Stripe

Barney Hussey-Yeo in conversation with John Collison

4919 summary words 22 min summary Watch video

Start with the signal

22 min read

Summary

At-a-Glance

  • Verdict: Watch fully
  • Core thesis: Deep vertical AI assistants that own full product stacks and take autonomous actions will beat frontier-model wrappers; organizational structure must flatten radically as LLMs enable CEOs to bypass middle management via direct code/Slack synthesis
  • Why it matters: Barney runs a £400M ARR AI fintech, describes dramatic operational changes from LLM tooling (flatter orgs, direct visibility into all PRs/Slack), and argues vertical depth + agentic action >> generic chatbots—directly relevant to Ken's agent systems and AI ops thesis
  • Best use: Study organizational redesign tactics (LLM-powered captain's log aggregating PRs/Slack/Notion), understand why proactive agentic products beat reactive chat UIs, extract contrarian takes on SaaS durability and venture capital irrationality

Executive Summary

Barney Hussey-Yeo, founder/CEO of Cleo (£400M ARR AI financial advisor, 7M+ users, Shoreditch-based unicorn), sat down with Stripe co-founder John Collison at a London event. Barney's background: funded university via online poker (pre-GTO solvers, 2012–14), now runs an AI agent that earns ~£200K/year in rake back at low stakes. He studied ML at master's level in 2016 when NLP was emerging but pre-Transformer, founding Cleo as an 'AI assistant for your money' when that concept was weird. The thesis: billions of people make wildly suboptimal financial decisions daily (60% of Americans have <$400 savings); an always-on, proactive agent can optimize spending, borrowing, saving, investing—and materially raise GDP by reducing waste (consumer credit interest, poor product selection, tax inefficiency).

Technical evolution: Cleo started with regexes and supervised intent classification (2016–18), then BERT-era NLP, now heavily LLM-based with a personalized 'knowledge base' per user (recursively structured via LLMs, storing goals/health/conversations) plus reinforcement-learning recommender systems optimizing for financial health + engagement. Personality matters hugely: Cleo hired 30+ female comedians to write humorous, approachable copy (finance is high-stakes/stressful for most users), offers a 'roast mode' that makes fun of spending, and tunes tone by income/location/user preference. Barney insists humor + simplicity is product contrarianism against boring/opaque incumbents (cf. Stripe making payments dev-friendly, Monzo making banking transparent).

Monetization: freemium subscription + financial products (credit cards, BNPL, paycheck advance, wealth/investing products). Barney spent first 4 years ignoring revenue to chase MAUs (worked until growth funds demanded monetization), advises founders to either bake one core margin primitive from day one or fully ignore revenue—no half measures. He expects stable coins/frictionless rails to shift subscriptions toward microtransaction models (Chinese-style) long-term. On advice: personalization is critical; generic index-fund advice (MSCI World, low fees) is fine for top 1% but 99% need help with daily spending, income volatility, rent/mortgage decisions, avoiding predatory credit. Cleo measures financial health via 0–100 score; recommender system now optimizes for health improvement + engagement, not just clicks.

Organizational transformation since Jan/Feb 2024: Barney now spends his day fundamentally differently. He uses LLMs (Karpathy-style recursive synthesis) to ingest every PR, Slack message, Notion doc, building a continuously updated 'captain's log' knowledge base. He knows who ships fastest, which engineers have highest code quality, what every team is building—unfiltered by middle management bias. Engineering managers 'hate him' but he argues orgs must flatten: middle layers had incentive-structure distortions; LLMs provide rational aggregation at scale. He envisions average manager span-of-control widening, fewer middle managers, more junior AI-native engineers (UCL/Cambridge ML grads who've used LLMs since A-levels) each managing fleets of agents to take '2,000 bets/year instead of 10.' Apprenticeship model likely returns: one senior + one junior is better than one senior + eight juniors when juniors have agentic leverage.

Key Takeaways

  • Claim: Deep vertical AI products with full stack ownership and autonomous action will beat frontier-model wrappers, contrary to 'SaaS apocalypse' narrative | Evidence: Barney's most contrarian take: 'There's so much craft in building great B2B/B2C products… someone wrote a prompt with a frontier model—wildly overblown. Investors have no technical knowledge, sit on fence, only fund frontier labs + Stripe.' He argues Cleo's 8 years of domain knowledge (financial products, behavioral nudges, regulatory navigation, recommender systems) cannot be replicated by a ChatGPT plugin. Cleo owns cards, lending, BNPL, wealth products; takes actions on behalf of users (moves money, applies for credit). Engagement requires proactive agentic pushes, not reactive chat. | Caveat: Barney admits 'there's a lot to do in financial advice' before expanding to health/education verticals. He's 'very AI-pilled' so struggles to be 'absolutely contrarian.' His critique of investors ('silliest, most irrational people') may reflect fundraising frustrations rather than universal truth—some VCs do have technical depth. | Implication: For Ken's agent systems work: vertical depth + agentic loops + proprietary data moats matter more than model size. Proactive recommendation engines (RL-optimized for long-term outcomes like financial health, not just clicks) are the competitive moat. Frontier labs will struggle to replicate domain-specific workflows without owning the full product stack. This validates building opinionated vertical agents over general-purpose assistants. | Timestamp: 37:30
  • Claim: Financial advice must be proactive and agentic, not reactive chat, because people don't wake up wanting to think about money—it's boring/stressful | Evidence: Barney: 'To engage people with financial advice, it's got to be proactive and agentic… People don't wake up and go, oh, how much did I spend? It's a boring thing. It has to be this agent that's always on, always pushing information… Frontier model companies are very reactive—replacing Google search. There's so much space for proactive, intelligent, always-on agentic things.' He contrasts Cleo (pushes notifications, takes actions autonomously) with ChatGPT/Plaid integration (low engagement, 'side project that goes nowhere'). | Caveat: He doesn't provide quantitative engagement data comparing proactive push vs. reactive query patterns. The claim rests on product intuition + observed ChatGPT/Plaid plugin failure, but no disclosed A/B test results. | Implication: For Ken's content/business/GTM work: users don't self-initiate for 'boring' workflows; agents must interrupt at optimal times. Recommender systems that learn when to surface information (time-of-day, context triggers, user state) are critical. Reactive tools (search, on-demand Q&A) lose to proactive schedulers. This also applies to AI ops: monitoring agents should alert/act before humans ask, not wait for queries. | Timestamp: 13:45
  • Claim: LLM-powered organizational visibility lets CEOs bypass middle management by synthesizing all PRs, Slack messages, Notion docs into a 'captain's log,' flattening hierarchies and increasing bet velocity | Evidence: Barney describes his workflow since Jan/Feb 2024: 'I have LLMs that go over every PR, summarizes what it does, grades it, tells me what's going on, builds a knowledge base. Goes over all Slack messages, all channels, builds up a knowledge base of what is happening. I know who's shipping, highest velocity, highest code quality, which teams are doing great work—unfiltered by leaders with different biases and incentive structures… Engineering managers hate me… Orgs have just become way flatter… much more efficient organizational flow.' He uses Karpathy's recursive-LLM-synthesis pattern (Karpathy tweeted about this a few months ago). He envisions '2,000 bets next year instead of 10.' | Caveat: Barney admits 'junior people need advice, time, nurturing'—pure flattening could hurt mentorship. He doesn't specify which LLM stack (OpenAI, Anthropic, internal fine-tunes) or how he handles privacy/permissions (does he read all employee Slack DMs?). Engineering managers 'hating him' suggests morale/trust risks. He hedges: 'only reason I'm caveating is you do need to give junior people advice.' | Implication: For Ken: This is a playbook for radical transparency + flat orgs at scale. Tools that aggregate disparate signals (code commits, chat, docs, dashboards) into queryable knowledge graphs are high-leverage. Risk: junior talent development suffers if seniors only interact via synthesized reports. Ken should explore whether this works better in engineering (objective metrics: velocity, test coverage) than in subjective domains (design, strategy). Also: privacy/security guardrails are critical—employees may rebel if CEO reads all Slack. | Timestamp: 31:00
  • Claim: UK/Europe must shift from 'conserving legacy' to embracing creative destruction and risk-taking; regulatory caution + centralized power stifles innovation vs. US federalism/checks-and-balances | Evidence: Barney: 'UK regulators: please come to our sandbox, work together for two years—but VC funding cycle is 18 months. Trying to be helpful constrains companies… I've seen angel investments try to engage FCA, run out of money before proving product-market fit. Founders need to take risk early… Read Federalist Papers—checks and balances, state vs. federal. We need more of that in UK. PM/Number 10 has so much power, regulators have complete control yes/no across 60–70M consumers. Doesn't happen in US… Why America is winning vs. Europe: checks/balances, culture of creative destruction vs. conservatism/risk-off.' He argues UK pension capital goes to US VCs (California pension funds LP into UK VCs), so GDP growth returns to California, not Britain. | Caveat: He doesn't engage with counter-argument that US regulatory fragmentation (50 state money-transmitter licenses, conflicting crypto rules) also creates massive compliance burden. His critique may reflect fintech-specific pain points (FCA sandbox friction) rather than universal regulatory failure. He admits 'Stripe took some risk with early clients'—regulatory arbitrage is part of startup playbook everywhere. | Implication: For Ken's investing lens: regulatory risk-taking is a feature, not a bug, of high-growth startups. Founders who wait for regulatory blessing lose momentum. This argues for investing in teams that will 'ask forgiveness, not permission' in ambiguous regulatory zones (AI content, synthetic data, crypto rails). Also: US institutional capital (pensions, endowments) willing to fund venture creates compounding advantage—Europe's pension conservatism is structural GDP drag. Ken should prioritize markets with deep risk capital + federalized checks-and-balances. | Timestamp: 18:30
  • Claim: Financial advice personalization requires income/location/goal-aware tone + humor to overcome money stress; generic 'buy index funds' advice ignores that 60% of Americans have <$400 savings and make irrational daily spending decisions | Evidence: Barney pushes back on John's 'just buy MSCI World, low-fee index fund' advice: 'Showing how out of touch you are… Your life is your money. Most stressful thing for majority of people… 60% of Americans <$400 savings. How they spend every day, income volatility, rent/mortgage, subscriptions—wildly unoptimized. Trillions of GDP left on table.' Cleo tunes tone by income/location ('You're in Boston, San Francisco'), hired 30 female comedians to write every line, offers 'roast mode' that mocks weekend bar spending. 'Humor + money go together weirdly well—no financial institution speaks that way (lawyers/executives).' Measures financial health 0–100; recommender optimizes for health + engagement, not just clicks. | Caveat: He doesn't quantify how much humor increases retention vs. conversion vs. financial outcomes. 'Roast mode' could backfire for users in genuine financial distress (risk of feeling mocked when vulnerable). Personalization based on income/location raises fairness questions (does lower-income user get patronized?). He admits 'humans are irrational'—using behavioral nudges ethically is tricky (cf. 'weaponizing irrationality for good'). | Implication: For Ken's content systems: tone-of-voice personalization is high-leverage in sensitive domains (money, health). Generic LLM outputs (overly formal, risk-averse, boring) lose engagement. Fine-tuning or prompt-engineering for humor/approachability may differentiate consumer AI products. Also: optimizing for long-term outcomes (financial health) vs. short-term engagement (clicks) requires RL with delayed rewards—hard but defensible moat. Ken should explore whether similar dynamics apply to other 'boring but important' verticals (tax, legal, HR compliance). | Timestamp: 10:00
  • Claim: Cleo's organizational shift to AI-native work includes hiring ML-trained grads (UCL, Cambridge) who are 'fully AI-pilled since A-levels' to manage fleets of agents, enabling '2,000 bets/year instead of 10' | Evidence: Barney: 'I hire grads from UCL, Cambridge who did ML masters—fully AI-pilled, been doing it from A-levels. Crazy how good they are with these tools. You want more junior people with learning experience in LLMs, really pilled… managing fleet of agents. In software you want more and more software, directed at right things. Great tastemakers run lots of parallel agents, recursion, more tokens. As many people as possible at lower level with good taste to run as many agents as possible.' He contrasts this with traditional 1:8 senior-to-junior ratio—prefers 1:1 apprenticeship or 1:many-agents. | Caveat: He doesn't define 'good taste' or how to assess it in junior hires. Risk of over-indexing on tool fluency (prompt engineering, agent orchestration) vs. domain expertise or product sense. 'Taking 2,000 bets' could mean low-quality experiments unless there's disciplined prioritization/evaluation. He acknowledges 'you need to nurture junior people'—pure agent delegation may not develop critical thinking. | Implication: For Ken: Hiring strategy should prioritize early AI exposure (university ML courses, hackathons, open-source contributions) over traditional SWE pedigree. Junior talent can have outsized impact if given agent tooling + clear taste criteria (what makes a good experiment?). This also implies Ken's content/research agents should be designed for 'tastemaker' oversight—humans set objectives/evaluate outputs, agents execute permutations. Risk: if everyone takes 2,000 bets, evaluation becomes bottleneck. Need robust automated evaluation harnesses. | Timestamp: 34:00

Detailed Brief

Cleo's Technical Evolution: From Regexes to LLM-Native Architecture

  • Claims: Founded 2016 as 'AI assistant for your money' when AI was weird; Barney did ML master's as NLP was emerging post-AlexNet vision era; Started with regexes, then supervised learning for intent classification + entity extraction; used RL-based recommender systems to push info at right time; Post-2018 Transformer/BERT adoption, then full LLM pivot post-ChatGPT 3.5 ('everyone in CS knew this was coming, everyone else woke up when 3.5 dropped'); Now builds per-user knowledge base (goals, health score 0–100, all conversations) via recursive LLM synthesis; recommender optimizes for financial health + engagement, not just clicks
  • Evidence: Barney: 'Did master's in ML, professors interested in NLP… worked at fintech as data scientist, fascination with ML… everyone's decisions wildly unoptimized, financial life most stressful thing'; Before LLMs: 'Big multi-class supervised learning for intent, entity classification, recommender systems, RL to push you information at right time'; Current: 'Karpathy did recursive LMs thing a few months ago—we've been doing that for years. Build structured knowledge base of every user: financial aspirations, health, goals, every conversation. Recommender thinks from knowledge base, what moves objective function most (financial health)?'; Financial health score: '0 to 100, all these variables/features. Know whether you're leveling up or degrading based on actions. Core primitive: spending less than you earn'
  • Caveats: He doesn't disclose which LLM providers (OpenAI, Anthropic, internal fine-tunes) or model sizes; Recursive synthesis can compound errors if initial extraction is wrong; no mention of human-in-loop validation; Optimizing for long-term financial health (2–3 month lag) via RL is hard—sparse rewards, attribution challenges
  • Implications: Vertical AI products need proprietary user knowledge graphs, not just frontier model API calls; RL-based recommenders that optimize multi-month outcomes (not clicks) are defensible moats but require tons of historical data; For Ken: structured knowledge extraction (not just RAG over flat docs) + time-series outcome tracking are critical infra for agentic products

Product Strategy: Proactive Agentic Nudges > Reactive Chat, Personality = Engagement

  • Claims: Financial advice must be proactive (pushes at right time) because users don't wake up wanting to think about boring/stressful money topics; Personality + humor critical: hired 30 female comedians to write all copy; 'roast mode' makes fun of spending ('cheeky pints'); tone personalized by income/location; Product contrarianism: incumbents are opaque/boring → Cleo is transparent/fun (cf. Stripe made payments dev-friendly, Monzo made banking simple); ChatGPT/Plaid integration will get 'very little engagement, side project that goes nowhere' because it's reactive, not proactive
  • Evidence: Barney: 'Humor and money go together weirdly well—no financial institution speaks that way (lawyers/executives). Real unlock for engagement. Roast mode makes fun of you for spending… people really like that'; On proactive: 'People don't wake up and go, oh how much did I spend? Boring. Has to be agent always on, always pushing… Frontier models very reactive, replacing Google search'; On contrarianism: 'Payments industry archaic/complicated, Stripe made it simple/dev-friendly. Banks bad UX, Monzo made it simpler/transparent. Financial advice complicated/boring, we make it simple/fun. Take inverse position, most extreme form'
  • Caveats: No quantitative engagement lift from humor/roast mode disclosed; could alienate users in genuine distress; Personalized tone by income risks feeling patronizing to lower-income users; Proactive pushes can be annoying if poorly timed (notification fatigue); Barney doesn't explain timing optimization beyond 'RL recommender'
  • Implications: For Ken's GTM/content agents: users won't query 'boring but important' workflows—agents must interrupt optimally; Personality fine-tuning (humor, conciseness, approachability) >> generic LLM tone in consumer products; Timing optimization (when to push) is as important as content (what to push)—RL or learned schedulers critical; Reactive tools (ChatGPT, search) lose in verticals where users avoid the topic; proactive agents win

Organizational Redesign: LLM-Powered Visibility Flattens Hierarchies, Increases Bet Velocity

  • Claims: Since Jan/Feb 2024, Barney uses LLMs to synthesize every PR, Slack message, Notion doc into 'captain's log' knowledge base—knows who ships, code quality, unfiltered by management bias; Engineering managers 'hate him' but orgs must flatten: middle layers distort via bias/incentives; LLMs provide rational aggregation; Envisions fewer middle managers, wider span-of-control; more junior AI-native grads (UCL/Cambridge ML masters, 'AI-pilled since A-levels') each managing agent fleets; Goal: '2,000 bets/year instead of 10' via parallel agent experiments; 1:1 apprenticeship model better than 1:8 when juniors have agentic leverage
  • Evidence: Barney's workflow: 'LLMs go over every PR, summarize, grade, build knowledge base. Go over Slack messages, all channels. I know highest velocity, code quality, which teams doing great work—not filtered through leaders with biases… Engineering managers hate me'; On flattening: 'Middle layer probably hates me. Orgs become way flatter, more efficient flow. I hope less middle managers… Only caveat: junior people need advice, nurturing—always that need'; On juniors: 'Hire UCL/Cambridge grads, ML masters, fully AI-pilled, been doing it from A-levels. So good with tools. Want more junior people with LLM experience, managing fleet of agents… Great tastemakers run parallel agents, recursion, more tokens'
  • Caveats: No detail on privacy/permissions: does he read all employee Slack DMs? Could violate trust/norms; Engineering manager resentment risks morale/retention issues; 'Good taste' undefined—how to assess in junior hires? Risk of over-indexing on tool fluency vs. domain sense; '2,000 bets' only valuable if evaluation/prioritization scales; otherwise noise
  • Implications: For Ken's AI ops: building 'captain's log' aggregators (PR summaries, chat synthesis, doc deltas) enables unprecedented CEO visibility—flatten reporting chains; Hiring: prioritize ML-trained generalists with agent orchestration skills over narrow specialists; Risk: pure flattening + agent delegation may hurt mentorship—need structured 1:1 apprenticeship; Evaluation becomes bottleneck at scale—automated harnesses + clear taste criteria (metrics, rubrics) essential; Cultural: transparency + accountability rise; managers lose information-asymmetry power—some will resist

Monetization & Financial Product Landscape

  • Claims: Cleo monetizes via freemium subscription + financial products (cards, BNPL, paycheck advance, wealth/stocks); spent first 4 years ignoring revenue to chase MAUs (worked until growth funds demanded it); Advises founders: either bake one core margin primitive (tokens, transaction fees) from day one or fully ignore revenue—no half measures; Subscription dominant due to payment rail friction (transactions fail, high cost); expects shift to microtransactions when stablecoins/frictionless rails emerge (Chinese model); Most consumers use credit cards inefficiently: max limit, pay minimums for 10 years at high interest; BNPL is better (fixed-term product) but still bad if just enables more borrowing
  • Evidence: Barney: 'First 4 years: financial health, no revenue, this is great—worked until growth funds: where's your revenue? I'd probably do same: build product people use/like, integrate monetization after'; On primitives: 'Make sure there's one core thing—Stripe takes margin, tokens, whatever—baked in from day one. Or completely forget it, get consumers. Don't do anything in middle'; On credit: 'Most consumers take credit card, max limit, stay at max paying minimums 10 years—most inefficient form. BNPL better: pay over actual time period, termed product. But if you use it to borrow more, bad'; On subscriptions: 'Dominant because payment rails—hard to do microtransactions, transactions fail. Recurring monthly works for software. Future: stablecoins, no friction, no cost → move to Chinese model'
  • Caveats: Freemium + financial products creates misaligned incentives (maximize product usage vs. maximize user financial health)—Barney claims they optimize for health + engagement, but doesn't disclose metric weighting; Stablecoin/frictionless rails future is speculative (regulatory hurdles, adoption inertia); BNPL 'better than credit cards' assumes users don't just layer BNPL on top of existing debt—no data on whether Cleo users reduce total borrowing
  • Implications: For Ken's business models: subscription + usage-based hybrid only works if payment rails frictionless; otherwise pick one; Consumer fintech: advice-only models struggle to monetize—must sell products (lending, wealth) but creates incentive tension unless outcome-optimized; Embedded finance: selling financial products via software layer (Stripe Issuing, Treasury) is high-margin but requires regulatory investment; Stablecoins may disrupt subscription SaaS if microtransaction costs drop to zero—Ken should monitor infrastructure buildout (Stripe crypto rails, Lightning, L2s)

Regulatory & Macro: UK/Europe Needs Creative Destruction, US Federalism Advantage

  • Claims: UK regulators (FCA) constrain via slow sandbox processes (2 years) misaligned with VC funding cycles (18 months); startups run out of money before proving PMF; Founders must take regulatory risk early (Barney: 'Stripe took some risk with early clients'); overly cautious approach kills innovation; US federalism (state vs. federal, checks/balances per Federalist Papers) enables experimentation; UK centralized power (PM, Number 10, regulators) stifles; UK pension capital (which should fund VC) goes to California LPs → GDP growth returns to California, not Britain; Europe lacks risk-taking culture; Need cultural shift from 'conserving legacy' to 'creative destruction' mindset
  • Evidence: Barney: 'FCA sandbox: please come work together for two years when funding cycle is 18 months… constrains companies. Seen angel investments try to engage FCA, run out of money… Need to take risk early'; On US: 'Federalist Papers great insight on running company/country. US has checks/balances, state/federal. UK: PM has so much power, regulators complete control yes/no across 60–70M consumers. Doesn't happen in US'; On capital: 'British VCs—California pension funds are the LPs. Not returning GDP growth to British state, going back to California. Really bad for Europe if we don't deploy institutional capital differently'; On culture: 'UK: culture of conserving legacy/past, greatness of Britain. Actually stagnating, declining, don't have much to protect. Need to go risk-on, creative destruction'
  • Caveats: Doesn't address US regulatory fragmentation costs (50 state money-transmitter licenses, conflicting crypto rules); UK pension conservatism is structural (fiduciary duty, political pressure) not just cultural—hard to change via exhortation; Some UK VCs do have domestic pension LPs (e.g., British Patient Capital), but scale far smaller than US CalPERS/CalSTRS; Federalism enables experimentation but also creates compliance burden—tradeoffs not discussed
  • Implications: For Ken's investing: favor jurisdictions with risk-tolerant regulators + deep institutional VC capital (US, Canada, Australia > UK/EU); Startups in regulated industries: 'ask forgiveness not permission' is viable strategy early (regulatory arbitrage) but requires legal reserves for later compliance; Europe's structural disadvantage (centralized regulation, pension conservatism) may worsen—US/China bipolarity increasingly likely; Cultural shift ('creative destruction' mindset) is necessary but insufficient—need policy changes (pension mandates, regulatory sandboxes with real teeth)

Notable Concepts & Terms

  • GTO (Game Theory Optimal): Poker strategy based on exact probability tables; only became mainstream 2016–17 with solver tools. Metaphor for 'optimal financial decisions' but Barney jokes it 'wouldn't sell' to consumers. Shows how recent many 'solved' problems are (cf. Black-Scholes for options pricing).
  • Recursive LLM synthesis (Karpathy workflow): Pattern where LLMs iteratively summarize/structure raw inputs (PRs, Slack, docs) into hierarchical knowledge base. Barney: 'Karpathy tweeted this a few months ago, we've been doing it for years.' Enables CEO-level visibility at scale without middle management filters.
  • Captain's log (organizational knowledge base): Barney's metaphor: LLM agent goes around org with clipboard taking notes, building up structured memory (goals, decisions, code changes, discussions). CEO queries the log instead of attending meetings. Flattens hierarchy by removing information asymmetry.
  • Financial health score (0–100): Cleo's composite metric tracking user's financial state (income, spending, savings, debt, investments). Used as RL objective function: recommender optimizes for improving score + engagement, not just clicks. Core primitive: 'spending less than you earn.'
  • Proactive agentic vs. reactive chat: Key product distinction: proactive agents push notifications/take actions at optimal times without user query (Cleo, health apps); reactive tools wait for user to initiate (ChatGPT, search). Barney argues boring/stressful domains need proactive; reactive fails to engage.
  • Product contrarianism: Strategy: identify incumbent's biggest pain point, deliver extreme inverse. Examples: Stripe (complex enterprise payments → simple dev-friendly), Monzo (opaque banks → transparent), Cleo (boring advice → fun/humorous). Barney: 'Such an obvious thing but how many founders actually do it?'
  • Creative destruction: Schumpeterian concept: innovation displaces incumbents, driving economic growth. Barney argues UK/Europe culturally resist this ('conserving legacy') while US embraces it (federalism, pension VC funding). Claims Europe's structural disadvantage without more risk-taking.
  • AI-pilled (grads): Slang for deeply immersed in AI/ML from early education. Barney hires UCL/Cambridge ML master's grads who've 'been doing it from A-levels'—they're fluent in LLM tooling, agent orchestration, prompt engineering. Contrasts with traditional SWE hires.
  • Tastemaker (in AI org): Role that sets experiment objectives, evaluates agent outputs, decides what 'good' looks like. In Barney's vision, junior AI-native engineers with 'good taste' manage agent fleets to take '2,000 bets/year.' Taste = prioritization + quality bar, not just tool fluency.
  • SaaS apocalypse (narrative): Investor/Twitter meme that frontier LLMs (OpenAI, Anthropic) will destroy vertical SaaS via generic wrappers. Barney's contrarian take: wildly overblown—deep vertical products with domain expertise, full stack ownership, agentic action will persist. 'Someone wrote a prompt'—not enough.

Operator Notes / Why Ken Should Care

  • Barney's 'captain's log' aggregation (LLMs over PRs/Slack/Notion) is a near-term playbook for Ken's AI ops work—tools that synthesize disparate org signals into queryable knowledge graphs enable radical CEO visibility without middle management. Risk: privacy/trust violations if not communicated transparently.
  • Proactive agentic products (push notifications, autonomous actions) beat reactive chat in 'boring but important' domains (finance, health, compliance). For Ken's agent systems: design for interruption optimization (RL-based schedulers, context triggers) not just query-response.
  • Vertical depth + domain expertise + full product stack > frontier model wrappers, per Barney's contrarian take. Ken should prioritize investments in companies that own user data, regulatory relationships, and end-to-end workflows—not thin LLM layers.
  • Hiring strategy: prioritize ML-trained generalists with early AI exposure (university courses, hackathons) over traditional SWE pedigree. Barney's UCL/Cambridge ML grads are 'AI-pilled since A-levels'—they're fluent in agent orchestration, can manage fleets. Ken's teams should similarly over-index on agentic literacy.
  • RL-based recommenders optimizing long-term outcomes (financial health, not clicks) are defensible moats but require massive historical data + sparse reward handling. For Ken's content/research agents: track delayed outcomes (user retention, quality metrics at 30/60/90 days) not just immediate engagement.
  • Regulatory strategy in ambiguous domains: 'ask forgiveness not permission' early to prove PMF, then invest in compliance. Barney admits Stripe 'took some risk with early clients.' Ken's portfolio companies in gray areas (AI content, synthetic data, crypto) should follow this playbook but with legal reserves.
  • Subscription + usage-based hybrid only works if payment rails frictionless; otherwise pick one. Stablecoins/L2s may disrupt subscription SaaS by enabling microtransactions—Ken should monitor Stripe's crypto infrastructure buildout as leading indicator.
  • Organizational flattening via LLMs risks morale if managers feel disintermediated. Barney's 'engineering managers hate me' is a warning: transparency + accountability require cultural buy-in. Ken should pair 'captain's log' tools with explicit mentorship structures (1:1 apprenticeships) to retain talent development.
  • Personality fine-tuning (humor, conciseness, approachability) is high-leverage in consumer AI products. Cleo's 30 female comedians writing all copy → MAU/revenue lift. Ken's content agents should explore tone-of-voice personalization (by audience segment, use case) beyond generic LLM defaults.
  • Europe's structural disadvantage (centralized regulation, pension conservatism) vs. US federalism + institutional VC capital is macro headwind for non-US startups. Ken should weight portfolio toward US/Canada/Australia; European investments need regulatory risk discount unless founder has 'ask forgiveness' mindset.

Watch Map

  • 00:00: Intro: Barney's poker background, funded university via online poker pre-GTO, now runs AI agent earning £200K/year rake back
  • 04:30: Clio origin story: founded 2016 as AI money assistant when 'weird,' pre-Transformer; Barney did ML master's in NLP era
  • 07:00: Personalization: generic 'buy index funds' advice out of touch; 60% Americans <$400 savings, need daily spending help
  • 10:00: Personality matters: hired 30 female comedians, 'roast mode' mocks spending, humor + money 'go together weirdly well'
  • 13:45: Proactive agentic > reactive chat: users don't wake up wanting to think about boring money; must push at right time
  • 15:30: Financial products landscape: BNPL better than credit cards (fixed-term) but bad if enables more borrowing; most consumers inefficient
  • 18:30: UK/Europe regulatory critique: FCA sandbox 2 years misaligned with 18-month VC cycles; need creative destruction, risk-taking culture shift
  • 22:00: Monetization: freemium + financial products; spent 4 years ignoring revenue for MAUs; advises either bake margin primitive day 1 or ignore fully
  • 25:00: Technical evolution: regexes → supervised learning → LLMs; now recursive synthesis builds per-user knowledge base (goals, health score)
  • 28:00: Cleo architecture: RL recommender optimizes financial health + engagement (not just clicks); measures 0–100 score with 2–3 month lag
  • 31:00: Org transformation: since Jan/Feb 2024, LLMs synthesize all PRs/Slack/Notion into 'captain's log'; CEO knows velocity/quality unfiltered
  • 33:00: Flattening orgs: engineering managers 'hate me'; middle layers had bias/incentive distortions; LLMs provide rational aggregation at scale
  • 34:00: Hiring: UCL/Cambridge ML grads 'AI-pilled since A-levels,' manage agent fleets; goal '2,000 bets/year instead of 10'
  • 36:00: Apprenticeship model: 1 senior + 1 junior better than 1:8 when juniors have agentic leverage; need 'good taste' to run parallel agents
  • 37:30: Contrarian take: SaaS apocalypse overblown; deep vertical products with domain expertise >> frontier model wrappers; 'someone wrote a prompt' not enough
  • 39:00: Investor critique: 'silliest, most irrational people, no technical knowledge, sit on fence, only fund frontier labs + Stripe'; over-extrapolating
  • 40:00: Closing: pro-abundance, creative destruction; economy driven by software for 30 years; need more software companies, not just frontier labs

Source/Metadata

  • Title: Barney Hussey-Yeo in conversation with John Collison
  • Transcript words: 11000
  • Duration seconds: 2370
  • Timestamp note: Timestamps manually estimated based on ~40-minute video duration and transcript flow; MM:SS format used as HH:MM:SS not needed
Search this video Resume transcript at 0%
Full transcript 7660 words · 49 min read
0:00

SPEAKER_01

Thanks for being here. Thanks for having me. So, Clio, you guys are pretty familiar with, a phenom in the consumer financial advice space, 400 million in ARR. As of last month? Yes. Unicorn and really rapidly growing one of the top AI startups and based right here in Shoreditch. But before you started Clio, I want to go back to your prior career. Yeah. You were a poker shark. Is that right? A shark, a fish. I played at university. I funded my university career through playing online poker. I only know what I've seen in Rounders and Molly's Game. So, how does actually making, professionally playing poker work?

0:13

SPEAKER_01

This was back before you had GTO and The Solvers. So, it's all online and you play on multiple sites and you play four to six hours a day, heads up. And actually, really interestingly, I've just built an AI agent that now plays for me and earns rake back. So, I make about, I shouldn't say this publicly, but I earn about 200K every year now just through my little agent that plays and barely beats the low stakes. But yeah, I used to play low mid stakes, heads up, and it was quite profitable.

0:20

SPEAKER_01

[SPEAKER_00] And how, so, GTO is game theory optimal, which is doing the exact right thing that the probability tables would tell you that you should do, right? And you didn't have that in 2012, 13, 14 when I was playing. Really? It was invented that recently? [SPEAKER_00] Only 2016, 17, you had The Solvers and we actually learned how you could beat The Game. It's still right at the edge. You can probably beat The GTO Solvers. But it's basically a solved game, which means online play is dead now. But you should play tournaments. You should go play a live tournament. They're always fun. No, because I'll be playing against your boss, probably.

0:40

SPEAKER_01

[SPEAKER_00] Yeah, don't play online against me. You're going to get rid of it. [SPEAKER_00] Yeah, that's very interesting. The Solvers, a lot of these things are discovered later than one would think. Like Black-Scholes is a relatively recent invention for valuing financial options. And I hadn't realized that GTO is such a recent invention in poker. [SPEAKER_00] Well, you don't want to tell people if you've got an edge, right? [SPEAKER_00] Yeah. [SPEAKER_00] So, there's probably all these things in hedge funds where you could have a massive economic benefit to GDP, which is not yet. No, but they keep it secret.

1:13

SPEAKER_00

[SPEAKER_01] Exactly.

1:21

SPEAKER_00

[SPEAKER_01] Yeah, this happened in cycling as well, in pro-cycling, where people used to inflate their tires up really high, you know, 120 PSI or something like that. And then one Tour de France team figured out that actually everyone was over-inflating their tires, and you don't need to over-inflate your tires. And they kept it as a secret from everyone for a long time. Anyway, there's still plenty of room for new inventions is what we're taking away here. But we are not here to talk about pro-cycling or pro-poker. We're here to talk about Clio. So, maybe just to orient people who aren't familiar or aren't using the product, what does it do? How did you guys get here?

1:26

SPEAKER_00

Yeah, well, we started in 2016, so it was an AI assistant for your money, which was weird in 2016, and now is the thing to be building. Because AI was not as hot in 2016.

1:34

SPEAKER_00

[SPEAKER_01] Yeah, building an AI system was completely weird. Natural language processing was barely working. But it was the really big research output at the time. So I did my master's in machine learning, and we'd just gone from the vision era of beating AlexNet, and then NLP became the thing all my professors were interested in. So I'd worked at FinTech as a data scientist, and I had this fascination with machine learning, and knew it was going in that direction. So the thesis was, everyone's decisions are wildly unoptimized, and your financial life is the most stressful and important thing in your life. So how could you make a billion people make better decisions every day, or make decisions on their behalf? And how could you change the economy, the world, everyone's life through doing that?

1:40

SPEAKER_00

[SPEAKER_01] GTO for your financial life, if you will.

1:44

SPEAKER_01

GTO, yeah, I don't think it would sell. But maybe to you, John. Yeah, for the, just free idea for the magazine article, you know, for the opening paragraph on Clio. Okay, but that's interesting, right? Because you were founded in 2016. That's before, it was obviously before ChatGPT, but that's, before LLMs.

1:48

SPEAKER_00

GTO for your financial life, two years before the Transformer. GTO for your financial life, but even back then, if you had the Meta Lab we were just talking about, they were really focused on it, and they were pushing out a bunch of good work. It was just intent classification back in the day. Then you had two years later, it was the Transformer, then you had BERT. Everyone then in computer science knew this was going to be a thing, and then everyone else just woke up when 3.5 dropped, but you could see it, right? [SPEAKER_01] But so what, I mean, I presume today you're heavily LLM-based. [SPEAKER_01] Yeah.

2:04

SPEAKER_00

[SPEAKER_01] Just how did this work? How were you providing the financial advice? The user input some data, and then what?

2:04

SPEAKER_00

[SPEAKER_01] It started with regexes. We had a lot of regexes to start with, but then it was just traditional supervised learning, right? So you had intent classification and entity classification. It's a bit nerdy, but intent is what you say in a sentence. So it was big, multi-class supervised learning, and then you had, and we still do, had the recommender systems and reinforcement learning to push you information at the right time. So you've gone from a supervised learning world to an LLM-based world, and that completely changed our architecture and our products, actually.

2:06

SPEAKER_00

Yep. How, the conceit of Clio is that you should have personalized financial advice. How personalized does financial advice actually need to be? Because, you know, there's a view that you could say you should have the right savings rate, and then you should invest in a low-fee index fund, and you shouldn't buy financial products. It's, you know, you shouldn't eat foods that are advertised on TV, because the good foods don't have an ad budget. And similarly, you shouldn't buy financial products that are aggressively sold. You should buy a low-fee index fund, and then you shouldn't have a home country bias. You shouldn't buy the FTSE 100. You should buy the MSCI World.

2:07

SPEAKER_00

Does financial advice actually need to be? Because there's a view that you could say you should have the right savings rate, and then you should invest in a low-fee index fund, and you shouldn't buy financial products. It's like you shouldn't eat foods that are advertised on TV, like the good foods don't have an ad budget. And similarly, you shouldn't buy financial products that are aggressively sold. You should buy a low-fee index fund, and then you shouldn't have a home country bias. You shouldn't buy the FTSE 100. You should buy the MSCI World Index. Is that not enough financial advice for everyone?

2:10

SPEAKER_00

Showing how out of touch you are, John, with the real consumer and the normal humans on the street. Your life is your money. You spend it, and it's the most stressful thing for the majority of people in their life. You spend so much boring admin time thinking about how you make financial decisions. From where are you going to live? Where's that mortgage? Where's that rent going to be? How big should that be? Every single thing you do every day is a financial decision in reality. You make all these transactions every single day. Lots of them on Stripe. But imagine if you had a world where everyone was making rational, sensible, more optimized decisions. If you think about how much money goes towards consumer credit or paying on interest when it doesn't need to, you could really change not just the top 1% of people that get financial advice. It's the 99% that actually need the financial advice. The people that are living one month, two month, three month savings. If they can have a better life and make better decisions, and you can do it for them, you shouldn't have to think about finance, right? It should just happen on your behalf. You need credit. It should just ask. It should go get it for you. If you have wealth, then it should just go get the best return. It should be tax free. There should be a world where you don't manage your money by looking at a balance and a list of transactions. And that's coming very soon, right?

2:12

SPEAKER_01

Where are you generally trying to steer people? Like I can imagine there's avoiding bad financial products, like the classic one of borrowing on a credit card. There's the maybe price comparison where if you're getting a mortgage, you can get somehow better rates here versus there. There's tax efficiencies. But just like...

2:13

SPEAKER_01

[SPEAKER_00] It's actually your life. You think about most of the consumers in the U.S. It's 60% of Americans have less than $400 of savings. It's how they spend their money every single day. How much do they earn? What's their income volatility? How much are they spending on rent, mortgages, subscriptions? What are they actually doing with their financial life? And you can really help someone just by laying it out to them, making better decisions for them. And then on the products, they are wildly unoptimized, especially at the lower end. They're paying incredibly high interest rates. They are not getting any returns on savings. Most people don't have investments when they just sit in a cash ISA in the U.K. It's crazy the amount of GDP growth and the amount of financial wealth you could be creating people that's left on the table. So I just think there's trillions and trillions of good that you can do for society and humans, but also just the economy as well. If you have this thing moving effectively through the system, you can dramatically increase GDP growth.

2:14

SPEAKER_01

How do you advise people spending? I mean, there's the meme online of the millennials with their avocado toast. Is it like you're trying to reduce the amount of avocado toast consumed? Or how do you actually steer people usefully in an actionable way on spending?

2:18

SPEAKER_00

Do you know what was funny back in the day? Because it wasn't as intelligent, we had to really rely on humor. So we had 30 female comedians on staff. Cleo was a female agent. So we had these female comedians. And they wrote every single line of copy, both what we pushed and what we responded with. And we found that humor and money actually go together weirdly well, which no financial institution spoke to with all your lawyers and your executives. Bankers generally not known for their humor.

2:27

SPEAKER_00

Yeah. So that was a real unlock for us in terms of engagement. And there's a roast mode in Cleo that makes fun of you for your spending. I don't know what it would make fun of you. Maybe the cheeky pints, John. But it roasts you for what you've done over the weekend at the bar, whatever it is. And people really like that. And it's a way to engage them in simple, actionable financial advice. And increasingly, it's just taking action on their behalf. It's moving accounts. It's moving money. It's getting them the best credit, whatever that is at that time for that user. [SPEAKER_01] Do you have a way of measuring how Cleo is steering people's financial health? Yeah.

2:38

SPEAKER_00

[SPEAKER_01] Yeah, we have measured. I'm actually really surprised that you probably do internally. But there is, we have constructed a view of financial health for every individual at Cleo. There's a score, 0 to 100, and all these variables and all these features. So we know whether you're leveling up your financial health or you're degrading it based on your actions that you do. So actually, the objective function of the recommender system, it used to just be engagement. So it was like, can we make you click this button? And it was fun. But now it is financial health and engagement. So we're optimizing for, did this bit of advice, did this action lead to in two months, three months time, their financial health getting better? And then spending less than they earn, which is the core primitive.

2:42

SPEAKER_00

[SPEAKER_01] Spending less? Spending less than they earn. [SPEAKER_01] Oh, yeah. It's the core primitive of every human in that low to middle income, right?

2:54

SPEAKER_01

Yes. Yes. Who do you compete with? Is this substituting for existing products? Is this just a new thing?

3:06

SPEAKER_01

[SPEAKER_00] Yeah, well, think about how people make financial decisions today. It's a vibe-based optimization game for most people. Maybe it's a fag packet and a napkin, but the vast majority of consumers don't budget. They don't think about their money in a very rational way, or they do it in a very late way. There's only a really specific number of nerds, probably everyone in this room, actually, that will have an Excel, an Opti... I'm sorry, I shouldn't make fun of all the audience. But have this set of advice that they'll construct themselves. So we're really in the advice game for people that don't have advice. And I think there's going to be loads of avenues like this.

3:11

SPEAKER_00

Maybe it's a fag packet and a napkin, but the vast majority of consumers don't budget. They don't think about their money in a very rational way, or they do it in a very late way.

3:21

SPEAKER_01

[SPEAKER_00] There's only a really specific number of nerds, probably everyone in this room, actually, that will have an Excel, an Opti... [SPEAKER_00] I'm sorry, I shouldn't make fun of all the audience. [SPEAKER_00] But have this set of advice that they'll construct themselves. [SPEAKER_00] So we're really in the advice game for people that don't have advice. [SPEAKER_00] And I think there's going to be loads of avenues like this. [SPEAKER_00] Imagine health, right? [SPEAKER_00] We all make terrible decisions for our health all the time.

3:46

SPEAKER_00

We've all got a whoop now.

3:53

SPEAKER_01

[SPEAKER_00] We're all getting into it.

3:57

SPEAKER_00

But there's going to be all these big verticals in your life, whether it's education or health or finance, where your decisions are wildly optimized today, and they're just going to be so much more thought through. Imagine how you deploy capital in people at Stripe.

4:01

SPEAKER_01

[SPEAKER_00] There's probably an army of nerds thinking about how you deploy that versus your strategy and where your goals are going.

4:06

SPEAKER_00

Imagine if you do that for every human, for every decision they ever have to make.

4:09

SPEAKER_01

[SPEAKER_00] So you think you go broad and beyond financial advice into... [SPEAKER_00] There's a lot to do in financial advice. [SPEAKER_00] Big time. [SPEAKER_00] So we're going to nail that first. [SPEAKER_00] And I think you're going to have these four or five agents that go really deeply vertical in these big, specific areas of your life. [SPEAKER_00] And I don't think they're going to be the frontier model companies, actually. [SPEAKER_00] I think they're going to be companies that are focused all on health. [SPEAKER_00] They go all the way to the MRI and the blood test, and they have a whole stack of advice.

4:35

SPEAKER_01

[SPEAKER_00] So I'm bullish on that, and I'm less bullish on OpenAI being the interface for everyone, for everything. [SPEAKER_00] Do you see that competition dynamic with users where anyone who's building AI tools, there's always the... [SPEAKER_00] Anyone who's building specific vertical AI tools, there's always the dynamic where people might just take a bunch of financial statements or connect their bank account to Claude and start asking questions. [SPEAKER_00] Do you practically see that behavior where an alternative to using Clio is wiring up the relevant data to an LLM? [SPEAKER_00] Well, you can connect your plan account now to ChatGPT and ask questions, right?

5:00

SPEAKER_01

[SPEAKER_00] But I would bet a huge amount of money they get very little engagement, and it becomes a side project that goes nowhere in reality for that business.

5:01

SPEAKER_00

Because to engage people with financial advice, it's got to be proactive, and it's got to be agentic. Yes. So people don't wake up and go, oh, how much did I spend on... Yes. Is it just a boring thing to think about? So it has to be this agent that's always on, always pushing information. If you think about all the frontier model companies, it's a very reactive thing. It's replacing Google search for most people today. [SPEAKER_01] You have something in your head that you want to research, you want to go into, but there's so much space in the product cycle for it to become a proactive, intelligent, always-on agentic thing. [SPEAKER_01] Yes.

5:51

SPEAKER_00

[SPEAKER_01] And that's going to require a lot of effort and a lot of training data and a lot of understanding of what you need to say to a user at what time that's going to work. [SPEAKER_01] It feels to me that we're early in personalization. When you're saying it'll be very personalized, it's just... I mean, if you look at how OpenClaw works underneath the hood, it's this static model and then a bunch of flat files that it's writing notes to itself. It's like the movie Memento where he's writing on his hand. And that's the only memory he has. That's basically the AI models today.

6:19

SPEAKER_00

And I'm curious, just for your experience trying to build personalized products, where the current AI technology is just fundamentally a bit resistant to personalization. I've never thought... There's definitely a meta point here. We just have text, right? And how does the human store memory? And when are we going to get to a different form factor for that? But in reality, and what we've been doing, you can build a knowledge base of every single human, like their financial aspirations, their financial health, their financial goals, every conversation they've ever had.

6:45

SPEAKER_00

So you remember Carpathie did that thing a few months ago where it's like, oh, I do this thing recursively with LMs. We've been doing that for years.

6:53

SPEAKER_01

[SPEAKER_00] That is such a powerful way to build a structured knowledge... [SPEAKER_00] I was also giving you a bit of shit for Stripe not being intelligent enough. [SPEAKER_00] And they should be that knowledge base at Stripe. [SPEAKER_00] They should be every single card connected to every single user, who this user is, how much income they have, where they spend. [SPEAKER_00] So we've built that over time for every user.

7:17

SPEAKER_00

And then we have a recommender system that is thinking from the knowledge base. There's all of these opportunities for us to say things to users. What is the thing that's going to move the objective function the most, [SPEAKER_01] that's going to move that financial health, deliver that information, take that action on the back end? [SPEAKER_01] So I think text works as long as it's stored in a thoughtful, clear schema and database. [SPEAKER_01] One thing we were talking about backstage is the concept of personality in AI is interesting.

7:43

SPEAKER_00

[SPEAKER_01] I think at the beginning of the field, when people were first introducing LLMs, people might not have predicted that personality would be such a big thing. But we're talking about the new anthropic model Fable and what its personality is like. Or famously, GPT-4-0 had this very controversial personality where some people found it too sycophantic.

7:57

SPEAKER_01

[SPEAKER_00] For some people, I guess they really liked how sycophantic it was. [SPEAKER_00] And they were annoyed when they tried to take it away from them. [SPEAKER_00] And I get the impression that, for you guys, the personality of Clio and how it interacts with you is a big part of the differentiation. [SPEAKER_00] So what goes into getting the personality right? [SPEAKER_00] And, for example, do you need to tune it for different users? [SPEAKER_00] Do you try to figure out what personality people want?

8:23

SPEAKER_00

Is it just what it is? How do you build that? Have you moved beyond the army of comedians? Or is that still how you engineer the personality? I think this is one of the biggest things at Frontier Labs. Maybe they don't need to. And I get the impression that, for you guys, the personality of Clio and how it interacts with you is a big part of the differentiation. And so what goes into getting the personality right? And, for example, do you need to tune it for different users? Do you try to figure out what personality people want? Is it just what it is? How do you build that? Have you moved beyond the army of comedians? Or is that still how you engineer the personality?

8:55

SPEAKER_00

I think this is one of the biggest things at Frontier Labs. Maybe they don't need to. Maybe they should be boring and it should be a generic, not even personalized thing and they should just be for the essence of intelligence. But I think for any brand or any company doing this, the personality and personalization to the user dramatically changes engagement. It dramatically changes what a user does. [SPEAKER_01] So being funny, being concise, writing well, all of these things move our MAUs.

9:05

SPEAKER_00

It moves our revenue. So we've known from the first three to six months of building this business that the brand, the personality, and the tone of voice mattered a lot. So it's been a linear supervised learning with one tone of voice. And then since LLMs, it is now a much more personalized tone of voice based on your income, your financial health, where you live. You're in Boston. You're in San Francisco. I don't know how it talks. I want to make fun of John, but I shouldn't make fun of John in front of a 15-minute group. But it will talk to you however it thinks you want to be spoken to, and you can craft that as well. But the general learning is that finance for a lot of people is a high-stakes topic.

9:10

SPEAKER_00

[SPEAKER_01] It's maybe they're nervous about it. It's a bit of a schlep.

9:15

SPEAKER_00

And so you're trying to make it more approachable is the fundamental personality. I think all these great companies take the inverse. It's like Stripe. The payments industry is completely archaic and complicated. I've got to go through enterprise sales to do anything in payments. Let's do the opposite. Let's make it really simple, dev-friendly. It's going to be a one-click thing. It's the exact same with Monzo. It was really bad UX to do anything with a bank. And they just made it simpler, more transparent, less opaque. And we've gone, okay, it's really complicated financial advice. We're going to make it really simple. It's really boring. We're going to make it really fun. And we've just taken the inverse position, which is such an obvious thing to say, but I think, how many founders do you see actually take that? So I think it's worked for us. It's product contrarianism or product differentiation.

9:20

SPEAKER_00

Yeah, it's like – and you always want it in the most extreme form, generally. What is the biggest pain point of your incumbent? And what am I going to do? I'm just going to go incredibly hard at their biggest pain point and make it the most delightful experience.

9:27

SPEAKER_01

So it's a hack for building a first product, right? Yes. How is the landscape of financial products that you advise people on changing? Borrowing is presumably changing somewhat at an industry level due to the emergence of BNPLs and things like this. Is the mortgage industry changing and how people are getting those here? Yeah, I'm curious how you think about that.

9:37

SPEAKER_01

[SPEAKER_00] There's only so many primitives with money, as you definitely know, right? You can borrow. You can spend. There's just only a finite set of things you can do. So all of the innovation is useful, but it's slow and it's guided by regulators. I think BNPL is a great example of credit cards with this really bad thing for all these years. So we're doing something better and something different. But it's not that interesting, in my opinion, generally. Hopefully, the biggest thing will be stablecoins and how do we transact and how do we move away from those rails? But I don't find building – we build a lot of financial products, so I shouldn't say this because all my teams work on them. I find them constrained and slightly boring because of the limited number of primitives. But I think the intelligence and the advice – if I was building Stripe, I would be like the intelligence, what I can give consumers from that would be the most important thing. But I guess that's also my wonk.

9:40

SPEAKER_01

[SPEAKER_00] Well, it's not just a limited number of primitives, right? It's also all these products are heavily regulated and so they end up in certain streams. But, for example, do you see different – should people be consuming a lot of BNPL or a little? Like, what's –

9:45

SPEAKER_01

[SPEAKER_00] This shouldn't be maximizing – what happens generally for consumers is they'll take a credit card out and they will max the limit and then they will stay at a max limit paying the minimums for 10 years. It's probably the most inefficient form of credit the vast majority of society has. So BNPL is a better way to pay over an actual time period and make it more like a termed product. For a fixed amount of borrowing, though if you just use it to borrow more, presumably that's bad. This has been my whole raison d'etre for a long period of time. Credit is such an important part of the economy and people's lives, but it is used so poorly by so many people.

9:50

SPEAKER_01

[SPEAKER_00] and then they will stay in a max limit [SPEAKER_00] paying the minimums for 10 years. [SPEAKER_00] It's probably the most inefficient form of credit [SPEAKER_00] the vast majority of society has. [SPEAKER_00] So BNPL is a better way to pay over an actual time period and make it more like a termed product. For a fixed amount of borrowing, though if you just use it to borrow more, presumably that's bad. This has been my whole raison d'etre for a long period of time. Credit is such an important part of the economy

10:29

SPEAKER_00

[SPEAKER_01] in people's lives, [SPEAKER_01] but it is used so poorly by so many people. [SPEAKER_01] And this is why I think the intelligence is the answer, right? [SPEAKER_01] It helps make people make better decisions [SPEAKER_01] about how much credit they consume, [SPEAKER_01] how much they spend. [SPEAKER_01] If you just provide financial advice [SPEAKER_01] or provide financial products without the advice, [SPEAKER_01] I don't want to bash Robin Hood, [SPEAKER_01] but you end up just selling products [SPEAKER_01] and trying to get the most margin [SPEAKER_01] and it leads you down this way of, [SPEAKER_01] ooh, let's do day trading, or let's do...

11:17

SPEAKER_00

None of these things are bad, right? And they can be used in the right way. I do think you need some sort of... You can't be purely revenue maximizing, is your view? 100%. I think it just leads to bad outcomes for consumers and bad outcomes for the business as a whole. So I think that... I think the intelligence and the decision-making and helping users do it for them is just so critical for everything. If you were in charge of regulating financial products in the UK, what would you change? Where do I start with the UK and its regulators? Innovation is good. Creative destruction is good.

12:30

SPEAKER_00

It's how you move every industry in the world along. It's what we've seen for the last 100 years. Creative destruction is what drives better outcomes for consumers, what drives better outcomes for businesses. So you need to enable creative destruction. You need to have a level of risk-taking to allow upstarts to take on incumbents. And all too often in the UK,

12:58

SPEAKER_01

[SPEAKER_00] it's, please come to our sandbox and let's work together for two years when a venture funding cycle is actually 18 months. And it actually... Them trying to be helpful normally constrains a company. I've seen so many of these little angel investments I've done

13:20

SPEAKER_00

[SPEAKER_01] that have tried to engage the FCA, but it's just taken so long to get there that they've run out of money and not been able to prove product market fit. So if you're a founder doing it, I actually do think you need to take some risk early on. I'm sure Stripe took some risk with its early clients and how it built the business. Don't want to get you in trouble, John. But taking some risk in this sector at the start is a way to build something that the incumbents aren't going to do as well. And you think there's not enough risk-seeking behavior in the UK or it's not enabled in the right way? Well, it's... Go back to the Federalist Papers, right?

14:13

SPEAKER_00

You have checks and balances. You have state and federal. It's probably the... Read those Federalists. [SPEAKER_01] They're great insight [SPEAKER_01] on how you should run a company and run a country. And we need more of that in the UK. Let's see, the Prime Minister has so much power. Number 10 has so much power. The regulators have complete control of yes or no across 60, 70 million consumers. It doesn't happen in the US.

14:47

SPEAKER_01

[SPEAKER_00] We have the cause. We have the regulators. We have state.

15:01

SPEAKER_01

There's all these checks and balances which make America and capitalism work,

15:11

SPEAKER_00

[SPEAKER_01] which is why I love building in America [SPEAKER_01] and why all my users are in America, [SPEAKER_01] because it's actually a better place to do business, [SPEAKER_01] and it leads to more rational outcomes. You get, obviously, dramatic polarization and fights and all these things, but the checks and balances of the US institutions is why America is winning versus Europe. So moving beyond financial regulation and more broadly to the business climate, you think more embrace of creative destruction is the answer? Yeah, if I was going to really synthesize it, it's we've got to go from a culture of trying to conserve our legacy and our past

15:54

SPEAKER_00

and the greatness of Britain and Ireland to a place where we go, we're actually all stagnating. We're all declining. We actually don't have much to protect at the moment. We actually need to take some risk. We need to go from a very conservative risk-off mindset to we're going to build, and we're going to build great stuff, and we're going to take some risk, and we're going to deploy capital, all these pension funds, all these ISAs,

16:22

SPEAKER_01

[SPEAKER_00] all these institutions that could be doing productive things in society.

16:25

SPEAKER_00

[SPEAKER_01] You need to move from a mode of conservatism [SPEAKER_01] to risk on creative destruction. [SPEAKER_01] Does that filter down into how you think about [SPEAKER_01] how people should invest, [SPEAKER_01] both at the individual level and at the institutional level? [SPEAKER_01] If you could tune how people are investing, [SPEAKER_01] what encouragement would you give? I think the individual is a largely solved problem, even if people aren't doing it. The solved problem is the World Index Fund, right, and that's investing in the S&P 500. Maybe it's just getting a little bit tech-heavy when Stripe goes public at some stage, but that's an almost solved problem

16:55

SPEAKER_00

if you had people making rational decisions there, and Clio could dramatically help that. I think the new fund that they've set up with Gerstner and the baby accounts or whatever, the Trump account.

17:09

SPEAKER_01

[SPEAKER_00] What encouragement would you give?

17:10

SPEAKER_01

[SPEAKER_00] I think the individual is a largely solved problem, even if people aren't doing it. The solved problem is the World Index Fund, right, and that's investing in the S&P 500. Maybe it's just getting a little bit tech-heavy when Stripe goes public at some stage, but that's an almost solved problem if you had people making rational decisions there, and Clio could dramatically help that. I think the new fund that they've set up with Gerstner and the baby accounts or whatever, the Trump account. It's always the Trump account, isn't it? That's a good thing, and get people investing from their age. But on an institutional level, there's a dramatic difference between the U.S. and the U.K. and all of Europe. You look at our pension capital, which should be the big pool of capital going into creative destruction. Tiny, tiny, tiny, tiny fraction in the U.K. goes towards venture capital. I've got British VCs and the tier one European VCs. It's like the California pension funds are actually the LPs into those VCs. So I'm not returning to the British state any GDP growth. It's all going back to California. And if you don't want a world which is completely bipolar, where you can either build tech in China or you can build tech in the U.S., it's really bad for Europe if we don't have this creative destruction and ability to deploy institutional capital in a different way.

17:11

SPEAKER_00

It's an interesting point that there's a lot of U.S., I mean, it's not universal, but a lot of U.S., Canadian pension systems [SPEAKER_01] are a little bit in Australia, yeah, are invested in venture in a big way and much less so in the U.K. and Europe broadly. And it's worked. If you look at Canadian pension funds, if you look at Australia, I would love to be an Australian pensioner. Yeah, there was a news article about famously one of these pension funds that bought a lot of SpaceX a few years back. And as you can imagine, they're feeling quite good this week.

17:18

SPEAKER_00

You run a very scaled consumer AI product. I'm curious to talk about that. Well, actually, first off, how do you make money? We make money through subscriptions and financial products. So we've got cards, we've got lending products, buy now, pay later, and wage access. We've got wealth products, we've got stocks and shares. So the financial products make money, the subscription makes money, value occurring, and it's a freemium model. What advice do you have for people who are building AI products that they sell to consumers? Like, what have you learned about willingness to pay, what works well, acquisition,

17:25

SPEAKER_00

[SPEAKER_01] this new space that's only existed over the past few years?

17:26

SPEAKER_00

I don't know if I did it the right way. I spent the first four years of the business, maybe this is Zerp, just doing MAUs. It was build a great product, just get MAUs. Not focused on revenue, just focused on revenue. No revenue. And everyone was saying, financial health, no revenue, this is great. That worked up until I got to the growth funds, and then they were saying, where's your revenue? So I don't know if I'd advise that, but I would, you know. I would probably go back and do the same thing, of just try and build a product that people are using and like, and then integrate the monetization afterwards.

17:28

SPEAKER_00

The inverse of this, though, is just make sure there is one core primitive, like Stripe, where you are taking a margin on something, [SPEAKER_01] or there is just one thing, tokens, whatever it is, [SPEAKER_01] and build that from day one. I wouldn't do anything in the middle, which is being half-hearted on the monetization strategy. Like, it's either baked into the product 100% from day one, or you just completely forget it and get consumers. On the consumer side, I think it works, probably to just get consumers first, then monetize, but I know, what do you think?

17:40

SPEAKER_00

Well, I'm curious, on the consumer side of things, it seems like many businesses decide that they're going to be one or the other, where they're going to sell on a usage basis, [SPEAKER_01] you know, a normal grocery store, just charging markup on what you buy,

17:42

SPEAKER_00

or in the U.S., Costco has this famous model where they basically make no money on the actual things you buy there, and then they just sell a $100 a year subscription, it's probably gone up, a subscription, membership subscription, and that is where they make all the money that Costco makes, and they've decided that they're not going to bother making money on anything you buy in the store, they're just a subscription business. And do you guys just view it as both, or are you in the business of selling subscriptions, or are you in the business of selling financial products? Or is it just both?

17:46

SPEAKER_00

It's probably gone up, a subscription, membership subscription, and that is where they make all the money that Costco makes, and they've decided that they're not going to bother making money on anything you buy in the store, they're just a subscription business. And do you guys just view it as both, or are you in the business of selling subscriptions, or are you in the business of selling financial products? Or is it just both?

17:48

SPEAKER_00

We are in both, but I think the interesting thing there is subscription has been the dominant form factor to monetize because of payment rails as well, right? Because it's actually hard to do microtransactions or lots of transactions and transactions fail, so having a recurring monthly thing and making this a phenomenon for the masses works really well for software businesses. But I can definitely see a world in the future where you have stable coins, there's no friction, the payment rails actually work, there's no cost. Yes.

17:54

SPEAKER_00

[SPEAKER_01] Then you move way more to the Chinese model, right? For self-improvement products, I feel like there's also sometimes a sunk cost dynamic in consumer psychology where people say, I am going to buy a gym membership and I am going to get in shape and now that I've spent this money on the gym membership, I have to use it because I'll feel bad if I spent the money and didn't. Is there any of that dynamic with Clio where people are investing in their own financial health?

17:57

SPEAKER_00

[SPEAKER_01] I think there's all this behavioral theory in all of subscriptions. Like we just added a third tier because that works and we got this middle tier and just because we have three, people will pick the middle. It just—there's all this weird behavioral stuff with payments and subscriptions that you have to optimize and learn about. So, yeah, humans are irrational actors and I think you have to learn where they're irrational to help them but also to build a business. Yeah, you guys are weaponizing people's irrationality for good rather than for evil. It's not right that in the press, but thanks, John. Harnessing.

18:05

SPEAKER_00

Yes, there you go. Yeah, there we go. Better. We're just workshopping here. Speaking of people's irrationality, you know, people in poker talk about this idea of going on tilt where people may mostly have a very solid betting strategy and then they just break at a certain point and they start making irrational decisions, maybe after a losing streak or something.

18:08

SPEAKER_01

Does that also happen in behavior you see in the app and you're trying to get people back on the straight and narrow? It's human psychology and it's the biggest thing, right? Money is such a personal, emotive, psychological thing. People will lose a job. People will have a breakup. We have all these user research interviews where you see someone go completely on tilt and do things that are slightly irrational. So, yeah, we're definitely trying to help in those circumstances, but humans are humans in a way as well.

18:14

SPEAKER_01

[SPEAKER_00] But there's all these interesting mechanisms and features that we build around that. You started Clio before the current wave of AI, but you're still a small and fast-growing company. How AI native is Clio in how it actually works? What ways of working do you guys have that we should be borrowing from you guys or people in the audience should be borrowing? [SPEAKER_00] Well, my day-to-day is dramatic. Since probably January, February, I spend my time completely differently.

18:18

SPEAKER_00

Oh, please do tell. [SPEAKER_01] Give us the before-after. [SPEAKER_01] Yeah, well, the before was 500 people and you'd have all these division leaders and you'd talk to the division leaders and you'd farm for information and give them your opinions and you'd have the pre-reads and the way of getting information in a company February, I've, I spend my time completely differently. Oh, please do tell. [SPEAKER_01] Give us the before-after.

18:33

SPEAKER_00

[SPEAKER_01] Yeah, well, the before was 500 people and you'd have all these division leaders and you'd talk to the division leaders and you'd go farm for information and give them your opinions and you'd have the pre-reads and the way of getting information in a company was to talk to all of these leaders and then give them your feedback and do that. You don't need to do that in an organization. I think one of the interesting things that you probably won't want to talk about is how you can understand what is happening in your organization day-to-day. You know their Slack messages. You know Notion docs. You know every single PR. You know what is built.

18:51

SPEAKER_00

So I understand if software is the output of all these companies, I know what we're shipping in a really visceral, clear way [SPEAKER_01] which I didn't know before LLMs.

18:55

SPEAKER_01

And just very practically, how are you synthesizing all this information across the company? [SPEAKER_00] Are you using a third-party tool? Do you guys build internal tools? [SPEAKER_00] How are you scouring the Slack channels and ingesting all the docs? That workflow that Carpathie described is what I use for so much stuff. I just have LLMs that go over every PR and it summarizes what that PR does and it grades it and it tells me what's going on in that PR and it just builds a knowledge base by going over all these things. Goes over all the Slack messages and all the channels, builds up a knowledge base of what is happening.

19:11

SPEAKER_01

[SPEAKER_00] So you take lots of disparate information across your entire org.

19:13

SPEAKER_00

Yes. You build it up into a higher level. And it's assembling a captain's log, like it's going around the organization with its clipboards taking notes. On this dramatic scale that you could never have done. And then you're querying the log. Well, you get humans also to input into that log. Their pre-reads, their dashboards that they're making, everything they do, all goes into logs and becomes more synthesized and aggregated over time. Are you doing that? I mean, I'm not doing specifically that, but I think one of the interesting things

19:33

SPEAKER_00

[SPEAKER_01] that all organizations are focused on right now is consumer AI tools are phenomenally useful and they are trained on the corpus of information that's out there and they have web search tools and everything. [SPEAKER_01] And the companies are trying to internally open up the data, but not too open, you know what I mean, with all their permissions and security tools and everything. And so I feel like that is the thing [SPEAKER_01] that all organizations are reckoning with right now. [SPEAKER_01] But it's really, as you say, I mean, does it drive your internal leaders up the wall? The fact that you are now going straight to the source for information?

19:48

SPEAKER_00

My engineering managers hate me. Honestly, I know who's shipping. I know who is the highest velocity, highest code quality. I know which teams, because it's a collection of engineers, are actually doing great work. Where before, I had this view filtered through all these leaders that had all their different biases and all these different incentive structures. You can actually get really rational, clear thought from these LLMs and build it up with context.

20:00

SPEAKER_01

[SPEAKER_00] So, yeah, I think the middle layer has, probably at Clio, definitely hates me, but I think orgs have just become way flatter because of this and it's just become, it's going to become a much more efficient organizational flow and information flow. Do you think the average number of people that a manager [SPEAKER_00] at Clio manages will be different in two years' time versus currently? [SPEAKER_00] Yeah. [SPEAKER_00] I've never been a massive believer of you have to do way flatter because of this and it's just become, yeah, it's going to become a much more efficient organizational flow and information flow. Do you think the average number of people

20:35

SPEAKER_01

that a manager [SPEAKER_00] at Clio manages

20:38

SPEAKER_00

will be different in two years' time versus currently? Yeah. I've never been a massive believer of you have to do a one-to-one with every single one of your directs every day. Yeah, but when you say becoming flatter, do you think that will show up in the numbers? I really hope you're going to have less middle managers and I think orgs are going to definitely become flatter. The only reason I'm caveating myself is you do need to give junior people

21:33

SPEAKER_01

[SPEAKER_00] advice and time [SPEAKER_00] and grow [SPEAKER_00] and develop these people. [SPEAKER_00] So I think [SPEAKER_00] there's always going to be [SPEAKER_00] that need for nurturing

21:47

SPEAKER_00

and maybe it's done in a different way, but I think those layers and layers of people and people and people, it's probably [SPEAKER_01] half the Stripe [SPEAKER_01] team [SPEAKER_01] fall into that category [SPEAKER_01] in the room, [SPEAKER_01] but I think [SPEAKER_01] that has to [SPEAKER_01] change [SPEAKER_01] and maybe [SPEAKER_01] we just get

22:27

SPEAKER_01

to do more stuff,

22:28

SPEAKER_00

[SPEAKER_01] right? [SPEAKER_01] Instead of [SPEAKER_01] more hierarchical [SPEAKER_01] fewer business units, [SPEAKER_01] maybe you just have [SPEAKER_01] way, way, way more bets, [SPEAKER_01] way, way more business units [SPEAKER_01] and it becomes

22:46

SPEAKER_01

much more agentic and much flatter. I think you plausibly have that because you can have more impact from a small number of people.

23:00

SPEAKER_00

[SPEAKER_01] I think the thing [SPEAKER_01] people have been [SPEAKER_01] wondering as well [SPEAKER_01] that I think is very interesting is when you talk about junior people, do you end up with more of an apprenticeship model because it's different if you have one senior person with eight junior people then it's maybe a classic org chart view whereas if you have one senior person with one junior person then you can do much more of a traditional knowledge transmission

23:59

SPEAKER_01

[SPEAKER_00] apprenticeship model. [SPEAKER_00] I think the inverse [SPEAKER_00] is one person [SPEAKER_00] with [SPEAKER_00] I hire all these grads [SPEAKER_00] from

24:17

SPEAKER_00

UCL and Cambridge that have done their machine learning masters. They are so they are fully AI-pilled. They've been doing it from [SPEAKER_01] their A-levels, [SPEAKER_01] right? [SPEAKER_01] It's crazy [SPEAKER_01] how good they are [SPEAKER_01] with these tools. [SPEAKER_01] So I think you're going to want more junior people that have had the learning experience in LLMs from these institutions and are really pilled. And then they are [SPEAKER_01] just dramatically [SPEAKER_01] they're managing [SPEAKER_01] a fleet of agents [SPEAKER_01] to do things. [SPEAKER_01] So in software you want more and more software, right?

25:32

SPEAKER_00

But you want it directed at the right things. So you need great tastemakers that can run lots and lots

25:47

SPEAKER_01

[SPEAKER_00] of these parallel agents [SPEAKER_00] and recursion

25:50

SPEAKER_00

just more and more and more tokens. So you want as many people, I think, as possible at a lower level with good taste to run as many agents as possible. I like the idea that I can take 2,000 bets next year instead of the 10 that I took last year. Yes. That's the world I would love to live in. Be cool. We'll see. Last question. What's your most contrarian AI take at the moment? I'm very AI-pilled, so it's hard to be absolutely contrarian at the moment.

26:45

SPEAKER_01

[SPEAKER_00] I think the [SPEAKER_00] SaaS apocalypse,

26:49

SPEAKER_00

Frontier Labs dominating everything, every tweet is [SPEAKER_01] they've just destroyed [SPEAKER_01] X industry Be cool. We'll see. Last question. What's your most contrarian AI take at the moment? I'm very AI-pilled, so it's hard to be absolutely contrarian at the moment. I think the SaaS apocalypse, Frontier Labs dominating everything, every tweet is, [SPEAKER_01] they've just destroyed X industry with their little wrapper. [SPEAKER_01] There's so much work that goes into building a great business like Stripe or anyone else. [SPEAKER_01] There's so much that you need to know. [SPEAKER_01] How much have you learned about payments that no one in the world needs to do?

27:07

SPEAKER_00

[SPEAKER_01] No one should have to learn. [SPEAKER_01] It is awful. [SPEAKER_01] Too much. [SPEAKER_01] Way too much. [SPEAKER_01] But there's so much of that and so much craft that goes into building a great consumer, business-to-business or business-to-consumer product, that you're going to have, you're still going to have great software companies that are completely vertical and really, really deep. And this idea that someone wrote a prompt and they've just got a frontier model and it's going to kill everything is wildly overblown and investors are just the silliest, most irrational people I've ever met in my life and I can't believe they have all this capital.

27:16

SPEAKER_00

So they are over-extrapolating from the... They have no idea what's going on. They don't have any technical knowledge at all and they are in a world they don't understand so they're in a world of we're going to sit on the fence and just put money into the frontier labs and Stripe apparently and that's all they're going to do. But I think that rationality will play out in a couple of years and you'll see valuations, multiples. The economy is driven by software. It has been for the last 30 years.

27:27

SPEAKER_01

[SPEAKER_00] We want more software companies. [SPEAKER_00] That is where GDP growth is going to come from. [SPEAKER_00] We want more creative destruction.

27:31

SPEAKER_00

That is what AI is going to lead to. The abundance is not a contrarian take but I'm pro-abundance. That's a good note to end on. You heard it here. Lots more creative destruction to come and the story is not yet written and it's not just a small number of firms. They'll be shipping everything useful. With that, that brings our programming for Tour London to a close. So thank you all for coming this year and we are going to roll into a reception which we are having out in the expo hall and we can't wait to see you guys next year. Thanks, Barney. Thanks, Luke. [SPEAKER_01] Thanks for being back. probably to just get consumers first, then monetize,

27:50

SPEAKER_00

but I know, what do you think? Well, I'm curious, like, on the consumer side of things, it seems like many businesses decide that they're going to be one or the other,

28:02

SPEAKER_01

where they're going to sell on a usage basis, like, you know, a normal grocery store, just like charging markup on what you buy,

28:11

SPEAKER_00

or in the U.S., you know, Costco has this famous model where they basically make no money on the actual things you buy there, and then they just sell a $100 a year subscription, it's probably gone up, a subscription, membership subscription, and that is where they make all the money that Costco makes, and they've decided that they're not going to bother making money on anything you buy in the store, they're just kind of a subscription business. And do you guys just view it as both, or are you in the business of selling subscriptions, or are you in the business of selling financial products? Or it is just both? We are in both, but I think, you know,

28:43

SPEAKER_00

the interesting thing there is subscription has been the dominant form factor to monetize because of payment rails as well, right? Because it's actually hard to do microtransactions or lots of transactions and transactions fail, so having a recurring monthly thing and making this a kind of phenomenon

28:58

SPEAKER_01

for the masses works really well for software businesses. But I can definitely see a world in the future where, like, you know, you have stable coins,

29:09

SPEAKER_00

there's no friction, the payment rails actually work, there's no cost. Yes.

29:13

SPEAKER_01

Then you move way more to the Chinese model, right? For self-improvement products, I feel like there's also sometimes a kind of using the sunk cost dynamic in consumer psychology

29:24

SPEAKER_00

where, you know, people say, I am going to buy a gym membership and I am going to get in shape and now that I've spent this money on the gym membership, I have to use it because I'll feel bad if I, you know, spent the money and didn't. Is there any of that dynamic with Clio where people are investing

29:35

SPEAKER_01

in their own financial health? I think every, there's like all this kind of behavioral theory in all of subscriptions like we just added a third tier because that, you know, that just works and we got this middle tier

29:49

SPEAKER_00

and just because we have three, people will pick the middle, like, it just, there's all this kind of weird behavioral stuff with payments and subscriptions

29:57

SPEAKER_01

that you have to optimize and learn about.

30:00

SPEAKER_00

So, yeah, humans are kind of irrational actors and I think you kind of got to learn where they're irrational to help them but also to build a business. Yeah, you guys are kind of like duolingo or something, right? You're like weaponizing people's irrationality for good rather than for evil. It's not right, that in the press, but thanks, John. Harnessing. Yes, there you go. Yeah, there we go. Better. We're just workshopping here. Speaking of people's irrationality, you know, people in poker talk about this idea of going on tilt where, you know, people, you know, may mostly have a very solid betting strategy and then they just, they break at a certain point

30:39

SPEAKER_00

and they start making irrational decisions, you know, maybe after a losing streak or something.

30:44

SPEAKER_01

Does that also happen in behavior you see in the app and you're trying to kind of get people back on the straight and narrow? It's human psychology and it's the biggest thing, right? Like, money is such a personal, such an emotive, such a, you know, psychological thing. People will lose a job. People will have a breakup. Like, we have all these user research interviews where you see someone, you know, go completely on tilt and, you know, do things that are, you know, slightly irrational there. So, yeah, we're definitely trying to help in those circumstances, but it's, yeah, that humans are humans in a way as well.

31:20

SPEAKER_00

But there's all these interesting mechanisms and features that we build around that. You started Clio before the current wave AI, but you're still a kind of small and fast-growing company. How AI native is Clio in how it actually works? What ways of working do you guys have that we should be borrowing from you guys or people in the audience should be borrowing? Well, my day-to-day is dramatic. Like, since probably January, February, I've, I spend my time completely differently. Oh, please do tell.

31:49

SPEAKER_01

Give us the before-after. Yeah, well, the before was 500 people and you'd have all these division leaders and you'd talk to the division leaders and you'd go farm for information and give them your opinions and you'd have the pre-reads and the way of getting information in a company was to talk to all of these leaders and then give them

32:03

SPEAKER_00

your feedback and do that. You don't need to do that in an organization. I think one of the interesting things that you probably won't want to talk about is, like, how you can understand what is happening in your organization day-to-day. You know their Slack messages. You know Notion docs. You know every single PR. You know what is built. So I understand if software is the output of all these companies, I know what we're shipping in a really visceral, clear way

32:26

SPEAKER_01

which I didn't know before LLMs. And just very practically, how are you synthesizing all this information

32:31

across the company? Like, are you using a third-party tool? Do you guys build internal tools?

32:35

SPEAKER_00

Like, how are you scouring the Slack channels and kind of ingesting

32:38

SPEAKER_01

all the docs? That workflow that Carpathie described is what I use for so much stuff. I just have LLMs that go over, you know, say every PR and it summarizes what that PR does and it grades it and it tells me what's going on in that PR and it just builds a knowledge base by going over all these things. Goes over all the Slack messages and all the channels, builds up a knowledge base of what is happening.

32:59

SPEAKER_00

So you take lots of disparate information across your entire org. Yes. You build it up into a higher level. And it's assembling a sort of captain's log, like it's going around the organization with its clipboards taking notes. On like this dramatic scale that you could never have done. And then you're querying the log. Well, you get humans also to input into that log. Their pre-reads, their dashboards that they're making, everything they do, all goes into logs and becomes more synthesized and aggregated over time. Are you doing that? I mean, I'm not doing specifically that, but I think one of the interesting things

33:31

SPEAKER_01

that all organizations are focused on right now is like consumer AI tools are phenomenally useful and they are trained on the kind of corpus of information that's out there and they have web search tools and everything. And the companies are trying to internally open up the data,

33:49

SPEAKER_00

but not too open, you know what I mean, with all their permissions and security tools and everything. And so I feel like that is the thing that all organizations

33:55

SPEAKER_01

are reckoning with right now. But it's really, as you say,

33:59

SPEAKER_00

I mean, does it drive your internal leaders up the wall? The fact that you are now going straight to the source for information? My engineering managers hate me. Like, honestly, like, I know. I know who's shipping. I know, like, who is the highest velocity, highest code quality. I know which teams, because it's a collection of engineers, are actually doing great work. Where before, I had this kind of, it was this weird view filtered through all these leaders that had all their different biases and all these different incentive structures. You can actually get really rational, clear thought from these LLMs and build it up with context. So, yeah, it's like,

34:33

SPEAKER_00

I think the middle layer has, probably at Clio, definitely hates me,

34:38

SPEAKER_01

but I think orgs have just become way flatter because of this and it's just become, yeah, it's going to become a much more efficient organizational flow and information flow. Do you think the average number of people that a manager

34:48

SPEAKER_00

at Clio manages will be different in two years' time versus currently? Yeah. I've never been a massive believer of, like, you have to do a one-to-one with every single one of your directs every day. Yeah, but when you say becoming flatter, like, do you think that will show up in the numbers?

35:06

SPEAKER_00

I really hope you're going to have less middle managers and I think orgs are going to definitely become flatter. The only reason I'm caveating myself is, like, you do need to give junior people advice and time and grow and develop these people. So I think there's always going to be that need for nurturing and maybe it's done in a different way, but I think those layers and layers of people and people and people, it's probably

35:32

SPEAKER_01

half the Stripe team fall into that category in the room, but I think that has to change and maybe we just get to do more stuff, right? Instead of, like, more hierarchical fewer business units, maybe you just have way, way, way more bets, way, way more business units and it becomes much more agentic and much flatter. I think you plausibly, yeah, have that because you can have more impact from a small number of people. I think the thing people have been wondering as well that I think

35:58

SPEAKER_00

is very interesting is when you talk about junior people, do you end up with more of an apprenticeship model because it's different if you have, like, one senior person with eight junior people then it's maybe kind of a classic org chart view whereas if you have one senior person with one junior person then you can do much more of a traditional kind of knowledge transmission apprenticeship model. I think the inverse is like one person with, like, like, I'm so, I hire all these grads from, like, UCL and Cambridge that have done their machine learning masters. They are so, they are, like, fully AI-pilled. They've been doing it from,

36:31

SPEAKER_01

like, their A-levels, right? Like, it's crazy how good they are with these tools. So I think you're

36:37

SPEAKER_00

going to want more junior people that have had the learning experience in LLMs from these institutions and are really pilled. And then they are

36:44

SPEAKER_01

just dramatically, you know, they're managing a fleet of agents to do things. So, like,

36:49

SPEAKER_00

in software you want more and more software, right? But you want it directed at the right things. So you need great tastemakers that can run lots and lots of these parallel agents and recursion just more and more and more tokens. So you want as many people, I think, as possible at a lower level with good taste to run as many agents as possible. I like the idea that I can take 2,000 bets next year instead of, like, the 10 that I took last year. Yes. That's the world I would love to live in. Be cool. We'll see. Last question. What's your most contrarian AI take at the moment?

37:28

SPEAKER_00

I'm very AI-pilled, so it's hard to be, like, absolutely contrarian at the moment.

37:35

SPEAKER_00

I think the SaaS apocalypse, like, Frontier Labs dominating everything, every tweet is,

37:40

SPEAKER_01

like, they've just destroyed X industry with their little wrapper. There's so much work that goes into building a great business like Stripe or like anyone else. Like, there's so much that you need to know. How much have you learned about payments that no one in the world needs to do? No one should have to learn. Ever. Like, it is awful. Too much. Way too much. But there's so much of that and so much craft that goes into building a great consumer, business-to-business or business-to-consumer product, that you're going to have, you're still going to have great software companies that are, like,

38:08

SPEAKER_00

completely vertical and really, really deep. And this kind of idea that someone wrote a prompt and they've just got a frontier model and it's going to kill everything is wildly overblown and investors are just, the silliest, most irrational people I've ever met in my life and I can't believe they have all this capital. So they are kind of over-extrapolating from the... They have no idea what's going on. They don't have any technical knowledge at all and they are. They are in a world of they don't understand so they're in a world of we're going to sit on the fence and just put money into the frontier labs and stripe apparently and that's all they're going to do. But yeah,

38:40

SPEAKER_00

I think that rationality will play out in a couple of years and you'll see valuations, multiples. The economy is driven by software. It has been for the last 30 years. We want more software companies. That is where GDP growth is going to come from. We want more creative destruction. That is what AI is going to lead to. The abundance is not a contrarian take but I'm pro-abundance. No, that's a good note to end on. You heard it here. Lots more creative destruction to come and the story is not yet written and it's not just a small number of firms. They'll be shipping everything useful. With that, that brings our programming for Tour London to a close. So thank you all

39:17

SPEAKER_00

for coming this year and we are going to roll into a reception which we are having out in the expo hall and we can't wait to see you guys next year. Thanks, Barney. Thanks, Luke.

39:28

SPEAKER_01

Thanks for being back.

Reading tools

Type to find a passage

Appearance
Ask this transcript

Notebook

Export Markdown

Add a note