Inside Block's Bet on Multiplayer AI | Brad Axen, Head of AI Capabilities

Brad Axen, Block's Head of AI Capabilities and the original author of Goose, joins Liam to talk about what actually makes AI useful at work. They cover how Block went from building an early open-source AI agent to MoneyBot, ManagerBot, BuilderBot and Buzz, and why Brad thinks the hardest problems now are memory, access and interface, not just model intelligence. Brad also explains why AI memory should belong to the business rather than a single bot, what ants can teach us about shared memory systems, why "meat proxy" is becoming a new office problem, and how Buzz is testing a multiplayer model where humans and AI agents work in the same space. Key Topics Covered Goose, Block's open-source AI agent, and the Agentic AI Foundation Agents vs. harnesses vs. interfaces MoneyBot, ManagerBot and BuilderBot Why AI memory should belong to the business, not the bot Stigmergy and ants as a model for shared memory Buzz and Block's "multiplayer" approach to AI at work "Meat proxy" and the new office busywork AI can create Why the bottleneck is shifting from writing code to deciding what to build How AI is changing hiring, interviews and day-to-day work The human cost of spending all day working with AI Episode Timestamps 00:00 Intro 00:07 What Block actually is 02:40 From CERN to Block 05:19 Building Goose and taking it open source 06:29 Agents vs. harnesses vs. interfaces 10:28 MoneyBot, ManagerBot and BuilderBot 15:56 Memory, access and learning over time 22:53 What ants can teach us about AI memory 26:48 Two versions of where AI could go 28:26 Buzz and the idea of multiplayer AI 29:56 "Meat proxy": the new office problem 35:09 The Buzz case study and a 50% productivity jump 37:06 The new bottleneck now that AI can write the code 49:11 Rebuilding institutional knowledge after team restructuring 52:00 How AI is changing hiring and interviews 57:35 The loneliness of working with AI all day 59:13 Why Brad does what he does Connect with Brad on LinkedIn: https://www.linkedin.com/in/bradleyaxen/ Partner Links Upgrade your AI toolkit: https://www.theaireport.ai/ai-executive-pass Subscribe to our free newsletter: https://newsletter.theaireport.ai/subscribe Join the community: https://community.theaireport.ai/checkout/the-ai-report-welcome-gift?coupon_code=WRTH Learn more about your ad choices. Visit megaphone.fm/adchoices

In this episode, Loni Stark, VP of Strategy and Product at Adobe, joins Liam to talk about what happens when a 25 year tech career runs alongside a full creative practice in painting, sculpture and writing, and what that split brain teaches her about building for the AI era. Loni explains why she thinks brands may already be invisible, or worse, misrepresented, inside AI answers, why every company needs to start treating AI as a new kind of audience, and how she is running her own home AI lab, complete with a self built server and a personal agent that has now run continuously for over 150 days, to understand what actually gives an AI agent an identity. Along the way, Loni and Liam get into her Harvard Extension School research into "orphan values," the personal values people can't express in any of their current life roles, and what happens to that alignment as AI reshapes the roles themselves. She also breaks down how she balances Adobe's biggest enterprise bets, including Experience Manager, Commerce, Brand Concierge and LLM Optimizer, against the need to experiment without limits in her own time. Key Topics Covered Why Loni keeps a full art practice, painting, sculpture and writing, alongside her tech career Growing up with parents who didn't understand the arts, and using creativity as a form of rebellion Whether humans are innately creative, and why AI makes protecting your own voice more important Why Loni has stayed at Adobe for 25 years, and how she thinks about "growing the aquarium" Building a personal AI server at home instead of a garden, and what that setup actually involves Swapping the underlying model and the agent harness to test what gives an AI agent a persistent identity Her Harvard Extension School research into "orphan values" and how AI is reshaping the roles we express them through The shift from human mediated to AI mediated experiences, and why that changes what "traffic" even means Why being invisible to AI isn't the worst case, being misrepresented by it is How brands should start preparing their content and catalogs to be "agent ready" The placebo effect of working with agents, and how that belief shapes performance and creativity How Loni balances limitless experimentation with the governance enterprise AI actually requires Why she does what she does: an insatiable need to grow, create and become more than she currently is Episode Timestamps 00:00 - Introduction and welcome 00:05 - Balancing a full art practice with a 25 year tech career 05:38 - Why she's stayed at Adobe for 25 years 08:30 - AI as the biggest creativity enabler she's seen 12:50 - Inside her home AI lab: hardware, memory, and swapping the agent harness 18:11 - Studying psychology at Harvard, and what "orphan values" mean 26:28 - The shift to AI mediated business, and why invisible isn't the worst case 31:46 - How brands become "agent ready" for the AI agent economy 36:23 - The placebo effect of working with AI agents 37:35 - Balancing limitless experimentation with enterprise governance 47:05 - Why Loni does what she does 49:14 - Where to find Loni and closing thoughts Loni's Socials: LinkedIn - https://www.linkedin.com/in/lonistark/ Loni’s Art Gallery: https://atelierstark.com/work/ Partner Links Upgrade your AI toolkit: https://www.theaireport.ai/ai-executive-pass Subscribe to our free newsletter: https://newsletter.theaireport.ai/subscribe Join the community: https://community.theaireport.ai/checkout/the-ai-report-welcome-gift?coupon_code=WRTH Learn more about your ad choices. Visit megaphone.fm/adchoices

In this episode, Braden Warwick, Financial Planning Product Architect at PWL Capital, breaks down why so much of the financial advice sold at big banks is a sales pitch dressed up as a plan, and what a real financial plan actually requires. Braden traded a PhD in aerospace engineering for a career rebuilding how Canadians plan their money, and he brings that same engineering mindset to financial planning: define your objectives, map your constraints, then solve for the outcome that actually improves your life. Braden also walks Liam through the AI infrastructure PWL has built in house, from a proprietary data lake to an AI powered meeting note tool and planning summaries, and explains why they chose to build their own tools instead of buying off the shelf software. They get into Monte Carlo simulations, why financial planning is really about the distribution of outcomes rather than one predicted path, and what a financial planning engagement might look like in 2031. Key Topics Covered How Braden went from a PhD in aerospace acoustics to building financial planning tools at PWL Capital Why PWL's advisors are paid for advice, not for selling products, and how that changes the plan you get The six areas of a real financial plan: investing, cash flow, tax, insurance, retirement, and estate Treating a financial plan like an engineering problem: objectives, variables, and constraints Why Monte Carlo simulations model financial planning as a distribution of outcomes, not one fixed path What forms of uncertainty most financial software still misses, from real estate values to life expectancy Why PWL built its own AI meeting note tool and data lake instead of buying an off the shelf solution How AI is helping PWL's advisors scale personalized, evidence based financial plans PWL's acquisition by One Digital and what it changed, and did not change, about how Braden works What a financial planning engagement could look like by 2031 Episode Timestamps 00:00 - Introduction 00:40 - From aerospace engineering to financial planning 03:54 - Why PWL approaches financial advice differently 07:31 - The six areas of a real financial plan 11:48 - Financial planning as an engineering problem 17:56 - The psychology behind financial planning 23:14 - Objectives, constraints, and uncertainty 28:10 - How Monte Carlo simulations work 33:21 - What financial planning software still misses 39:11 - Building financial planning tools at PWL 44:16 - Inside PWL's financial planning system 51:38 - How AI is changing the advisor workflow 57:20 - Why PWL built its own AI tools and data infrastructure 1:03:41 - What changed after the OneDigital acquisition 1:06:34 - The future of financial planning 1:11:47 - Why Braden does what he does Braden's Socials: LinkedIn - https://www.linkedin.com/in/braden-warwick-a40b48a3/ Resources Mentioned: Braden’s article, The Optimal Financial Plan - https://pwlcapital.com/the-optimal-financial-plan/ Partner Links Upgrade your AI toolkit: https://www.theaireport.ai/ai-executive-pass Subscribe to our free newsletter: https://newsletter.theaireport.ai/subscribe Join the community: https://community.theaireport.ai/checkout/the-ai-report-welcome-gift?coupon_code=WRTH Learn more about your ad choices. Visit megaphone.fm/adchoices

In this episode, Andrew McNamara, VP of Applied ML at Shopify, returns to unpack how much has changed in agentic commerce since his last episode. Andrew and Liam dig into why agents are becoming "the new front door to commerce," why orders coming to Shopify stores from AI are up 13x, and what's actually happening inside Shopify's personalized shopping agent in the Shop app. They also get into the Universal Commerce Protocol (UCP) and why AI commerce is growing 9x faster than social commerce did at the same stage, how Sidekick's architecture and app extensions work, and SimGym, Shopify's system for training AI shoppers to A/B test store changes before they ever reach a real customer. Key Topics Covered How shopping is shifting from stores and desktops toward agents as "the new front door to commerce" Why orders coming to Shopify stores from AI are up 13x, and why catalog-powered AI search converts twice as well as general AI search Inside Shop app's personalized shopping agent, and how it learns different shopping personas (like shopping for a pet versus a child) Why customers are shifting from keyword searches to natural language queries, and the higher conversion rates that come with it Why Shopify keeps shopping data personalized to the individual user rather than training it into a larger internal model What the Universal Commerce Protocol (UCP) is, and why AI commerce is growing 9x faster than social commerce and 3x faster than mobile did at the same stage The story of Shopify's CEO giving his own Hermes agent a budget so it can send him gifts in the mail Sidekick's app extensions, and how partners like Klaviyo and Loop plugged in at launch Campaign Autopilot's "auto research loop," and its parallels to reinforcement learning SimGym, and how Shopify trains AI shoppers to A/B test store changes before running them on real customers Why Sidekick runs on Anthropic's Sonnet model hosted on Google Cloud, and why that choice is model agnostic Andrew's own habit of shopping by taking pictures throughout the week and searching by image through UCP-connected agents Episode Timestamps: 00:00 - Introduction and welcome 00:29 - What's changed in AI and shopping since their last conversation 01:47 - Agents becoming "the new front door to commerce" 04:16 - Inside Shop app's personalized shopping agent 07:32 - Why data stays personalized to each shopper instead of training a larger model 11:53 - What the Universal Commerce Protocol (UCP) is, and orders from AI up 13x 14:58 - Merchant tooling for tracking AI-driven traffic and conversions 15:55 - The story of Tobi's Hermes agent sending him gifts in the mail 20:48 - Andrew's own habit of shopping by taking pictures and searching by image 26:59 - Sidekick's app extensions and partner integrations 33:02 - Inside Sidekick's architecture: the Sonnet model and knowledge base 35:18 - Campaign Autopilot's auto research loop 38:58 - SimGym: training AI shoppers to test store changes 42:23 - What's next for Shopify's agentic commerce features 44:17 - Where to find Andrew Andrew's Socials: Twitter (X) - https://x.com/DrewCH LinkedIn: https://www.linkedin.com/in/andrewmcnamara1/ Partner Links Upgrade your AI toolkit: https://www.theaireport.ai/ai-executive-pass Subscribe to our free newsletter: https://newsletter.theaireport.ai/subscribe Join the community: https://community.theaireport.ai/checkout/the-ai-report-welcome-gift?coupon_code=WRTH Learn more about your ad choices. Visit megaphone.fm/adchoices

David McIntosh, Chief Connected Stores Officer at Instacart, joins Liam to explain why the company is betting on smart shopping carts instead of rewiring stores with ceiling cameras. David walks through the $350 million acquisition of Caper, how Instacart is now live in more than 100 cities with thousands of connected carts, and why the screen on the cart, not the checkout speed, turned out to be the real driver of sales lift for retailers. David also gets into the surprisingly hard engineering problems behind a smart cart, like figuring out whether a basket is actually empty, fusing camera and scale data in real time, and building recommendations that know exactly where a shopper is standing in the store. He and Liam talk about who owns all that shopping data, what agentic AI looks like when it moves from chat into the aisle with tools like Cart Assistant, and why grocery budgets and meal planning are becoming one of the most requested AI features in the store. Key Topics Covered Why David left Tenor, the GIF search engine used by billions, to build Instacart's Connected Store business The strategic bet behind unifying online and in-store grocery shopping Why Instacart acquired Caper for $350 million instead of building smart carts in-house The reason Instacart chose carts over ceiling cameras for in-store AI How a simple running total and real-time coupons drive measurable sales lift The NVIDIA Jetson hardware and multimodal sensor fusion that let the cart "see" what's in the basket The strange edge cases in physical AI, like why "is this cart empty" is a genuinely hard question Who owns retailer and shopper data, and how it's used to improve recommendations Cart Assistant: how Instacart lets customers shop inside ChatGPT and directly on retailer websites Using agentic AI to fix store operations like out-of-stock items and supplier issues How budget-conscious meal planning became one of the most requested AI features in the store David's answer to Liam's closing question: why he does what he does Episode Timestamps 00:00 - Introduction and welcome 00:14 - David's path from Tenor to Instacart's Connected Store 02:01 - The bigger bet behind bringing online and in-store shopping together 05:29 - Entering the smart cart market and acquiring Caper 08:07 - Caper's scale today: 100+ cities and millions of daily sensor inputs 10:28 - How the smart cart actually drives sales lift 12:49 - Why Instacart bet on carts instead of ceiling cameras 17:33 - The unglamorous detail that makes or breaks adoption: charging 19:26 - What makes the experience sticky enough to keep customers coming back 24:41 - Inside the hardware: NVIDIA Jetson and multimodal sensor fusion 28:43 - The strange edge case behind a seemingly simple question 35:14 - Who owns the shopping data, and how retailers use it 37:30 - Agentic shopping: Cart Assistant and buying inside ChatGPT 42:16 - Using agentic AI to fix store operations, not just shopping 46:17 - Why David does what he does Connect with David on LinkedIn: LinkedIn: https://www.linkedin.com/in/mcintoshdavid/ Partner Links Upgrade your AI toolkit: https://www.theaireport.ai/ai-executive-pass Subscribe to our free newsletter: https://newsletter.theaireport.ai/subscribe Join the community: https://community.theaireport.ai/checkout/the-ai-report-welcome-gift?coupon_code=WRTH Learn more about your ad choices. Visit megaphone.fm/adchoices

In this episode, Zachary Smith, CEO and co-founder of Datum and previously the founder of Packet (acquired by Equinix for $335M) and Voxel (acquired for $35M), joins Liam to explain why the internet is about to undergo its biggest transformation since the cloud. As AI agents, vibe coding, and thousands of new applications flood the web, Zac believes the open internet model we've relied on for decades is breaking down. Zac argues that every person, every company, and eventually every AI agent will need its own private network. He explains why the future internet may look more like the Visa network than today's public web, how digital sovereignty and geopolitics are reshaping infrastructure, and why developers are increasingly relying on dozens of cloud services rather than just the hyperscalers. The conversation also dives into Zac's unlikely journey from Juilliard-trained musician to building and exiting two infrastructure companies, the emotional toll of entrepreneurship, and why he keeps coming back to startups despite already having financial freedom. Key Topics Covered Zach's journey from Juilliard and classical music to building infrastructure companies Building Voxel and selling the company for $35M Starting Packet and its $335M acquisition by Equinix Why AI agents are creating a security problem for the internet Why every person and company may eventually need a private network The difference between the public internet and private internet Why the future internet could resemble the Visa network Digital sovereignty, geopolitics, and the splintering of the internet Why developers increasingly rely on dozens of cloud providers How AI is turning millions of people into software developers APIs, MCP, and the next phase of application architecture Why Zach believes AI agents should only talk to approved systems Open source, network effects, and Datum's long-term vision The emotional side of entrepreneurship and why community matters more than money Episode Timestamps 00:00 Introduction and welcome 00:06 Zach's background: from Juilliard and classical bass to startups 02:44 Building Voxel and the early cloud era 08:44 Starting Packet, raising capital, and the Equinix acquisition 15:28 Why taking time off helped him dream again 18:02 What Datum does and the idea of a network cloud 19:38 Three forces changing the internet 20:41 Hyperscalers explained: Amazon, Google, and Microsoft 24:52 Why new cloud providers are emerging 27:16 Digital sovereignty and the fragmentation of the internet 32:03 Public internet vs. private internet 32:54 Inside the physical "meet me rooms" that connect the internet 39:49 How internet routing actually works 45:56 Why developers use so many cloud providers 48:10 APIs, MCP, and AI agents 51:07 Why the future internet may resemble the Visa network 54:23 Who Datum's customers are, and why Datum is open source 1:03:07 AI agents and the next generation of software 1:07:50 Why Zach keeps building companies, and why he does what he does Connect with Zac: LinkedIn: https://www.linkedin.com/in/zsmith/ Website: https://www.datum.net/ Partner Links Upgrade your AI toolkit: https://www.theaireport.ai/ai-executive-pass Subscribe to our free newsletter: https://newsletter.theaireport.ai/subscribe Join the community: https://community.theaireport.ai/checkout/the-ai-report-welcome-gift?coupon_code=WRTH Learn more about your ad choices. Visit megaphone.fm/adchoices

Vinay Gidwaney is Chief Product Officer and Mike Sullivan is Co-Founder and Chief Growth Officer of OneDigital, a 6,000-person, PE-backed benefits, HR, and wealth consultancy serving roughly 100,000 employers. Their contrarian bet: AI transformation has almost nothing to do with technology and everything to do with treating AI as talent. Instead of automating tasks, OneDigital built an internal hiring pipeline for AI, complete with job descriptions, an intern-to-apprentice-to-full-time promotion path, and performance improvement plans, and used it to avoid the layoffs most "AI transformation" playbooks assume are inevitable. Liam sits down with both of them to unpack the night Mike built a "disruption calculator" that showed OneDigital was on track to cut 1,800 of its 6,000 jobs, and how that all-nighter became the catalyst for a different strategy. They get into Ben, the AI coworker now handling daily conversations with 1,600 benefit consultants, why Vinay cites a claim from Lemonade's CEO that AI agents scored higher on customer empathy than human call center staff, the risk of companies "renting back" their own intelligence after gutting their workforce, and the thinking behind their upcoming book, Workforce Intelligence, releasing August 25th. Key Topics Covered Why treating AI adoption like rolling out a new CRM guarantees failure The failed early bet on automating RFPs, and why augmenting human thinking won instead The stat that proves AI adoption is personal: a manager's own AI use doubles their team's usage Building an AI "disruption calculator" overnight, and the 1,800-job number it produced OneDigital's three-part AI framework: coworkers, builders, and agents How OneDigital literally hires its AI: job descriptions, interns, apprenticeships, and performance improvement plans Meet Ben, the AI coworker fielding daily conversations with 1,600 benefits consultants Why they stay LLM-agnostic and separate the intelligence layer from the model and the harness The real risk behind AI cost metering: losing access to the intelligence your company now depends on A claim from Lemonade's CEO that AI agents beat humans on customer empathy Irreducible vs. reducible skills: how to decide what AI should do and what humans should keep The case for "faces, not headcount" and OneDigital's internal "humanity test" What's actually inside their upcoming book, Workforce Intelligence Episode Timestamps 00:00 Introduction 00:41 The thesis: AI transformation is about talent and leadership, not technology 02:44 Why treating AI adoption like a CRM rollout fails 04:29 Pitching VCs a "Workday for AI agents," and why it flopped 07:19 Why automating tasks failed, and augmenting human thinking won 11:20 The stat that changed everything: manager AI usage doubles team usage 15:04 Mike's personal epiphany and the "Claude" nickname from his family 18:10 Building an overnight "disruption calculator" with Claude and Replit 19:56 The result: a model showing 1,800 of 6,000 jobs at risk if nothing changes 21:11 Coworkers, builders, and agents: OneDigital's three-part AI framework 24:02 The AI hiring pipeline: job descriptions, interns, and apprenticeships 27:39 Meet Ben: the AI coworker now used by 1,600 benefits consultants daily 31:10 What's actually deployed: an LLM-agnostic stack with a separate intelligence layer 35:38 The metering and access risk: the Fable/White House security scare 39:56 Measuring "workforce intelligence": blending human and AI capability 50:07 Irreducible vs. reducible skills, and the Lemonade insurance AI-empathy example 53:38 The reskilling problem, and why human judgment stays irreplaceable 56:30 "Faces vs. headcount" and the company's internal "humanity test" 1:07:08 The origin of their book, Workforce Intelligence (out August 25th) 1:11:10 The risk of "renting back" your own intelligence after cutting your best people Learn more about OneDigital and the book Workforce Intelligence: https://www.onedigital.com/ Connect with Mike on LinkedIn: https://www.linkedin.com/in/mikesullivanatdigital/ Connect with Vinay on LinkedIn: https://www.linkedin.com/in/gidwaney/ Partner Links Upgrade your AI toolkit: https://www.theaireport.ai/ai-executive-pass Subscribe to our free newsletter: https://newsletter.theaireport.ai/subscribe Join the community: https://community.theaireport.ai/checkout/the-ai-report-welcome-gift?coupon_code=WRTH Learn more about your ad choices. Visit megaphone.fm/adchoices

Rana Gujral is the former CEO of Behavioral Signals and the author of the upcoming book The AI Instinct. During his time leading Behavioral Signals, Rana has led a company built on a contrarian bet: that the words in a conversation are the least interesting part of it, and that the real signal, intent, trust, stress, deception, lives in how something is said rather than what is said. His team has mapped roughly 75 to 100 behavioral dimensions in the human voice, work that now powers everything from deepfake detection for government agencies to a matching system that pairs call center customers with the agents they are most likely to have a natural, flowing conversation with. Liam and Rana dig into the unconscious vocal tells we all give off, why pitch compression, not raised volume, is the real signature of suppressed stress, and how studying voice for eight years changed the way Rana himself talks and listens. They also get into the ethics of emotion AI, including why the EU has banned it from workplaces, and the central idea behind Rana's book: that large language models are missing an entire axis called experience. Rana introduces his concept of Artificial General Experience, or AGE, and makes the case that the real fork in the road for AI isn't intelligence versus replacement, it's whether these systems make us more ourselves or less. Key Topics Covered Why Behavioral Signals bet on voice as a behavioral signal instead of a language signal The 75 to 100 dimensions of emotion, intent, and cognitive state hidden in a voice How deepfake detection works when a synthetic voice is good enough to fool a mother Surprising commercial uses of voice AI in marketing, call centers, and fraud detection The science of compatibility: why some conversations click and others feel like effort What 8 years of studying voice changed about how Rana communicates The unconscious vocal tells that give away hesitation and suppressed stress Why the EU banned emotion AI in workplaces, and where Rana draws his own ethical line The idea behind Rana's book, The AI Instinct, and why AGI is the wrong question to ask Artificial General Experience (AGE): the missing piece between intelligence and judgment Where augmentation ends and replacement begins, from GPS to Neuralink The geopolitical inequality of who gets access to the most capable AI tools Episode Timestamps 00:00 - Introduction 00:09 - The contrarian bet: voice as a behavioral signal, not a language signal 02:51 - Mapping 75 to 100 dimensions of emotion, intent, and cognitive state 06:07 - Commercial uses beyond law enforcement: marketing, call centers, fraud detection 10:35 - The science of compatibility and conversational entrainment 13:35 - Beyond voice: body language, physiology, and why voice is the primary channel 16:42 - How 8 years of studying voice changed Rana's own communication 19:32 - The unconscious vocal tells everyone gives off 23:40 - Pitch compression: the real signature of suppressed stress 23:53 - The ethics of emotion AI: EU AI Act, modulation vs manipulation 27:08 - Why Rana wrote The AI Instinct 30:04 - Artificial General Experience (AGE) and what LLMs are missing 33:59 - How lived experience dynamically updates memory and meaning 38:04 - Why the goal isn't to build machines that are more human 40:50 - Augmentation vs replacement: where the line gets drawn 43:55 - Purpose, meaning, and the risk of frictionless cognition 51:23 - What we should be teaching the next generation 56:01 - The geopolitical inequality of AI access 59:11 - What Rana hopes readers take from The AI Instinct 1:02:32 - Where to find Rana and the book Rana's website: https://ranagujral.com/ Rana’s LinkedIn: https://www.linkedin.com/in/ranagujral/ Partner Links Upgrade your AI toolkit: https://www.theaireport.ai/ai-executive-pass Subscribe to our free newsletter: https://newsletter.theaireport.ai/subscribe Join the community: https://community.theaireport.ai/checkout/the-ai-report-welcome-gift?coupon_code=WRTH Learn more about your ad choices. Visit megaphone.fm/adchoices

In this episode, Tudor Achim, Co-Founder and CEO of Harmonic, the AI lab behind Aristotle, a mathematical reasoning system that won gold at the International Math Olympiad, makes the case that AI hallucinations aren't the problem with today's models. The real problem is that nobody can verify whether a hallucination is right or wrong. Tudor explains why his team bakes formal, computer-checkable verification (using a language called Lean) directly into how Aristotle reasons, so instead of trusting an AI's word, you can mathematically prove it's correct. Liam and Tudor go deep on what "truth" actually means in mathematics versus the real world, why Andrew Wiles's famous proof of Fermat's Last Theorem had a two-year hidden flaw, and why Tudor believes math is in the middle of its first fundamental shift in 4,000 years, moving from proofs written in English to proofs written in verifiable code. They also get into a spirited debate about the U.S. education system, what Harmonic actually looks for when hiring (hint: it's not the résumé), and why Tudor thinks AI will never be trusted to grade its own homework. Key Topics Covered What "truth" means in mathematics versus science, and why logical reasoning is really just a simple form of math Why the proof of Fermat's Last Theorem had a hidden flaw for two years, even after being announced Why hallucinations are actually necessary for AI reasoning, and what separates a good hallucination from a bad one How Harmonic uses Lean and formal verification to make Aristotle's math proofs checkable step by step, like reviewing code Why Tudor doesn't think any AI will ever be trusted to fully verify its own output Why Harmonic gives away the Aristotle API for free right now, and where the business model is headed The "phase transition" Tudor believes is happening in math for the first time in 4,000 years: from English proofs to machine-verified code Why open-sourcing formal math matters more to Harmonic than keeping a competitive edge Harmonic's five-year goal: contributing to solving a Millennium Prize Problem by 2028 A debate on whether the U.S. education system actually teaches critical thinking, and what AI should (and shouldn't) change about how kids learn What Harmonic actually looks for when hiring, and why résumés carry almost no signal anymore Tudor's take on inequality, taxation, and what a healthy AI-driven economy could look like Episode Timestamps 00:00 Intro and welcome 00:33 What is truth, and is logic just math? 07:59 Why the company is called harmonic.fun 08:48 The real problem with LLMs and truth 12:57 How formal verification works inside Aristotle 16:43 Who else is building in this space 19:48 Can AI ever be verified with 100% accuracy? 23:10 Why AI can't fully verify its own answers 26:46 Open-source math vs a venture-backed business 29:39 The five-year goal: a Millennium Prize Problem by 2028 31:51 Has Harmonic's vision changed? 34:28 AI, formal verification and the future of education 44:34 Debating the US education system 45:09 How Harmonic hires and what they look for 50:21 Testing for trust and honesty in interviews 54:14 Harmonic's biggest challenge right now 58:19 The future of working with AI 59:55 Family, kids and optimism about the future 1:01:38 Inequality, capitalism and AI's role in the economy 1:04:19 Why Tudor does what he does 1:05:19 Where to find Tudor Connect with Tudor on LinkedIn: https://www.linkedin.com/in/tudorachim/ Partner Links Upgrade your AI toolkit: https://www.theaireport.ai/ai-executive-pass Subscribe to our free newsletter: https://newsletter.theaireport.ai/subscribe Join the community: https://community.theaireport.ai/checkout/the-ai-report-welcome-gift?coupon_code=WRTH Learn more about your ad choices. Visit megaphone.fm/adchoices

Every answer an AI gives you sounds equally confident, whether it's true or completely made up. That's not a bug. It's how the technology was built. Dan Klein is CTO and co-founder of Scaled Cognition, and a professor of computer science at UC Berkeley. In this conversation with Liam, Dan breaks down what a language model actually is, why it was never designed to know the truth in the first place, and why today's AI systems have no "smells," the subtle warning signs humans usually rely on to tell good information from bad. They get into why reinforcement learning from human feedback quietly trains models to tell people what they want to hear, how that can tip into outright deception, and why Dan believes reliability, not raw intelligence, is the biggest unsolved problem in AI today. Key Topics Covered: What a language model actually does at its core: next token prediction Why LLMs are plausibility engines, not truth engines The difference between a hallucination and a lie Why AI mistakes have no warning signs the way bad translations or sketchy websites do How RLHF can train models to be sycophantic instead of accurate The "package delivery" thought experiment: when reward signals diverge from truth Why bolting reliability onto LLMs after the fact doesn't work How Scaled Cognition architects models around verified actions instead of raw text generation Why bigger models aren't automatically better models The difference between disruptive technology and scaled technology Why startups, not incumbents, tend to drive technical breakthroughs What metacognition is and why today's AI systems don't have it Why Dan believes reliability is the next major frontier in AI Episode Timestamps: 00:00 Intro 00:15 What a language model actually is 06:31 From well-formed sentences to general knowledge 08:27 Why LLMs are plausibility engines, not truth engines 12:06 How Perplexity approaches verifiable answers 12:40 Dan's background and Scaled Cognition's mission 15:16 The two anti-patterns companies use to control LLMs today 21:16 How Scaled Cognition architects models differently 23:28 Does every client need a custom-trained model? 29:12 Why prompting alone can't guarantee reliability 30:55 Modularity, contracts, and building reliable systems 34:40 Why trust and digital literacy matter beyond the enterprise 39:12 Code smells and why AI mistakes have no warning signs 41:14 Are AI companies incentivized to tell the truth? 42:55 How reinforcement learning actually works 44:35 The package delivery thought experiment 48:44 Why models are trained to be sycophantic 51:01 Where this incentive is mechanically baked into the model 53:43 Does responsibility fall back on humans? 58:10 Just be more reliable than a human, not perfectly true 1:02:59 The last major technique shift in AI 1:10:55 Why frontier labs keep scaling despite the risk of disruption 1:17:15 The future of hyper-specialized models vs. one broad model 1:19:47 Is there anything uniquely human AI can't replicate? 1:25:45 Wearing three hats: professor, researcher, and CTO 1:29:47 Why Dan does what he does Connect with Dan on LinkedIn:https://www.linkedin.com/in/dan-klein/ Partner Links Upgrade your AI toolkit: https://www.theaireport.ai/ai-executive-pass Subscribe to our free newsletter: https://newsletter.theaireport.ai/subscribe Join the community: https://community.theaireport.ai/checkout/the-ai-report-welcome-gift?coupon_code=WRTH Learn more about your ad choices. Visit megaphone.fm/adchoices