Cliff Weitzman on Speechify, AI Tools and Learning Through Iteration

Speechify founder and CEO Cliff Weitzman joins Harry Stebbings to discuss learning, advertising, AI engineering, management and investing. Weitzman’s account includes personal health experiences and company figures; those statements are not independent medical evidence, audited results or individualized advice.

Weitzman recalls moving to the US at 13 without speaking English and applying to 26 colleges instead of the six suggested by his counselor. He wrote 48 drafts of his main essay. He describes the experience as an early expression of his belief in taking many shots and doing repeated work.

Weitzman says dyslexia made reading and spelling difficult for him as a child. He describes building a text-to-speech system and says Speechify now reads the equivalent of millions of books annually. His usage figure is a company claim, and his personal experience does not establish that the product treats or cures dyslexia or ADHD.

He recalls writing the same brief message in classmates’ yearbooks to avoid spelling mistakes, only to discover that he had misspelled it. Years later, he says, listening to his college paper with Speechify helped him identify spelling errors. He describes that moment as personally meaningful rather than a clinical outcome.

Stebbings asks what Weitzman learned from seeking out leaders of consumer subscription companies. Weitzman says he reads widely, identifies experts and contacts them repeatedly but politely. He says several people he met through this process became early investors in Speechify.

When a CEO did not respond, he says he would reach out to a marketing or growth leader, request introductions and follow up until he received a reply or was asked to stop. He used shared interests such as books to build a connection, then sometimes traveled to observe a team’s work firsthand.

Weitzman describes growth as a competition for attention in which creative content and distribution both matter. He says he learned by observing the people directly buying ads and editing content, rather than relying only on senior leaders. His examples reflect his experience at Speechify, not a universal recipe.

He recalls advice from a Blinkist founder to concentrate on Meta before expanding to other paid channels. Weitzman followed it at the time, but the quoted spending threshold is a specific heuristic, not a guarantee that Meta will be the right channel or budget for another business.

Weitzman tells the story of a growth leader who had spent weeks planning hires without producing ads. He asked the person to source candidates and make and buy ads directly. He values leaders who can still perform the work themselves, rather than waiting for a larger team to execute.

For hiring, Weitzman says he looks for enthusiasm for the product, loyalty to the team and an ability to learn and ship. He describes rapid replies and fast review of contributions across teams as ways to unblock colleagues. These are company expectations and personal heuristics, not validated predictors of individual performance.

Weitzman compares a company’s growth and profitability phases to bulking and cutting. He says Speechify took about four and a half years to reach product-market fit and that conversion and renewal metrics rose around 2020. The timeline and metrics are his recollection, not independently audited results.

During a period when growth fell short of his hopes, he tracked his time in half-hour increments and decided to shift from reading about growth to executing. Later, he says, the focus moved toward margins and profitability. He believes pursuing rapid growth and cutting costs simultaneously can divide attention.

Weitzman says Speechify had run profitably for roughly four and a half years before entering another growth phase. He distinguishes direct customer acquisition cost from broader investment in creative experimentation and engineering. His profitability and spend figures are self-reported and do not establish present margins.

He says he will fund experiments in creative, channels and engineers but dislikes spending on ads he cannot attribute. The goal is to learn which expenditure works, not simply accept higher acquisition costs. The example is his operating philosophy; attribution itself can be imperfect.

Weitzman describes creator whitelisting: ask a creator for multiple videos, test them as paid ads and potentially distribute the successful versions through the creator’s account. He says different audiences respond to different presentations and considers the creative format a variable to test.

He reports testing nearly 1,000 AI-generated ads a day and making thousands of pieces with human creators each month. He says his team changes a performer’s apparent age, identity or background to produce variants. These are the founder’s figures and practices; the interview does not evaluate how the ads are labeled or perceived.

Weitzman says he built an internal ad-generation platform after studying the field and hiring a dedicated team. He describes automating distribution and reporting while trying different models as they change. His characterization of the tools and their results is a company account rather than an independent comparison.

For each batch, Weitzman says the team looks at click-through rate, cost per action and cost per thousand impressions. Ads that acquire paying customers below an internal threshold may enter a larger campaign; others are dropped. Some days, he says, no candidate passes the test.

Weitzman describes using AI tools to organize information, including family medical records and research, as well as business research and a map of school districts. These are his personal uses and do not verify the tools’ outputs or establish the privacy protections of any particular workflow.

Weitzman recalls studying hundreds of successful advertisements and rewriting their structures for Speechify. He describes a video featuring red headphones and a suit in a hot tub, followed by dozens of edits and tests. He says one version performed for years; the impression and revenue figures are his company’s account.

He says another video, a rap about his own dyslexia and ADHD, generated about $3 million, while an ad he expected to perform well did not. His lesson is to test rather than assume a creative idea will convert. The revenue attribution is self-reported and the example does not establish likely results for others.

Weitzman recommends reusing a successful format across topics and platforms while changing the hook. He says paid distribution permits many variations to be tested, and Stebbings asks whether a video with views but little conversion is worthwhile. Weitzman says his priority is conversion when the goal is funding paid acquisition.

He says Speechify offers a free experience as well as paid subscriptions, and that paying customers fund advertising. He describes the company’s accessibility mission alongside its commercial model. Claims about user numbers and the prevalence of different conditions are his estimates, not verified clinical or population statistics.

Weitzman says growth activities sit under one organization with specialized subteams. He describes word of mouth as the largest source of users and cites Apple features, search and referrals from ChatGPT as additional channels. The percentages and channel performance are his company’s claims, not externally verified data.

He predicts advertising in AI assistants could become substantial and says Speechify is testing a new channel. He speculates about platforms’ knowledge of users and targeting. The interview does not establish what private messages a platform uses for advertising or validate his claims about its data practices.

Asked about Meta stock, Weitzman expresses a bullish personal view based on leadership, distribution and data. He later gives an opinion on Figma. These are speculative investment opinions from the recording, not verified valuations or recommendations tailored to a reader’s circumstances.

Weitzman clarifies that by design becoming easier he means some production work in design tools, not the disappearance of taste or user research. He says Speechify still hires designers and uses them to improve products after an initial prototype. He regards selecting and refining an experience as important.

He says Speechify moved from heavy Cursor use toward Claude Code and asks engineers to use AI coding tools aggressively. He predicts that next year token spending could exceed engineering salaries at his company and that others may eventually follow. This is a forecast for engineering, not a reported company-wide result.

Weitzman describes internal demonstrations and asks team members to explain why they are not using the tools. He says he sometimes sets extreme expectations to encourage adoption. These are his management practices, not proof that a particular usage quota improves work or is appropriate for every employee.

He lists Speechify products he says include text-to-speech, dictation, podcast and teaching tools, and a voice assistant. He says users need education and repeated iterations before adopting features. Those product and adoption claims come from the founder and are not independently evaluated in the interview.

Weitzman describes using AI-assisted research to organize information about his brother’s health and identify possible genetic findings. This is an unverified family account, not evidence that an AI system can diagnose or interpret a person’s condition reliably.

He also recounts his father’s prostate-cancer history, a disagreement about scan timing and his search for a higher-resolution scan. He says he contacted a research center, his father was scanned and later underwent surgery. The interview does not include medical records, comparative evidence or enough information to judge the recommended care.

Weitzman says language models gave him confidence to ask questions and seek another scan. He believes patients should learn and participate actively in decisions. His account does not establish that earlier imaging or surgery was medically superior in general, or that AI output can replace qualified clinical judgment.

Returning to hiring, Weitzman calls persistence through difficulty “adversity quotient.” He says he looks for projects candidates actually shipped and an ability to keep working through hard problems. This is his personal hiring framework, not a validated test or a reason to infer someone’s capability from a single trait.

He asks candidates to describe systems they have altered or projects they have built, and looks for evidence of learning and initiative. He says misleading statistics and long, unfocused answers concern him. His preferences reflect how he evaluates his company’s roles and do not establish a general measure of honesty or talent.

Weitzman prefers candidates who can explain a connection to Speechify’s mission, such as personal use or a family experience, over those motivated primarily by money. He reasons that mission may support retention. That is his hiring preference, not evidence that candidates seeking better pay are less loyal.

Discussing dyslexia and ADHD, Weitzman describes his own experience and speculates that repeated obstacles can shape resilience. He cites striking percentages about incarcerated people and billionaires without supplying sources. The interview does not verify those figures or establish that any diagnosis predicts entrepreneurial success.

Weitzman says some people with ADHD may find a broad leadership role stimulating while early individual work can be harder to focus on. He also offers a high-speed listening example. These are his observations and hypotheses, not clinical guidance or findings that apply to all people with ADHD.

Asked about emerging roles, Weitzman highlights coaching colleagues to use AI. He reports application and technical-challenge numbers for Speechify and says he regrets not hiring a particular coach after failing to build internal consensus. Those recruiting figures are his account and not independently audited.

Weitzman says he would pay more for someone he trusts and is willing to hire quickly, then part ways if the role does not work. Stebbings similarly discusses trusting early interview signals. These are individual hiring judgments; neither offers evidence that rapid decisions or intuition reliably predict performance.

Weitzman says his own experience hiring from Google and different geographic labor markets has shaped his preferences about hunger and retention. His broad claims about former employees and regional loyalty are unverified generalizations, not evidence about individuals or a fair basis to assume any candidate’s behavior.

He recalls an employee considering leaving because of a lack of local friendships. Weitzman says he organized dinners and introductions, after which the person stayed. He presents this as an example of looking for an underlying need, not proof that such an approach will retain everyone.

He describes helping another employee with immigration paperwork and visiting an employee in Ukraine before the war intensified. He says he tries to understand the reason someone wants to leave and address it. These are his personal accounts; the episode does not independently verify the events or his broader attrition claims.

Asked about layoffs, Weitzman says he worries about unequal distribution of AI’s gains and does not know how to reskill everyone. He believes entrepreneurship and practical projects may help some workers. His claim about the availability of jobs for graduates is his impression, not a labor-market measurement.

He says he values experienced engineers who understand systems and can adopt new AI tools, while also hiring talented younger candidates. Speechify’s engineering challenge, he says, tests the use of coding agents as well as whether the candidate can find and repair breakage.

Weitzman describes Speechify as remote-first and gives team-size and country counts from the time of recording. He asks employees to acknowledge requests quickly so colleagues know whether to wait or find another solution. The precise 60-second expectation is his company’s practice, not proof that response time predicts engineering success.

He says a reply can simply provide a time when help will be available, and encourages people to communicate priorities to their leaders. Stebbings describes his own three-hour response rule, including weekends. Both frame rapid communication as valuable; they do not measure the impact on focus, rest or work-life boundaries.

To increase speed, Weitzman proposes explicit deadlines on shared calendars and early updates when a deadline will slip. He asks for a short demonstration when work ships. He believes leaders should model the same behavior, though the approach may have different trade-offs across teams.

Weitzman recalls two hires from large technology companies whom he believes did not fit Speechify’s culture. He regrets not addressing the situation sooner. This is his view of an unnamed personnel episode, not evidence that leaving work at a particular time or coming from a particular employer indicates poor performance.

He prefers a brief call to a long Slack discussion when immediate clarification is needed, while recognizing that text is useful asynchronously. He also describes an internal voice agent. His preference for synchronous communication reflects his style and does not establish that it is faster for every task or team.

Weitzman criticizes formal performance reviews and favors clearly communicating an expected outcome and deadline. His example asks a developer to reduce the scope of a feature and bring it through review and quality assurance. His proposed timetable is a management preference, not a universal evaluation standard.

He argues that quality assurance becomes more important when AI can generate code quickly. He wants people to test across devices, connectivity conditions and user journeys, and to report bugs with recordings. His claim that QA is the most valuable skill is an opinion rather than a measured ranking.

Weitzman says an engineer who struggles to own a feature might still contribute in QA and be coached before a decision to part ways. He values identifying edge cases and fixing issues before release. That description does not mean every feature can or should be shipped on a fixed short timetable.

He dislikes long scheduled meetings and says the team shortened recurring performance conversations. He prefers leaders who remain hands-on and receive brief updates about priorities. The approach is his description of Speechify’s culture and may affect different roles and employees differently.

Weitzman says he wants employees to identify blockers and ask leaders for help, rather than wait for directions. He describes giving talented people additional responsibility and believes encouragement can help them develop. He also acknowledges that sometimes people need further support when they do not initially succeed.

On fundraising, Weitzman says Speechify grew slowly for four and a half years and had passed $5 million in revenue by that point. He names early investors but declines to disclose the amount raised or the company’s valuation. His figures and investor list are his account, not audited financial statements.

Weitzman cites large user, review and market-share figures for Speechify and says the company is launching an enterprise business. The market-share assertion lacks a defined denominator or independently verified source in the interview. He says he prefers discussing customer acquisition to publicizing financing.

Weitzman offers a favorable assessment of Logan Paul, whom he says he knows personally. He describes creator careers as moving between exploration and exploitation, citing entertainment, consumer products and other ventures. His assessment reflects a personal relationship and does not independently verify the ventures’ results.

He suggests Stebbings might reach new audiences by participating in an established show, television format or book distribution network. He describes this as using the reach of a larger institution to amplify work, while acknowledging that building independently is another option.

Weitzman recounts spending time with MrBeast and his team, observing shoots and their attention to visual details. He believes simple concepts that can be understood across languages, repeatable formats and careful hiring help the creator reach audiences. These are his observations, not evidence of future performance.

Asked about changes in views, he says MrBeast’s team has also invested time in streaming, games and consumer products. He is optimistic about its next phase, but the performance figures and projections in the discussion are not independently verified.

Weitzman says his brother Tyler, not he, purchased a leveraged Nvidia option after the family’s experience seeking GPUs. He reports a large gain and says he bought Nvidia stock directly. This is a personal, self-reported trade, not evidence of repeatable returns; leverage and options can magnify losses as well as gains.

Weitzman describes investing when he believes he has context that others may lack, while Stebbings stresses diversification before concentrating on winners. Their views differ in emphasis. The episode’s stock valuations and growth expectations are opinions from the recording, not current prices or personalized investment guidance.

On energy for AI, Weitzman expresses optimism about solar and eventual fusion, while saying he does not know when fusion will become practical. He offers broad claims about hydroelectric capacity, solar efficiency and regulation without evidence in the interview. These are his views, not established limits or investment forecasts.

He describes SolarCity as an example in which financing and customer acquisition mattered alongside technology. He says the firm arranged long-term financing to make installations possible. The precise financing details are his recollection and are not independently documented in the conversation.

Weitzman returns to product-market fit: he recalls watching users navigate early versions of Speechify and repeatedly fixing confusing interfaces. He says he put a help button in the app that reached him directly and called users who had left. The feedback helped him iterate, in his account.

His lesson is to distinguish a weak implementation from an invalid underlying idea, while continuing to talk to users and improve the product. He values a clear product vision but says distribution still matters after launch. Persistence can create opportunities, though his story does not guarantee that every idea will find a market.

Stebbings describes offering a shorter audio edit and a fuller video edit of the show because audio can be cut more precisely. Weitzman suggests tools that might synchronize separate versions. Their exchange illustrates an editorial trade-off and an untested product suggestion.

Weitzman recounts a security incident in which he says a vulnerability let someone use Speechify models without paying, but did not expose user data. He says he had prioritized features over the fix and estimates the loss in the hundreds of thousands of dollars. The scope and loss are his account, not an independent incident investigation.

He says he found the young person involved, asked them to stop and eventually hired them first for personal projects and then for company work after building trust. His story is an anecdote about his response; it does not establish the facts independently or recommend overlooking security responsibilities.

Asked about fast-growing consumer AI businesses, Weitzman says investors should examine inference costs alongside revenue. He recalls that Speechify focused on reducing the cost of serving heavy usage, and claims its own models became much cheaper. The figures and comparisons he gives are unverified and may not be directly comparable across products.

He argues that a compelling product can grow quickly while much of its revenue goes to model providers, making gross margin an important question. He also says growth, community and product differentiation matter. His description of Cursor and other companies’ economics is an interpretation without access to their full financials.

Weitzman suggests that a product serving a recurring need may retain users better than a novelty. He compares Speechify to assistive tools, but his statement that users do not churn is not supported by cohort or retention data in the interview and should not be treated as a universal outcome.

He predicts Apple would build a competing dictation product rather than buy another company, and discusses his experience with platform constraints. His views about Apple’s leadership, plans and technology are opinions and speculative claims, not verified announcements.

Weitzman favors OpenAI in a hypothetical choice at equal valuations, while saying he currently uses Anthropic’s tools more. He praises leadership and distribution, and Stebbings challenges parts of that assessment. The comparison is a speculative preference, not a substantiated investment recommendation or a verified valuation.

Weitzman recounts spending 11 days on a proposed MrBeast project with limited resources. He says the video did not come together and that he initially felt resentful about production problems. His lesson was to notice gratitude and optimism even when a project goes badly; the account is his perspective on the experience.

The experience also made him think about the difficulty of navigating without a phone or stable resources. He says he now gives phones to some people and believes public programs should make access easier. This is a policy preference informed by a short personal experiment, not a comprehensive study of homelessness.

Weitzman worries that AI could concentrate economic gains and says reskilling and education matter. He hopes tools could enable more entrepreneurship and imagines adults returning to education in new forms. He acknowledges that he does not know what the system should look like or whether everyone will adopt the tools.

He says university was personally valuable for friendships, an open curriculum and self-directed learning, and that he would encourage his children to consider it. He also listens to audiobooks and AI-generated research, while imagining future teaching agents. Those aspirations do not establish the accuracy of generated educational material.

Asked about a purchase, Weitzman describes his own protein supplement routine and exercise habits. His comments on protein intake and absorption are personal views, not dietary evidence or individualized health advice. The interview does not evaluate whether that routine is appropriate for anyone else.

Weitzman describes seeking coaching in gymnastics and using tools to generate exercise ideas. He also gives his personal views about body composition and nutrition. The claims about AI assessment and dietary quantities are not independently validated and should not be treated as a health assessment.

Looking ahead, Weitzman hopes people will use AI to learn and act with greater agency. He draws on his own family’s experience while acknowledging the risk that people could instead surrender decisions to the tools. His closing vision is a hope about how technology may be used, not evidence that it can solve every problem.