AI 2030

Laura Stapleton joined Foley as Chief People Officer with a mandate she had never been handed: lead an AI transformation, not just run HR. Her answer is stewardship, assigning each agent to whoever owns that domain, regardless of title, and writing it into job descriptions.

She walks host Keith Jensen through Foley's Access, Usage, Impact maturity curve, why a shared ontology decides whether agents work, and how AI drafted performance reviews in Lattice that people did not hate. Her through-line: jobs change, headcount logic breaks, accountability stays human.

Topics discussed:

• Writing agent stewardship into formal job descriptions

• Pushing past AI access and usage to real impact

• Creating a context-enablement role that PMs AI context

• Building a shared ontology so humans and agents align

• Drafting performance reviews in Lattice without the usual dread

• Running a Value Creation Office for in-work AI support

• Setting the hiring gates humans, not agents, decide

• Protecting domain expertise as org chart layers collapse

Devi Parikh ran generative AI as a senior director at Meta for eight years before leaving to build Yutori, where she's making AI agents proactive rather than reactive. The design choice that stands out: the first product, Scouts, asks for zero access to your context and only reads the public web. Trust gets earned before an agent touches your money, your logins, or your calendar.

Devi takes Keith through Yutori's "trust staircase" and why read-only was the deliberate starting point, not a limitation. She details the near-term shift from notifying you to completing tasks, where you sandbox specific sites behind auth walls one at a time instead of handing over a browser extension with access to every logged-in state. For the long tail of the web with no API or MCP server, Scouts spin up a remote browser and use a browser-use model Yutori trained in-house to click through and complete the task.

Topics discussed:

• Yutori's "trust staircase" for earning permission to act

• Read-only as a deliberate trust-building wedge, not a constraint

• Per-site sandboxed auth access versus browser-extension exposure

• In-house browser-use model for the API-less long tail of the web

• "Inspect work" transparency to build trust in agent actions

• Prosumer chief-of-staff spanning personal and work context

• Selling to the employer to reach work context, individual as the user

• Why connected email, Slack, and Notion unlock proactivity without always-on listening

Most AI apps treat ads as a tax on the product. Nic Baird, co-founder of Koah Labs, is betting the opposite: that at $2-3 per thousand queries, an AI app that refuses ads either paywalls its best features or dies. He tells Keith why Koah only serves an ad when it's unit-economic positive against token cost, and why relevance is the only defensible moat in this category.

Nic goes into the mechanics: how Koah generates ad copy from product data (with static images coming from the advertiser when brand control matters), why standard click attribution breaks when nobody transacts inside a chat interface, and the brand-lift and topic-lift proxies he's testing instead. He also lays out where he thinks websites go next, and why a personal agent handing off context between sites is the protocol someone needs to build.

Topics discussed:

• The $2-3 per thousand queries cost floor forcing AI apps to paywall features

• Serving ads only when unit-economic positive against token cost

• 99% approval when a publisher polled users about dropping $5 subscriptions for ads

• Character AI's interstitial video ad backlash as the anti-pattern

• Generating ad copy from product data, with static images from the advertiser when brand control matters

• Mobile CTR at 3-5%, spiking to 14-15% in early tests

• Extracting intent from emotional context, not stated purchase signals

• Brand lift and topic lift as substitutes for direct response attribution

• The rabbit hole ad format that enriches first-party retargeting cohorts

• Subsidizing marketplace liquidity Uber-style with the $20M raise

• Personal agents handing off user context as the missing web protocol

What does it take to build an AI consultancy that's tracking for 3x revenue growth in year two, with a founder who thinks he's got a ten-year head start before anyone can catch up?

Chad Lohrli, Co-Founder and Chief AI Officer at Cadre AI, walks Keith Jensen through the specific operating mechanics behind that growth: a goal-tracking system built around paper on the wall with plans to add a public Hall of Shame for missed goals, the 8,000-row prompt it took to make an internal chatbot actually reliable, and the exact point in the AI workflow where Chad says things start to go sideways.

Topics discussed:

• Printing goals on paper walls and planning a public Hall of Shame

• A 200% jump in hours tracked from visibility alone

• An 8,000-row JSON prompt that made an HR chatbot accurate

• Why Cadre treats every prompt as an SOP, not a shortcut

• A code knowledge base keeping AI-built client products consistent

• Why reviewing AI output is the real bottleneck now

• The ten-year model-freeze math behind Cadre's growth bet

• The contract mistake that taught Cadre where AI shouldn't write

Most attribution setups reward the click and the form fill. Alex Lirtsman, founder and CEO of CorralData, built his platform to reward the person who actually shows up and pays, then feed that signal back to the ad platforms so bidding optimizes toward revenue instead of cheap leads.

He and Keith get specific: how Corral sits between a company's data and its LLM of choice as a governance and semantic layer, so permissions set once carry into Claude or ChatGPT, and every metric answers to the company's own definition. Now serving 120 brands, Alex lays out why he thinks the software moat disappears by 2030, leaving proprietary data and service quality as the only defensible advantages.

Topics discussed:

• Why automated bidding rewards clickers, not buyers

• Feeding offline conversions back to optimize on revenue

• Matching ad click IDs to EMR records for closed-loop attribution

• Setting permissions once, inheriting them into any LLM

• Encoding company-specific definitions of LTV and rebooking rate

• AI interrogating every feature ticket before writing the PR

• Running ad campaigns through Google, Meta, and LinkedIn agents

• Why proprietary data and service become the only 2030 moat

Rohit Sharma spent his first decade in Silicon Valley building a company from two employees to 750 before it got acquired, then landed at True Ventures, where he's stayed for almost fourteen years without ever planning the next move. He tells Chad and Dhruv why he calls AI a new form of labor rather than a technology, why LLMs are Shannon's machines and not magic ones, and why founder vision and distribution, not technical execution, are the only things still scarce.

Topics discussed:

• Reframing AI as a new form of labor, not a technology

• Envisioning 150 customers each getting a critically different product

• Evolving forward deployed engineers into armies of specialized agents

• Explaining why LLMs are Shannon's machines, not magic machines

• Arguing AGI rests on the wrong standard entirely

• Calling the GPU-transformer fit a historical accident, not design

• Predicting the same fate for tokens as transistors and servers

• Locating scarcity in founder vision and distribution, not execution

Inbound stopped being a signal the moment LLMs let candidates fire the same resume at a thousand companies at once. Ishan Gupta, CTO and Co-founder of Juicebox, argues the fix isn't better filtering of that flood but a shift back to outbound, where his agents read what someone has actually done to infer adjacent skills, tech stacks, and likely employers rather than matching keywords. He's just as clear about what Juicebox won't do: it replaces your sourcing software, not your ATS, never emails a candidate's work address, and stops the moment a recruiter takes over the relationship. 

With over 4,000 customers and a fresh $80M round at an $850M valuation, the roadmap is narrow on purpose, going deep on search and a memory layer that learns who actually succeeds at your company instead of building a broad platform.

Topics discussed:

• How inbound volume became noise, pushing hiring outbound

• Inferring adjacent skills and tech stacks over keyword matching

• Bolt-on positioning: replacing sourcing software, not the ATS

• The contact-data waterfall and why work emails stay off-limits

• Protecting sender-domain reputation through narrow targeting

• Internal talent rediscovery: mining your team's existing network

• A company-specific memory layer built on who actually succeeds

• The outbound-first 2030 view of talent acquisition

Most enterprise AI projects don't fail because the AI can't perform. They fail because nobody documented what the AI is actually supposed to do, or how it should behave when things get ambiguous. Binny Gill, former CTO of Nutanix and now CEO of Kognitos, spent five years solving that by making English the programming language for business process automation, with a proprietary interpreter that extracts precision from ambiguity rather than demanding it upfront, the same way a parent figures out what a toddler wants.

His architecture deliberately avoids using LLMs during execution. For deterministic workflows like order-to-cash or vendor onboarding, injecting a creative model at runtime is a liability. The LLM earns its place at design time. He also makes a pointed argument about what AI governance actually requires that most companies are completely unprepared for: with AI, every model is a blank slate, which means the tacit knowledge that's always lived in people's heads has to be explicitly captured or it won't exist. And his 2030 prediction reframes the whole conversation away from technology capability and toward organizational architecture.

Topics discussed:

• Why the English interpreter was always the problem, not the language itself

• Neuro-symbolic architecture: separating creative AI at design time from deterministic execution at runtime

• Why the toddler model of ambiguity resolution beats forcing precise inputs upfront

• Why zero behavioral assumptions with AI models demand a new documentation standard

• How AI autonomously authors tribal knowledge rather than relying on humans to write it down

• The three enterprise buyer personas (CFO, CIO, Chief AI Officer) and what each one is actually evaluating

• Tailor shop vs. assembly line as the framework for assessing organizational AI readiness by 2030

• Why Binny predicts legacy company architecture, not technology, is what fails in the transition

Most VCs are chasing AI deals right now. Will Quist, Partner at Slow Ventures, thinks most of them are chasing the wrong thing. With somewhere around $1B deployed across pre-seed and seed, Will makes a direct case that the shift from client-server to SaaS was more economically disruptive than anything AI is doing to software today, and that building a defensible company in this environment requires the same fundamentals most founders skip.

Will walks through the logic-chain framework Slow Ventures uses to decide whether anything is actually venture-backable: is the hypothesis novel, falsifiable, and objectively valuable if true? He applies that same test to the current AI wave, separating structural arbitrage from deals that are just riding a spread until it closes. He closes with a point almost nobody in the room is making: why selling software is sometimes the worst financial decision a software company can make.

Topics discussed:

• Novel, falsifiable hypothesis framework for venture-backable ideas

• Why AI SaaS is incremental, not transformational disruption

• Structural vs. spotted arbitrage and why only one compounds into enterprise value

• "Leverage is leverage" mental model for evaluating startup valuations

• Valuation gravity: pricing rounds so 24-month milestones are actually achievable

• Fast boil vs. slow boil deals and where early-stage returns actually come from

• Why complex customer relationships and proprietary data sets may outlast pure AI plays

• Growth buyouts and vertical integration as higher free cash flow alternatives to selling software

Francois Chaubard co-created CS224D at Stanford with Richard Socher and others, the course that replaced the entire NLP curriculum two years after launch and still carries his name. He then spent nine years as a founder before closing a $60M enterprise deal at Focal. He is now a Visiting Partner at Y Combinator and a PhD researcher working on alternatives to backpropagation. He is not a generalist with opinions.

In this conversation with Chad, Francois opens with a provocation: in terms of fluid intelligence, humans still dominate. The ARC-AGI 3 benchmark makes this concrete. The best LLMs are 50x less efficient than humans on novel tasks. Nucleus sampling structurally prevents LLMs from ever being funny, not as a solvable limitation but as a mathematical consequence of how the word "pretend" sits at 2-in-1,000 token probability. Stanford's entire compute cluster for 4,000 CS students is 250 H100s, what a single OpenAI new hire gets on day one. Each of these points has an implication, and Francois spells them out.

Topics discussed:

• The Stanford vs. MIT founder model and why novelty is structurally a startup liability

• YC's "agency and taste" funding criteria and what it looks like in a three-week window

• Why vibe coding and Claude Code collapse activation energy for international founders

• GPU-per-student as a research moat metric and the compute gap between academia and industry

• Why in-context learning does not monotonically improve and what that means for agents

• The batch size one problem: why humans learn stably from a single example and models cannot

• The token probability argument for why nucleus sampling structurally prevents LLM humor

• Cartridges: learning a compressed KV cache from random initialization using SGD, with the model frozen

• Auto Research with good ideas: a priority queue over ideas.md ranked by expected value, with a Gemini Pro reflection loop

• EVO's DNA base-pair prediction model: capped at 80B parameters due to funding, and what scaling to 800B might mean