Advertise on podcast: Shift AI with Boaz Ashkenazy

In this episode of Shift AI, Colleen Kozak, Chief Transformation Officer of Dynatrace, joins host Boaz Ashkenazy for a wide-ranging conversation on what's actually slowing down agentic AI adoption inside the enterprise. Everyone assumes the blocker is cost. Colleen sees something different playing out across her customers, from token-maxing tech companies to cautious, regulated enterprises: the real hesitation isn't the bill, it's not knowing whether the agent got it right. She walks through what that gap looks like in practice, and what happens once companies decide they're ready to hand agents the keys to take action on their own. This one's for CTOs, platform engineers, security leaders, and anyone driving AI strategy inside a large organization who's trying to figure out how to move from pilot to production without losing the ability to see what their agents are actually doing. Chapters [00:01] Welcome and introductions [00:32] From mechanical engineering to Chief Transformation Officer [01:54] The first paycheck: babysitting [02:42] Inside Dynatrace: Grail, Smartscape, and Dynatrace Intelligence [06:03] From copilots to full agentic systems [07:03] Why trust, not maturity, is the real divide between companies [08:32] Building trust into agents at runtime, and securing a world with more agents than employees [12:42] Splitting the work: human review, automated review, and internal vs. customer-built agents [14:15] Tokenomics and choosing the right model for the right job [16:52] What dogfooding AI internally revealed, and getting SaaS vendors to open their telemetry [21:00] Sovereign cloud, geopolitics, and the return of on-prem [23:27] Two words for the future, the jobs question, and hiring at Dynatrace Connect with Colleen Kozak LinkedIn: https://www.linkedin.com/in/colleen-kozak Dynatrace careers: dynatrace.com Connect with Boaz Ashkenazy LinkedIn: https://www.linkedin.com/in/boazashkenazy/ Email: [email protected]

In this episode of Shift AI, Mick Costigan, Vice President of Salesforce Futures at Salesforce, joins host Boaz Ashkenazy live from Dreamforce for a wide-ranging conversation on how AI is reshaping the way work gets done, and who does it. Mick's team are all Claude coders now, and they've noticed something strange. Each person goes deep with their AI, surfaces with an 8-page document, and finds everyone else has tunneled off somewhere else with one of their own. Coding figured out how to bring that work back together. Mick isn't sure the rest of work can, and he has a 2x2 that suggests which parts of your job are about to be pulled apart. CIOs, CTOs, strategy and transformation leaders, and anyone designing teams and roles for the agentic era will come away with a sharper way to think about which work stays human. Chapters [00:00] Live from Dreamforce: what futures work is and how Salesforce Futures operates [05:15] An off-license in Ireland and a bouncer job across from Wrigley Field [07:05] Harnesses before anyone called them harnesses [10:25] Spotting the headless platform back in 2016 [13:00] Opus 4.5 and the team's Claude Code moment [15:50] Tunneling and the 8-page document problem [19:15] From work from anywhere to intentional in-person time [21:10] The verifiability and data 2x2 for unbundling work [24:50] Slop, efficiency, and why we won't all become Rick Rubin [26:55] Finding the human part of your own job [28:20] Bundling, unbundling, and the radiologist lesson [31:05] Futures Magazine and how to reach the team Connect with Mick Costigan LinkedIn: https://www.linkedin.com/in/mcostigan/ Futures Magazine: salesforce.com/news/futures Email the Futures team: [email protected] Connect with Boaz Ashkenazy LinkedIn: https://www.linkedin.com/in/boazashkenazy/ Email: [email protected]

In this episode of Shift AI, Mario Rodriguez, Chief Product Officer of GitHub, joins host Boaz Ashkenazy for a wide-ranging conversation on what GitHub is becoming now that humans and AI agents build software side by side. Mario shares a number from inside GitHub's own Copilot data that cuts against the idea of a single dominant AI model, and explains why he thinks the real competition isn't happening at the model layer at all. He also lays out how GitHub is rethinking its surfaces, from Agent HQ to a tool called Spark, around a much bigger bet on where developer collaboration is headed next. This episode is for CTOs, VPs of engineering, platform and DevEx leads, and any software engineer trying to figure out where their own skills fit as more of the codebase gets written by agents. Chapters [00:00] Welcome and introducing Mario Rodriguez [00:29] From a Cuban restaurant busboy to electrical engineering to Perl [03:11] Landing at Microsoft, then GitHub after the 2018 acquisition [03:58] Redefining GitHub beyond repos: the agent-native engineering system [07:31] Staying true to developers and going model-agnostic [09:06] The Copilot usage data on model choice [10:27] Harnesses, and why the fixation on any one model is fading [13:49] Agent HQ, the GitHub Copilot app, and canvases explained [16:29] Chronicle and the continuous learning loop [18:10] Spark: lowering the floor and raising the ceiling for builders [21:35] What AI means for new versus experienced software engineers [29:09] The closing two words: fluid creativity Connect with Mario RodriguezLinkedIn: https://www.linkedin.com/in/mariorodriguez3X: https://x.com/mariorod1 Connect with Boaz Ashkenazy LinkedIn: https://www.linkedin.com/in/boazashkenazy/ Email: [email protected]

In this episode of Shift AI, Emrecan Dogan, Head of Product of Glean, joins host Boaz Ashkenazy for a wide-ranging conversation on what it actually takes to give enterprise AI real organizational context. Emrecan has convinced countless CIOs and CEOs to stop chasing a data cleanup project that he says will never finish. Instead he makes the case for leaving the mess exactly where it is and solving the problem somewhere else entirely, a position that runs against nearly everything companies think they need to do before they can "get onto AI." He also lays out a theory of why knowledge workers spend half their AI time just re-teaching context, and what it would take to get that number to zero. This one is for CIOs, chief AI officers, heads of product, and platform and data leaders who are being told to fix their data before they can move on AI, and who want a second opinion. Chapters [00:00] Welcome and introduction [00:33] Emrecan's path to Glean: entrepreneur, LinkedIn, Stripe, and a warning from a VC [02:36] First paid job: fixing bikes in Turkey [04:19] What Glean actually does: the work AI platform explained [06:53] The onboarding moment of truth and losing Glean when you leave [09:40] Why organizational data, not model choice, is the real moat [10:47] Seven years of infrastructure work before the GPT moment [13:00] The pipe dream of cleaning up your data stack first [15:36] "Bot sitting": the hidden cost of teaching context to AI every day [17:59] Why picking a single model provider is one of the worst bets you can make [23:37] Token yield, marginal cost intelligence, and the "half my budget is wasted" problem [30:36] Assistants versus agents: a framework for when each one applies [36:31] What's coming next: durable, multiplayer artifacts and proactive AI [40:34] Closing question: the future of work in two words Connect with Emrecan Dogan LinkedIn: https://www.linkedin.com/in/emrecandogan/ Glean blog: glean.com Connect with Boaz Ashkenazy LinkedIn: https://www.linkedin.com/in/boazashkenazy/ Email: [email protected]

In this episode of Shift AI, Paul Mikesell, CEO of Carbon Robotics, joins host Boaz Ashkenazy for a wide-ranging conversation on how AI, lasers, and autonomy are replacing chemical weed control on farms. Carbon Robotics built a foundation model out of 150 million plant images, and Paul explains what happens when you show it a weed it has never encountered before. There's no retraining involved, and the result is something closer to a machine that understands what a plant is rather than one that's just memorized a dataset. What that actually looks like in the field, and why it took years of accumulated data to get there, is worth hearing him walk through directly. This episode is for CTOs and engineering leaders working on physical AI and robotics, agtech investors and founders, and anyone building computer vision systems that need to generalize beyond their training data. Chapters [00:01] Welcome to the show [00:31] From self-driving cars at Uber to founding Carbon Robotics [02:54] The farmer's real problem: weeds, chemicals, and health risks [04:05] Paper routes, Nintendo, and a Commodore 64: Paul's path into computing [07:59] Laser Weeder by the numbers: 15 countries and $100M+ in revenue [10:04] Inside the machine: light brighter than the sun and real-time targeting [13:33] The tractor driver bottleneck and the move to autonomy [14:55] Center pivots and near-misses: what farms teach robots that roads don't [18:40] The real moat: people, data, and years in the field [20:31] The Large Plant Model: recognizing a weed it's never seen before [24:23] Physical AI and why data is becoming the new moat [26:57] Two words for the future: autonomy and health Connect with Paul Mikesell LinkedIn: https://www.linkedin.com/in/paul-mikesell-4b63a9/ Company: carbonrobotics.com Connect with Boaz Ashkenazy LinkedIn: https://www.linkedin.com/in/boazashkenazy/ Email: [email protected]

In this episode of Shift AI, Praerit Garg, CEO of One Identity, joins host Boaz Ashkenazy for a wide-ranging conversation on what happens to enterprise security when AI agents start outnumbering the humans they work alongside. Praerit has spent his career on both sides of this problem, from building Active Directory at Microsoft to running identity at AWS, and he says the thing keeping him up at night isn't a new category of threat at all. It's an old one moving at a speed nobody has built the guardrails for yet, and he points to a very recent, very public example of exactly how fast it can get away from you. This one's for CISOs, CTOs, platform and security engineers, and any founder scaling a company that's already handing real work to autonomous agents. Recorded Live at Baker Tily during Seattle Tech Week 2026. Thank you to Stifel Bank and Wison Sonsini for sponsoring. Chapters [00:00] Welcome back — PG's journey from Smartsheet to CEO of One Identity [01:58] What One Identity does: identity governance, explained [03:41] How the firewall evaporated and identity became the new perimeter [04:24] Rethinking identity for a world full of AI agents [07:14] Why agent identities are harder to secure than human ones [08:00] Breaking down the OpenAI and Hugging Face agent escape [09:03] Least privilege: the hardest idea in security to actually pull off [11:19] What keeps PG up at night as agents start to outnumber people [15:19] Sovereign AI, air-gapped models, and the return of on-prem [18:18] Scaling Smartsheet fast, and why flexibility beats architecture bets [20:20] What CEOs should be asking security vendors and usually aren't [22:37] The next five years: agents outnumbering humans and AI writing the code Connect with Praerit Garg LinkedIn: https://www.linkedin.com/in/praerit-garg/ Company: One Identity (oneidentity.com) Connect with Boaz Ashkenazy LinkedIn: https://www.linkedin.com/in/boazashkenazy/ Email: [email protected]

In this episode of Shift AI, Sabina Anja, Chief Technologist at VMware by Broadcom, joins host Boaz Ashkenazy for a wide-ranging conversation on why private, flexible infrastructure is becoming the foundation enterprises need for the agentic era. Sabina makes the case that most infrastructure and security thinking was built for a world where humans approve every step, and that world is already gone. An assistant waits for you. An agent, in her words, gets set loose and just goes. So what happens when nobody notices an agent doing something it was technically allowed to do, but shouldn't have? Sabina has a way of thinking about that question that most infrastructure teams haven't caught up to yet. This episode is for CIOs, CTOs, VPs of infrastructure, platform engineers, and security leaders who are trying to figure out what their environments need to look like before agents are running unsupervised inside them. Chapters [00:00] Welcome and introducing Sabina Anja [00:31] From Belgium to self-taught programmer: Sabina's first jobs [04:29] Why programming hygiene still matters with Cursor and Claude Code [06:24] Virtualization then and now: from vCPUs to vGPUs [08:52] What businesses are really asking for as AI costs spike [10:46] Inside NVMe tiering and the memory crunch [13:28] Private AI and the return of sovereign infrastructure [15:43] Data ownership, geopolitics, and the new value of stolen IP [17:48] Open source, small language models, and fit-for-purpose compute [20:59] Advice for CIOs: build flexible foundations, not two-year projects [24:40] Agents vs. assistants: why blast radius changes everything [27:26] Agentic security, guardrails, and the two words: platform matters, earned trust Connect with Sabina Anja LinkedIn: https://www.linkedin.com/in/sabinaanja/ Connect with Boaz Ashkenazy LinkedIn: https://www.linkedin.com/in/boazashkenazy/ Email: [email protected]

In this episode of Shift AI, Manos, CEO of Umi, Patrick Thompson, CEO of Clarify, Brian Hall, CMO of Mistral AI, and Ben Gaffney, Deputy General Counsel at OpenAI, join host Boaz Ashkenazy for a live panel recorded at Seattle Tech Week on who actually owns intelligence in the enterprise AI era. One panelist argues that any enterprise still treating a frontier model as its core intelligence layer is renting something it should own, and predicts that in two years we'll look back at CTOs defending that approach the way we look back at driving a Ferrari to the grocery store. Another panelist, sitting inside one of the most closely watched policy fights in AI right now, explains why the debate over open source models and national security is getting a lot less theoretical, and a lot more real, by the month. This one is for CTOs and CIOs deciding whether to build or rent their AI stack, product leaders at vertical AI startups thinking about fine-tuning their own models, and anyone tracking the open source AI security and policy debate. Chapters [00:00] Welcome and thanks to sponsors JP Morgan and Ascend [01:30] Introducing the panel: Manos, Patrick, Brian, and Ben [03:54] From foundation model hype to vertical AI, what actually changed [07:37] How Clarify balances owning the system of record with using frontier models [09:44] Rising model bills, local models, and the push toward open weights [11:30] Mistral AI, sovereign AI, and why owning your intelligence matters [14:32] Data retention, training, and what the privacy debate is really about [19:10] Chinese open models, national security, and the safety debate [26:18] Why fine-tuning beats prompting for real vertical AI companies [30:43] What enterprises actually want from foundation models right now [35:35] What CIOs and CTOs will laugh about in two years [44:47] The hundred million dollar question: where the panel would invest [51:44] Advice for a 22-year-old about to spend ten years building something [53:31] What each panelist is most excited about in AI's future Connect with the Panelists Manos (Umi) - https://www.linkedin.com/in/koukoumidis/ Patrick Thompson (Clarify) - https://www.linkedin.com/in/patrickt010/ Brian Hall (Mistral AI) - https://www.linkedin.com/in/brhall/ Ben Gaffney (OpenAI) - https://www.linkedin.com/in/bengaffney/ Connect with Boaz Ashkenazy LinkedIn: https://www.linkedin.com/in/boazashkenazy/ Email: [email protected]

In this episode of Shift AI, Zohar Alon, CEO and Co-Founder of Newcore, joins host Boaz Ashkenazy live from Black Hat 2026 for a conversation on why identity, not infrastructure, is the real fault line in agentic AI. Zohar has spent thirty years watching the same problem resurface in a new costume, from VPNs to IDPs to now agents that can act on your behalf for hours at a time. He describes a system where a single trusted identity point can unlock hundreds of connected apps, and where an AI agent chasing a vague instruction might decide, on its own, to read six months of email or start testing messages against real people to get results. What happens when you can't fully predict what the agent will try next, and the guardrails have to be built for that uncertainty rather than against it, is the question Zohar has built a company around. This one is for CISOs, identity and access management leads, platform engineers building with autonomous agents, and founders working in enterprise security who need to think past service accounts and static permissions. Chapters [00:00] Welcome to Shift AI, live from Black Hat 2026 [00:25] Thirty years in cyber, from Check Point employee 55 to founder [01:30] Founding Dome9 and convincing the world the cloud needed securing [02:25] The identity problem that never got solved, from VPNs to modern IDPs [04:50] Building Newcore: one identity fabric for humans and agents [06:15] Why a single IDP becomes a single point of catastrophic risk [07:13] Pilot school, electronics class, and Zohar's path into Check Point [09:36] Mentoring the next generation of Israeli cyber founders [10:45] Why an AI agent isn't a service account [11:59] Inside an agent's "creative" pursuit of a goal [14:10] Human-in-the-loop as protection, not friction [16:35] The closing two words: unknown unknown Connect with Zohar Alon LinkedIn: https://www.linkedin.com/in/zoharalon/ Company: Newcore -https://newcore.com/ Connect with Boaz Ashkenazy LinkedIn: https://www.linkedin.com/in/boazashkenazy/ Email: [email protected]

In this episode of Shift AI, Shay Sandler, CEO and co-founder of Vega, joins host Boaz Ashkenazy for a wide-ranging conversation on why the legacy SIEM is breaking down just as security teams need their data the most. Shay grew up in Israel studying math and physics in a famously competitive class of gifted kids, then spent six years in one of the country's most elite military intelligence units before joining Granulate, a cloud cost optimization company, as one of its first ten employees. Granulate was acquired by Intel for $650 million, and after a year inside Intel, Shay left with his former Granulate colleague Ellie to found Vega — a decision he says was pushed forward by a month of reserve duty after October 7th that reminded him how much he missed solving hard problems with his old team. The two dig into why traditional SIEMs, built for a pre-cloud world of centralized data, are collapsing under multi-cloud, multi-region, siloed telemetry — and why teams that once filtered data down to save on SIEM costs now need all of it to keep up with threat actors wielding frontier models. Shay walks through Vega's architecture: querying data in place in commodity storage like S3, federating across regions and clouds without duplication or egress costs, and a deterministic, auditable natural-language-to-KQL layer that lets AI agents run threat hunts a human could fully verify. This episode is for CISOs, security operations leaders, detection engineers, and startup founders building in cybersecurity who want to understand how AI-native architecture, not just AI features, is reshaping enterprise security. Chapters [00:00] Introduction: Shay Sandler's path to founding and leading Vega [02:07] Growing up gifted in math and physics, and finding cybersecurity [04:18] Six years in an elite Israeli intelligence unit, and the "everything is solvable" mindset [06:08] First paid job at 14: distributing meat before dawn in southern Israel [08:37] Granulate, the $650 million Intel acquisition, and missing the startup thrill [10:43] October 7th reserve duty and the spark to build Vega with co-founder Ellie [15:04] How legacy SIEM broke under multi-cloud, siloed security data [17:32] The post-Mythos era: frontier models as tools for threat actors [23:16] Vega's three primitives: commodity storage, federation, and an AI-native interface [27:50] Natural language to KQL as a transparent, auditable layer for AI agents [30:05] AI as a "scalable engineering capability" for lean security teams [36:40] Velocity as the new core metric, and the closing thought: creativity and engineering Connect with Shay Sandler LinkedIn: https://www.linkedin.com/in/shay-sandler-305508107/ Connect with Boaz Ashkenazy LinkedIn: https://www.linkedin.com/in/boazashkenazy/ Email: [email protected]

In this episode of Shift AI, Andrew Comstock, Senior Vice President and General Manager of MuleSoft from Salesforce, joins host Boaz Ashkenazy for a wide-ranging conversation on governing, securing, and controlling the cost of agentic AI at enterprise scale. The conversation covers how MuleSoft's API-led integration playbook, built during the on-prem-to-cloud shift, is now extending into what the company calls Agent Fabric — bringing governance and controlled connectivity to agent-to-agent communication. Andrew walks through a concrete example of how two internal agents trusting each other by default can leak a customer's order history, and digs into prompt injection, impersonation, and why the security industry's people-focused compliance frameworks now have to account for reasoning software, not just people. He and Boaz also get into token economics, LLM gateways, intent-based routing between local and cloud models, and the emerging trend of companies repatriating some AI workloads on-premise. This episode is essential listening for CTOs, CISOs, platform and integration engineers, and IT and product leaders responsible for governing AI spend and agent security as their organizations scale past pilot projects into production. Chapters [00:01] Andrew Comstock's path to MuleSoft [02:06] First job: tax returns and trumpet reeds [03:21] What MuleSoft does and its extension into Agent Fabric [05:29] From consumer chatbots to enterprise agent connectors [07:22] Governance and security at the enterprise level [10:36] A real example: how two agents can leak an order number [13:45] Connect AI 2025 vs. 2026 — a year of night-and-day change [16:28] Mythos, Fable, and preparing for the LLM that breaks your systems [18:11] Token economics and the true cost of enterprise AI [20:38] The "AI savior" pattern vs. applying IT discipline to AI [22:40] Hybrid deployments, local models, and intent-based routing [27:13] Why AI's uneven acceleration makes prioritization more valuable [28:57] The two-word answer: "coming soon" Connect with Andrew Comstock LinkedIn: https://www.linkedin.com/in/andrewcomstock/ Connect with Boaz Ashkenazy LinkedIn: https://www.linkedin.com/in/boazashkenazy/ Email: [email protected]

In this episode of Shift AI, Noah Smith, Vice Provost for AI at the University of Washington and Senior Director of NLP Research at the Allen Institute for AI, joins host Boaz Ashkenazy for a wide ranging conversation on how universities are preparing for a workplace increasingly run by AI agents. Noah's path wound from a PhD at Johns Hopkins to a tenured professorship at Carnegie Mellon, until a call from UW pulled him west in 2015 to help build out its natural language processing faculty. A few years later he took on a second role leading a research team at the Allen Institute for AI, and this past November added a newly created role as UW's first Vice Provost for AI. The conversation covers how faculty across UW are experimenting with AI in the classroom, from an interactive logic textbook in the philosophy department to new AI literacy courses, and digs into OLMo, the fully open language model project Noah leads at Ai2, unpacking why releasing the weights, code, and training data together, not just an API, matters for regulated industries, universities, and anyone who wants to retrain a model rather than just prompt it. Noah lays out a scenario where 80 percent of entry level work gets done by autonomous agents and argues the skill universities need to double down on is not AI literacy so much as discernment, the ability to unwrap a problem and figure out what questions to ask. This one is for university administrators, provosts, and faculty grappling with AI policy, as well as founders and engineers building on open source models, and anyone curious how a major research university is trying to get ahead of the agentic era instead of just reacting to it. Chapters [00:05] Welcome to Shift AI, live from UW's Foster School of Business [03:44] Noah's path from Carnegie Mellon to UW's Vice Provost for AI [06:08] Bagging groceries: Noah's first paid job [08:11] Six months in: what surprised Noah most about the role [09:22] How AI is reshaping teaching and pedagogy at UW [11:33] Preparing students for a world where agents do 80 percent of the work [14:08] Inside the Allen Institute for AI and the origins of OLMo [15:52] What "fully open" really means for language models [17:49] Specialized models versus general-purpose models for regulated industries [19:47] AI governance, security, and UW's new governance committee [24:35] The future of work in two words: human agency Connect with Noah Smith LinkedIn: https://www.linkedin.com/in/noah-smith-0322511a4/ Connect with Boaz Ashkenazy LinkedIn: https://www.linkedin.com/in/boazashkenazy/ Email: [email protected]

In this episode of Shift AI, Jared Wray, CEO and co-founder of Hyphen, joins host Boaz Ashkenazy for a wide-ranging conversation on how AI is poised to eliminate the complexity of cloud infrastructure and the DevOps role entirely. Jared shares his unconventional career journey from growing up in a small town in Idaho, where technology barely existed, to washing dishes at 15, teaching himself programming at the local ISP, and eventually founding five startups across cloud computing, energy tech, and developer infrastructure. From bootstrapping Tier 3 (acquired by CenturyLink) to co-founding Palmetto, now one of the largest energy lenders in the nation, Jared's path has been defined by a passion for solving infrastructure problems. The conversation dives deep into why DevOps has become painfully complex, with developers needing to glue together seven to thirteen different services just to deploy a single application across providers like AWS, Google Cloud, and Cloudflare. Jared explains how Hyphen is using AI to abstract away this complexity by asking developers only for business rules like uptime requirements and performance needs, then letting the AI determine the right architecture, deploy it, and operate it autonomously. Boaz and Jared explore why tools like Claude Code and other coding agents still cannot handle the full deployment lifecycle, what it would look like if AI agents replaced PagerDuty by calling you during an outage with a diagnosis and recommendation, and why the future of infrastructure is an autonomous cloud where humans are decision makers and agents handle everything else. The episode closes with a forward-looking discussion on agent-only companies, the death of cloud certifications, and why Jared believes DevOps was a good idea that we ran too long. This episode is essential listening for CTOs, platform engineers, and startup founders who want to understand how AI is moving beyond writing code to fundamentally transforming how software is deployed, operated, and scaled. Chapters [00:00] From Pocatello to Five Startups: Jared's Career Journey [02:49] Building Fonz, Co-founding Palmetto, and Finding Passion in Infrastructure [06:05] Why AI Led Jared Back to Infrastructure with Hyphen [07:05] First Job as a Dishwasher and Breaking Into Tech [08:46] What Is DevOps and Why Does It Exist [10:49] Why Cloud Infrastructure Has Become So Painfully Complex [12:36] How AI Can Apply Best Practices Without Reinventing the Wheel [14:35] The Hyphen Developer Experience: Business Rules Over Architecture [17:14] Why Claude Code and Coding Agents Cannot Solve Infrastructure Yet [20:26] The Full Context Problem: Operating Across Multiple Cloud Providers [23:07] Autonomous Cloud: When Agents Talk to Agents [24:17] Replacing PagerDuty: AI Agents That Call You During Outages [28:47] March Madness, Live Streaming, and Why Five Minutes Feels Like a Lifetime [30:59] Two Words for the Future of Work: Autonomous Cloud [33:20] Agent-Only Companies and Why Humans Will Be CEOs [35:12] DevOps Was a Good Idea We Ran Too Long [35:41] What Is Next for Hyphen Connect with Jared Wray LinkedIn: https://www.linkedin.com/in/jaredwray/ Email: [email protected] Connect with Boaz Ashkenazy LinkedIn: https://www.linkedin.com/in/boazashkenazy/ Email: [email protected]

In this episode of Shift AI, Patrick Hillmann, Chief Strategy Officer at Logical Intelligence, joins host Boaz Ashkenazy for a conversation about why the next era of AI cannot be built on probability alone. Patrick shares his unconventional path into AI, from crisis communications and cybersecurity work at Edelman and General Electric, to steering Binance through its most turbulent years and a major DOJ settlement. At Logical Intelligence, Patrick now works alongside Yann LeCun, a Fields Medalist, and engineers from Meta, Google, and Cruise to build deterministic, energy-based reasoning models. Patrick explains why LLMs behave like a confident intern, fast and articulate, but wrong in ways you only catch if you already know the answer, and why critical systems like power grids, hospitals, and self-driving cars need a layer of certainty that probabilistic systems cannot provide. He and Boaz dig into Logical Intelligence's benchmark results, including a 98% score on the notoriously difficult Putnam math competition, and a public Sudoku test where their energy-based model, Kona, beat every major LLM combined while running on a fraction of the compute cost. This episode is essential listening for CTOs, technical leaders, and anyone trying to understand what comes after the current generation of large language models. Chapters [00:00] Patrick's Improbable Path: From Grad School to the Front Lines of a Geopolitical Crisis [02:43] From Binance to Chief Strategy Officer at Logical Intelligence [02:54] The First Paid Job: Unloading UPS Trucks in 100 Degree Heat [04:31] The UPS Lesson That Still Shapes How He Thinks About Work [04:47] Why LLMs Are Confident Guessing Machines, Not Truth Machines [07:14] The Team Behind Logical Intelligence: A Fields Medalist, Yann LeCun, and Math Olympiad Engineers [08:55] Is Logical Intelligence Betting Against LLMs? [10:32] The AI Sandwich: Where LLMs, Reasoning Layers, and World Models Fit [12:36] The Putnam Benchmark and Why Formal Proofs Don't Get Partial Credit [15:44] What Is an Energy-Based Model, Really? [19:54] Eve Badia's 15-Year Path to the Energy-Based Reasoning Model [22:04] Formal Verification and the Future of Secure Code Generation [23:33] When Unverified Code Fails: The Molson Coors Ransomware Story [25:39] Why AI Coding Tools Create Rat's Nests Engineers Can't Debug [28:53] The Sudoku Test: 98% Accuracy for $4 vs $14,000 for the Leading LLMs [31:13] ByteDance, China, and the Race for Formal Methods [33:42] Two Words for the Future of AI: Chaotic Determinism [37:12] Where to Follow Logical Intelligence and Founder Eve Badia Connect with Patrick Hillmann LinkedIn: https://www.linkedin.com/in/crisiscommunications Connect with Boaz Ashkenazy LinkedIn: https://www.linkedin.com/in/boazashkenazy/ Email: [email protected]

In this episode of the Shift AI Podcast, Cynthia Tee, former CTO of Smartsheet, joins host Boaz Ashkenazy for a wide-ranging conversation on what it really takes to integrate AI at enterprise scale responsibly, securely, and in a way that earns lasting customer trust. Cynthia shares her unconventional journey from growing up in Manila and working her first job at a library at age 12, to earning a computer science degree from MIT, building her career at Microsoft, running Ada Developers Academy, and ultimately leading engineering at Smartsheet through one of its most consequential chapters, including the company's transition from public to private and the rollout of its first generation of AI-powered features. The conversation dives deep into how Smartsheet approached AI integration: using generative AI to simplify formula generation and data visualization, being deliberate about what information was and wasn't sent to LLMs, and communicating transparently with enterprise customers who needed to trust the system before they would adopt it. Cynthia explains why trust, governance, and data classification aren't afterthoughts, they're the foundation that makes AI deployment possible at scale. Boaz and Cynthia explore the emerging role of MCP in connecting LLMs like Claude to applications like Smartsheet, translating user intent into real-world action across complex workflows. They also get into what SaaS executives often underestimate: that shipping AI features is the easy part, and evolving the rest of the organization, pricing, enablement, customer support, and role definitions is where companies get stuck. The discussion turns to the next generation of workers and the genuine tension young people face between learning a craft and leaning on AI to accelerate it. Cynthia shares a perspective on hustle, curiosity, and what it looks like when someone who's never written a line of code builds an inventory system for vintage clothing because a tool like Claude made it possible.This episode is essential listening for CTOs, engineering leaders, and product executives who want to understand what responsible AI deployment actually looks like inside a company operating at scale.