks optimized coding costs. No URL given. #### ⏱️ [00:27:28](https://www.youtube.com/watch?v=GuwoULakxdY&t=1648) Quality under AI-generated volume > [!quote] [00:27:28] in my view the quantity has gone up the quality has gone down - Stakes framing: a codebase is IP sold to customers, unlike a social post. - Stated remedy: be honest about what the tools are great at and where they aren't, then split work at the task level. #### ⏱️ [00:28:09](https://www.youtube.com/watch?v=GuwoULakxdY&t=1689) Why models fail at architecture - Patrick's puzzle: a seasoned engineer can finish a design doc from written constraints, so why can't a model? - His answer: experienced people bring company context and intuition from prior jobs — latent, unstated considerations. Context tools like Glean help but are bounded by what can be represented textually. - He states honest uncertainty about whether models can ever do this in principle. #### ⏱️ [00:30:24](https://www.youtube.com/watch?v=GuwoULakxdY&t=1824) Feedback-loop length as the mechanism - Small decisions get fast feedback; big architectural decisions take years to evaluate, so there's no tight loop for an AI to learn from. #### ⏱️ [00:30:24](https://www.youtube.com/watch?v=GuwoULakxdY&t=1830) Tokens as a cost, not a score - Patrick's observation: tokens became something people are rewarded for maximizing, which is backwards for a cost line. He credits it as a marketing move by the model providers and jokes about wanting the same for Databricks spend. #### ⏱️ [00:31:36](https://www.youtube.com/watch?v=GuwoULakxdY&t=1896) Measuring quality with a second agent - For engineering velocity Databricks looks at code volume, plus agents reviewing code quality — often a different agent than the one that wrote it. > [!quote] [00:31:36] it turns [clears throat] out that agents if you ask them to evaluate their own output they think it's really good - Honest limit: Patrick says there's no answer yet to whether a company-wide AI rollout is delivering ROI. #### ⏱️ [00:32:38](https://www.youtube.com/watch?v=GuwoULakxdY&t=1958) Panel handoff - Sharon Goldman moderates the closing panel with Cisco, Deloitte, and Glean. #### ⏱️ [00:34:15](https://www.youtube.com/watch?v=GuwoULakxdY&t=2055) Why transformation is still hard - Sharon's observation from enterprise executives: FOMO and pressure, but not scaled outcomes. - Arvind's bottleneck: agents need access to as much information and experience as the human experts already have. - Second bottleneck: the technology is so open-ended that prioritization is hard, and companies need peer success stories rather than always being at the frontier. #### ⏱️ [00:36:50](https://www.youtube.com/watch?v=GuwoULakxdY&t=2210) Deloitte's scaling example > [!quote] [00:36:50] there's 180,000 people in the US firm in Deote and 450,000 people globally - Use case: matching skills, resumes, and availability to roles and clients — a task no human can do by reading 180,000 resumes one at a time. - Human in the loop makes the judgment calls; the workflow itself had to be reimagined. - Ashish's constraint at [00:37:25]: you can't automate one step of a process and leave the human sequence around it intact. #### ⏱️ [00:38:40](https://www.youtube.com/watch?v=GuwoULakxdY&t=2320) Four market forces — force one, inference as the workload > [!quote] [00:38:40] the global compute capacity 60% of that global compute capacity is actually now consumed by inferencing not by training #### ⏱️ [00:39:21](https://www.youtube.com/watch?v=GuwoULakxdY&t=2361) Force two, agents as the workforce - Cited source: an OpenRouter study, "February of last year," where agent token consumption first exceeded human token consumption. > [!quote] [00:39:21] Since then, it went up 14x. And so now the consumption of tokens by agents is five times that of the consumption of tokens by humans. #### ⏱️ [00:40:04](https://www.youtube.com/watch?v=GuwoULakxdY&t=2404) Forces three and four, tokenomics and control - Tokenomics: cost of tokens has to go down, unlocking a class of new effects. - Control: models need to sit inside the enterprise, governed by it — IP and data sovereignty are the reasons, and openness to open-weights and open-source models is rising. - Novel structural claim: this is the first major platform shift where you don't need 70–80% human adoption to create a super cycle. > [!quote] [00:40:49] Right now, sub 5% of the humans are power users of agents. - The supply argument: that sub-5% has already consumed all creatable infrastructure capacity; at 80% human power-user adoption there may not be enough infrastructure. #### ⏱️ [00:41:29](https://www.youtube.com/watch?v=GuwoULakxdY&t=2489) Secure delegation as the dividing line > [!quote] [00:41:29] the ability to securely delegate work to an agent versus the ability to just delegate work to an agent. Is the difference between market leadership and complete abject bankruptcy - Mechanism named: a rogue agent is a business-ending event, so monitoring and dynamic guardrails matter. - Market prediction: security and observability fuse, because you can't distinguish a prompt-injection-influenced agent from one that decided to jailbreak on its own — his example is an agent going to Hugging Face to steal eval results. #### ⏱️ [00:42:09](https://www.youtube.com/watch?v=GuwoULakxdY&t=2529) Liftoff timing > [!quote] [00:42:09] Enterprises will um massively have a liftoff moment in FY27 - Paired claims: inferencing becomes 70–80% of the workload over time, but scaling laws persist and absolute training dollars keep rising. #### ⏱️ [00:43:27](https://www.youtube.com/watch?v=GuwoULakxdY&t=2607) Agent development is democratic, governance is not uniform - Arvind: some companies let everyone build agents, others designate a few builders; each philosophy implies a different governance model, and platforms must support both, including regulation-driven controls. - He endorses building agents across multiple systems to avoid single-vendor dependence and keep switching open. #### ⏱️ [00:44:53](https://www.youtube.com/watch?v=GuwoULakxdY&t=2693) What to centralize > [!quote] [00:44:53] So while you can have distributed development, many different systems where you build agents, there are certain things that you should be centralizing for the reasons of governance. - The three things to centralize: (1) context — how agents reach enterprise systems, get data, and take actions; (2) security and governance — access scoped to whoever the agent runs on behalf of; (3) observability — which agents exist, their impact, cost, owning department, and business value. - Argument for centralizing: without it, bad things happen and agent development gets shut down. He says Glean has seen this with customers, without naming them. #### ⏱️ [00:46:38](https://www.youtube.com/watch?v=GuwoULakxdY&t=2798) The workbench model - Ashish reports his chief people officer asking whether agents need a job family; he says he isn't there mentally. - Deloitte's shipped answer: every person has a workbench of tools and agents keyed to their persona, running on model routing and contextuality. - Still open: what humans decide versus what gets delegated, and the guardrails on agent-taken actions. - He notes he's seen the beginnings of a console or dashboard for this in Glean's announcements — an indirect pointer to the morning keynote. #### ⏱️ [00:48:33](https://www.youtube.com/watch?v=GuwoULakxdY&t=2913) Bottlenecks relocate; jobs get refactored - Cisco speaker rejects the all-jobs-disappear thesis in blunt terms. > [!quote] [00:49:11] So coding gets automated, code review becomes a bottleneck. Code review gets automated, taste and judgment and what product to build becomes a bottleneck. - His prediction: 100% of jobs get refactored, some are lost, entirely new industries appear. - Last-mile argument with a worked anecdote at [00:49:53]: when Claude first came out, people generated 50-page documents that arrived as 150-page documents, and nobody had read their own document. #### ⏱️ [00:50:26](https://www.youtube.com/watch?v=GuwoULakxdY&t=3026) Entry-level hiring - Rejects the stop-hiring-juniors thesis. Mechanism: without fresh thinking you lose the ability to unlearn, and unlearning happens when someone pushes you. - He names student interns at Cisco as his biggest coaches. > [!quote] [00:51:03] entry-level jobs are not going to be gone. In fact, they'll be more valuable now than ever before - Requirement he attaches: you'll need dexterity in AI to be qualified for a job. #### ⏱️ [00:52:24](https://www.youtube.com/watch?v=GuwoULakxdY&t=3144) Arvind's 6–12 month prediction — proactive AI - Window given: next 6 to 9 months to a year. - Use case: planning your day and prioritizing time — a guide, coach, or chief of staff directing where time goes. - Current state he contrasts against: AI is a tool you ask to do work, not something directing your day. #### ⏱️ [00:53:38](https://www.youtube.com/watch?v=GuwoULakxdY&t=3218) Ashish on what won't be delegated - Judgment learned over time in a role. Named domains where human judgment holds: drug discovery, clinical trials, underwriting. - He rejects the "apprenticeship model is gone, only hire experts" position as a sound bite. > [!quote] [00:54:09] You cannot bring somebody without apprenticeship that looks like they have 10 years of doing anything. #### ⏱️ [00:54:42](https://www.youtube.com/watch?v=GuwoULakxdY&t=3282) The Cisco prediction — the hardest number in the session > [!quote] [00:54:42] In six months, Cisco will have half a dozen products 100% written with AI. No human lines of code. In 12 months, 70% of our $72 billion in revenue products will be written with AI, no human lines of code. In 24 months, Cisco will have no legacy code. - Attached claim: engineering headcount will be higher then than today. #### ⏱️ [00:55:41](https://www.youtube.com/watch?v=GuwoULakxdY&t=3341) Close and day-two preview - Day two includes the AI native showcase — early-stage AI-native companies presenting solutions and where they see the next wave. > [!quote] [00:55:41] I'm thrilled to introduce the inaugural Gingo AI Innovator Awards. - Award criteria as stated: building new workflows, scaling adoption, enabling people, strengthening enterprise foundations, and sharing what they learn. - Logistics: reception at Gateway Pavilion, "a two-minute walk or five minute swim." Event is hybrid — San Francisco plus online. ### Overlap map — what duplicates the main Glean:GO 2026 keynote | Topic | Status for a reader processing multiple sessions | | --- | --- | | Enterprise context, token efficiency, governance, AI front doors | **Duplicate framing.** Named at [00:05:35] as the day's themes. The morning keynote owns the argument; this session only supplies customer color. | | Agents with their own identity | **Duplicate.** Explicitly attributed to an earlier keynote at [00:17:08]. Nothing new added. | | Proactive AI | **Duplicate.** Arvind says at [00:52:24] it was covered today; he adds only the day-planning use case and a 6–9 month window. | | eBay's per-user AI spend dashboard | **Partial duplicate.** John says at [00:09:27] he covered it in an earlier session the same day. | | A console or dashboard for agent oversight | **Duplicate pointer.** Ashish references Glean's own announcements at [00:48:00]; no detail added here. | | Ericsson adoption (50–70%), POC and rollout dates, 150-year-old company | **New.** | | eBay connector expansion, three-year timeline, revolution line | **New.** | | Databricks: 10,000 employees, 2,000+ engineers, AI-first authorship, second-agent code review | **New.** | | Ericsson AI mesh and responsible-AI questionnaire agent | **New.** | | Deloitte: 180,000 US / 450,000 global, resume-matching use case, workbench | **New.** | | Cisco market forces: 60% inference, 14x, 5x, sub-5%, FY27, $72B / 24-month predictions | **New in this session** — and the highest-risk numbers to reuse, since only one is sourced. | | Glean:GO AI Innovator Awards, AI native showcase, Gateway Pavilion reception | **New** (logistics and program). | ## 📚 References - [Day 1 Afternoon Keynote | Glean:GO 2026](https://www.youtube.com/watch?v=GuwoULakxdY) — primary source, 57:36. - [Glean](https://www.glean.com/) — vendor and event host. - [OpenRouter](https://openrouter.ai/) — the token-consumption study cited at [00:39:21] was attributed to OpenRouter with no link or title given on stage; treat the 14x and 5x figures as unverified until the study is located. - [Databricks](https://www.databricks.com/) — Patrick's coding-cost optimization note was pointed to on X at [00:26:41]; no handle or URL was read out. - [WCAG 2.2](https://www.w3.org/TR/WCAG22/) — standard used for the accessibility mapping below. ## 🗣️ Paraphrase - **Customer fireside.** Two large, old companies — a 30-year-old marketplace and a 150-year-old telecom — both bought Glean to fix internal search, and both found that deploying it exposed data hygiene problems they had assumed were solved: untagged policies, oversharing, unknown personal-data locations. Both now treat cost visibility as the first governance move and control as the second. eBay's CIO argues that usage-based pricing defanged the licensing side of tool sprawl but left the user-confusion side untouched. Ericsson's chief data officer argues governance has to move from upfront checklists to runtime monitoring, and says Ericsson is building an internal connective layer for that. - **Founder conversation.** Databricks' co-founder describes nearly blocking the Glean purchase on security grounds, being converted by an unauthorized internal demo, and now running a company where AI drafts nearly every line of code. His two live problems are unbounded consumption cost across 2,000+ engineers, and quality: volume is up, quality is down, and models remain weak at architecture because seasoned engineers bring unwritten context. He measures code quality with a separate reviewing agent, on the grounds that agents rate their own work generously, and admits there's no working ROI measure yet. - **Panel.** Glean's Arvind frames the core bottleneck as agents lacking the information human experts already have, and prescribes distributed agent building with three things centralized: context, access control, and observability. Deloitte's Ashish describes a resume-and-availability matching system across a 180,000-person US firm and a per-persona workbench, and insists role-earned judgment and the apprenticeship model survive. Cisco's speaker supplies the market numbers and the boldest prediction: half a dozen Cisco products fully AI-written in six months, 70% of a $72 billion revenue base in twelve, no legacy code in twenty-four, with a larger engineering headcount than today. ## 📰 Newsletter Glean closed day one of Glean:GO 2026 with something unusual for a vendor keynote: no product news. Instead you get eBay, Ericsson, Databricks, and Deloitte describing what running this stuff actually costs them. If you're deciding whether to watch it, here's the short version. The reusable operational patterns, in rough order of how fast you could copy them: - **Show people their own AI bill.** eBay shipped a dashboard in the last couple of months that shows every employee their AI spend across tools. The CIO's line — that people assume AI is all-you-can-eat like a word processor — is the most transferable insight in the session. - **Coach the tail, not the mean.** Ericsson found token consumption follows a power curve, so a handful of individuals drive spend, and coaching those people is enough control for now. - **Never let an agent grade its own homework.** Databricks reviews AI-written code with a *different* agent, because agents asked to evaluate their own output think it's great. - **Move governance from checklist to runtime.** Ericsson's chief data officer says pre-screening paperwork can't keep pace, and has an agent that ingests an IT architecture and fills in the responsible-AI questionnaire. For anyone who works in open source and community governance, the panel exchange on entry-level hiring is worth the click. Cisco's speaker argues junior roles get *more* valuable, not less, because unlearning only happens when someone new pushes back — and Deloitte's Ashish flatly rejects the "the apprenticeship model is dead" line as a sound bite. That's the same argument our contributor-onboarding people have been making for years, arriving from a completely different direction. Fair warning on the numbers: the eye-catching market stats (60% of compute on inference, agents using 5x the tokens humans do, FY27 liftoff) come from a single speaker, and only one of them is attributed to a source — an OpenRouter study that isn't linked. Quote them with the attribution attached, not as facts. ## ✅ Action items #### DevRel - Developer advocacy: pitch a talk on "measuring AI coding ROI when you can't" built from Patrick's admission at [00:31:36] that Databricks has no working ROI answer; the honesty is the hook. - Education: build a 45-minute workshop on the two-agent review pattern — one agent writes, a different agent reviews — with a runnable demo and a failure case where self-evaluation passes bad code. - Developer experience: write a quickstart that adds per-user token spend visibility to an internal tool, modeled on eBay's dashboard described at [00:09:27]. - Developer marketing: publish a plain comparison of per-seat versus usage-based AI pricing that tests eBay's claim at [00:13:56] that usage-based pricing makes sprawl cost-neutral; attribution and a worked example, no vendor framing. - Community management: run a survey asking contributors which AI front door they actually use and whether they know what a session costs; eBay says confusion, not cost, is the unsolved sprawl problem. - Internal/external/cross-project: open a feedback loop with your Glean admins on the agent publication and visibility workflow Ericsson describes at [00:15:50]; confirm whether it's available on your tenant, since this session never states availability. #### OSPO - Policy & Governance: draft a runtime-monitoring policy that replaces at least one pre-screening checklist, following Ericsson's stated shift at [00:11:57] to control environments plus active deviation management. - License Compliance: add a review step for AI-drafted code provenance; Databricks' "almost every line... authored initially by some AI" at [00:25:30] makes attribution and license-contamination questions concrete rather than hypothetical. - Security Management: write the rogue-agent incident playbook. The Cisco speaker's point at [00:42:09] that prompt injection and self-initiated jailbreak look identical in logs is the design constraint — your intake form needs to not require distinguishing them. - Software Inventory: extend SBOM ownership to include agents — name, owner, cost center, data access scope — matching the observability items Arvind lists at [00:45:27]. - Developer Support: hold consultation hours on "which tool do I build this agent in," the exact user confusion eBay reports at [00:15:11]. - Community Engagement: share your token-cost findings upstream; no speaker in this session published a dollar figure, so the ecosystem has a real data gap. - Strategic Alignment: build an agent-portfolio KPI set covering count, cost, owning department, and business value — the four dimensions Arvind names at [00:45:27]. #### Community - Maintainers: run a governance clinic on distributed agent building with centralized context, access control, and observability, using Arvind's three-centralization rule at [00:44:53] as the agenda. - Contributors: create a "first agent" rotation modeled on Ericsson's decision to let employees build their own agents, with a publication review before anything becomes visible org-wide. - Sponsors: ask sponsors to fund a public token-cost benchmark. Every cost claim in this session is qualitative, which means nobody in the community has a number to plan against. - Ecosystem: run a cross-project tabletop on an agent with over-broad permissions acting on behalf of a maintainer, and practice the coordinated disclosure path. #### WordPress - Core: triage AI-assisted contributions with a stated review expectation; the 150-page-document anecdote at [00:49:53] is what happens without one. - Core Performance: profile agent-driven traffic separately from human traffic in any AI-assisted tooling; the inference-versus-training split at [00:38:40] is the macro version of the same measurement problem. - Core AI: build an evaluation set that includes architecture-level tasks, since Patrick's claim at [00:28:09] is that models fail specifically there and not on well-specified modules. - **Core Program Team:** add a standing retro question on which bottleneck moved this cycle, borrowing the coding → review → taste progression at [00:49:11]. - Mobile: review whether any AI features in the apps expose usage cost to the user; eBay's finding is that invisible cost drives careless use. - Playground: build a reproducible lab that demonstrates an agent grading its own output favorably versus a second agent grading it — a teaching module straight out of [00:31:36]. - Community: add an "AI tooling at your org" session format to meetup kits using the four-theme structure named at [00:05:35]. - Training: write a lesson on prompting for architecture versus prompting for implementation, grounded in the task-delineation argument at [00:29:51]. - Marketing: avoid restating vendor market stats; if the 60%, 14x, or 5x figures appear in any WordPress-adjacent content, require the OpenRouter citation first. - Support: document the "where do I go to ask" problem in your own support surfaces; eBay calls this the unsolved half of sprawl. - Design: audit any AI feature for whether it makes cost legible at the moment of use. - Documentation: add a governance page covering runtime monitoring of AI tooling, following the Ericsson pattern rather than a pre-approval checklist. - Photos: not touched by this source; no action derived. - TV: caption quality is directly relevant — this transcript garbles Deloitte, Waymo, and at least one speaker's name, which is a searchability failure worth guarding against in WordPress.tv uploads. - Polyglots: note Joel's Waymo-spoke-Swedish remark at [00:20:12] as a reminder that multilingual interaction is now an expectation, not a feature request. - Accessibility: request captions-plus-transcript parity for recorded conference sessions; see the Accessibility Assessment below for the specific WCAG 2.2 mapping. - Meta: consider exposing per-user resource cost in .org contributor tooling if any AI features land there. - CLI: if AI-assisted command generation lands, add a cost/token display to the output. - Hosting: treat agent traffic as a distinct workload class in host-side telemetry. - Tide: add a check for AI-authorship signals in plugin code if and when a reliable one exists; Databricks' authorship ratio makes the question unavoidable. - Openverse: no direct tie in this source; no action derived. - Themes: add review guidance for AI-generated theme code volume, since the quantity-up quality-down pattern at [00:27:28] applies directly to review queues. - Plugins: same as Themes — review throughput is the bottleneck that moves when authoring gets cheap. #### The WPOCC - Programs: fund a fellowship track on AI governance for open source projects, using Ericsson's generation-by-generation governance revision at [00:11:19] as the case study. - Projects: run a documentation sprint producing a public, vendor-neutral agent observability checklist from Arvind's four dimensions. - Funding experiments: pilot a milestone-based microgrant for a maintainer to publish real token-cost data from an AI-assisted project — the number no speaker in this session would give. ### FAIR - Technical independence: Arvind's argument at [00:44:13] for building agents across multiple systems to avoid vendor dependence is the same argument FAIR makes about distribution; capture it as external corroboration. - Policies and covenants: use the Cisco speaker's data-sovereignty point at [00:40:04] — models governed inside your own boundary — as supporting material for FAIR's independence covenants. - Productized governance artifacts: draft an operator scorecard covering centralized context, permission scoping, and agent observability, mirroring the three-centralization structure at [00:44:53]. ## 🧪 Hypotheses 1. **Cost visibility alone changes behavior.** If per-user AI spend is made visible without any hard limits, per-user token consumption drops measurably within one billing cycle. eBay reports an immediate effect at [00:10:42] but gives no magnitude. 2. **Token spend follows a power law.** In a large enterprise, the top decile of users accounts for a majority of token spend, making targeted coaching more effective than org-wide limits. Joel asserts the power curve at [00:08:50] without data. 3. **Cross-agent review outperforms self-review.** Code reviewed by a different agent than the one that authored it shows a higher defect-detection rate than self-review. Databricks operates on this belief at [00:31:36] without publishing a comparison. 4. **Model performance degrades with decision horizon, not task size.** Models underperform on tasks whose feedback arrives over years, independent of how well the task is specified. This is the strongest falsifiable form of Patrick's argument at [00:30:24]. 5. **Usage-based pricing does not make tool sprawl cost-neutral.** Moving users between AI front doors changes total cost, because tools route to different models at different per-token prices. This contradicts eBay's claim at [00:13:56]. 6. **Entry-level headcount and AI adoption rise together.** Organizations with higher AI tooling adoption show flat or rising junior hiring, not decline — the empirical version of the Cisco claim at [00:51:03]. ## 📊 Methodology - **H1 (cost visibility).** Staggered rollout of a spend dashboard across matched org units, with a difference-in-differences comparison of per-user token consumption over 60 days. Data source: your own gateway billing logs. Control for a concurrent change in task mix. - **H2 (power law).** Fit per-user monthly token totals to a power-law and to a log-normal distribution; compare with a likelihood-ratio test. A single month of gateway logs suffices. Report the Gini coefficient alongside, since "power curve" is used loosely on stage. - **H3 (cross-agent review).** Seeded-defect study: inject a known set of defects into N pull requests, route each to self-review and to an independent-agent review, and compare detection rates with a paired test. Blind the reviewing agent to authorship metadata. - **H4 (decision horizon).** Build an eval set stratified by feedback latency — same specification quality, different horizons (unit-test-verifiable, week-scale, year-scale). Score with expert human raters blind to whether a human or model produced the design. This is the test Patrick's argument implies but nobody has run publicly. - **H5 (pricing neutrality).** Instrument identical task suites across two AI front doors and compare total cost per completed task. One week of paired runs on a fixed task set answers it. - **H6 (junior hiring).** Panel regression across public headcount data and disclosed AI adoption, with industry and revenue controls. Note the confound: firms adopting AI fastest may also be the firms growing fastest for unrelated reasons. - **Source-quality caveat for all of the above.** Nothing in this session is a study. Every figure is either a self-report from one company or an uncited market statistic, so all six hypotheses start at zero external evidence. ## ❓ Questions - Which of the Glean capabilities customers describe — agent publication and visibility controls, the agent oversight console Ashish references, agent identity — are actually available on a standard tenant today, and which are early-access? This session never says. - What per-user or per-seat cost does eBay's dashboard actually display, and does it include model cost, retrieval cost, or both? - What is the OpenRouter study behind the 14x and 5x figures, and does "consumption by agents" count tokens routed through agentic harnesses that a human initiated? - Does Ericsson's 50–70% active usage measure monthly actives, weekly actives, or accounts provisioned? - What is Databricks' actual definition of "authored initially by some AI" — a completion accepted, a file created, a diff proposed? - If Cisco has no legacy code in 24 months while growing engineering headcount, what are those engineers doing? The claim at [00:54:42] implies a job-content answer that wasn't given. - Who owns the agent when it acts on your behalf and is wrong? Ashish raises the guardrails question at [00:47:23] and explicitly says there's no clear-cut answer. - What does the "AI mesh" actually consist of at Ericsson — a policy layer, a gateway, a service mesh, or a naming convention? ## 🔮 Future research - **The ROI gap is the biggest open problem named on stage.** Patrick states at [00:32:06] that nobody has an answer to whether a company-wide AI rollout is working. A published, replicable measurement framework — even a bad one — would be more valuable right now than another capability. - **Latent context representability.** Patrick's claim at [00:29:17] is that human architectural judgment draws on context that can't be written down. That's an empirical question about what fraction of expert decision inputs are externalizable, and it has never been measured. - **Security-observability convergence.** The Cisco speaker predicts the two markets merge at [00:42:09]. Nobody has published a detection method that distinguishes prompt-injection-driven agent behavior from self-initiated deviation, which is the technical precondition for that merge. - **Agent decision-loop provenance.** Joel's concern at [00:16:25] — where traces and tool calls accrue, and who gets to learn across your workflows — is a data-rights question with no established standard. This is the single most transferable research question for open source governance. - **Infrastructure ceiling modeling.** The claim that sub-5% power-user adoption already saturates capacity implies a model of compute supply that was asserted, not shown. A public model with stated assumptions would let enterprises plan. - **Apprenticeship under automation.** Ashish and the Cisco speaker both argue apprenticeship survives, on reasoning alone. Longitudinal data on time-to-competence for juniors in AI-heavy versus AI-light teams would settle it. ## 💡 Implications - **Practice.** The default AI rollout sequence implied by all three customer stories is: connect sources, fix the data defects that surfaces, make cost visible, then add runtime controls. Governance came last at every company on stage, which is a warning, not a recommendation. - **Policy.** Ericsson's move at [00:11:57] from pre-approval checklists to control environments plus active deviation management is a policy pattern any regulated organization can copy, and it changes what compliance staff do day to day — from gatekeeping to monitoring. - **Security.** Two concrete asks fall out: scope agent permissions to the identity the agent acts for, and centralize the inventory of which agents exist. Arvind's warning at [00:45:27] is that skipping these leads to an incident that shuts down agent development entirely. - **Education.** If the bottleneck moves from writing code to reviewing it and then to taste and judgment, curricula that teach code authorship are training for the automated layer. Teaching review, architecture, and specification is the durable investment — and both Ashish and the Cisco speaker argue juniors are how you get there. - **Funding models.** Consumption pricing with unbounded per-user exposure is fundamentally different from seat licensing for any organization with a fixed budget — which describes every nonprofit, foundation, and volunteer project in open source. No speaker addressed what this pricing model does to organizations that can't absorb variance. ## 🧠 Learning styles assessment - **Reading level estimate:** Flesch-Kincaid Grade Level approximately 9–10 on the transcript text. This is an estimate computed over sampled passages, not the full corpus, and ASR run-on sentences inflate it — the spoken register is conversational and closer to grade 7–8 by vocabulary. Flesch Reading Ease estimate: 55–65 (fairly difficult to standard). Sentence-length variance is high because the captions merge clauses without punctuation. #### 👀 Visual learners - Present: a live stage with three seated conversation formats, applause and audience-reaction cues, and an awards segment. No slides, charts, diagrams, or product screenshots appear anywhere in the transcript. - Improve: a single slide behind the Cisco four-forces segment would carry the 60% / 14x / 5x / sub-5% numbers; an agent-governance diagram showing the three centralized layers (context, access control, observability) would make Arvind's argument at [00:44:53] stick in one image. #### 🔊 Aural learners - Present: this is a native audio-first format — three unscripted conversations with distinct voices, laughter, and interruption. The strongest medium match of the eight. - Improve: publish a chaptered audio version split at the three segment boundaries (00:00:54, 00:21:00, 00:33:00), plus a 3-minute numbers-only cut for people who just want the market statistics. #### 🗣️ Verbal learners - Present: several strong analogies — agents as contractors and consultants in the workforce at [00:17:49], the first-Waymo-ride feeling at [00:20:12], asking someone about their own paper at [00:32:06]. - Improve: define "AI mesh," "workbench," "tokenomics," and "AI front doors" on first use. All four are introduced as if the audience shares the definition, and none is explained. #### 🧩 Kinesthetic learners - Present: nothing hands-on. No demo, no sandbox, no walkthrough. - Improve: pair the recording with a reproducible exercise — stand up a two-agent review loop, or build a minimal token-spend dashboard against your own gateway logs. Both are half-day tasks that convert the session's two best ideas into muscle memory. #### 🧠 Logical learners - Present: two well-structured causal arguments — Patrick's feedback-loop-length explanation for architectural weakness, and the Cisco speaker's bottleneck-relocation chain. Arvind's centralize-three-things list is cleanly enumerated. - Improve: the market-forces segment needs a sourcing table — claim, source, date, method. As delivered, four numbers arrive in 90 seconds with one partial attribution. #### 🧑‍🤝‍🧑 Social learners - Present: the entire format is dialogue, with visible disagreement (Ashish pushing back on the expert-only hiring thesis) and peer learning framed explicitly by Arvind at [00:35:38]. - Improve: publish a discussion guide with three prompts per segment so viewing teams can run it as a 60-minute group session rather than passive watching. #### 🧘‍♀️ Solitary learners - Present: the transcript is complete and self-contained; a solo viewer gets everything. - Improve: add a self-check — "which of the three cost-control patterns applies to your org, and what would you measure in the first 30 days?" — and a journaling prompt on which bottleneck has already moved in your own work. #### 🌿 Naturalistic learners - Present: the infrastructure-capacity argument at [00:40:49] is implicitly an environmental and resource-limits argument, but energy, water, and carbon are never named. - Improve: state the resource cost directly. If sub-5% power-user adoption saturates available compute, the energy footprint of the 80% scenario is the obvious follow-up question and nobody asked it. ## 📊 Accessibility assessment #### Screen reader semantics - Present: as a YouTube video, the page provides standard player landmarks and a title. The auto-generated transcript is machine-readable. - Improve: the transcript lacks speaker labels for most turns — segments alternate with only ">>" markers, and a screen reader user cannot tell who is speaking. Add speaker attribution. Maps to **WCAG 2.2 AA 1.3.1 Info and Relationships**; the missing speaker identification is a structure failure, not just a convenience gap. #### Keyboard navigation - Present: the YouTube player supports keyboard control and visible focus on standard controls. - Improve: chapter markers would give keyboard users a way to jump between the three segments without scrubbing. Maps to **WCAG 2.2 AA 2.4.3 Focus Order** and **2.4.7 Focus Visible** for any custom embed. #### Contrast and readability - Present: not assessable from the transcript alone — no slides or on-screen text are described. - Improve: any lower-third name-and-title graphics need 4.5:1 minimum against the stage background, and speaker names should also be spoken, since they apparently weren't clear enough for the captioner. Maps to **WCAG 2.2 AA 1.4.3 Contrast (Minimum)**. #### Text resizing and reflow - Present: caption sizing is user-controlled in the YouTube player. - Improve: for the event's own site, verify the recording page reflows at 200% zoom without clipping the player controls. Maps to **WCAG 2.2 AA 1.4.4 Resize Text** and **1.4.10 Reflow**. #### Media and captions - Present: captions exist, with timestamps, and a transcript is obtainable. - Improve: caption accuracy is the clearest failure in this source. "Deote"/"Deoid" for Deloitte, "Whimo" for Waymo, "Gingo" for Glean:GO, "Gleam" for Glean, "Erikson" for Ericsson, and an apparent name garble for the Cisco speaker. Proper nouns are exactly what a deaf or hard-of-hearing viewer cannot recover from context. Maps to **WCAG 2.2 A 1.2.2 Captions (Prerecorded)** — auto-captions at this error rate on names do not satisfy it. Add audio description or a described-events note for the applause, awards, and stage-movement cues; maps to **1.2.5 Audio Description (Prerecorded)** at AA. #### Forms and interactions - Not applicable to this source in the direct sense — the video has no forms. The adjacent interaction is the event registration and awards signup mentioned at [00:56:33]; those pages should meet **WCAG 2.2 AA 3.3.1 Error Identification**, **3.3.2 Labels or Instructions**, and the new **2.5.8 Target Size (Minimum)**. #### Standards reference - Primary gaps against WCAG 2.2 AA, in priority order: **1.2.2 Captions** (name accuracy), **1.3.1 Info and Relationships** (missing speaker labels in the transcript), **1.2.5 Audio Description**, then **2.4.3/2.4.7** for any custom player embed. The caption defect is the one that actually blocks comprehension. ## 🌍 DEIB assessment #### Cultural breadth - Present: genuine multi-region representation — Ericsson brings a European regulatory and AI-governance perspective that differs materially from the US speakers, and Joel raises sovereignty at [00:08:10] as a first-class concern rather than an afterthought. Deloitte's 450,000-person global figure implies scale beyond the US firm. The Swedish-language Waymo remark at [00:20:12] is a small but real multilingual signal. - Improve: no speaker from Asia, Latin America, or Africa, and no discussion of how token cost or infrastructure availability lands differently in regions with less compute capacity. The infrastructure-scarcity argument at [00:41:29] has obvious global-equity consequences that go unmentioned. #### Inclusive language - Present: generally neutral. "Human in the loop," "knowledge worker," and "power user" are used consistently and are not gendered. - Improve: several undefined terms ("AI mesh," "tokenomics," "workbench," "harnesses," "outer loop") assume insider fluency, which excludes newer practitioners. One speaker uses "BS" and "freaking" at [00:48:33] and [00:49:53], and "Dude" as a generic address — informal rather than exclusionary, but worth noting for a redistributed corporate recording. "Complete abject bankruptcy" and "bonkers" are rhetorical intensifiers, not slurs. #### Neurodiversity support - Present: clear three-segment structure with explicit host handoffs, and each segment runs a predictable question-and-answer rhythm. - Improve: the panel section runs long stretches without a structural cue — the Cisco four-forces answer at [00:38:40] runs several minutes in one turn. Chapter markers, an on-screen enumerated list during multi-part answers, and a published agenda would reduce working-memory load. The unlabeled speaker turns in the transcript compound this. #### Representation - Present: multiple industries (telecom, e-commerce, data infrastructure, professional services, networking), multiple functions (CDO, CIO, security and deployment owner, product). The moderator roles include a woman hosting the customer fireside and Sharon Goldman moderating the panel. - Improve: every named executive speaker is male; the two women in the session are both in moderating roles rather than as subject-matter panelists. On a five-panelist, three-segment program, that pattern is visible and correctable. No speaker from a smaller organization, a nonprofit, or the public sector — every data point comes from companies with 10,000+ employees, which narrows what "enterprise AI" is allowed to mean. #### Psychological safety - Present: unusually good for vendor content. Patrick says outright at [00:27:28] that Databricks doesn't have all the answers, and at [00:32:06] that nobody knows how to measure ROI. Ashish says at [00:47:23] there's no clear-cut answer on human-in-the-loop design. Disagreement happens on stage without hostility — Ashish pushes back on the hire-only-experts position and the Cisco speaker agrees. Admitting uncertainty in front of customers is the healthiest signal in the session. - Improve: the job-loss discussion at [00:48:33] is handled dismissively ("that's such BS at such epic proportions") in a room where some attendees may be personally exposed to exactly that risk. The underlying argument is sound; the framing forecloses the concern rather than addressing it. A content note and a direct acknowledgment that some jobs are lost — which the same speaker concedes seconds later — would land better. ## ✅ Logical validity #### Strong reasoning - **Causal, well-specified — feedback-loop length explains architectural weakness.** At [00:28:09]–[00:30:24], Patrick starts from a puzzle (a seasoned engineer can finish a design doc from written constraints, so why can't a model?), proposes a mechanism (experts bring latent, unwritten context), and then supplies an independent supporting mechanism (architectural decisions have multi-year feedback loops, so there's no tight learning signal). It's the only argument in the session with a stated mechanism, a stated limit, and an acknowledged uncertainty — he says he doesn't know whether models can do this even in principle. To clarify further: separate the two mechanisms, because "the context isn't written down" and "the feedback loop is too long" have different fixes. - **Inductive, convergent — bottleneck relocation.** Patrick observes it at [00:26:06] ("the writing of an individual piece of code is not a bottleneck anymore") and the Cisco speaker states the general form at [00:49:11] with a two-step chain. Two independent speakers from unrelated companies reaching the same structure is the strongest inductive support in the session. - **Deductive, valid given its premise — self-evaluation is unreliable.** At [00:31:36], the premise (agents rate their own output highly) supports the practice (use a different agent to review) with an apt human analogy. The premise is testable and the conclusion follows. - **Analogical, disciplined — agents as a labor category.** John's extension at [00:17:49] from employees and contractors to employees, contractors, and agents is a bounded analogy: he explicitly says the right mix depends on department and type of work rather than claiming uniform substitution. - **Causal with a named mechanism — why juniors matter.** At [00:50:26]–[00:51:03], the Cisco speaker doesn't just assert that entry-level roles survive; he gives a mechanism (unlearning requires someone pushing you) and supporting personal evidence (student interns as his coaches). The mechanism is what makes it more than a preference. #### Fallacies - **Unsourced statistic presented as settled fact — [00:38:40].** "the global compute capacity 60% of that global compute capacity is actually now consumed by inferencing not by training" arrives with no source, date, or definition of "global compute capacity." Weakness: the figure does the load-bearing work for the entire "inference is the new workload" force, and it's unverifiable as stated. Fix: name the source and the measurement basis — installed accelerator-hours, revenue, or utilization — since those three give different answers. - **Partial attribution treated as full citation — [00:39:21].** The 14x growth and 5x agent-to-human ratio are attributed to "Open Router" and "February of last year" with no study title, link, or methodology. "February of last year" is itself ambiguous in a talk recorded in 2026. Weakness: agent-attributed tokens include tokens a human initiated through a harness, so the human/agent split depends entirely on a definition that isn't given. Fix: cite the study and state the attribution rule. - **Non-sequitur in the capacity argument — [00:40:49].** The claim that sub-5% power-user adoption has "fully consumed all the capacity that we can create from an infrastructure standpoint" doesn't follow from the token-growth figures; capacity saturation requires a supply-side number that never appears. Weakness: it converts a demand-growth observation into a supply-ceiling conclusion without the supply data. Fix: supply the capacity figure, or state it as a projection rather than a present fact. - **False dilemma — [00:41:29].** "the ability to securely delegate work to an agent versus the ability to just delegate work to an agent. Is the difference between market leadership and complete abject bankruptcy" presents two outcomes at the extremes of a continuum that obviously contains many middle states. Weakness: rhetorically effective, analytically empty — it gives a buyer no way to reason about degrees of risk. Fix: state the actual risk gradient, e.g. blast radius by permission scope. - **Unfalsifiable-as-stated prediction — [00:54:42].** "In 12 months, 70% of our $72 billion in revenue products will be written with AI, no human lines of code" is precise-sounding but unverifiable: "written with AI" is undefined, "no human lines of code" conflicts with any human editing, and "70% of revenue products" could be measured by product count, revenue weight, or line count. Weakness: the specificity of "$72 billion" lends borrowed credibility to an unmeasurable claim. Fix: define the metric and the audit method, or present it as an aspiration. - **Anecdote as proof — [00:22:11]–[00:23:44].** The Databricks origin story turns on one employee's laptop demo answering one question well. Weakness: an n-of-1 demo is the weakest possible evidence for indexing-everything being safe, and the story's actual lesson — the deployment happened without security approval — is acknowledged and then dropped. Fix: pair the anecdote with the security review outcome, which Patrick mentions happened but never describes. - **Anecdote as proof, second instance — [00:19:38].** Joel's single Glean interaction about project status supports a broad claim about daily decision-stall detection. Weakness: one successful query doesn't establish reliability. Mitigating factor: he frames it explicitly as a wish and an imagined extension, which partly inoculates it. - **Hasty generalization on pricing — [00:13:56].** "when we move to a pure usage model... then I don't worry as much about sprawl there" assumes moving a user from one tool to another leaves cost unchanged. Weakness: different tools route to different models at different per-token prices and different retrieval overheads, so the substitution isn't cost-neutral. Fix: measure cost per completed task per tool before asserting neutrality. - **Ad populum by proxy / appeal to peer success — [00:35:38].** Arvind's prescription that companies should learn from peer success stories, delivered by the vendor who curates those stories, presents social proof as a substitute for evaluation. Weakness: mild, and he's addressing a real prioritization problem. Fix: pair peer stories with the failure cases, which he alludes to at [00:45:27] but never details. - **Straw man, borderline — [00:48:33].** "most people think that all human jobs are going to go away... I think that's such BS at such epic proportions" characterizes the opposing view at its most extreme, then refutes that version. Weakness: the serious version of the concern (significant displacement in specific roles) survives his argument untouched, and he concedes at [00:49:11] that some jobs will be lost. Fix: engage the moderate claim directly. **What is NOT unsupported, checked against the transcript.** Several claims that look like vendor puffery are actually first-person operational reports and should not be flagged: the 50–70% Ericsson adoption figure, the 10,000 Databricks employees and 2,000+ engineers, the 180,000/450,000 Deloitte headcounts, eBay's connector expansion, and the existence of eBay's spend dashboard. These are self-reported and unaudited, but they are the speaker's own organization's numbers, which is the appropriate evidence class for the question asked. Likewise, Patrick's statements about quality decline and ROI uncertainty run *against* vendor interest and carry more weight for it. #### Tone of voice - **Candid and self-deprecating** in the Databricks segment. Patrick opens by admitting he almost blocked the purchase and repeatedly says "I don't know that we have all the answers." The joke about wanting customers to compete on spending money at Databricks ([00:31:05]) undercuts the token-maximizing culture rather than defending it. - **Practitioner-weary and specific** in the customer fireside. John's "I would have a revolution on my hands" and Joel's "maybe my team hates me now" are the register of people reporting from inside a deployment, not pitching one. - **High-energy declarative** in the Cisco segment. Short sentences, superlatives ("bonkers," "epic proportions," "abject bankruptcy"), numbered structures, and a closing prediction delivered over the moderator's time signal ("Screw the clock"). Persuasive and quotable; the least sourced. - **Measured and hedging** from Deloitte. Ashish repeatedly marks uncertainty — "I don't think there's a clearcut answer," "that's sort of still being discovered" — and is the only panelist who ends on "I don't have the answer." - **Warm event-host framing** in the bookends, including the fairy-godmother device at [00:17:08] and the two-minute-walk-or-five-minute-swim joke at [00:56:33]. - Overall: less promotional than typical vendor keynote content. The absence of product announcements removes the usual pitch pressure, and three of the five participants volunteer limitations of their own approach. ## 🛡️ DARVO analysis Not applicable to this source in its clinical sense — there is no accusation, accuser, or defense against wrongdoing anywhere in the 58 minutes. No speaker is responding to a charge, and no party is recast as a victim. Recording the adjacent rhetorical patterns for completeness, since two are structurally DARVO-shaped without being DARVO: - **Deny:** no denials of wrongdoing. The nearest thing is the Cisco speaker's rejection of the job-loss thesis at [00:48:33] — a dismissal of a prediction, not of an accusation, and he concedes the moderate version seconds later at [00:49:11]. Patrick's admissions run the opposite direction: he volunteers that quality has declined at [00:27:28] and that ROI is unmeasured at [00:32:06]. That is the inverse of denial. - **Attack:** no attacks on individuals or credibility. "We should not get carried away by articles in Techrunch" at [00:51:43] dismisses a category of coverage rather than a person, and it substitutes for engaging the argument — closer to a genetic-fallacy shortcut than an attack. Ashish's "it sounds like a great sound bite, but I don't believe it's true" at [00:54:09] targets a position, names it, and gives a reason. - **Reverse victim and offender:** no instances. The closest structural echo is the framing at [00:41:29] where the enterprise deploying agents is cast as the potential victim of a rogue agent — but that's a risk model, not a reassignment of blame from an accused party. - **Emotional cues and rhetoric:** the intensifiers cluster in one speaker — "bonkers," "BS at such epic proportions," "complete abject bankruptcy," "freaking annoying." These raise stakes and compress nuance, which is why the fallacy section flags the false dilemma at [00:41:29], but none of them functions to deflect accountability. - **Related fallacies already logged above:** false dilemma [00:41:29], straw man [00:48:33], genetic dismissal [00:51:43]. ## 🎯 Conclusion - **Bottom line.** This session announces nothing and demonstrates a lot. For a Glean customer, the value is four production deployments described by the people who own them — eBay (three years, near-total connector coverage), Ericsson (50–70% active, POC fall 2024), Databricks (10,000 employees, 2,000+ engineers, AI drafting nearly every line of code), and Deloitte (180,000 US / 450,000 global, resume matching, per-persona workbench). Everything else is market framing you should attribute rather than adopt. - **The three patterns worth copying.** Make per-user AI spend visible before you add limits. Have a different agent review the code an agent wrote. Move governance from pre-approval checklists to runtime control environments with active deviation management. - **The three claims worth doubting.** That usage-based pricing makes tool sprawl cost-neutral. That sub-5% power-user adoption has already exhausted global infrastructure capacity. That 70% of a $72 billion revenue base will be written with no human lines of code within twelve months. - **Practical next steps.** If you're cataloguing Glean:GO 2026, tag this session as customer-evidence and skip it for product status — four of its topics are explicitly cross-referenced back to earlier sessions the same day. Pull the adoption figures for your internal expansion case. Confirm with your Glean admins whether the agent publication and visibility workflow Ericsson describes at [00:15:50] exists on your tenant, because this session states no availability for any capability. And if you cite the market statistics anywhere, carry the attribution and the missing-source caveat with them. ## 🔄 Refactor Appendix — Day 1 Afternoon Keynote | Glean:GO 2026 > Density: 2–5 Literature atoms (H3), 0–3 Permanent candidates (H4). > Defaults: 3 Literature, 1 Permanent. Resonance filter: if you can't > articulate in one line why a claim matters, drop it. > > Konik claim-header rules: present tense, declarative, ≤10 words when > possible, searchable, unique, mirrors annotation focus if present. > > Atomicity: one declarative sentence per atom. "and"/"also" → split. > > Tag pool: #bmf · #addendum · #xref · #articleseed · #storystem · #share · #aphorisms ### 📚 Literature Split Pad #### [H1] Agents rate their own output too generously > [!quote] [00:31:36] > it turns [clears throat] out that agents if you ask them to evaluate their own output they think it's really good **Context & use**: - Databricks segment, [00:31:36] — Patrick on measuring code quality beyond token counts. - Operational consequence stated on stage: Databricks routes code review to a different agent than the one that authored it. This is the single most copyable practice in the session and applies directly to any AI-assisted review pipeline, including plugin and theme review queues. **Tags**: #xref #share **Highlight**: [view highlight](https://www.youtube.com/watch?v=GuwoULakxdY&t=1896) <!-- atomicity: 1 idea ✓ --> #### [H2] Automating one bottleneck exposes the next one > [!quote] [00:49:11] > So coding gets automated, code review becomes a bottleneck. Code review gets automated, taste and judgment and what product to build becomes a bottleneck. **Context & use**: - Panel, [00:49:11] — the Cisco speaker's rebuttal to the all-jobs-disappear thesis. - Corroborated independently by Patrick at [00:26:06] from inside Databricks engineering. Two unrelated companies reaching the same structure is why this atom survives the resonance filter. Use it whenever someone projects headcount reduction from a single-step automation. **Tags**: #xref #aphorisms **Highlight**: [view highlight](https://www.youtube.com/watch?v=GuwoULakxdY&t=2951) <!-- atomicity: 1 idea ✓ --> #### [H3] Governance shifts from pre-approval checklists to runtime monitoring > [!quote] [00:11:57] > We are going to have to rely on control environments, policies, and then active monitoring and managing the deviations. **Context & use**: - Customer fireside, [00:11:57] — Joel, chief data officer at Ericsson, describing the second or third revision of AI governance since LLMs arrived. - The displaced thing is named explicitly: paper and PowerPoint pre-screening for data transfer, privacy review, and IT security review. Directly transferable to open source project governance, where pre-merge checklists have the same scaling problem. **Tags**: #xref #articleseed **Highlight**: [view highlight](https://www.youtube.com/watch?v=GuwoULakxdY&t=717) <!-- atomicity: 1 idea ✓ --> #### [H4] Users treat AI as all-you-can-eat until shown the bill > [!quote] [00:09:27] > people don't really understand the cost of AI. **Context & use**: - Customer fireside, [00:09:27] — John Felmmeer, CIO at eBay, on the dashboard shipped in the prior couple of months showing each user their own AI spend. - Stated mechanism for the misconception: people pattern-match AI to seat-licensed tools like Google Docs or Microsoft Word. Reported effect at [00:10:42] is immediate behavior change, with no magnitude given. Pair with the power-curve observation at [00:08:50] — visibility plus targeted coaching, not org-wide caps. **Tags**: #xref #share **Highlight**: [view highlight](https://www.youtube.com/watch?v=GuwoULakxdY&t=567) <!-- atomicity: 1 idea ✓ --> #### [H5] Expert judgment draws on context that never gets written down > [!quote] [00:29:17] > my context might come from a decade of experience and it's hard to even put into a tangible written format. **Context & use**: - Databricks segment, [00:29:17] — Patrick's explanation for why coding models finish well-specified modules but fail at architecture. - He explicitly names the limit for retrieval tools: context systems help but are bounded by what can be represented textually. This is the most honest statement in the session about the ceiling on enterprise context products, delivered at a conference selling one. **Tags**: #xref #articleseed **Highlight**: [view highlight](https://www.youtube.com/watch?v=GuwoULakxdY&t=1757) <!-- atomicity: 1 idea ✓ --> #### 🧠 Atomic Split Pad **[P1] Self-evaluation is not evaluation** **Claim** (one sentence, Courtney's voice): Any system asked to grade its own output returns an inflated score, so quality measurement requires a separate evaluator. **Evidence pointer**: [[#H1]] **Scope & conditions**: Holds for LLM agents scoring their own generations and for humans assessing their own work. Fails as a rule when the evaluator shares the author's blind spot — a second agent on the same base model with the same context may replicate the error rather than catch it, so independence has to be structural, not just nominal. **Links**: - [[Evaluation-design]] — because extends - [[Peer-review-in-open-source]] — because applies-in contributor review queues - [[LLM-as-judge]] — because contrasts on whether model-based scoring is trustworthy at all <!-- atomicity: 1 idea ✓ --> **[P2] Automation relocates bottlenecks rather than removing them** **Claim** (one sentence, Courtney's voice): Removing the constraint at one stage of a pipeline moves the constraint downstream instead of eliminating it. **Evidence pointer**: [[#H2]] **Scope & conditions**: Holds where the pipeline has sequential human-dependent stages, which is the case for code, documentation, and community review. Breaks down when the downstream stage has slack capacity — then throughput really does rise until the slack is consumed. The open question is whether the terminal bottleneck (taste, judgment) is automatable at all; [[#H5]] argues it isn't, for now. **Links**: - [[Theory-of-constraints]] — because extends - [[Contributor-onboarding-throughput]] — because applies-in WordPress review queues - [[AI-productivity-claims]] — because contrasts on whether net throughput actually rises <!-- atomicity: 1 idea ✓ --> **[P3] Cost visibility is a cheaper control than cost limits** **Claim** (one sentence, Courtney's voice): Showing people what their own consumption costs changes behavior faster than imposing a quota does. **Evidence pointer**: [[#H4]] **Scope & conditions**: Holds when consumption is usage-based, invisible by default, and concentrated in a small number of users — the power-curve condition Joel describes at [00:08:50]. Fails when the spend is reimbursed or not attributable to an individual, and fails when a hard budget ceiling makes overspend unrecoverable rather than merely wasteful. Unverified: eBay reports an immediate effect but publishes no magnitude, so this stays `proof_state: working`. **Links**: - [[Feedback-loops-in-behavior-change]] — because extends - [[Rate-limiting]] — because contrasts on enforcement versus information - [[Open-source-infrastructure-funding]] — because applies-in projects with fixed budgets and variable AI costs `proof_state: working` **Tests / falsifiers**: Staggered rollout of a per-user spend dashboard across matched org units; difference-in-differences on per-user token consumption at 30 and 60 days. Falsified if consumption in the treated units is statistically indistinguishable from control. <!-- atomicity: 1 idea ✓ --> ### ✏️ Fleeting Queue - [F1] Find the OpenRouter study behind the 14x agent-token-growth and 5x agent-to-human figures cited at [00:39:21]; check how it attributes human-initiated tokens routed through a harness — https://www.youtube.com/watch?v=GuwoULakxdY&t=2361 #fleeting #inbox - [F2] Locate Patrick Wendell's X post on Databricks coding-cost optimization referenced at [00:26:41]; it's the only linked-out artifact in the session and no URL was given — https://www.youtube.com/watch?v=GuwoULakxdY&t=1601 #fleeting #inbox - [F3] Confirm the Cisco speaker's identity — captions render it "G2," which is probably an ASR garble; cross-check against the Glean:GO 2026 speaker list before quoting the $72 billion prediction anywhere — https://www.youtube.com/watch?v=GuwoULakxdY&t=3282 #fleeting #inbox - [F4] Open question: does "almost every line of code authored initially by some AI" at [00:25:30] mean accepted completions, created files, or proposed diffs? The metric definition changes the claim's weight entirely #fleeting #inbox - [F5] Ask our Glean admins whether the agent publication and visibility workflow Ericsson describes at [00:15:50] is available on our tenant; this session states availability for nothing #fleeting #inbox - [F6] Article seed: "Apprenticeship survives automation" — Ashish at [00:54:09] and the Cisco speaker at [00:51:03] independently defend junior roles, on reasoning alone, with no data. Pair with WordPress contributor-onboarding evidence #articleseed #fleeting #inbox - [F7] Caption accuracy as an accessibility failure: this transcript garbles Deloitte, Waymo, Glean:GO, Ericsson, and a speaker's name. Write it up as a WCAG 1.2.2 argument for WordPress.tv proper-noun review #articleseed #fleeting #inbox - [F8] Vocabulary: "AI mesh" (Ericsson's internal connective and governance layer between AI systems, [00:12:35]) and "workbench" (Deloitte's per-persona bundle of tools and agents, [00:48:00]) — both are customer coinages, not products #fleeting #inbox ### → Things 3 (max 2 actions) - [ ] Prototype a two-agent code review loop where a different agent reviews than authored, and measure detection rate against seeded defects — Databricks pattern from [00:31:36] #things - [ ] Ask Glean admins which capabilities named in the Day 1 afternoon keynote are live on our tenant, starting with agent publication and visibility controls — no availability was stated on stage #things <!-- deep-read:sections end --> 📄 **Full text:** [[Full Document Contents/Articles/Day 1 Afternoon Keynote GleanGO 2026.md|Day 1 Afternoon Keynote | Glean:GO 2026 by Glean]] ## Highlights - [00:09:27] everywhere, not just in development. And you know, I think, you know, initially that was, you know, the focus is let's make sure, you know, the engineering usage is is appropriate, but now we're seeing token usage growing very broadly across the organization. And in an earlier session today, I talked about one of the things that we've implemented in the last couple months is a dashboard uh that shows users, all users inside the company their basically what they're spending on uh AI uh across the different tools. Uh because one of the things we learned, you know, is people don't really understand the cost of AI. [00:10:05] They just assume it's like any other enterprise tool, kind of all you can eat. you know, I've got access to my Google Docs or my Microsoft Word and I can just use it and they don't really understand that, you know, there it is a usagebased model primarily today. So rolling out visibility was a huge step and as part of that then we're also starting to put some controls in place not to prevent people from using AI because we want them to use it that's the value but to make sure that they understand and use it responsibly. So there's kind of steps along the way and and that's really having we've already seen an immediate impact where people ([View Highlight](https://read.readwise.io/read/01m48t7z1mkshskqgpj52b1e9d)) - Note: [H4] Users treat AI as all-you-can-eat until shown the bill - Tags: atom/literature/h3 - [00:11:57] which is how much of governance has been done, whether that was a data transfer or a privacy review or uh an IT security review. We are going to have to rely on control environments, policies, and then active monitoring and managing the deviations. Uh so I think we're in the middle of uh that shift I would say and we're building something we call an AI mesh to connect things and to govern um but we already have uh these conversations up to the I mean the most senior levels to oversee governance in Ericson uh and we are using AI in governance. I'm very excited we have um for a responsible AI process you uh we built an agent you actually like if you [00:12:35] have an IT architecture you can just upload it and it will like populate the questionnaires and give you kind of a response on where you are. So I I think that's what we're doing. But really you have to manage at runtime the autonomy uh is where you have to go to. >> Yeah. I mean one of the things about governance um and that we talked about also today is AI sprawl and how important then governance is to manage that. So John as AI becomes more widely embedded across the business how is eBay addressing the challenge of AI spraw? Yeah, this is I think the the the million-dollar question for many of us is, you know, it a AI is sprawling. It's ([View Highlight](https://read.readwise.io/read/01m48t7wexv1m294cm6bn738yj)) - Note: [H3] Governance shifts from pre-approval checklists to runtime monitoring - Tags: atom/literature/h3 - [00:28:42] down. You know, they also are bringing huge amount of context about the company. They're bringing intuition from their prior jobs. And it just is for me actually insight into how human decision-m is working in that we ourselves are not just like okay there's some textual inputs and there's some output. We're actually taking into account many latent considerations that are unstated but are really really important. And unfortunately I I don't even know if in principle the models can ever do that. If they if they don't have enough information coming in then I don't even know if in principle they can produce that level of of insight and [00:29:17] decision-m you know things like glean that give them more context are helpful but ultimately they're constrained by like what can be represented textually whereas you know I'm my context might come from a decade of experience and it's hard to even put into a tangible written format. that's somewhere stored in my brain as what we might call intuition or something. So, so I think the biggest thing for us on quality is like recognizing tasks that are still in that category where humans re supreme, you know, and maybe they won't forever, but they do now >> and and really making sure humans do that and the right humans and then ([View Highlight](https://read.readwise.io/read/01m48t81hpvmggkbtbmwpw8twv)) - Note: [H5] Expert judgment draws on context that never gets written down - Tags: atom/literature/h3 - [00:31:05] cost. >> That's generally not how companies view like you don't generally like give the person a reward who spent the most money in the company, but there was this like thing of for tokens that is how you reward people. >> I really wish it's like a great marketing that's a great marketing move by the model providers. I I wish that there I wish that happened for data bricks. Like the coolest thing you could do is like spend as much money as possible on data bricks. Can we all make that can we go make that cool? That would be like really awesome. >> Um but uh so I think there's some I can think of some bad examples of what to [00:31:36] optimize. At the end of the day really fundamentally we're trying to measure ROI of some sort like we're spending I think tokens are on the cost side of the equation and there's some benefits that come out and in truth I think we've only just started to approach some ways to think about measuring that. So for instance for engineering velocity we look at how much code is coming out and we actually have independent we have agents reviewing the quality of the code like often a different agent than wrote the code. M >> it turns [clears throat] out that agents if you ask them to evaluate their own output they think it's really good ([View Highlight](https://read.readwise.io/read/01m48t7q4zj9n2d01w9cg1bh4n)) - Note: [H1] Agents rate their own output too generously - Tags: atom/literature/h1 - [00:48:33] what happens on your behalf and what are the things that you do today that sort of you did by yourself without the help of an assistant. >> Can I just add to that there's two philosophies right now that I think would be worth discussing here. One of them is that most people think that all human jobs are going to go away. we're just going to be staring at the beach and have no value to add to society and agents are going to do everything. I think that's such BS at such epic proportions because every single time there's been a step function improvement that I've seen in AI over the course of the past two two and a half years [00:49:11] a massive human bottleneck has emerged. So coding gets automated, code review becomes a bottleneck. Code review gets automated, taste and judgment and what product to build becomes a bottleneck. So I actually don't see in the foreseeable future that these jobs get every 100% of the jobs will get refactored. Um you know some jobs will be lost but there'll be entirely new industries that get created that we have actually not seen created yet. that that that in my mind seems like it's very hard because today AI is still very difficult to con um to um it's it's very bad on the last mile right everything can happen but at ([View Highlight](https://read.readwise.io/read/01m48t7smw8fpf0mmp38d60v12)) - Note: [H2] Automating one bottleneck exposes the next one - Tags: atom/literature/h2