Summary

Cluely is betting that the winning consumer-AI interface is an always-on desktop layer, not a destination chatbot. Its translucent overlay has access to system audio and the microphone, then supplies notes, definitions, questions, and suggested replies; native Mac screenshots omit it, although Mac screen recordings capture it. The strategic objective is a “land grab” for the Command-Backslash habit, eventually accumulating enough context for a multimodal model to reason over “one year of everything” and become a personalized Jarvis.
Roy Lee’s doctrine is AI maximalism: use AI whenever it improves output because today’s models are “the stupidest models we will ever use.” He rejects a blanket rule requiring AI to identify itself, calling it unenforceable and arguing that AI-assisted economic work should generally be judged by results. He makes a narrower exception for cases where human effort itself is the intended output, such as a sentimental gift.
The company’s mission rests on a highly binary view of AI progress. In Lee’s upside case, superintelligence eliminates cancer and Alzheimer’s, extends lives from roughly 80 to 800 years, makes output effectively infinite, and frees people to pursue intrinsically meaningful work; in the fallback, AI merely automates “10–20% of white-collar jobs” and leaves life about 20% better. He admits these projections deserve “less than a grain of salt,” yet argues that even a 10% productivity lift would equal roughly “700 million human beings’ worth of productivity.”
“Cheat on everything” resonates because Lee sees young people losing faith in the school-to-prestige-job bargain. He claims 95–99%, “if not fully 100%,” of Columbia students have used AI to cheat on an exam or assignment, while graduates face what he described as a 30% unemployment rate. His own conversion story—publicly using an invisible assistant to get an offer through Amazon’s interview process, being suspended by Columbia, and saying he “would have rather died than apologize”—turns institutional conflict into a founding story for Cluely.
Lee wants employers to replace proxy tests and one-page résumés with AI analysis of demonstrated work. An AI should inspect a candidate’s portfolio, verify claims, identify precise deficiencies, and perhaps score someone as a “76 out of 100 candidate”; any assignment AI can complete should disappear. His qualification is important: “cheating” cannot remove every learning requirement, so people should take every AI-enabled jump, then backfill whatever mathematics, coding, or domain knowledge the final unsolved step demands.
Erik Torenberg’s strongest pushback is that output alone cannot govern romance, gifts, public discourse, or trust. Lee ultimately concedes that some outputs are valuable precisely because of “human input”: an AI-generated Father’s Day drawing passed off as handmade would fail because effort was the intended product. His narrower conclusion is that undisclosed automation should generally be acceptable for economic production, while sentimental connection remains an enduring exception.
Cluely is pairing countercultural positioning with unusually deliberate distribution and talent spending. Its in-house team uses human videographers, lighting, cameras, and editors because AI cannot yet match their output, while using AI for scripts and storyboards; separate sub-10-second UGC clips can reach 5 million views among 16–20-year-olds. Lee advertised packages up to $1 million plus equity, reasoning that premium compensation wins attention first and an intense “frat house” culture earns loyalty later.

Deep dive

1. Cluely makes AI a pane of glass over every application

Lee defines Cluely as “the new way that humans will interact with AI”: a desktop app with system-audio and microphone access that appears as translucent glass over other applications. During meetings it takes notes, surfaces definitions, proposes questions, and supplies suggested replies without requiring a trip to chat.openai.com.
Torenberg’s hands-on verdict preserved both promise and limitation. Cluely produced a useful “running diary” during his conversation with Waymark’s CEO, and repeatedly refreshing its suggested replies worked “often enough,” but the model lacked the unspoken context accumulated across their relationship of more than 10 years.
The “undetectable” feature has a precise boundary: a native Mac screenshot does not show the overlay, but a Mac screen recording does because it operates lower in the OS. Lee calls undetectability “just a for-fun feature,” while acknowledging its utility for a salesperson who wants live assistance during a shared-screen demo.
On retention, Lee says Cluely normally stores an AI summary or notes rather than the MP3; enterprise customers may specifically request audio storage. It also saves triggered questions and responses, but he says an encryption layer prevents the company from knowing which user made a particular request.

2. The initial moat is interaction habit; the endgame is total context

Lee expects a natively multimodal model eventually to ingest not two minutes but “one year of everything you’ve done on your computer.” Reasoning across that history would turn a generic assistant into a hyper-personalized Jarvis that understands relationships, work, and preferences.
Cluely’s present objective is therefore behavioral lock-in: “instead of going to chat.openai.com, press Command-Backslash.” Lee calls this a “land grab” to establish the desktop-assistant interaction pattern before models acquire the context and intelligence required for the full vision.
Lee presents Apple’s translucent Liquid Glass interface as evidence that the market is converging on this UX. His distribution thesis is that once users are locked into the interface and Cluely has their data, it can eventually distribute a superintelligent open-source model even if it does not own the frontier model itself.

3. Superintelligence is meant to end necessity, not human striving

Lee’s analogy is a blacksmith shown a steam engine 600 years ago. The blacksmith can foresee job loss and the erosion of craft but cannot imagine future occupations; likewise, today’s workers may see automation while missing accelerated science, new activities, and expanded life.
In Lee’s optimistic case, an intelligence that performs cognition better than humans cures cancer and Alzheimer’s and changes death from roughly age 80 to age 800. Capital creation and working for economic value then recede because anything—from information to a hamburger—can be supplied almost immediately.
That abundance does not make people inert. Lee points to chess becoming more popular after machines surpassed humans: people undertake difficult activities for the experience, not merely the output. Coffee shops, woodworking, literature, conversation, nature, and gardening remain because people choose their processes.
He goes further, imagining biological control over reward itself: people might disable dopamine from four hours of doomscrolling and redirect it toward running or reading English text from the 1800s. Lee says “99% of humans” would choose healthier, effortful pleasures, though he concedes that if superintelligence never arrives, AI might automate only 10–20% of white-collar jobs.

4. Cluely’s theory of change is accelerated adoption

Torenberg presses the apparent entrepreneurial paradox: why build a generational company if superintelligence for everyone will wash away conventional moats and legacies? Lee answers that the future is inevitable only because people work “tirelessly to make that future an eventuality.”
Even popularizing the translucent overlay would feel like a worthy contribution. Lee contrasts that mission with the life he might otherwise have led—interning at Meta as a “brain-dead software engineer,” fearing for his career, and preparing for junior-year summer classes at Columbia.
His quantitative shorthand is that making everyone 10% more productive would inject the equivalent of “700 million human beings’ worth of productivity” into the world. Greater usage could also funnel more money toward AI R&D, normalize adoption, and increase optimism about the technology.
Lee nevertheless tells listeners to take his five-, 10-, and 20-year forecasts with “less than a grain of salt.” Nobody knows the path; his operating rule is simply to move toward the future he believes could cure disease and release people from unwanted work.

5. Institutional disillusionment powers “cheat on everything”

Lee argues that young people do not want to become “a cog in the wheel” producing capital without meaning. The old promise—attend a good school, earn good grades, secure a safe job—looks broken to a generation he says sees college graduates confronting a 30% unemployment rate and accelerating AI disruption.
He casts himself as “the path not trodden,” visibly rejecting both Ivy League prestige and the respectable Big Tech route. The appeal is partly aspiration, partly frustration that years of memorization purchased neither stability nor a convincing social purpose.
From his “report from the front,” 95–99%, “if not fully just 100%,” of Columbia students have used AI to cheat on an assignment or exam. He distinguishes AI-assisted take-home essays—which he calls functionally over—from tightly supervised classroom exams, which remain difficult to cheat on but may disappear within five years.
Universities have largely delegated policy to individual professors, while the “AI hive mind” among students has already normalized usage. Lee’s prediction is that institutions are unprepared for these students to enter high-profile jobs five years from now as a genuinely AI-native generation.

6. The Amazon stunt converted a hiring critique into a founding myth

Interview Coder, Cluely’s precursor, photographed programming questions, sent them to ChatGPT, and displayed answers in an overlay invisible to screen sharing. Lee used it throughout Amazon’s interview process, got an offer, recorded the deception, and posted the footage publicly.
His substantive critique is that programming riddles are already online, turning interviews from tests of critical thinking into contests over who memorized the most puzzles. He considered the stunt “the equivalent of how a computer science major conducts a protest.”
Lee says an Amazon executive threatened that Amazon would stop hiring from Columbia unless it expelled him; Torenberg says he would need Amazon’s comment before accepting that characterization. Columbia subjected Lee to disciplinary hearings that ended in suspension.
Although offered opportunities to recant, Lee says he “would have rather died than apologize.” His justification was that engineers would use AI for the real work anyway—he cited Google’s reported 90% figure—and an apology would have diluted both the story and his ability to lead the change.

7. AI maximalism judges economic work by output

Lee’s personal code is simple: use AI at every opportunity where it helps. “The models we use today are the stupidest models we will ever use,” so refusing present assistance makes even less sense if the same task will plainly be automated within five years.
That framework makes overemployment acceptable. If someone can satisfy several companies’ output expectations across multiple remote jobs, Lee says the companies “themselves don’t care”; employers must define sufficient output rather than police hours. Cluely similarly sets demanding deliverables without strict schedules.
Torenberg extends the logic to eight laptops running eight Cluely agents and eventually removing the employee from the loop. He then tests Yuval Noah Harari’s proposed rule that an AI must always identify itself as AI.
Lee rejects the rule as unenforceable and conceptually unstable: brainstorming, outlining, drafting, and final editing form a continuum with no clean point where work becomes “AI-generated.” An automated customer-support call should be judged by its output, not whether a human performed a job that few people intrinsically enjoy.

8. Human effort remains valuable when effort is the product

Torenberg’s hardest counterexample is intimacy: would Lee accept a date secretly taking every conversational line from an AI, or spouses speaking through assistants? He invokes a child’s imperfect Father’s Day art and a Google Olympics advertisement criticized for using AI to perfect a child’s letter—the “thought” matters independently of polish.
Lee concedes the distinction. A child’s drawing can remain meaningful even when AI can instantly make a superior image, because the desired output was never visual quality; it was the child’s effort. Passing off an AI image as handmade would therefore be “poor output” under the real criterion.
The resulting boundary is narrower than “cheat on everything” first implies. Economic production can be automated without disclosure when recipients care about the result, but sentimental acts and human connection survive because “the actual output that you’re wanting there is human input.”

9. Assessment should inspect real work and learning should chase goals

Lee would retire any assignment an AI can perform. Instead of a programming riddle, employers should inspect everything a candidate has built; an AI could review the code, verify proficiency, identify gaps, and conclude that someone is, for example, a “76 out of 100 candidate.”
He applies the same critique to one-page résumés, which compress careers into gameable claims filtered through a reviewer’s mood and attention. A better process might let candidates speak extensively with an AI that asks personalized questions, verifies experience, and evaluates the cadence, persuasion, or technical skills an employer actually values.
“Cheat on everything” does not eliminate foundations. If the goal is building Google, AI may bridge many steps, but the final unsolved jump could require backfilling addition, matrix multiplication, linear algebra, or PageRank. Lee’s rule: take every available shortcut, then learn whatever prerequisites technology cannot cross.
For children, he prefers goal-driven exploration to isolated drills. A seven-year-old inspired by Harry Potter could use AI to generate and compare story ideas, discover plot preferences, and complete a larger creative project; the child might know less about semicolons but learn more through sustained experimentation.

10. Trust may migrate from factual fluency to human intention

Torenberg worries that universalized cheating lowers trust across political discourse, Fiverr, Upwork, and other remote markets: a participant may be an AI, a hybrid, or a troll farm optimizing for wasted attention rather than shared understanding.
Lee counters with real-time English-Spanish translation, then a conversation with a coal-mining expert where Cluely supplies unfamiliar definitions and historical context. Neither assistance necessarily deceives; both can remove memory and language barriers so the participants reach the subject they actually wanted to discuss.
His claimed limit is desire itself. AI can recommend or inform, but cannot replace the human impulse behind “You think this girl is pretty; go talk to her.” Conversations should increasingly strip away fact retrieval and reveal preferences, opinions, and reasons two people chose to engage.
Torenberg tests this against his own excitement about live translation: whether Lee raised the example randomly, researched Torenberg’s Twitter history, or received a runtime prompt mattered less than expected. Lee says effort signals will simply relocate, as beautiful lettering ceased implying calligraphic labor once typing became commonplace.

11. Distribution, cultural escape velocity, and talent are Cluely’s bets

Cluely’s polished videos remain mostly human-made: an in-house studio supplies cameras, lighting, videographers, and editors because AI cannot yet match the desired output. AI helps storyboard shots and draft comedic scripts, while a team “sufficiently brain-rotted” by four daily hours of Instagram and TikTok knows what makes viewers stop.
Its UGC engine is different: fresh accounts, a random 20-year-old, roughly two sentences, and clips under 10 seconds. Lee says these videos can generate 5 million views and convert strongly because Cluely has refined hooks for 16–20-year-old students—the reason some users cannot escape the campaign while Torenberg rarely sees it.
Long-term brand posture depends on whether this marketing changes corporate culture. If it helps kill buttoned-up professionalism, Cluely will preserve its “do whatever I want” identity; if it merely supplies escape velocity, Lee allows that the company may eventually “button up.” He insists the durable ingredient is honesty, not underdog status.
Lee advertised compensation up to $1 million plus equity. His logic is that a two-month-old startup cannot demand belief before candidates experience it: top-end pay creates the funnel, then a steak-eating, gym-going “frat house” culture and ambitious equity story retain the specific full-stack engineers Cluely needs. His closing advice: “the riskiest move is probably what you determine to be the safest move.”

The Fractured Entangled Representation Hypothesis (Intro)Big Beautiful Bill, Elon/Trump, Dollar Down Big, Harvard's Money Problems, Figma IPO