Suspended from Columbia, He Raised $15 Million to Cheat on Everything

Nathan Labenz opens with a guest who has been hard to miss. Roy Lee, founder and CEO of Cluely, had just announced a $15 million investment from Andreessen Horowitz after a run of viral videos inviting people to "cheat on everything." Labenz, who often talks about existential risk on this show, says he is usually a fan of edgy marketing and of people who challenge outdated norms. What pulled him into the conversation was not the fundraise. It was a reply Lee wrote to Y Combinator president Garry Tan after Tan tweeted that he was muting the word Cluely.

Lee told Tan he had big dreams for the company. He knew the messaging might make him look like "some stupid kid who's doing something immature and unethical," but he said the end goal was "a giant positive for humanity." That goal, as Labenz frames it, is a utopian world in which people do what they intrinsically want rather than what economic necessity demands. The conversation that follows is Labenz stress-testing that claim, from the overlay product to Columbia, overemployment, and a post-AGI social contract. a16z, he notes, had no role in making the episode; it came together over Twitter DMs before the investment was announced.

Once they are on the mic, Labenz quotes a16z's own blog post, which asked, "What if breaking the rules was the unlock?" Lee's motto is blunter. Asked what Cluely cheats on, he says the company is "the new way that humans will interact with AI." Instead of going to ChatGPT.com, people will use Cluely: a desktop app wired into system audio and the microphone, sitting as a pane of glass over every other window, feeding relevant information in real time.

In a meeting, Cluely takes notes, suggests questions, and floats definitions of terms as they come up, all without the other person knowing AI is in the room. Lee demonstrates live: when Labenz asks what "cheat on everything" means, the panel fills in a definition and even prompts him to ask the question. Labenz had tried the product with the CEO of Waymark, a longtime friend, refreshing a "what should I say now" prompt throughout. Often enough, he says, it was pretty good.

Undetectability is part of the pitch. A Mac-native screenshot will not capture the overlay; a screen recording will, because that path runs lower in the OS. Lee calls it a "for fun" feature that matters when a sales rep is demoing a product and does not want anyone to see the assistant. If nothing else, it is a marketing hook. Labenz's harder product question is context. Most of Cluely's famous use cases are shallow and episodic: a coding test, an interview, a first meeting. Helping in a ten-year relationship, full of unspoken common knowledge, is a different problem, because the model has never had a chance to learn it.

Lee's answer is to wait for the model that can take a year of computer use, reason over it, and become a hyper-personalized Jarvis. Every consumer AI company, he says, is building toward that. Cluely's job is to win the interaction pattern first. Apple's Liquid Glass looks to him like the same translucent overlay, which he reads as the world catching up. Force people off chatgpt.com and onto Command-backslash, collect the data, and when an open-source superintelligence arrives, Cluely is the one that distributes it.

On what is actually stored today, Lee is more cautious than the Jarvis story. Meeting audio is not saved as an MP3 unless an enterprise customer asks for it; the product keeps an AI summary, because a running recording of someone can be illegal. Queries and responses are saved, but under encryption, so the company says it cannot see which user made which request. Labenz treats that as both a technical discussion and a sociological one. The overlay is a product. The slogan is a theory of how norms change.

One of Lee's more interesting lines, Labenz says, is that in the future it will all be legal. Asked which rules and expectations will look anachronistic, Lee starts with data privacy. The moment a technology massively improves efficiency in exchange for less privacy, he thinks people will run toward it. Humans have already shown they value the efficiency of their own work more than almost anything else. If giving up data makes a model more personalized, they will not hesitate.

Copyright and training-data restrictions go the same way. America is in an AI arms race with China, and China, in his telling, already gives its model providers full access to the data it has. The United States will have to ease legal restrictions to keep up. If the result is a much better model for everyone, he thinks people will be happy. Patents and the habit of protecting one's own technology follow. He imagines a world where everything is open source, fully democratized, because that is the only way coding agents improve: they need production codebases from the best companies, and it is the only way to train models that significantly beat China's.

Labenz says a positive vision for the future is the scarcest resource in AI talk, and he asks Lee to spend as long as he needs. Lee reaches for a 600-year-old blacksmith handed a steam engine. The smith can only see his job vanishing, people losing the craft of hammer on metal, critical thinking draining away. He cannot imagine a data analyst. Most people, Lee says, are equally shortsighted about superintelligence. If AI can do everything a human can do, only better, scientific progress speeds up. Cancer is no longer a thing. Alzheimer's is no longer a thing. "You and I don't die at 80 years old, we die at 800 years old."

Capital creation, and the idea of working to add capitalist value, probably gets removed. If you can snap your fingers for a hamburger, what is left? Lee thinks that is when true instinct shows. You will not become a mindless puppet. You walk to get coffee because you want the walk. You run a coffee shop because you want to. You meet interesting people, philosophize, explore nature. Chess is his evidence: when the first bot solved the game, people declared it dead, and chess is now more popular than ever. Default human desire does not devolve into output for the sake of output.

He is also optimistic that science might let people flip what they find joy in. Doomscrolling four hours a day is, to him, a societal danger that decays the brain. If you could switch off dopamine from scrolling and take it from nineteenth-century English literature instead, he thinks every person would want that, and that 99 percent would choose literature, painting, running, gardening — not the spa. That is as close as he can get to utopia. If superintelligence never comes, he adds, everyone has been telling a joke. AI defaults to a tool, maybe 10 or 20 percent of white-collar jobs go away, and life is the same, 20 percent improved.

Labenz finds entrepreneurship almost paradoxical in that world: where are the moats, and who is a generational company for if everyone ends up living their best philosophical life? Lee says the future is not automatic. It arrives because people work to make it an eventuality, and he is excited to help. If his only contribution is planting the translucent glass overlay as the way people use AI, he will be happy. The theory of change is adoption. If the tool makes everyone 10 percent more productive, he claims that is 700 million human beings' worth of productivity. More usage means more money into research, more optimism, and every step required before the utopian state.

Lee's shots at the default prestige path — Ivy League, then a big-tech internship — are not an accident. People root for the underdog, he says, and nobody wants to be a cog producing capital until they die. His generation was told that a good school and good grades would buy a good job. Then college graduates face something he puts at a 30 percent unemployment rate, AI arrives, and a seed of doubt takes hold: maybe the older generation was wrong. He presents himself as a living example of the path not taken. At Columbia, professors set their own AI rules, but he will say with "100 percent confidence" that at least 95 to 99 percent of students have used AI to cheat on an exam or assignment. Claiming you never have would make you the odd one out.

Labenz draws a distinction from his own school years: homework was already collaborative, so consulting AI is not a huge leap, whereas in-room exams were hard to cheat on. Lee splits the difference. Take-home essays are functionally over; most people brainstorm, draft, outline, and edit with AI. Sit-down classroom exams are still hard to cheat on, but he thinks the format is outdated and will be gone from America in five years. The people who will grow into high-profile jobs, he says, have already used AI to massively improve school output. "People are not ready for what happens when these kids get five years older."

The expulsion story is the one already all over the internet, so they keep it brief. In the fall semester Lee had entertained a big-tech job. Those interviews ask the programming equivalent of how many hairpins fit in the Empire State Building, riddles that are all online, so the test becomes hours of memorization. He built Interview Coder, a glass pane almost identical to Cluely, invisible to screen share, that photographs the question and returns the answer. He recorded himself using it through the Amazon process, fooled the interviewers, posted the video, and it went very viral. An Amazon executive, in his telling, told Columbia that if they did not expel him the company would never hire from the school again. Columbia pulled him into disciplinary hearings that ended in suspension.

He was offered chances to apologize. He says he would rather have died. Using AI on a programming assignment while Google reports employees writing 90 percent of code with AI struck him as absurd, and as the computer-science version of a protest. Bowing his head would have diluted the story. He thought he was one of the few people in a position to charge the change in the industry, and he would not have been proud of himself if he had recanted.

Lee's personal code is AI maximalism. Every chance he can use AI where it helps, he should. If a model can do a task today, you would have to be "really stupid" not to think it could do the task in five years. "The models we use today are the stupidest models we will ever use" for the rest of our lives. Adopt that, and a lot of the work he is doing is work he should not have to do. Labenz shares the instinct: he is always trying both to finish the object-level task and to learn how much AI can help. The limit case he wants to test is the Overemployed subreddit, which predates AI, where people hold multiple remote jobs, do the minimum not to get fired, and collect several packages without telling anyone.

Lee is fine with it, 100 percent. Companies hire for a level of output. If you hit that level, why should they care about input? Cluely itself sets no strict hours and strict output requirements. If the output can be done in an hour, you should not have to stay. Overemployment is the company's fault; the onus is on the employer to make sure people perform at a level it is happy with. Labenz pushes further: eight Cluelys on eight laptops, the human barely in the loop, and Yuval Noah Harari's proposed rule that AI must always identify itself as AI. Lee does not think that is a solution. AI and human output are already converging. If you use a model to draft an essay, or only to outline, or only for final edits, at what point is it AI-generated?

A better frame, he says, is that AI is a tool for more output. You do not announce that a hat was made by a Chinese worker, a stitch machine, and silicon; blame infinitely regresses. The only thing that ought to matter is output. Even on a date, if you had a great time, that is the output. Forcing an AI to announce itself is unenforceable and, to him, ridiculous. Customer support is the clean example: he doubts anyone wakes up excited to take those calls, and an AI should do them. The world focuses too much on input and too little on what happens when output 10xes.

Labenz puts the romantic hat on. If a date was later revealed to have an assistant feeding every line, is that fine? Spouses saying what the model tells them starts to feel dystopian. People still believe the thought counts: a child's imperfect Father's Day drawing matters because the child made it. Google's Olympics ad, in which a dad wants a child's letter to an Olympian to be perfect, was widely read as missing the point. Lee agrees there are conversations valued only for sentimental effort, and that those can coexist with AI automating tasks. If a kid passes off an AI card as hand-drawn, the output being judged was never the drawing. It was the input. That, he says, is poor output. Shipping a T-shirt with robots instead of factory workers is the other case: nobody needs a disclosure at every step.

Where does cheating pay, and where does it fail? Labenz speaks for university administrators and people hiring junior developers. The coding test is a bad proxy and everyone knows it, but they cannot hire you for a year to find out. If candidates cheat, the signal dies. Lee has two replies. First, assignments should be obsolete; competency should be judged on output generated in the past. Second, the programming riddle should go. Open-source everything you have ever done. A human or an AI should read the code and the tasks and say, with some exactness, that this person is a proficient TypeScript developer, lacking here and here, a 76 out of 100 candidate. Any assignment an AI can already do should not be asked.

The second half of the answer is where maximalism still requires school. Suppose the goal is to make Google, with every AI tool allowed. A code editor will not take you all the way. At some point the model lacks the knowledge, and "cheating" looks like manually backfilling what it cannot do — maybe as far back as addition, on the way to the linear algebra in PageRank. Even under cheat-on-everything, there are jumps technology cannot make. Completing the last one might take four years of learning. You still take every jump you can, then backfill the prerequisite skills for the jump you cannot. Cheating and learning foundational skills, he says, coexist in the world he is proposing.

Labenz offers his own mantra: you want to learn how to do hard things, but not necessarily the hard way. Do kids still need to write unassisted essays, the way we still run and swim even though we have cars and boats? Lee wants people to be output-oriented. Kids building a sandcastle see the castle in their heads, enjoy the work, and pick up random skills along the way. He remembers being seven, having just read Harry Potter, wanting to write like that. Pen and paper would have produced two pages of seven-year-old prose and the realization that he was not J.K. Rowling. With AI he could generate a million story ideas, draft a million versions, notice where the plot was cooler, and, he guarantees, learn more than he would have by a fire with a pencil — different skills, maybe not the semicolon, but more of them, aimed at the actual goal.

Labenz worries about a longer trend toward lower trust, which predates AI and might get worse. Online, you already wonder whether you are arguing with a model. On Fiverr or Upwork you can still sort of trust a back-and-forth with someone halfway around the world; if they are cheating, is the market even worth entering? He still likes Harari's identify-yourself rule, even while admitting a troll farm will not follow it. Lee answers with a different example. You are speaking with someone in Mexico who only has Spanish, you only have English, and AI live-translates. Is that a loss of trust? Probably not.

Take a small step further. You are talking to a coal-mining consultant and know nothing about the field. Cluely pulls up definitions, so when they mention the Amherst mining incident of 1982 you know what they mean and the conversation can go deeper. That, too, is probably not a loss of trust. It is a way past the random need to memorize facts. People fear a time when AI is more human than humans, romancing their girlfriend or telling them the pool will be more fun than the beach. Lee thinks that fear is wrong. Preferences and opinions are the point. "The ultimate meaning of what it is to be human is to decide the things that you want to do." AI cannot make those decisions. It can only make them infinitely easier.

As usage grows, he expects conversations to boil down to the human essence: what are you actually trying to say, and why are you here talking to me? Two people still made the decision to engage. Discourse about what year World War II started may go away, and he does not think that was discourse. Society will expect people to use AI. It will never replace "you think this girl is pretty, go talk to her." Labenz then notices that Lee's translation example happens to be one he has praised himself: Google's I/O demo of real-time simultaneous translation in Meet, which he called one of the most aspirational product implementations he has seen. Would he rather Lee had found that by coincidence, by scrolling his Twitter, or by Cluely popping it up live? In this case he is not sure he cares, and might even prefer the coincidence.

Lee's read is that Labenz still maps knowledge onto manual effort, because AI lookup is not yet commonplace. A letter that looks like calligraphy used to mean someone spent time making letters beautiful; today it means they typed. When overlays can do what Cluely does, people will disconnect that particular link and reconnect effort to something else — maybe a personal fact you only told them in person. Breaking one link does not undermine "I appreciate it when people put effort into me." On hiring more broadly, Lee would rather an AI go deep on a person's history than trust a one-page resume read in a bad mood. Labenz worries that total memory could trap people on track A. Lee says if the model is that smart it can also decide what we should be allowed to know; if it is not, almost anything beats the current gameable process.

Cluely's content is cinematic, and Labenz assumes an AI maximalist would use AI to make it. Lee says they almost have an in-house film studio: videographers, a serious camera, lighting, editors, everything needed to produce a movie. They do not use AI for much of the actual videography and editing, because the models are not yet at the level they can hit themselves. Output is still the score. What AI is good for is storyboarding each video and helping draft scripts that will be funny. The ideas themselves come from being 20- and 21-year-olds who spend four hours a day on Instagram and TikTok, "sufficiently brain-rotted" to know what lands.

Most tech bros, he says, are too nerdy to scroll. They pride themselves on being intellectual and not funny. Lee has spent the last 20 years on Instagram, so he claims he knows what is engageable better than they do. Money then lets him make it controversial and cinematic, and cinematic has a second benefit: people pause because someone put effort into the thing. Asked if he is not himself a tech bro, he says he is more bro than tech bro — more human, more tech-agnostic, the kind of person who would not be mistaken for one by actual tech bros, even if the outside world files him that way.

On influencer and UGC campaigns, the split is familiar: influencers on one side, user-generated clips on the other. The UGC they run is often not even ten seconds, from fresh accounts, aiming only at "this feels like a real person talking." A viral hook, two sentences about the product, ridiculously simple, and those are the videos that generate five million views and convert best. They have the hooks for 16-to-20-year-olds in school down to a science. Labenz does not see them because he is not the intended customer. Some people cannot get Cluely off their timeline. He has never seen this particular campaign. Lee's promise is that they are going to take over the internet anyway.

Labenz's long-term read is that you cannot stay an insurgent forever. If Lee is right, things that count as cheating now will not later, and the slogan will stop resonating. Is the plan to blitz-scale on today's asymmetric advantage, push past polite companies like Anthropic, Google, and OpenAI, then mature alongside them as norms resettle? Lee is less certain about this future than the others. Everything depends on how much of corporate culture the marketing can change. He already sees tremors: launch videos getting more cinematic, more people trying to crack the controversial thing.

If that becomes gigantic and corporate professionalism dies, Cluely will carry the rebellious, do-whatever-I-want attitude into maturity. If the wave is only a bump that gets them to escape velocity, and the standard is still to be buttoned up, then perhaps there is a future where they button up. All of that is downstream of whether the marketing reaches escape velocity and changes the game for everyone. He also rejects the idea that "cheat on everything" only works because they are punching up. The same marketing from Sam Altman or Elon, he thinks, would be just as controversial. After the raise, many people no longer see them as underdogs anyway.

What actually resonates, in his account, is transparency. He seems controversial, but he is just being honest, including on days he posts something heartfelt. The polarizing material is a slice of his life and thoughts; he happens to have polarizing opinions. "We don't aim to be super polarizing. We just aim to be honest."

Lee has said online that compensation packages go up to a million dollars plus equity. He thinks the market is getting this wrong. Revenue timelines are compressed, VC money flows freely, and if you are actually swinging big, someone will fund you. Everyone agrees the way to swing big is the best talent, and yet people still pay at the bottom of the range because candidates are supposed to be "bought in." His response is blunt: nobody knows your company, you were founded two months ago, why would anyone die for it? Give them a reason to walk in. Convince them you will win after they arrive. At Cluely, he says, there is not an employee who would sell a dollar of equity, because they are convinced the company will grow massively.

Top-of-range pay builds a strong top of funnel. Even the most money-hungry hire, the one who says they are only there for cash, will, he promises, get captivated by the culture and not bring themselves to leave. The company is more fun than any other in the world, existing "like a frat house": late-night conversations, steak, the gym, talk about girls and dudes. Stay because you cannot do anything other than Cluely. What has surprised him in hiring is the noise among programmers. They want really competent full-stack engineers. Plenty of people are great at research or low-latency optimization, with impressive backgrounds in the wrong thing.

Labenz thanks him for an earnest, minimally AI-assisted conversation, though he would not have minded Cluely in the mix and knows it was on at the beginning. Asked for a last thought, Lee does not recap the product. "This is the most interesting time in human history. I think you should take more risks, and right now the riskiest move is probably what you determine to be the safest move."