20VC: What I Learned from 100 of the Best CEOs in the World | What I Learned from Staying with Mr Beast for 3 Weeks | How We Will Spend More on Tokens than Salaries with Cliff Weitzman, Speechify

What It Covers

Cliff Weitzman, founder and CEO of Speechify, details how he scaled a voice AI platform to 60 million users by applying volume-based testing across ads, hiring, and product development. He covers AI token spend surpassing salaries, adversity quotient as a hiring filter, bulking and cutting cycles for companies, and lessons from embedding with top consumer subscription CEOs and MrBeast.

Key Questions Answered

• •Meta-First Ad Spend Threshold: Do not allocate budget to any advertising platform other than Meta until monthly Meta spend reaches $100,000. Below that threshold, diversification dilutes learning and wastes capital. Once Meta is saturated, test AppLovin, TikTok, and OpenAI ads — Speechify is among 200 companies provisioned to test OpenAI ads, and early adoption builds a compounding skill advantage before broader rollout increases competition and CPMs.
• •AI-Generated Ad Volume at Scale: Speechify tests approximately 1,000 AI-generated ads daily alongside 8,000 human-produced creatives monthly, using a proprietary in-house platform built in four days. Ads are auto-posted to Meta, TikTok, and YouTube, with performance tracked by click-through rate and cost per acquisition. Winners enter a main campaign bracket against top historical performers, receiving increased spend only if they outperform existing benchmarks.
• •Token Spend Surpassing Salaries: Speechify is approaching a point where annual spend on AI tokens through tools like Claude Code will exceed total salary expenditure across engineering. Weitzman mandates engineers spend a minimum of 1,000 Claude Code credits daily, publicly tracks usage via screenshots, and runs live screen-share sessions to demonstrate workflows. He predicts most high-quality companies will reach this ratio within three years.
• •Adversity Quotient as Primary Hiring Filter: AQ — how a person performs under sustained difficulty — outranks IQ and EQ as a hiring signal. Weitzman screens for it by asking candidates to show side projects shipped to production, not just built. Engineers who grapple with hard problems for five-plus hours rather than quitting at 30 minutes produce the breakthroughs that move companies. A useful screening question: "What non-CS system have you hacked to your advantage?"
• •Bulking and Cutting Cycles for Companies: Companies should commit to either a growth phase (bulking) or a profitability phase (cutting) for six-month minimum cycles, not oscillate between both simultaneously. Speechify ran profitably for four and a half years before entering a current hypergrowth phase. Attempting both simultaneously is equivalent to pressing the accelerator and brake at once. Revenue growth requires singular focus; any Harvard MBA can cut costs, but scaling revenue demands concentrated obsession.
• •Whitelisting Ads as an Arbitrage Play: Whitelisting involves sourcing niche creators, having them produce 15–20 videos without posting organically, then running those as paid ads to identify top performers by CTR and CPA. Only the highest-converting video gets posted organically from the creator's account with increased spend. This method allows demographic-specific testing — Speechify reskins the same ad with different ages, ethnicities, and backgrounds — without brand consistency constraints, since performance ads run in the shadows.
• •QA as the Highest-Value Skill in an AI-Native Stack: When software engineering and design are commoditized by tools like Claude Code, QA becomes the primary differentiator between a product and a great product. AI coding agents cannot self-QA across devices, network conditions, and edge cases. Weitzman personally finds production bugs and calls engineers immediately with screen recordings. Engineers who cannot ship outcomes but demonstrate strong QA instincts should be reassigned to QA rather than terminated.