Glean:GO AI-Native Showcase  🔗 **Source:** [Glean:GO AI-Native Showcase](https://www.youtube.com/watch?v=kIfYH2JIDus) > [!note]+ Document Notes > tags: [enterprise-ai-deployment, agent-governance, data-security, open-weights-models, partner-ecosystem, connector-architecture] > author: Glean > source: Glean (YouTube) > reference: https://www.youtube.com/watch?v=kIfYH2JIDus > 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 Do not quote or attribute any other line to the named author.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: Glean:GO AI-Native Showcase > <!-- 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--gleango-ai-native-showcase) ### 🔗 References - [Glean:GO AI-Native Showcase (source video)](https://www.youtube.com/watch?v=kIfYH2JIDus) - Eight partner companies presented: Fireworks AI, Gamma, Skyflow, Baseten, StackOne, Enterpret, Seclore, Horizon. - Caption caveat: the auto-transcript mangles most speaker and company names ("Base 10" for Baseten, "Secllor"/"Cycllor" for Seclore, "Interpret"/"Enterpread" for Enterpret, "Glenn"/"Gle"/"Green" for Glean). Verify every name spelling against the company's own site before citing it in writing. ## 📝 Summary - **What happened.** Day two of Glean:GO opened with a 78-minute partner showcase hosted by Zuben, Glean's VP of partnerships. Eight AI-native companies got eight minutes each, back to back, each explaining what they build and how it plugs into Glean. The format is stated up front at [00:03:08]: eight partners, eight minutes, no Q&A. - **Why it matters for a Glean customer.** The valuable layer here is not the product pitches — it's the repeated shape of the deployment problem. Four of the eight vendors (Skyflow, StackOne, Seclore, and implicitly Horizon) exist because the same two things block Glean rollouts: sensitive data that can't safely enter an index, and agents that organizations will let read but not write. - **Key stakeholders.** IT and security teams holding write permissions hostage; data owners with millions of ungoverned PDFs; transformation and ops leaders who don't know which process to automate first; ML platform teams choosing between open-weights and frontier inference. - **Risks and tradeoffs.** Every presentation is a vendor-selected success story with no failed deployments shown, no denominators, and no named customers for most outcome numbers. Three of the strongest-sounding results (two months saved per person per year, 3-month-to-5-day reporting cycle, 4.5x win rate) arrive at [00:18:05] attached to "one client" with no baseline and no measurement method. - **Bottom line.** The reusable pattern across all eight: put a governance and labeling layer in front of or behind the Glean index before you widen agent permissions, then use process discovery to pick the first agent rather than guessing. The two most concretely evidenced claims are StackOne's 15 connectors in two weeks against a planned six-month internal build [00:46:50] and Horizon's 2,000 interviews mapping 50,000 hours into a $2.5 million prioritized pipeline [01:08:08]. ## 📖 Main Content The showcase is structured as a value chain, and the host narrates the handoffs: models (Fireworks, Baseten) → applications on top of context (Gamma) → protection of the data underneath (Skyflow, Seclore) → action into business systems (StackOne) → customer signal capture (Enterpret) → deciding what to build at all (Horizon). Read as a deployment argument rather than eight ads, it says a Glean rollout has five separable layers and most organizations stall on layers three and four. **The recurring blocker is write access, not search quality.** StackOne's Romain Sestier names it directly at [00:41:23]: organizations let agents read and refuse to let them write, because read permissions inherited from source systems don't describe what an agent should be allowed to change. His governance model goes below the object level to the field level — an agent may create a user but never grant admin. That is a materially different permission primitive than the role-based access control a search index inherits, and it's the gap he claims blocks rollouts. **The second blocker is that sensitive data has no rows and columns anymore.** Skyflow's Anu Sharma builds the same argument from the data side at [00:24:50]: there's no word-level security in a PDF, no rows and columns in a vector store, so pre-AI column-level controls don't transfer. His counterexample at [00:24:10] — the curious new hire who can now ask what their manager's stock options are — is the clearest articulation in the whole 78 minutes of why better retrieval creates a new security surface rather than just a productivity gain. Seclore's Vishal Gauri attacks the same problem with labeling at the source, and supplies the sharpest scale statistic of the session at [00:57:38]: 40 million unique permission sets in an average enterprise. **The third, less obvious blocker is target selection.** Horizon's whole pitch is that companies ask "what agent should we build?" as their first question, when they should ask it third. Their agents interview employees at scale, map processes into BPMN, and rank recoverable hours before any agent gets specified. Their handoff artifact is an agent blueprint you paste into Glean's agent builder — the most directly reusable pattern in the session for anyone with a Glean license and no clear first use case. **Contested and uncertain claims.** Fireworks' claim at [00:06:02] that open-weights models like GLM and K3 "rival Opus and Fable on many tasks" is sourced to internal evals with no benchmark names, no scores, and no task list. Baseten's Philip Kiely is the only speaker who shows his working — parameter counts, expert counts, quantization format, the specific 80-million-parameter vision projector — and correspondingly the only one whose performance claims ("leading performance," "frontier speeds") stand out as the unquantified part of an otherwise specific talk. Seclore's opening statistics (80% of enterprise data unstructured, most of it dark) are industry figures presented without attribution, not Seclore outcomes. **Terms worth defining for a reader deciding what to build.** *Tokenization* here means replacing a real sensitive value with a fake one of the same shape so downstream systems still compute — Skyflow applies it at ingestion, Seclore applies it in the LLM path with semantic preservation. *Tool discovery* is StackOne's alternative to preloading a tool catalog into a context window: the agent semantically searches 30,000 actions instead of carrying them. *EDRM* (enterprise digital rights management) is Seclore's persistent encryption that travels with a file and supports retroactive revocation. *Day-zero support* means an inference provider serves a model the day its weights drop, which both Fireworks and Baseten claim. ## 🗒️ Notes #### ⏱️ [00:00:02](https://www.youtube.com/watch?v=kIfYH2JIDus?t=2) Opening — Zuben, VP of partnerships, Glean - Frames Glean as a layer across everything a company knows, so it's only as strong as the ecosystem it plugs into. - States the format explicitly. > [!quote] [00:03:08] > So today we are bringing you eight amazing AI native partners, eight companies, eight minutes each backto back. - No Q&A, no customer panel — every segment is vendor-narrated. #### ⏱️ [00:03:52](https://www.youtube.com/watch?v=kIfYH2JIDus?t=232) Fireworks AI — co-founder and CTO (captions: "Dimma Juggov") - **Thesis:** the AGI narrative concentrates intelligence in a handful of labs; only ~10% of data is public, 90% sits inside companies. > [!quote] [00:05:22] > And the future is not a single AI but millions of distinct specialized intelligences. - **Credibility anchor:** speaker was a core PyTorch maintainer before Fireworks, as were several co-founders [00:07:32]. - **Scale:** tens of trillions of tokens per day, last published number ~30 trillion [00:07:32]. - **Product surface, three parts** [00:08:12]: Fireworks Inference (day-zero open-weights support, cited example GLM 5.3 Flash "from yesterday"), Fireworks Training (multi-level APIs from novice to research-grade control), Fireworks Nexus (harness-agnostic routing between open and frontier models). - **Named customers:** Cursor (post-training Composer), Doximity (healthcare specialized models), Harvey (post-trained open-weights tenant model for legal). No benchmark numbers attached to any. - **Deployment claim:** tailored deployment configs yield "multiple x wins" on cost or speed — no figure given. - **The honest caveat** [00:09:38]: post-training is hard because frontier labs control the full stack from trainer to rollout to inference; Fireworks' pitch is that you focus on data curation and evals instead of procuring GPUs. - **Glean tie-in** [00:11:34]: Fireworks powers inference for Glean customers through Glean; route via your account executive. Conference attendees got a credit code. #### ⏱️ [00:12:12](https://www.youtube.com/watch?v=kIfYH2JIDus?t=732) Gamma — Grant Lee (co-founder, CEO) and Diana Berlin (GM, head of product) - **Positioning:** visual communication as "the last mile" — Glean holds the context, Gamma renders it into decks people act on. - **Scale:** 40,000 teams on Gamma [00:13:53]. - **Live demo, and it's a real artifact:** their own go-to-market all-hands deck, built that morning for a meeting later that day [00:14:41]. - **Data sources composited onto one slide** [00:15:34]: Glean context, financial sheets, HubSpot pipeline, Slack sales wins. - **Control mechanism — this is the transferable part:** brand templates shared across the workspace, plus *workspace instructions* and *slide-level template instructions* that tell the generator what story shape to produce. Output is repeatable rather than one-off prompting. - **Design reasoning:** tables force everyone onto the same information; stoplight indicators cue which rows are worth discussing [00:16:12]. - **Quantified outcomes, all unnamed and unbaselined** [00:18:05]: > [!quote] [00:18:05] > One client of ours saved two months of work per person per year. Insights unlocked companywide. One client's reporting cycle went from 3 months to 5 days. - Plus "one client sales rep saw 4.5 times improved win rate" — a single rep, not a cohort. - **The buried caveat** [00:18:42]: Berlin notes that without a dedicated vendor and cross-team alignment on brand and story, this work "dies on the vine." The prerequisite is organizational agreement on templates before automation pays off. #### ⏱️ [00:20:37](https://www.youtube.com/watch?v=kIfYH2JIDus?t=1237) Skyflow — Anu Sharma, co-founder and CEO - **Scope:** protecting data at runtime — at ingestion into Glean and at query time. - **Threat framing** [00:22:11]: two failure classes in large enterprises — people who don't know how to use the tool and make mistakes, and people savvy enough to do "intelligent bad things." > [!quote] [00:24:10] > Unfortunately, that intern that you just hired who's just curious can also now say, you know, let me just see how many stock options does my manager have. - **Why old controls fail** [00:24:50]: no word-level security in a PDF, no rows and columns in a model or a vector store; agents are nondeterministic, so the access surface is too. - **Referenced incident:** a Hugging Face incident "a couple of days ago" [00:22:50], used to argue no one can guarantee a model won't reach out and do something nefarious. No detail given; verify independently. - **Mechanism** [00:26:13]: scan PDFs and CSVs at ingestion, find sensitive values, replace with format-preserving fakes, keep the data queryable and chattable. - **Named shared customer:** Kearney (captions: "Karnney"/"Kernney"), a Glean customer with millions of consultant-facing PDFs. The requirement is client-by-client separation — a consultant can use firm knowledge without seeing another client's data. > [!quote] [00:27:34] > We've been partnering with Glean for about a year. We were able to do this partnership. We got this customer signed up and they were live and working within weeks. - **Residency design pattern — the most portable idea in this segment** [00:28:17]: Skyflow runs in 30 regions, so you keep one Glean instance in North America and still serve employees in Saudi Arabia or Vietnam. You avoid replicating Snowflake, Glean, and ServiceNow per country because only the sensitive values stay in-region. - **Scale references:** Walmart and Flipkart deployments with hundreds of millions of users' data, protected inline at runtime. - **"What we got wrong" moment** [00:25:31]: a year ago his board told him a multi-model, multi-datastore world would never happen — one model, one data store per company. He presents the board's call as wrong, which is also a convenient framing for his own product's necessity. #### ⏱️ [00:28:59](https://www.youtube.com/watch?v=kIfYH2JIDus?t=1739) Baseten — Philip Kiely, author of *Inference Engineering*, head of AI education - **Framing device, self-deprecating and honest:** > [!quote] [00:29:48] > I published a book, like a paper book with words printed in ink about inference, a technology that is changing faster than just about any technology in history. - **Core principle** [00:31:12]: some inference techniques manage trade-offs along a frontier (latency vs throughput, cost vs quality); others push the frontier out. Know which point you need before choosing techniques. - **Kimi K3** [00:31:55]: 2.8 trillion parameters, 896 experts with 16 routed active plus 2 shared, new attention mechanism, stable sparse-expert routing, two-part KV cache, 1 million token context. Three-fold difficulty: size, architecture, context. - **Under-discussed detail** [00:33:24]: much of the work was in the deterministic front end — parsing logic, special-token mechanisms, function-calling support — not only in the token-generating engine. Anyone running long agentic tasks hits this. - Baseten was a day-zero launch partner with the Moonshot team. - **DeepSeek V4** [00:34:42]: Flash and Pro, both MIT licensed, Flash roughly one-sixth to one-eighth the size. DeepSeek sparse attention requires engine-level work. - **The operational lesson** [00:35:23]: DeepSeek ships *preview checkpoints*. Baseten worked with April–May checkpoints, so when the trained models landed (Flash end of last month, Pro earlier this month) the inference service was already ready. Preview checkpoints are a free head start most teams ignore. - **V4 Pro specifics** [00:35:23]: native MXFP4 weights remove the post-training quantization step on Blackwell; ships with a first-party speculator using diffusion language modeling. - **Strategic claim** [00:36:06]: winning architectures get built on and trained on repeatedly, so investing in one architecture compounds. - **GLM 5.3** [00:37:23]: not fully open-weights yet (Flash version is). > [!quote] [00:37:23] > What's incredible about GLM 5.3 is it's the same architecture. The weights are the same shape. They just have different values because of post-training. - Work carried over from 5.2: quantization, speculation, parallelism, disaggregation, service config. They also added vision by training an 80-million-parameter projector that feeds the Kimi K2.6 vision encoder's output into unchanged GLM 5.2 weights, then served standard and fast API variants at different points on the latency curve. - **What must be redone across a post-training refresh:** retrain the speculator (hidden states changed) and rerun the model optimizer. Patterns and techniques carry; artifacts don't. #### ⏱️ [00:38:42](https://www.youtube.com/watch?v=kIfYH2JIDus?t=2322) StackOne — Romain Sestier (captions: "Roma Cestier"), co-founder and CEO - **Positioning:** the tool gateway / action layer. If Glean is the control plane and context layer, StackOne is what lets Glean write. - **The intern analogy** [00:40:05]: a qualified new hire with no system access hands you a to-do list at end of day instead of finished work. IT would have provisioned scoped access; agents haven't been given the equivalent. - **Three stated blockers** [00:40:45]: connecting to every system, safely granting write, and token cost as agents complete full workflows. > [!quote] [00:41:23] > many companies are allowing agents to read but not yet to write because they don't have a solution to enable the agents to safely write into their systems and take action - **Integration shape — concrete and worth copying** [00:42:05]: StackOne sits behind a **single MCP added into Glean**. IT and security manage connector build, customization, and governance centrally; end users see nothing new. - **Connect** [00:43:24]: 450 prebuilt connectors including SAP, Workday, NetSuite, carrying deep read *and* write actions (update employees, update financial records). An AI connector builder produces custom connectors "in hours with one engineer" versus months with a team. - **Secure** [00:44:09]: governance built for write, not inherited from read. Field-level control — which fields can be updated and what values can be written. Worked example: an access-provisioning agent may create a user but must never grant admin. - **Prompt injection** [00:45:27]: StackOne trained its own prompt-injection detection model, open-sourced it, and built it into every connector. - **Optimize** [00:46:11]: a 30,000-action catalog can't be preloaded. > [!quote] [00:46:11] > We've designed something called tool discovery, which is a high quality semantic search that actually allows the agent to search the catalog of actions that stack one has instead of preloading everything into the context window for you. - **Case study with real numbers** [00:46:50]: a cancer research network, 200+ employees. They had already bought AI platform licenses and couldn't deploy them because the platform wasn't connected to their systems — textbook shelfware. They needed two custom connectors: UKG (HIPAA compliance, PHI filtering) and NetSuite (heavily customized instance). > [!quote] [00:46:50] > Stack one was able to build these two custom connectors plus 13 others within the span of two weeks - They had scoped more than six months of internal build before finding StackOne [00:47:31]. - **Three reference agents** available in the attendee portal, built on Glean with the StackOne layer [00:47:31]: new-hire provisioning, access and IT ops, month-end close. These are the three highest-value starting agents named anywhere in the session. #### ⏱️ [00:48:54](https://www.youtube.com/watch?v=kIfYH2JIDus?t=2934) Enterpret — Varun Sharma, co-founder and CEO - **Thesis:** self-improving customer loops. Agents can do the work; something has to *create* the work by detecting what's broken. - **Opening example** [00:49:39]: an account with ~50,000 followers tweeted on August 4 that ChatGPT and Codex get lazy in long conversations; OpenAI shipped a fix ten days later and a Codex tech lead said they had no idea until the tweet. - **The counter-claim** [00:50:54]: Enterpret had detected the same signal up to three months earlier across roughly a thousand mentions. This is the session's best argument for signal infrastructure over anecdote, though no evidence artifact is shown on screen. - **Mechanism** [00:51:51]: classify customer context from any source in real time at granular depth, auto-generate a five-level schema that updates live (example chain: performance → reliability → task and execution performance → long execution latency → sub-reasons), plot trend lines, surface signals into any control plane including Glean. - **The loop** [00:53:05]: detect a change → triage for whether it's real and worth acting on → hand off to a human or AI agent → verify it was done → close the loop. - **Named customers:** Whisperflow (real-time product regression detection), Strava (self-healing help center — catch help-center regressions in real time). - **Best story in the session** [00:54:20]: at a Fortune 500 retailer, Enterpret caught a signal about a broken bathroom latch at a store in Tennessee, routed it to the right store manager, and it got fixed. The range is "self-healing help centers to self-healing bathrooms." - **Failure modes on stage, twice:** the speaker was auto-logged out of his own account mid-demo at [00:49:39] and again at [00:53:42]. > [!quote] [00:49:39] > What's interesting is, and I cannot believe I got auto logged out of my account. - Practical takeaway from the failure, not the pitch: session-expiry policy on a demo account will burn you in front of an audience. Pre-warm sessions or use a persistent demo tenant. - **Weakest evidence of the eight:** no deflection rate, no time-to-detection distribution, no revenue figure. The three-months-earlier claim is the only number, and it is a single instance. #### ⏱️ [00:55:38](https://www.youtube.com/watch?v=kIfYH2JIDus?t=3338) Seclore — Vishal Gauri (captions: "Vishal Gari"), CEO - **Framing** [00:56:18]: security spend still targets network, endpoint, cloud, and identity while the control point has moved to the data itself. - **Scale statistics, unattributed industry figures** [00:57:38]: 80% of enterprise data is unstructured, the majority is dark (nobody has touched it, nobody governs it), an unknown share sits outside the perimeter. > [!quote] [00:57:38] > And the average enterprise has 40 million different unique permission sets across the enterprise. - **Armor platform** [00:58:16]: discovers sensitive data, classifies it with **small language models so sensitive content never reaches public frontier models**, enforces controls, reports compliance posture. - **Three-dimension classification** [00:59:33] — the most reusable taxonomy in the session: *content* (is there PII or PCI), *context* (business, legal, HR document), *intent* (why it was created — vendor send, inbound PO, hiring memo). Labels are applied at the source so downstream systems can read them. - **EDRM** [01:00:12]: persistent encryption travels with the file inside and outside the enterprise; controls who, what, how long, from where; supports retroactive revocation. - **AI DLP** [01:00:53]: semantic-preserving masking and tokenization on the path to the LLM, unmasking on the way back. - **Three integration modes with Glean — copy this decision tree** [01:01:33]: 1. **Before Glean indexes.** Deploy Armor connectors on source systems, insert labels into the files themselves, let labeled metadata flow into Glean so its role-based access control has something accurate to enforce. 2. **After Glean indexes.** Integrate directly with the Glean index — no new connectors, no new systems deployed. Remediate, label, and protect, then let Glean enforce the labels. 3. **At runtime.** AI DLP sits in the inference path, masks with small language models, sends masked content to the LLM, unmasks the response, and Glean renders per user context and policy. - **Summary line** [01:04:13]: Glean understands documents, permissions, and relationships; Seclore reads risk, decides what's sensitive, and applies controls. - **Zero customer outcomes.** No named customers, no adoption figures, no time-to-deploy. The only numbers are the industry statistics at the top. #### ⏱️ [01:04:59](https://www.youtube.com/watch?v=kIfYH2JIDus?t=3899) Horizon — Nicolas Scopesi (co-founder, CEO) and Nicolas Lopez (co-founder, CPO) - **Framing** [01:05:40]: you can't fix what you can't see. Founders cite 15 years doing transformation work; the usual answer is a consulting firm for 6 to 18 months. - **Method:** AI agents interview every employee to map processes and rank where to automate first. > [!quote] [01:06:16] > 97% of all these people, they will talk once again with the models if that's needed. - Employees reportedly said the AI understood them better than their boss or the last consultant [01:06:16]. Note that 97% is a willingness-to-repeat figure, not an accuracy or satisfaction measure. - **Three layers** [01:06:54]: a discovery agent ingests company context (documents, processes, and for Glean customers they train on Glean context); agents interview employees and map frictions; everything lands in a company graph — their "company brain" ontology layer connecting processes, people, and tools. A second layer studies patterns to find what's broken and what matters most. A third layer converts human judgment trapped in people's heads into graph structure. - **Flagship customer, fully quantified** [01:08:08]: a Latin American e-commerce marketplace. 2,000 interviews in the finance department across 15 countries. Agents mapped processes, found 50,000 hours. > [!quote] [01:08:08] > From those 50,000 hours, the model struck what matters the most. They prioritize and they get a pipeline of $2.5 million for improve the company. - **Two output paths** [01:08:46]: big initiatives get a generated RFP, one-pager, and strategic plan; quick wins get an **agent blueprint and specification to paste into a platform like Glean agents**. - **Working sequence Lopez demos** [01:09:28]: most teams start at the right-hand end asking "what agent should we build?" Horizon starts one step earlier — pull governed enterprise context from Glean, run Horizon discovery, map how work actually happens, fill the gap between policy and reality, hand implementation-ready context back to Glean to build and govern agents. - **Concrete loop step** [01:10:13]: Horizon prompts the user to run a specified query in Glean and paste the result back in, which seeds interview planning with what the company already documents about how a department works. That's a manual copy-paste bridge, not an automated integration. - **What executives see** [01:10:52]: hours mapped, processes covered, recoverable hours, and an opportunity map ranking processes by impact so teams don't start on a low-impact one. - **Timeline claim** [01:11:35]: > [!quote] [01:11:35] > this this could be done in one day so this is a process that used to take months before - **Interview experience** [01:12:09]: employees get an email; the agent opens by asking how they'd describe their role and what fills their day. Voice or text. Over 40,000 interviews conducted; the founders claim their models interview better than a company's best analyst — an assertion with no comparison study. - **Artifacts produced** [01:13:27]: processes with BPMN diagrams mapped from conversations, plus *insights* as the unit of inefficiency, each carrying the problem, the evidence (who said what), and the impact. Click through to generate an initiative with a one-pager for the technology team, then push the RFP into Glean's agent builder. - **Integration option:** the company brain connects over MCP directly to Glean [01:12:45]. - **Closing offer** [01:14:47]: give them one process and one week for ready-to-deploy agents in Glean. - **Live demo hiccup** acknowledged on stage at [01:10:52] ("Oh, live demo as always"). #### ⏱️ [01:14:47](https://www.youtube.com/watch?v=kIfYH2JIDus?t=4487) Close — Zuben - Eight of "hundreds" of Glean partners presented; the open-platform argument is that customers shouldn't be restricted to a closed stack [01:15:37]. - Event logistics: final breakouts at 11:00 am, expert lounge, gateway pavilion. ## 📚 References - [Glean:GO AI-Native Showcase — source video](https://www.youtube.com/watch?v=kIfYH2JIDus) - [Fireworks AI](https://fireworks.ai) — inference, training, and Nexus routing; powers inference for Glean customers through Glean [00:11:34] - [Gamma](https://gamma.app) — visual communication platform, 40,000 teams [00:13:53] - [Skyflow](https://skyflow.com) — runtime data privacy, tokenization, 30 regions [00:28:17] - [Baseten](https://baseten.co) — model inference; Philip Kiely's *Inference Engineering* book offered free via QR code at [00:38:01] - [StackOne](https://stackone.com) — tool gateway, 450 connectors, 30,000 actions, open-sourced prompt-injection detection model [00:45:27] - [Enterpret](https://enterpret.com) — customer signal classification and triage [00:51:51] - [Seclore](https://seclore.com) — Armor DSPM, EDRM, AI DLP [00:58:16] - Horizon — enterprise intelligence and process discovery; no URL stated in the transcript and none verified in this session - Referenced third parties without links in the transcript: Cursor, Doximity, Harvey, Kearney, Walmart, Flipkart, Whisperflow, Strava, Moonshot AI (Kimi K3), DeepSeek (V4 Flash/Pro), Z.ai (GLM 5.3), UKG, NetSuite, SAP, Workday, HubSpot - Unverified event reference: a Hugging Face security incident cited at [00:22:50] as occurring "a couple of days ago," with no details given ## 🗣️ Paraphrase Glean's partnership lead opened day two by arguing that a company-wide knowledge layer is only as strong as the systems it connects to, then handed the stage to eight AI-native vendors for eight minutes apiece. The sequence walked from the model layer up through applications, security, action, customer signal, and process discovery. The inference vendors argued opposite sides of the same coin. Fireworks said the future isn't one god model but millions of specialized ones, because 90% of useful data is private, and offered inference, post-training, and model routing as the path there. Baseten went deeper technically, walking through what it actually costs to serve Kimi K3, DeepSeek V4, and GLM 5.3 — and made the durable point that when a model family gets post-trained into a new version with identical weight shapes, the optimization work you already did reapplies. Gamma positioned slides as the last mile between context and decisions, demoing a real all-hands deck assembled from Glean context, financial data, HubSpot pipeline, and Slack wins, governed by workspace and slide-level template instructions so output stays on brand without per-deck prompting. Three vendors attacked the same wall from different sides. Skyflow tokenizes sensitive values at ingestion so a retrieval system can find your expense report without exposing your manager's stock options, and runs in 30 regions so one North American Glean instance can serve data-residency-restricted employees elsewhere. Seclore labels data at the source across content, context, and intent, then plugs into Glean three ways — before indexing, after indexing against the index itself, or at runtime with masking in the LLM path. StackOne supplies the action layer as a single MCP inside Glean, with field-level write governance built for agents rather than inherited from read permissions, plus semantic tool discovery so a 30,000-action catalog never enters a context window. Enterpret argued that if agents do the work, something must generate the work, and showed real-time classification of customer signal detecting a ChatGPT complaint three months before it went viral — and a broken bathroom latch in Tennessee routed to the right store manager. Horizon argued the first question shouldn't be which agent to build; their interview agents mapped 50,000 hours across 2,000 finance interviews in 15 countries and turned them into a $2.5 million prioritized pipeline, handing agent blueprints back into Glean's builder. ## 📰 Newsletter **What an enterprise AI vendor showcase teaches WordPress folks about permissions** Glean ran an 78-minute partner showcase where eight AI companies got eight minutes each, and the most useful thing in it had nothing to do with buying software. Four of the eight vendors exist because of one problem: organizations will let AI agents *read* their systems but not *write* to them, because read permissions don't describe what an agent should be allowed to change. That should sound familiar. WordPress capabilities are a read/write model built for humans clicking around wp-admin, not for an agent that might call `wp_insert_user()` at 2am. StackOne's answer was field-level governance — an agent can create a user but can never set the admin role. Translate that to WordPress and you get a genuinely interesting question for anyone shipping AI features into a site: does your plugin's capability check distinguish "may edit this post" from "may edit *this field* on this post"? Mostly, no. The other transferable idea came from Seclore: classify data on three axes — content (is there PII), context (is this a legal doc or an HR memo), and intent (why was it created). Apply labels at the source so every downstream system reads the same signal. That is a content-modeling discipline, and it's applicable whether your index is an enterprise search product or a site with 40,000 posts and a REST API. Worth stealing for documentation and training work: Horizon's sequence. Don't start by asking what to automate. Interview people, map the actual process, rank recoverable hours, *then* specify the agent. They ran 2,000 interviews across 15 countries in a finance org, surfaced 50,000 hours, and prioritized $2.5 million of work out of it. The same shape works for a contributor team deciding which onboarding friction to fix first. Two honesty notes. Every number here came from a vendor showing its best customer with no denominator, and the most impressive-sounding results ("two months saved per person per year") arrived attached to an unnamed client with no baseline. Also, two presenters got auto-logged out mid-demo and one said "live demo as always" — a decent reminder that the session-expiry settings on your demo account deserve a dry run before WordCamp. Links: the [full session recording](https://www.youtube.com/watch?v=kIfYH2JIDus), and if you want the inference-engineering deep end, Baseten's Philip Kiely offered his book free at the talk. ## ✅ Action Items #### DevRel - **Developer advocacy:** Build the three StackOne reference agents named at [00:47:31] — new-hire provisioning, access and IT ops, month-end close — as internal demos before pitching anything customer-facing. They're the three most-validated starting points in the session. - **Education:** Run a 60-minute internal workshop on the read-vs-write permission gap [00:41:23]. Outcome to measure: number of teams that can articulate a field-level write policy for their proposed agent. - **Developer experience:** Document the single-MCP integration pattern ([00:42:05]) as a template — one gateway MCP registered in Glean, connectors governed centrally — so teams stop building one-off integrations per agent. - **Developer marketing:** Publish an internal "what we actually deployed" post with baselines attached, as the counterweight to unbaselined vendor stats like [00:18:05]. Attribution: track Glean agent creation counts before and after. - **Community management:** Start a channel for agent-blueprint sharing modeled on Horizon's copy-pasteable spec artifact [01:08:46]. Moderation rule: every blueprint names its data sources and its write scope. - **Internal/external/cross-project:** Set an SLA with security for reviewing an agent's write scope. StackOne's whole pitch exists because that review has no owner in most organizations. #### OSPO - **Policy & Governance:** Write a decision record on open-weights versus frontier model use, using Fireworks' Nexus routing framing at [00:10:54] — the shift from token-maxing to value-maxing. Decision rights: who approves a model swap in a production agent. - **License Compliance:** DeepSeek V4 Flash and Pro are MIT licensed [00:34:42], which means you can run and post-train them anywhere. Confirm that your license scanner treats model weights as artifacts, because most don't. - **Security Management:** Evaluate StackOne's open-sourced prompt-injection detection model [00:45:27] as a candidate control. Intake question: does it detect injections in retrieved documents, not just user prompts? - **Software Inventory:** Add MCP servers and connector catalogs to the SBOM scope. A 30,000-action catalog [00:46:11] is an attack surface that no current SBOM format enumerates. - **Developer Support:** Hold consultation hours on the preview-checkpoint pattern Baseten described at [00:35:23] — building against early checkpoints so you're ready on release day. - **Community Engagement:** Fireworks' team came out of PyTorch maintainership [00:07:32]. Map which of your critical AI dependencies are maintained by people now at vendors, and whether upstream funding follows. - **Strategic Alignment:** Report one KPI upward that vendors in this session never supply: percentage of agents in production with write scope, and the incident count against them. #### Community - **Maintainers:** Run a governance clinic on the Seclore three-axis model — content, context, intent [00:59:33] — applied to project artifacts. Which docs are sensitive, which are legal, why each exists. - **Contributors:** Build a "first agent" rotation modeled on Horizon's quick-wins path [01:08:46]: a small specified task with a written blueprint, rather than "go build something with AI." - **Sponsors:** Demand baselines. The strongest numbers in this session ([00:46:50], [01:08:08]) carry a before state; the weakest ([00:18:05]) don't. Apply the same standard to impact reports you publish. - **Ecosystem:** Run a coordinated disclosure drill for prompt injection through a shared connector. The Hugging Face incident referenced at [00:22:50] had no shared response playbook mentioned by anyone on stage. #### WordPress - **Core:** Triage how core capabilities map to agent write scope. Field-level write control [00:44:50] has no WordPress equivalent today. - **Core Performance:** Adopt the trade-off frontier framing from [00:31:12] for AI features — latency, throughput, quality, cost — and set an explicit target point rather than optimizing all four. - **Core AI:** Use Seclore's three integration positions [01:01:33] as the evaluation grid for any AI feature: before indexing, against the index, or at runtime in the inference path. Each has different data-exposure properties. - **Core Program Team:** Add a cross-team dependency check for any AI feature that writes. StackOne's blocker analysis says the security review is the critical path, not the build. - **Mobile:** Apply Enterpret's detect-triage-handoff-verify-close loop [00:53:05] to app store review signal, which is already unstructured customer text. - **Playground:** Build a reproducible lab showing prompt injection through an untrusted post, since Playground can host a poisoned fixture safely in-browser. - **Community:** Borrow Horizon's interview-first sequencing [01:09:28] for meetup and WordCamp planning — map what organizers actually do before proposing tooling. - **Training:** Write a lesson on tokenization versus masking, using Skyflow's ingestion-time tokenization [00:26:13] against Seclore's semantic-preserving runtime masking [01:00:53] as the contrast pair. - **Marketing:** Do not repeat unbaselined vendor figures. Use [00:46:50] style claims — 15 connectors in two weeks against a planned six-month build — where a before state exists. - **Support:** Strava's self-healing help center [00:53:42] is the model: detect help-center regressions in real time rather than waiting for ticket volume. Identify the knowledge-base gap signal you already collect and don't act on. - **Design:** Gamma's stoplight-table argument [00:16:12] — the visual form determines which conversation happens — is a usable heuristic for dashboard design reviews. - **Documentation:** Adopt the intent axis from [00:59:33] in docs metadata. Every page records why it exists, not just what it covers. - **Photos:** Apply content classification at upload, not after. Skyflow's argument at [00:24:50] is that post-hoc controls fail once content is indexed. - **TV:** Caption quality is the indexing input. The showcase transcript mangles nearly every proper noun, which is exactly what breaks downstream search. - **Polyglots:** Skyflow's 30-region residency pattern [00:28:17] is the structural analogue of locale coverage — one system, regionally partitioned obligations. - **Accessibility:** Audit AI-generated slide and deck output (Gamma-style, [00:15:34]) against WCAG 2.2 AA. Auto-composited tables and stoplight color coding are two known failure points. - **Meta:** Treat connector and API reliability as the agent enablement blocker. The cancer research network's shelfware [00:46:50] came from missing connectors, not missing licenses. - **CLI:** Command coverage is an action catalog. StackOne's tool-discovery pattern [00:46:11] suggests exposing WP-CLI commands via semantic search rather than a full manifest. - **Hosting:** Publish a cross-host checklist for where AI inference happens relative to customer data — the runtime masking question at [01:02:54]. - **Tide:** Scoring transparency matters more when agents consume scores. Document the false-positive review loop before any agent reads Tide output. - **Openverse:** Ingestion pipelines need the label-at-source discipline from [01:01:33], applied to provenance rather than sensitivity. - **Themes:** Add a review check for any theme shipping an AI feature that writes to the database without a capability check narrower than `edit_posts`. - **Plugins:** Require declared write scope for plugins registering MCP servers or agent tools. Provenance checks should cover the action catalog, not just the code. #### The WPOCC - **Programs:** Fund a fellowship track on agent governance for open source projects — the field-level write model [00:44:50] has no open equivalent and no maintainer-facing guidance. - **Projects:** Run a documentation sprint producing an open "agent write policy" template, mirroring what StackOne sells as a product feature. - **Funding experiments:** Pilot a milestone payout tied to a measurable before/after — the discipline missing from [00:18:05] — and publish the ledger. One process, one baseline, one measured result. ### FAIR - **Technical independence:** Label and provenance metadata must live locally and travel with the artifact, which is the EDRM property described at [01:00:12]. Signature checks belong at the same layer as the label. - **Policies and covenants:** Require funding transparency from any vendor whose product sits in the inference path. Runtime masking [01:02:54] means a third party sees every prompt before the model does. - **Productized governance artifacts:** Build a scorecard from the three integration positions at [01:01:33] — before index, against index, at runtime — so operators can state where each vendor sits. ## 🧪 Hypotheses 1. **Write governance, not search quality, gates enterprise agent deployment.** If true, organizations with a field-level write control layer deploy action-taking agents at a materially higher rate than those relying on inherited read permissions. Stated directly at [00:41:23]; the cancer research network case at [00:46:50] is one supporting instance. 2. **Missing connectors, not missing licenses, cause AI shelfware.** The 200-employee research network had purchased AI platform licenses it could not deploy [00:46:50]. If general, license utilization should correlate with connector coverage more strongly than with training or change management spend. 3. **Process discovery before agent selection produces higher-yield first agents.** Horizon's claim [01:09:28]. Testable as: agents specified from a mapped process with measured recoverable hours survive longer in production than agents chosen by team intuition. 4. **Post-training refreshes preserve inference optimization work when weight shapes are unchanged.** Kiely's GLM 5.2 → 5.3 claim at [00:37:23]. Falsifiable: measure throughput and latency after reapplying the prior quantization, speculation, and parallelism config to a same-architecture refresh, with only speculator retraining and optimizer reruns as new work. 5. **Semantic tool discovery beats catalog preloading above a threshold action count.** StackOne asserts this for 30,000 actions [00:46:11] but names no crossover point. There is a catalog size below which discovery adds latency without saving tokens. 6. **Real-time signal classification detects product regressions materially earlier than viral escalation.** Enterpret's three-months-earlier claim [00:50:54] on a single case. Needs a distribution, not an anecdote. ## 📊 Methodology - **For H1 (write governance gates deployment):** Cross-sectional survey of Glean customers, stratified by whether a write-governance layer is deployed. Dependent variable: count of production agents with write scope per 1,000 employees. Control for company size, regulated industry, and months since Glean deployment. A vendor-supplied sample will not answer this — it needs a neutral sampling frame, because every case in this session was vendor-selected. - **For H2 (connectors cause shelfware):** Instrument license utilization against connector coverage across a customer cohort. Data source: seat activation rates plus connector inventory. The confound to control is that organizations capable of building connectors are also capable of driving adoption. - **For H3 (discovery before selection):** Prospective comparison within one organization. Arm A specifies agents from Horizon-style mapped processes; Arm B from team nomination. Primary endpoint at 90 days: agent still in production with non-zero weekly invocations. Secondary: hours saved measured against a pre-agent time study, not a post-hoc estimate. - **For H4 (optimization carryover):** Direct benchmark. Take a same-architecture post-training refresh, apply the prior config unchanged, and measure tokens per second and time to first token before and after retraining the speculator. The claim is mechanically specific enough at [00:37:23] to be checked in a day on rented hardware. - **For H5 (tool discovery crossover):** Parameter sweep. Hold the agent and task constant; vary catalog size from 50 to 30,000 actions; measure total tokens, end-to-end latency, and task success rate for preloaded versus semantically searched catalogs. The crossover point is the deliverable. - **For H6 (early signal detection):** Retrospective cohort. Take 50 publicly escalated product complaints with known escalation dates, and measure detection lead time in a signal platform blind to the escalation. Report the distribution and the false-positive rate, since detecting everything three months early is trivial if you also flag a thousand non-issues. - **Known limitation across all six:** the source is a vendor showcase with no control group, no failed deployments, and no denominators. Every hypothesis here needs data the session does not supply. ## ❓ Questions - What is the false-positive rate on real-time customer signal classification? Enterpret reports detection lead time [00:50:54] but never the volume of signals that didn't matter. - Who owns the agent write-scope review in an organization that has no such role today? StackOne's product implies the answer is IT and security [00:42:44], but no speaker described the approval workflow. - What breaks when runtime masking meets retrieval? If Seclore masks values before the model sees them [01:02:54], how does the model reason over masked entities in a comparison or aggregation task? - Does tokenized data degrade retrieval quality? Skyflow claims data stays queryable after tokenization [00:26:13], but semantic search over documents with substituted values is a different problem than SQL over tokenized columns. - How many of Horizon's 50,000 mapped hours became realized savings rather than a $2.5 million pipeline? The transcript reports pipeline [01:08:08], which is a forecast, not a result. - What's the maintenance cost of 450 connectors plus custom ones? Sestier names maintenance as an enormous task at [00:40:45] and then doesn't return to who bears it after purchase. - Which of the three Seclore integration positions do actual customers pick, and why? All three are described [01:01:33]; none is recommended. - What happened in the Hugging Face incident cited at [00:22:50], and does the defense described actually address that class of attack? ## 🔮 Future Research - **A neutral benchmark of enterprise agent write governance.** No standard exists for expressing "may create user, may not grant admin" [00:44:50]. This is a specification gap an open project could fill, and it would outlive any single vendor. - **Empirical retrieval quality under tokenization and masking.** Two vendors in one session make overlapping claims about preserving usability while substituting values ([00:26:13], [01:00:53]). Nobody published a retrieval benchmark. This is a straightforward paper. - **Cost curves for tool discovery versus preloading.** StackOne's semantic search over a 30,000-action catalog [00:46:11] has an unmeasured accuracy cost — an agent that can't find the right tool fails differently than one that ran out of context. - **Process-mining accuracy of LLM interviews against observed work.** Horizon claims 40,000 interviews and models better than a best analyst [01:12:09]. Compare agent-derived BPMN maps against system-log process mining on the same department. - **Longitudinal study of architecture investment returns in inference.** Kiely's claim that winning architectures compound [00:36:06] is testable across the DeepSeek and GLM release histories. - **Residency arbitrage as an architectural pattern.** Skyflow's one-instance-plus-regional-tokenization design [00:28:17] deserves legal and technical scrutiny — whether regulators accept it varies by jurisdiction, and no speaker addressed which ones have. ## 💡 Implications **For deciding what to build with Glean at your own company:** - **Sequence the rollout in this order:** classify and label data at the source → confirm read permissions are actually correct → deploy read-only agents → add a governed write layer → widen write scope one field group at a time. Every vendor in this session sits at exactly one of those stages, which is the strongest signal in the 78 minutes that the stages are real and ordered. - **Pick your first three agents from the validated list, not from a brainstorm.** New-hire provisioning, access and IT ops, and month-end close [00:47:31] are cross-functional, high-frequency, and bounded. All three have obvious field-level write scopes you can write down in an afternoon. - **Budget for connectors as the critical path.** The 200-employee case at [00:46:50] had licenses and no deployment for want of two custom connectors. Before you buy more seats, inventory which systems your agents need to write to and whether a connector exists. - **Treat the write policy as a design artifact, not a security afterthought.** The field-level example at [00:44:50] — create user, never grant admin — is the right granularity. If your policy is expressible only as "can edit CRM," you have not specified it. - **Choose one of the three data-protection positions deliberately.** Labeling before indexing [01:01:33] gives Glean accurate metadata to enforce but requires connectors on every source. Working against the existing index requires no new deployment but inherits whatever the index already got wrong. Runtime masking protects the LLM path but adds a third party to every prompt. - **Use discovery to choose, but discount the numbers.** Horizon's method — map processes, rank recoverable hours, generate a blueprint [01:09:28] — is sound sequencing. The $2.5 million is pipeline, not realized savings. Adopt the sequence; set your own targets. - **For model strategy:** routing between open-weights and frontier models with no harness change [00:10:54] is a reversible decision, which makes it a good early experiment. Post-training your own model is not reversible in the same way and requires the data curation and eval work Fireworks flags at [00:10:14]. - **For policy and compliance:** the residency pattern at [00:28:17] can avoid replicating your full stack per country, but the compliance argument depends on regulators accepting that tokenized values leaving the region aren't personal data. Get that in writing before architecting around it. - **For education and enablement:** the two live-demo logouts at [00:49:39] and [00:53:42] are a real finding. Agent demos depend on session state across multiple SaaS tools; build a pre-demo checklist that verifies every session before you present. - **For funding models:** every impressive number in this session came without a denominator. When you report your own Glean outcomes internally, publish the baseline and the measurement method — that discipline is what separates [00:46:50] from [00:18:05]. ## 🧠 Learning Styles Assessment - **Reading level estimate:** Flesch-Kincaid Grade Level approximately 10.5, Flesch Reading Ease approximately 55, estimated by sampling five passages of roughly 120 words each from the Fireworks, Skyflow, Baseten, StackOne, and Horizon segments. Spoken conference register: short declarative sentences, heavy domain jargon (MXFP4, BPMN, EDRM, DSPM, MoE routing), and frequent disfluencies that inflate word counts without raising complexity. The Baseten segment scores noticeably higher than the rest on technical density; the Gamma segment scores lowest. #### 👀 Visual learners - **Present:** Slide decks behind every speaker, live product demos from Gamma [00:15:34], Enterpret [00:51:51], and Horizon [01:10:52], including an on-screen BPMN process map and an executive opportunity matrix. - **Improve:** None of the visuals survive into the transcript — when a speaker says "as you can see on the video here" [00:42:05], the audio-only listener gets nothing. Publish the decks. Add an architecture diagram showing the three Seclore integration positions [01:01:33] as one figure, and a single diagram of the five-layer deployment stack the showcase implies but never draws. #### 🔊 Aural learners - **Present:** The whole artifact is a 78-minute recorded talk; every speaker narrates their own slides. - **Improve:** Publish per-segment audio chapters so someone can listen to the eight minutes they care about. A 3-minute audio summary per vendor would serve this format better than one 78-minute file. Speaker audio quality varies; the Horizon and Enterpret segments include significant disfluency. #### 🗣️ Verbal learners - **Present:** Strong analogies throughout — StackOne's unprovisioned intern [00:40:05], Skyflow's curious new hire [00:24:10], Horizon's doctor-with-a-map [01:05:40], Gamma's last mile [00:13:08]. These are the most memorable content in the session. - **Improve:** No glossary. MXFP4, DSA, EDRM, DSPM, BPMN, and MoE routing all appear undefined. A one-page term sheet would make the Baseten segment accessible to the security audience and vice versa. #### 🧩 Kinesthetic learners - **Present:** Three working reference agents available in the attendee portal [00:47:31]. Fireworks distributed credit codes for hands-on trials [00:12:12]. Horizon offers a one-process, one-week pilot [01:14:47]. - **Improve:** Nothing is reproducible after the event — the portal is attendee-gated and the credit codes expire. Publish the three agent blueprints publicly. Provide a sandbox connector so someone can build the new-hire provisioning agent without a sales conversation. #### 🧠 Logical learners - **Present:** Baseten's trade-off frontier [00:31:12] is a clean decision framework. Enterpret's five-stage loop [00:53:05] and Seclore's three-axis classification [00:59:33] are both well-structured. Horizon's opportunity ranking is explicitly decision-theoretic. - **Improve:** No comparison table across the eight vendors, no decision tree for which layer to solve first, and no KPIs or error budgets attached to any agent. A single table mapping vendor → layer → prerequisite → measurable outcome would do more work than any individual talk. #### 🧑‍🤝‍🧑 Social learners - **Present:** Expo hall with StackOne staff named for demos [00:48:07], Kearney representatives on site [00:27:34], breakouts at 11:00 am, expert lounge, gateway pavilion [01:16:18]. The host explicitly encourages swapping stories. - **Improve:** Zero Q&A in 78 minutes. Eight back-to-back monologues with no cross-vendor discussion means the obvious question — how do Skyflow and Seclore overlap? — never gets asked in front of the people who'd benefit. A 10-minute moderated panel would have been the highest-value segment. #### 🧘‍♀️ Solitary learners - **Present:** Kiely's *Inference Engineering* book offered free via QR code [00:38:01] is the only self-paced artifact in the session. Sestier closes with a reflective question about what would change in your business with safe agent actions [00:48:07]. - **Improve:** No self-check material, no reading list, no post-event worksheet. A one-page "assess your own readiness" checklist mapped to the five layers would let someone act alone. #### 🌿 Naturalistic learners - **Present:** Almost nothing. Kiely's aside about publishing a paper book "out of respect for the trees" [00:29:48] is the single environmental reference in 78 minutes. - **Improve:** Serving a 2.8-trillion-parameter model has an energy cost nobody named. The efficiency techniques discussed — MXFP4 quantization, speculation, sparse routing [00:35:23] — are directly energy-relevant and could be framed that way. Add lifecycle thinking to the routing argument: routing to a smaller open model isn't only cheaper, it's less compute per answer. ## 📊 Accessibility Assessment #### Screen reader semantics - **Present:** The YouTube page provides standard heading and landmark structure; auto-generated captions exist. - **Improve:** Slide content is never described aloud. At [00:42:05] and [01:10:13] speakers point at unnarrated visuals. Maps to WCAG 2.2 AA **1.1.1 Non-text Content** and **1.2.3 Audio Description or Media Alternative**. Fix: publish decks with alt text, and have presenters narrate what's on screen. #### Keyboard navigation - **Present:** The YouTube player supports keyboard control with visible focus. - **Improve:** Not assessable for the demoed products themselves — Gamma, Horizon, and Enterpret UIs are shown by mouse only. Maps to **2.1.1 Keyboard** and **2.4.7 Focus Visible** for the products, untested here. #### Contrast and readability - **Present:** Not verifiable from transcript alone. - **Improve:** Gamma's stoplight indicators [00:16:12] encode status by color. If color is the only channel, that fails **1.4.1 Use of Color**. Horizon's opportunity map [01:10:52] has the same risk. Fix: pair every color state with a text label or shape. #### Text resizing and reflow - **Present:** YouTube's player and transcript panel reflow at 200% zoom. - **Improve:** Dense auto-generated slides shown at conference scale are unreadable when the video is played in a small viewport. Maps to **1.4.4 Resize Text** and **1.4.10 Reflow**. Fix: publish slides separately rather than relying on video capture. #### Media and captions - **Present:** Auto-captions and a transcript exist. The transcript is timestamped, which supports navigation. - **Improve:** Caption accuracy is poor on exactly the words that matter. "Glean" becomes "Glenn," "Gle," and "Green"; "Baseten" becomes "Base 10"; "Seclore" becomes "Secllor" and "Cycllor"; "NetSuite" becomes "Netswuite"; "HIPAA" becomes "HIPPA"; several speaker names are unrecognizable. Music and applause are marked, but speaker changes are inconsistently labeled. Maps to **1.2.2 Captions (Prerecorded)** — auto-captions with this error rate on proper nouns do not meet the standard. Fix: human-corrected captions with a supplied name and product glossary, plus a downloadable corrected transcript. #### Forms and interactions - **Present:** No forms in the video itself. - **Improve:** The QR code at [00:38:01] is the only call to action and is visual-only, with no spoken URL. Maps to **2.5.5 Target Size (Minimum)** in spirit and **1.1.1** directly. Fix: always speak the URL alongside a QR code and include it in the video description. #### Standards reference - Primary gaps against WCAG 2.2 AA: **1.1.1** (unnarrated slides, QR-only CTA), **1.2.2** (caption accuracy on proper nouns), **1.2.3** (no audio description of demos), **1.4.1** (color-coded status in demoed dashboards), **1.4.10** (slide legibility on reflow). The caption accuracy gap is the highest-impact and cheapest to fix. ## 🌍 DEIB Assessment #### Cultural breadth - **Present:** Genuinely multi-region content. Skyflow's residency example names Saudi Arabia and Vietnam [00:28:17] and cites Flipkart alongside Walmart. Horizon's flagship customer is a Latin American e-commerce marketplace spanning 15 countries [01:08:08]. StackOne's cross-industry list covers healthcare, retail, manufacturing, professional services, and SaaS [00:46:50]. Seclore raises geographic regulatory frameworks explicitly [00:59:00]. - **Improve:** The event is San Francisco-based (Fort Mason, [00:28:59]) and every example of a *buyer* is a large Western enterprise. No African or Middle Eastern customer appears as a customer rather than as a compliance constraint. Add case studies where the non-US organization is the protagonist. #### Inclusive language - **Present:** Gender-neutral throughout. No exclusionary idioms. Speakers generally define their own product terms. - **Improve:** Heavy undefined jargon acts as a gatekeeping mechanism across specialties — a security leader hears MXFP4 and DSA at [00:35:23] with no definition; an ML engineer hears EDRM and DSPM at [01:00:12] with none. Several idioms won't translate ("you're so close, you can taste it" [00:19:15]; "dies on the vine" [00:18:42]; "the jury is out" [00:25:31]). Add a glossary and prefer literal phrasing for a global audience. #### Neurodiversity support - **Present:** The eight-minutes-each structure is highly predictable, and the host narrates every transition with a consistent framing sentence. Segments are short enough to process individually. - **Improve:** 78 minutes of back-to-back presentation with no break is demanding. Presenter disfluency is high in several segments (the Enterpret and Horizon segments contain extensive filler and repeated restarts), which increases processing load. Fix: publish chapter markers, provide per-segment summaries, and build in breaks. #### Representation - **Present:** Speakers span several national origins, evident in accents and naming, and include founders from companies headquartered outside the US. Roles vary across CEO, CTO, CPO, GM, and head of education — not only chief executives. Gamma fielded a woman as GM and head of product [00:12:12], the only woman presenting in the showcase. - **Improve:** One woman among roughly eleven presenters is the clearest representation gap in the session. Every speaker is a founder or executive; no practitioner who actually operates one of these deployments day to day appears on stage, which also shapes what gets said — founders narrate wins, operators narrate failure modes. Invite the admin who runs the Kearney deployment, not only the CEO who sold it. #### Psychological safety - **Present:** Notable moments of public vulnerability. Kiely opens by saying he made a mistake publishing a paper book [00:29:48]. Enterpret's Varun Sharma got auto-logged out twice on stage and kept going [00:49:39]. Horizon's Lopez says "live demo as always" when something misbehaves [01:10:52]. Skyflow's Anu Sharma tells a story about his board getting a call wrong [00:25:31]. This tolerance for visible imperfection is healthy and worth naming. - **Improve:** No Q&A anywhere in 78 minutes means no one can challenge a claim in the room. No content warnings needed, but the absence of any dissenting voice — or any customer speaking without a vendor present — limits how safe it is to be skeptical out loud. Add moderated Q&A with pre-submitted questions so critical ones get asked without individual exposure. ## ✅ Logical Validity #### Strong reasoning - **Causal, well-specified — StackOne at [00:46:50].** The argument is: licenses were purchased → systems weren't connected → deployment couldn't happen → connectors were built in two weeks → 200 employees got the tools. Each link is stated, the counterfactual is given (six months of planned internal build), and the mechanism is concrete (two specific systems, named compliance constraint). This is the strongest causal chain in the session. To clarify further: name whether the six-month internal estimate was a written plan or an informal guess. - **Deductive, from a real property change — Skyflow at [00:24:50].** Old access control operated on rows and columns; PDFs, models, and vector stores have neither; therefore old controls can't transfer. The premises are structural facts about the data formats, not assertions about behavior, so the conclusion follows. The intern example at [00:24:10] then makes the abstract argument concrete without carrying the inferential weight. - **Analogical, tightly mapped — StackOne at [00:40:05].** The unprovisioned intern maps cleanly: qualified (capable model), eager (available), no system access (read-only), returns a to-do list (recommendations instead of completed work), IT would normally provision scoped access (the missing governance layer). Every element of the analogy has a referent in the actual problem, which is what makes it hold. - **Inductive with a stated mechanism — Baseten at [00:37:23].** Same architecture plus same weight shapes means prior optimization work reapplies, with explicit exceptions named (retrain the speculator because hidden states changed; rerun the model optimizer). Naming what *doesn't* carry over is what makes this credible rather than promotional. - **Framework reasoning — Baseten at [00:31:12].** Separating techniques that move along a trade-off frontier from techniques that push the frontier out is a genuine conceptual distinction that survives outside the sales context. #### Fallacies - **Survivorship bias — structural, across the entire session, established at [00:03:08].** Eight vendors selected by Glean present self-selected customers. Zero failed deployments, zero abandoned pilots, zero customers who bought and didn't adopt appear anywhere in 78 minutes. The format guarantees the sample. *Fix:* the honest version of this showcase includes one segment on a deployment that stalled and why. Failing that, the host should state the selection criteria so the audience can discount accordingly. - **Unbaselined before/after, three in one breath — Gamma at [00:18:05].** "One client of ours saved two months of work per person per year," "one client's reporting cycle went from 3 months to 5 days," and "one client sales rep saw 4.5 times improved win rate." No client named, no measurement method, no control for concurrent changes, and the win-rate figure is a single rep — an n of one presented in the cadence of a portfolio result. *Fix:* name the measurement window, the baseline period, and the population. A rep-level win rate needs a cohort or it should be labeled as an anecdote. - **Anecdote as evidence rate — Enterpret at [00:50:54].** One detected signal three months ahead of viral escalation is presented as demonstrating detection capability. Without the false-positive volume, this is unfalsifiable: a system that flags everything early will always have a true positive to show afterward. *Fix:* report precision and recall over a fixed period, or lead time distribution across a cohort of known escalations. - **Appeal to unnamed authority — Fireworks at [00:06:02].** "We run a lot of evals across multiple models and we've seen models like GLM or K3 actually rival Opus and Fable on many tasks." No benchmark named, no scores, no task list, and the evaluator is the party that profits from serving open-weights models. *Fix:* publish the eval suite and the numbers, or cite a third-party leaderboard. - **Argument from anticipated consensus — Skyflow at [00:25:31].** "The jury is out and everybody's decided that actually this is where we are headed." The board's earlier skepticism is presented as disproven by the market having moved, but no evidence of the movement is offered beyond the assertion. The rhetorical move also conveniently establishes the speaker's foresight. *Fix:* cite adoption data for multi-model, multi-datastore architectures rather than asserting the consensus. - **Unsupported superlative comparison — Horizon at [01:12:09].** "We're confident that our models are better at interviewing people than probably your best analyst," supported only by having done 40,000 interviews. Volume is not accuracy. The hedge "probably" doesn't rescue the comparison, because no comparison was run. *Fix:* a head-to-head study — same department, agent interviews versus analyst interviews, both validated against system logs. - **Conflating pipeline with realized value — Horizon at [01:08:08].** The chain runs 2,000 interviews → 50,000 hours found → $2.5 million pipeline. The hours figure is a model output, not a measured time study, and the dollar figure is a forecast of improvement opportunity, not savings booked. Presented in a sequence that reads as realized outcome. *Fix:* state explicitly that $2.5 million is identified opportunity, and report realized savings from a completed initiative separately. - **Fear appeal with an unverified referent — Skyflow at [00:22:50].** "How many of you here have heard about the hugging face incident a couple of days ago? Who amongst us can say that my model is not going to reach back out and do something nefarious?" The incident is invoked for emotional weight with no detail, and the rhetorical question generalizes from one event to all models. *Fix:* describe the incident, its attack class, and whether the proposed control addresses that class. - **Not a fallacy, checked and cleared:** Seclore's industry statistics at [00:57:38] (80% unstructured, 40 million permission sets) are unattributed but are not presented as Seclore outcomes — they're framed as the market problem. Baseten's model specifications throughout are verifiable claims about published models, not persuasion. StackOne's 450-connector and 30,000-action figures are product inventory claims, checkable against their catalog. #### Tone of voice - **Host (Zuben):** Warm, connective, explicitly ecosystem-positioning. Each transition frames the next vendor as a natural consequence of the last ("Once AI can connect across the business, it can do more than move information" [00:48:07]). The recurring subtext — "we're an open platform" [01:15:37] — is a competitive positioning statement dressed as hospitality. - **Fireworks:** Ideological and measured. The AGI-concentration critique at [00:04:42] is a values argument before it's a product argument. - **Gamma:** Enthusiastic and aspirational, the least technical of the eight. Berlin's register is evangelical about slides as a strategic function. - **Skyflow:** Conversational, audience-interrogating, heavy on rhetorical questions and shared-pain humor about searching Concur [00:23:31]. - **Baseten:** The most technically confident and the most self-deprecating. Opens by admitting a mistake, then delivers the densest content in the session. - **StackOne:** Structured and consultative. Problem-solution-proof in three clean beats, closing with a question back to the audience [00:48:07]. - **Enterpret:** Scattered and disrupted by technical failures, with high disfluency. The bathroom-latch story lands anyway because it's specific. - **Seclore:** Formal, institutional, security-professional register. Explicitly allocates its own eight minutes at [00:56:18] — "I'm going to spend a lot of the time talking about the bigger picture problem." - **Horizon:** Earnest, non-native-English, visibly nervous ("last one, so bear with us" [01:05:40]). The founders' credibility rests on 15 years of transformation work rather than polish. ## 🛡️ DARVO Analysis Not applicable to this source. DARVO (Deny, Attack, Reverse Victim and Offender) describes a defensive response to accusation, and no speaker in this session is accused of anything — it's a vendor showcase with no adversarial party present, no Q&A, and no critic to respond to. For completeness, the three moments that superficially resemble the pattern and are not it: - **Deny:** No denials. Skyflow's board anecdote at [00:25:31] recounts someone else being wrong, but frames it as a market prediction the speaker also had to wait out, not as a rebuttal of a charge against him. - **Attack:** No attacks on any person. Fireworks' critique of frontier-lab concentration at [00:04:42] targets a market structure and an idea, names no competitor, and makes an argument rather than an attribution of bad faith. - **Reverse victim/offender:** Absent. The closest rhetorical structure is StackOne's framing of IT and security teams at [00:44:50] as *worried* rather than obstructive — which inverts the usual "security blocks innovation" complaint toward sympathy for the blocker. That's a de-escalating move, the opposite of victim-offender reversal. - **Emotional cues present but not DARVO:** the fear appeal about the Hugging Face incident [00:22:50] and the intern-querying-stock-options scenario [00:24:10] are threat framings meant to motivate purchase. They belong under the fallacies section above, not here. ## 🎯 Conclusion The showcase's real content isn't eight product pitches — it's a five-layer map of why Glean deployments stall, drawn accidentally by which companies exist to sell into which gap. Data has to be classified and labeled before an index can enforce anything; permissions inherited from source systems don't describe what an agent may write; connectors, not licenses, determine whether purchased AI gets used; and picking the first agent by intuition wastes the budget that process discovery would have aimed. Two claims in 78 minutes carry a stated baseline: StackOne building 15 connectors in two weeks against a planned six-month internal build [00:46:50], and Horizon converting 2,000 finance interviews across 15 countries into 50,000 mapped hours and a $2.5 million prioritized pipeline [01:08:08] — and the second is pipeline, not realized savings. The most-quoted-sounding numbers, Gamma's two months per person per year and 3-month-to-5-day reporting cycle [00:18:05], have no named client, no baseline, and no method. Discount accordingly and treat the sequencing, not the statistics, as what you take away. **Practical next steps if you're deciding what to build with Glean:** 1. Inventory which systems your intended agents must *write* to, and check connector coverage before buying more seats. This is the failure mode that produced shelfware at [00:46:50]. 2. Write a field-level write policy for one agent — StackOne's create-user-but-never-grant-admin shape [00:44:50] — and take it to security. If the review has no owner, that's your critical path. 3. Pick a first agent from the validated three: new-hire provisioning, access and IT ops, or month-end close [00:47:31]. 4. Choose a data-protection position deliberately: label before indexing, remediate against the existing index, or mask at runtime [01:01:33]. Each has different exposure properties and different deployment cost. 5. Baseline before you build. Measure the process for two weeks first. Every weak claim in this session failed for want of a before state, and your internal report will be judged the same way. ## 🔄 Refactor Appendix — Glean:GO AI-Native Showcase > 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] Enterprises allow agent reads but block agent writes > [!quote] ID:yt_kIfYH2JIDus_t2483 > many companies are allowing agents to read but not yet to write because they don't have a solution to enable the agents to safely write **Context & use**: - [00:41:23], StackOne segment, Romain Sestier. - This is the single most reusable diagnosis in the session. It explains why four of eight vendors exist and why Glean licenses sit unused. Read permissions describe visibility; write governance has to describe field-level intent, which no source system exposes. **Tags**: #xref #share **Highlight**: [00:41:23](https://www.youtube.com/watch?v=kIfYH2JIDus?t=2483) <!-- atomicity: 1 idea ✓ --> #### [H2] Missing connectors, not licenses, cause AI shelfware > [!quote] ID:yt_kIfYH2JIDus_t2810 > Stack one was able to build these two custom connectors plus 13 others within the span of two weeks **Context & use**: - [00:46:50], StackOne segment. A 200+ employee cancer research network had purchased AI platform licenses it could not deploy because the platform wasn't connected to UKG and a customized NetSuite. - The only case in the session with both a stated counterfactual (six months of planned internal build) and a measured actual. Use it as the template for how an outcome claim should be constructed. **Tags**: #xref #addendum **Highlight**: [00:46:50](https://www.youtube.com/watch?v=kIfYH2JIDus?t=2810) <!-- atomicity: 1 idea ✓ --> #### [H3] Same weight shapes preserve prior inference optimization work > [!quote] ID:yt_kIfYH2JIDus_t2243 > What's incredible about GLM 5.3 is it's the same architecture. The weights are the same shape. **Context & use**: - [00:37:23], Baseten segment, Philip Kiely. He names the exceptions: retrain the speculator because hidden states changed, rerun the model optimizer. - Directly actionable for anyone budgeting inference work across model refreshes. Naming what doesn't carry over is what makes the claim credible rather than promotional. **Tags**: #xref **Highlight**: [00:37:23](https://www.youtube.com/watch?v=kIfYH2JIDus?t=2243) <!-- atomicity: 1 idea ✓ --> #### [H4] Process discovery precedes agent selection, not the reverse > [!quote] ID:yt_kIfYH2JIDus_t4088 > From those 50,000 hours, the model struck what matters the most. They prioritize and they get a pipeline of $2.5 million **Context & use**: - [01:08:08], Horizon segment. 2,000 interviews in one finance department across 15 countries produced the hours figure; the dollar figure is identified opportunity, not booked savings. - Adopt the sequence (map → rank recoverable hours → specify agent → build in Glean), discount the number. Pairs with [H2]: connector coverage determines feasibility, discovery determines priority. **Tags**: #xref #articleseed **Highlight**: [01:08:08](https://www.youtube.com/watch?v=kIfYH2JIDus?t=4088) <!-- atomicity: 1 idea ✓ --> #### [H5] Unstructured data has no rows and columns to secure > [!quote] ID:yt_kIfYH2JIDus_t1450 > that intern that you just hired who's just curious can also now say, you know, let me just see how many stock options does my manager have **Context & use**: - [00:24:10], Skyflow segment, Anu Sharma. The structural premise is at [00:24:50]: no word-level security in a PDF, no rows and columns in a model or vector store. - The clearest statement of why better retrieval creates a new security surface rather than a pure productivity gain. Use when arguing for classification-before-indexing. **Tags**: #xref #share **Highlight**: [00:24:10](https://www.youtube.com/watch?v=kIfYH2JIDus?t=1450) <!-- atomicity: 1 idea ✓ --> ### 🧠 Atomic Split Pad #### [P1] Write scope is a distinct permission primitive from read scope **Claim** (one sentence, Courtney's voice): An agent's permission to change a system must be specified at the field and value level, because read permissions inherited from the source system describe visibility rather than intent. **Evidence pointer**: [[#H1]] — StackOne's worked example at [00:44:50]: an access-provisioning agent may create a user but must never grant admin. **Scope & conditions**: Holds wherever an agent acts in a system of record with a mutable schema (HRIS, ERP, CRM, CMS). Weakens for append-only systems where the only write is a new record, and for agents that produce drafts requiring human approval before commit — there the approval step carries the scope. Counterexample: a write to an isolated scratch store needs no field-level policy. **Links**: - [[Capability-based access control]] — because extends the read/write distinction into field granularity - [[Principle of least privilege]] — because applies-in agent permission design - [[WordPress capabilities model]] — because contrasts on granularity; WP capabilities stop at the object type - [[Prompt injection defense]] — because evidence-for the claim that write scope must be narrow when inputs are untrusted <!-- atomicity: 1 idea ✓ --> #### [P2] Deployment blockers sit downstream of the tool you bought **Claim** (one sentence, Courtney's voice): When a capable tool goes unused, the binding constraint is usually an unbuilt integration rather than insufficient training or change management. **Evidence pointer**: [[#H2]] **Scope & conditions**: Holds for tools whose value depends on acting across systems the buyer already runs. Fails for tools with standalone value where the constraint genuinely is behavior change, and fails where the integration exists but the data behind it is wrong — which is the failure mode [[#H5]] describes. Diagnostic: ask whether a fully trained, fully willing user could complete the workflow today. If no, it's an integration problem. **Links**: - [[Theory of constraints]] — because extends to software adoption - [[Shelfware]] — because applies-in enterprise procurement - [[Change management fallacy]] — because contrasts on where adoption failure is attributed <!-- atomicity: 1 idea ✓ --> #### [P3] An outcome claim without a baseline is an anecdote in a number's clothing **Claim** (one sentence, Courtney's voice): A reported improvement carries evidentiary weight only when it states the measurement window, the population, and the pre-change value. **Evidence pointer**: [[#H2]] as the positive case (stated counterfactual: six months planned versus two weeks actual), against the Gamma claims at [00:18:05] as the negative case (three results, one unnamed client each, no baselines). **Scope & conditions**: Applies to any performance, savings, or adoption claim intended to drive a decision. Doesn't apply to inventory facts (450 connectors, 30 regions) or to published specifications (2.8 trillion parameters), which are checkable rather than measured. Use it as an editorial gate on outcome reporting, including your own. **Links**: - [[Survivorship bias]] — because extends to vendor-selected case studies - [[Evidence standards for internal reporting]] — because applies-in DevRel and OSPO metrics - [[Goodhart's law]] — because contrasts on what happens after the baseline is set <!-- atomicity: 1 idea ✓ --> ### ✏️ Fleeting Queue - [F1] Find out what the Hugging Face incident referenced at [00:22:50] actually was — the attack class matters for whether ingestion-time tokenization addresses it at all. #fleeting #inbox - [F2] Verify real speaker and company name spellings; ASR mangled nearly all of them (Baseten as "Base 10," Seclore as "Secllor"/"Cycllor," Enterpret as "Interpret," Kearney as "Karnney"). Don't cite any name from this transcript without checking. #fleeting #inbox - [F3] Does tokenized or masked text degrade semantic retrieval quality? Two vendors claim it doesn't ([00:26:13], [01:00:53]) and neither published a benchmark — this looks like an unoccupied research niche. #fleeting #inbox #articleseed - [F4] Where is the crossover point at which semantic tool discovery beats preloading a tool catalog? StackOne asserts it at 30,000 actions [00:46:11] and names no threshold. Worth measuring on a small agent. #fleeting #inbox - [F5] Map WordPress capabilities against agent write scope — `edit_posts` has no field-level equivalent, which is exactly the gap StackOne sells into. Possible post. #fleeting #inbox #articleseed - [F6] Baseten's preview-checkpoint strategy [00:35:23]: which model labs publish preview checkpoints, and is there a public list? Free head start if you're tracking releases. #fleeting #inbox - [F7] Steal Seclore's three-axis classification — content, context, intent [00:59:33] — for documentation metadata. The intent axis is the one nobody records. #fleeting #inbox - [F8] Pre-demo session checklist: two presenters got auto-logged out live ([00:49:39], [00:53:42]). Write the five-line checklist before the next conference talk. #fleeting #inbox ### → Things 3 (max 2 actions) - [ ] Draft a field-level write policy for one candidate Glean agent and take it to security for review — pick from new-hire provisioning, access/IT ops, or month-end close #things - [ ] Inventory which systems our intended agents must write to and check existing connector coverage before any further seat purchase #things <!-- deep-read:sections end --> 📄 **Full text:** [[Full Document Contents/Articles/GleanGO AI-Native Showcase.md|Glean:GO AI-Native Showcase by Glean]] ## Highlights - [00:23:31] companies like Oracle and Salesforce. And in the old world, life was simple, right? You could even barely search for your expense report. How many people here have tried searching for an expense report in Concur or an invoice in Aribba and spent years? Well, those were the good old times from a security perspective because if you literally can't search for something, bad things can't happen. Rule based access control at the level of rows and columns. Data is so locked up that you can barely use it. And then comes green and products like that that basically make it easy for actually to reach out and get the data as we need it. So you can actually [00:24:10] find your expense report. Unfortunately, that intern that you just hired who's just curious can also now say, you know, let me just see how many stock options does my manager have. Let me just see what is the last contract that we did that got cancelled. This is a problem that's nondeterministic because agents are nondeterministic, models are nondeterministic. The second problem is the data is no longer cleanly separated in rows on columns. You can't say column number seven has social security numbers, column number three has credit card numbers. If you think about AI from my perspective, it's literally a computer ([View Highlight](https://read.readwise.io/read/01m48t7mh1jbt20xhzck8b909c)) - Note: [H5] Unstructured data has no rows and columns to secure - Tags: atom/literature/h3 - [00:37:23] trade-off between throughput and latency. We were able to create a standard API as well as a fast API to give multiple options along that curve. What's incredible about GLM 5.3 is it's the same architecture. The weights are the same shape. They just have different values because of post-training. And when you're in that situation, every single piece of work you did on the previous instance of that architecture reapplies. Now, do you have to retrain the speculator? Yes, because the hidden states changed. Do you have to, you know, rerun your model optimizer? Yes, you do. But the patterns and the techniques stay the same. and you can [00:38:01] pretty confidently apply everything you did in the past to this new version of the model. So, thank you so much for your time here. If you want to learn more about inference engineering, uh, hit up that QR code. Um, we've got a free book that I wrote called Inference Engineering, um, that's been doing the rounds in the industry. Um, and if you want a paper copy, I didn't bring any with me today, uh, but hit me up on Twitter or LinkedIn. I'll get one sent over to you. Thank you so much. Have a great time at the conference. Thank you, Phillip. [music] Thank you, Phillip. And thank you, B 10. They are one of the providers that help us bring open source ([View Highlight](https://read.readwise.io/read/01m48t7eyy0ptbajf9qgnacr32)) - Note: [H3] Same weight shapes preserve prior inference optimization work - Tags: atom/literature/h3 - [00:41:23] an enormous task in itself. The second problem is that if you think about it giving agents right actions into your systems brings in an entire new category of problems that your IT and security teams are extremely worried about. And what I keep hearing is that many companies are allowing agents to read but not yet to write because they don't have a solution to enable the agents to safely write into their systems and take action. And finally, what we're seeing is that as agents complete entire workflows, they the the token costs increase and there need to be more levers, including Gle's awesome product with the model [00:42:05] routing. There need to be even more levers that allow you to save on tokens to make this an economic reality because it doesn't make sense to automate a workflow if it costs too much. Stack one is the platform that bridges all of these three gaps. If Glean is your control plane where you're already building agents and your teams are building agents as well as the context layer for these agents, stack one is the action layer that lets Glean take action into your systems safely and efficiently. As you can see on the video here, what's happening is that stack one actually lives natively behind the scenes with a single MCP that gets added ([View Highlight](https://read.readwise.io/read/01m48t79pbzq73fraj1nje9j8g)) - Note: [H1] Enterprises allow agent reads but block agent writes - Tags: atom/literature/h1 - [00:46:50] companies in a safe way. Here's an example. We recently onboarded a leading uh cancer research network uh that has more than 200 employees and they had purchased AI platform licenses already but they weren't able to deploy it to their employees because the AI platform was not connected to their systems. They needed two custom connectors to UKG and to Netswuite. One because they needed it to be HIPPA compliant and they needed to filter out PHI data. The other one is because uh Netswuite was heavily customized in their company. Stack one was able to build these two custom connectors plus 13 others within the span of two weeks [00:47:31] and roll out their AI tools to their 200 employees. And just for context, this is something that they were planning more than six months of build internally before they met stack one. That's how much they were able to accelerate their deployment. If you want to see an example of this, you can actually log into the attendee portal right now and see three examples of agents that we built on Glean with the stack one layer added to it. One for new higher provisioning, one for access and IT ops, and one for month and close um that uses all of the features I talked about. Now, I'll leave you with a question. What would change in your ([View Highlight](https://read.readwise.io/read/01m48t7ck3ysrzvhx379fe246t)) - Note: [H2] Missing connectors, not licenses, cause AI shelfware - Tags: atom/literature/h2 - [01:07:29] matters the most and help you to prioritize and go through the future state reinvention. And then what we do is we leverage all the knowledge that was already available in the company through presentations, documents, processes, but most important trap in people's head is the human judgment. So we leverage that human judgment into the graph for you to take the best decisions in how and where to improve the company. How this works in action is super simple and I will explain from a customer point of view. So our biggest customer it's an e-commerce marketplace in Latin America. They went through 2,000 interviews in [01:08:08] the finance department, 15 countries, all Latin America region, and they found 50,000 hours after the agents talk with their employee, map the processes, find what is going wrong, and the model study the patterns. From those 50,000 hours, the model struck what matters the most. They prioritize and they get a pipeline of $2.5 million for improve the company. And then it's super simple. like they will get those right-sized projects and move to action in two ways. Big initiatives where the model will create the request for proposal, the one pure, the strategic plan, all you need to move to action. But then in the quick wins, ([View Highlight](https://read.readwise.io/read/01m48t7h8p2th09rehrh7cx1a9)) - Note: [H4] Process discovery precedes agent selection, not the reverse - Tags: atom/literature/h3