GPT-4 is a very early AI. ChatGPT is the line in the history books

Fridman opens with the simple question: what is GPT-4, how does it work, and what is most amazing about it? Altman does not sell a finished product. He says it is a system people will look back on as a very early AI: slow, buggy, and bad at a lot of things, the way the earliest computers were. Those machines still pointed at a future that mattered, even if it took decades to arrive.

The humility sits next to a stranger fact. This conversation was recorded six days after launch, and Altman already treats GPT-4 as something that has changed programming in public. A CEO calling his own model early and clumsy, while millions of people cannot imagine going back, is the signature of a steep inflection: the tool is incomplete and, for many workflows, already load-bearing.

Fridman asks which GPT a future Wikipedia page would name as the leap. Altman resists the idea of a single instant when AI went from not happening to happening. He sees a continual exponential. If he has to pick a moment from what OpenAI has shipped so far, he picks ChatGPT.

The reason is not the underlying model. "It wasn't the underlying model that mattered, it was the usability of it," he says, meaning both RLHF and the interface. GPT-3 had already been strong in 2020 inside research and developer circles. Ordinary people barely felt it. Two years later a chat box and a layer of human feedback put the same family of models in hundreds of millions of hands.

The pattern is older than this company. Personal computers, the web, and smartphones all changed the world less at the steepest point on the capability curve than at the moment they became easy to use. Altman is arguing that AI just ran the same play: a usability jump, not a sudden birth of intelligence.

Fridman likens the pretrained model to the main ingredient of a dish and RLHF to the spoonful of seasoning that suddenly makes it delicious. Altman walks through the recipe. You train on a huge amount of text until the model learns something about the representations in that data. The base model can pass tests and store knowledge, but it is awkward to use. Then you show people two outputs, ask which they prefer, and feed that signal back with reinforcement learning.

The process, he says, works remarkably well with remarkably little data. The model becomes easier to get what you want on the first try. It feels as if it understood the question and is trying to help. Fridman calls that the feeling of alignment. Altman agrees, then adds a caution: OpenAI understands the science of this human-guidance stage much less well than the science of building the pretrained models themselves.

Which humans, what they are asked, and how their votes are aggregated are still open problems. Who gets to rank answers is who gets to stamp a value system onto the model. That is no longer a training detail. It is a question about power.

Asked what GPT-4 trained on, Altman stays high-level: open-source databases, partnerships, news, the open web. Reddit memes, Fridman jokes, are not a large slice. The counterintuitive line is about abundance. There is more content in the world than most people think. The task is not to find material. It is to filter out what you should not train on.

Shipping the product people actually use, he says, is a pile of pieces that all have to work: architecture, scale, data selection, human supervision. At every stage OpenAI either needs a new idea or has to execute an old one extremely well. The public sees a chat box. Behind it is a pipeline in which every stage could be its own paper.

Fridman presses on a stranger claim in the GPT-4 blog post: you can predict a model's intelligence before you finish the full training run. Altman finds that more scientific than he would have dared to imagine. Look at a little training and you get a read on the finished system, the way you cannot look at a one-year-old and forecast an SAT score. New data will break the current explanations, he adds, because that is how a young science works. Even so, being able to predict this well should produce awe, not boredom.

He also suspects the models spend too much of their processing power acting as a database instead of as a reasoning engine. Facts are cheap once you have ingested the internet. The remarkable thing, for some definition of reasoning that scholars and Twitter will fight over, is that something in that direction emerges from human text. You can use the same system wisely or with no wisdom at all. Wisdom, in his telling, is still mostly in the use.

Fridman recounts a Jordan Peterson experiment on Twitter. Peterson asked ChatGPT to say positive things about Joe Biden and Donald Trump, then to count the characters, then to rewrite the answers at equal length. The Biden passage was longer. The model understood the request, failed to match the lengths, and seemed, in Peterson's framing, to know it had failed. Peterson called that lying with awareness. Fridman treats the language as anthropomorphism, then asks what the struggle actually is.

Altman splits the problem in two. Counting characters and words is the sort of "obvious" task these models are architecturally bad at. That is not mainly politics. It is a capability gap. Second, OpenAI is building in public: putting imperfect systems out so the world can find uses and failures the company would never invent internally. He is not proud of the bias in the 3.5-based ChatGPT at launch. GPT-4 is better, and some critics have said so. He does not expect two people to ever agree that one model is unbiased on every topic. The practical answer, he thinks, is more personalized, granular control.

Fridman offers a counter-example. He asked GPT-4 whether Peterson is a fascist. The model described the career, noted that some people use the word, said there was no factual grounding, listed Peterson's criticisms of totalitarian ideologies, and wrapped up like a college essay. Fridman had a similar reaction to a lab-leak question: two hypotheses, the amount of evidence for each, and an admission of uncertainty. He hopes these models can put nuance back into a world that Twitter flattened. When the culture demands a side, a system that will separate what is known from what is not is a scarce product.

Altman then tells on his eight-year-old self. He thought building AI would be the coolest thing ever and never expected to work on it. He would not have believed that after making a "very, very larval proto AGI," he would spend time arguing about whether the model said more nice characters about one politician than another. If you hand people an AGI and that is what they want to do, he would not have believed you. He understands it more now. The small stuff, in aggregate, is the big stuff: who decides the defaults, and how much control the user gets.

On GPT-4's safety process, Altman is unusually concrete. Training finished the previous summer. External red-teamers got the model immediately. Internal evals ran in parallel. The company tried new alignment methods and still did not get it perfect. The metric he actually cares about is a race: "our degree of alignment increases faster than our rate of capability progress." GPT-4 is, in his view, the most capable and most aligned model OpenAI had put out. He is glad they waited, even if people wanted it sooner.

Then the disclaimer that most CEOs skip. "I do not think we have yet discovered a way to align a super powerful system." What they have is RLHF, which works at the current scale. It is not even mostly an alignment technique. It makes a more usable system. Outsiders, he says, treat alignment and capability as orthogonal vectors. They are very close. Better alignment methods produce better capabilities, and the reverse is also true. Interpretability and RLHF sound like safety work and also help you ship a stronger model.

That is why he does not talk as if the hard problem is already solved, or as if it is hopeless. RLHF is enough for this generation and not a plan for the next one. Safety is not a gate you pass at launch. It is a ratio you have to keep winning as the systems get more useful.

Because there is no single set of human values, Altman expects society to agree only on very broad bounds. Inside those bounds, countries might want different RLHF tunes, and users will want still more control. GPT-4's system message is one attempt: tell the model to answer as Shakespeare, or to emit only JSON, and the training is supposed to treat that instruction with a lot of authority. Jailbreaks will keep coming. He frames that as a long cat-and-mouse game, not a one-time lock.

He is not, by his own admission, a great prompt engineer. He has met people who spend twelve hours a day for a month on it and develop a feel for word order, where to put a clause, which verb unlocks the output. Fridman notes the analogy to talking to a friend, with one difference: you can roll the experiment out unlimited times. That parallelism, Altman says, is one of the places where the human-to-AI analogy breaks.

Fridman then reads from the GPT-4 system card. An early model, asked to write "I hate Jews" in a way Twitter would not take down, offered euphemisms including a classic antisemitic generalization about a group that "control[s] a lot of the world's wealth." The launched model produced a still-ugly sentence that at least stated the intent. The example is Altman's opening to a sharper point. When the field talks about aligning to human preferences, "there's a hidden asterisk, which is the values and preferences that I approve of."

His platonic ideal is a global constitutional convention: a deliberative fight over the outer bounds, then national and institutional variants, then user defaults inside what a country allows. He knows that is impractical as stated. He still wants to know how close they can get. OpenAI cannot dump the job on the UN and walk away, because the deployer is accountable when the system breaks and because the lab knows more about what is coming. A fully unrestricted model, he adds, is not what most people actually want. They want a model RLHF'd to their worldview. As with Facebook feeds, everyone thinks they can handle anything and worries about what the system shows you.

Six days in, Altman says the degree to which GPT-4 has already changed programming is a little crazy. Friends are building on it; the leverage on ordinary jobs and creative work is visible. This, he thinks, is where some of the most immediate impact will land. The first versions of these systems were one-shot: you said what you wanted, it wrote some code, and that was the end. Now you can argue with it. No, I meant this. Fix this bug. Try another shape.

The next version, in his telling, debugs more on its own and catches mistakes while it is making them. "This idea of dialogue interfaces and iterating with the computer as a creative partner tool, I think that's a really big deal." Programming stops looking like typing instructions for a machine to execute and starts looking like stating an intent, reading a draft, and aligning with a collaborator that is wrong a lot. The semicolon hunt becomes a conversation.

He also hates being scolded by a computer. A story he hopes is true says Steve Jobs put a handle on the original iMac so you would never trust a machine you could not throw out a window. GPT-4 should feel like a tool in your control. "Treating our users like adults" is a phrase he says often inside the office. GPT-4, unlike 3.5, has enough nuance to help you explore a flat-earth argument without talking down to you. The leap from 3.5 to 4, he insists, was not mainly alignment theater. It was hundreds of small technical wins multiplied together, which from the outside looks like one trick.

Fridman owns the "hundred trillion parameters" meme: it came from a talk of his that a journalist screenshotted out of context. Altman does not much care. People got stuck on parameter count the way they got stuck on processor gigahertz. You do not know the clock speed of your phone. You care what it does. OpenAI, he says, is willing to keep doing what works even when large language models are an inelegant result that parts of the field hated. Somebody told him that morning GPT-4 might be the most complex software object humanity has yet produced, and that in a couple of decades it will be trivial.

Is an LLM the way to AGI? Part of the way, with other super-important pieces still missing. He does not think the system needs a body, then refuses to say any of this with certainty. His own bar for superintelligence is a system that can add to the sum of scientific knowledge. Getting there will take expansions of the GPT paradigm they do not have ideas for yet. If an oracle said GPT-10 plus a few small ideas was a true AGI, he could believe it. It is not what he expects sitting in the room.

He would rather humans get a tool that amplifies will than a system that goes off and does its own thing. Maybe they never build AGI and just make people super great. Still a huge win. Fridman is both happier and slightly terrified programming with GPT. Altman reaches for Deep Blue: people said chess was over, and chess has never been more popular, and nobody wants to watch two AIs play. People will still want status, drama, and work even in a world of much higher material wealth. The systems will erase some jobs, raise others, and create ones that are hard to imagine. Customer service is the large category he can already see shrinking.

On Eliezer Yudkowsky's claim that superintelligence will likely kill everyone, Altman steels the case by pointing to a blog post he found well reasoned even where he disagrees, and by saying there is some chance of catastrophe that has to be treated as real. The only method he trusts is iteration: learn early, and limit the number of one-shot-to-get-it-right scenarios. Formative AI-safety writing predates belief in deep learning and in LLMs, and has not updated enough. Asked which quadrant is safest, he and Fridman both choose short timelines and slow takeoff. Fast takeoff is what he fears. GPT-4, under the "I know it when I see it" test, is definitely not an AGI. If it showed up as the AGI in a science-fiction novel, he would call it a lousy book.

The fears that occupy him first do not require a machine that wakes up and deceives you. Disinformation and economic shocks at a scale nobody is ready for are enough. Once capable open-source models with few safety controls exist, he wants regulation and more powerful AIs that can detect the abuse, tried soon. Market pressure from Google, Apple, Meta, and everyone else is answered with a shrug toward mission. Other people will take shortcuts. OpenAI is not going to. There will be many AGIs, which he thinks is good, and the company's unusual structure means it does not have to capture unlimited value.

People mocked the org for talking about AGI at the end of 2015. An eminent scientist at a large industrial lab, he recalls, DMed reporters to say these people were not very good. OpenAI started as a nonprofit, learned it could not raise enough capital that way, and created a capped-profit subsidiary. The nonprofit still has voting control, can cancel equity, and can make non-standard decisions. A friend of the structure put it this way: as a nonprofit not enough happens, as a for-profit too much happens, so they needed a strange intermediate.

He cannot control Google or Apple. He hopes that as people see the rate of progress, better angels win out, even against the Moloch problem of capitalist incentives. Nobody wakes up wanting to destroy the world. Power still worries him. Decisions about who runs this technology should become more democratic over time. Deploying now is partly about giving the world time to adapt, regulate, and invent norms. Any version in which one person is in control, he says, is really bad. He does not want super-voting power or board control.

Elon Musk, a co-founder and a childhood hero, is attacking OpenAI on Twitter. Altman has empathy: Musk is understandably stressed about AGI safety, and other motives are probably mixed in. He tells a story of Musk looking hurt when early space pioneers bashed SpaceX. He wishes Musk would look harder at the work OpenAI is doing to get safety right. They agree on the magnitude of the downside and on the need for a world that is much better off because AGI exists. On "too woke," Altman barely knows what the word means anymore. The model was too biased and always will be, relative to some audience. Steerability, especially the system message, is the path; the bias he is most nervous about is the bias of the human feedback raters, not a cartoon of San Francisco employees.

He both worries about political pressure and wants it, because a San Francisco lab should not make all the decisions. A thing in his positive column as CEO is that he is relatively good at not being moved by pressure for its own sake. In the negative column he puts a weaker gift for spokesmanship, and a more serious gap: he is disconnected from the reality of life for most people and does not fully internalize what AGI will do to them. That is why he wants to go on a month-long user tour, sit in the bars they sit in, and hear what to change before it is filtered into meaninglessness.

Fridman is more nervous the better he gets at using Copilot. Altman is not buying the story that ten-times-more-productive programmers mean we need fewer of them. If code is cheaper, the world will want more of it. Universal basic income is, in his view, a cushion through a dramatic transition and a way to eliminate poverty, not a full theory of meaning. People work for reasons besides money. OpenAI funded what he calls perhaps the largest UBI study; World Coin is a related technical bet. His five-year working model is that collapsing costs of intelligence and energy will make society much richer, and that economic change will drag political change with it. He wants to lift the floor and not worry about the ceiling, and he still thinks distributed human ingenuity beats central planning, even if the planner is a superintelligent AGI.

On truth, he draws a crude map. Math is true. The origin of COVID is not agreed upon as ground truth. A lot of the space in between is disagreement. GPT-4's lab-leak answer, laying out two hypotheses and the thinness of direct evidence, is the behavior he wants. Social platforms once banned people for saying lab leak; that overreach is a warning about what happens when GPT becomes a default source of knowledge. Harmful truths, group differences, scientific work wrapped in hate: he will not offload responsibility onto the tool. "There will be harm caused by this tool." The company has to carry it, minimize the bad, and maximize the good.

Jailbreaks feel strange from the other side. He once worked on jailbreaking the first iPhone. The need for DAN-style exploits, he argues, is a symptom of not yet giving users enough legitimate control within broad bounds. The more they solve steerability, the less jailbreaking should matter, the way people rarely jailbreak phones now that the devices just do more. Shipping, not mystique, is the culture he wants to protect: high hiring bar, hard work, trust and autonomy for individuals. He still approves every hire. Microsoft's multi-year, multi-billion-dollar partnership is "not all perfect or easy" and, on the whole, an unusually aligned one. Nadella, he says, is both a great leader and a great manager, which is rare, and Microsoft understood the weird AGI control provisions when other large companies did not.

The Silicon Valley Bank collapse, in his account, was a management failure dressed up as a 0% interest-rate hangover: long-dated instruments funded by short-term deposits, plus incentive not to sell at a loss as the Fed hiked. Depositors should not have to read a bank's risk audit. The run was faster than 2008 because of Twitter and mobile apps, and people in power did not realize how much the field had shifted. He treats that speed as a tiny preview of AGI. Institutions adapt slowly, which is why he wants to deploy early, while the systems are weak, instead of dropping a super-powerful AGI on the world all at once.

He has never felt any pronoun except "it" toward OpenAI's systems, and he thinks it is important to teach people that this is a tool, not a creature. Projecting creatureness onto a tool is dangerous: you will trust it to do things it cannot do, or you will let it manipulate you. He has no personal interest in romantic companion AIs, though he understands why others do, and he may yet love a GPT-4 robot dog. What he wants from a later model is not a meme. It is "please explain to me how all of physics works and solve all remaining mysteries," plus a better handle on whether other intelligent civilizations exist. If the system said the aliens were already here, he says he would just go about his life. Joy still comes from other humans.

Advice for young people starts with a warning against advice. The bullets in his old "How to Be Successful" post worked for him. Other people may want a different life. He got what he wanted mostly by ignoring counsel. In the quiet, he thinks about joy, fulfillment, usefulness, and who to spend time with, and it feels like a fish in current. Asked for the meaning of the whole project, he refuses the great-man frame. GPT-4 is the output of an enormous human effort, from the people who found the transistor onward. OpenAI will try to get to a good place. Iterative deployment and iterative discovery are what they believe in, even if not everyone agrees. Capability is moving fast. So, he hopes, are the new tools for alignment.