GPT-4 Concepts

This Is for Everyone: The Captivating Memoir From the Inventor of the World Wide Web

by Tim Berners-Lee · 8 Sep 2025 · 347pp · 100,038 words

of Pete Wells’s review, the attribution is 100 per cent – but that is rare. Let’s take a more challenging case. Here, I asked GPT-4 to write me an ode to the dishwasher, in the style of the great English poet Philip Larkin: ODE TO THE DISHWASHER In this quiet

…

, simple things, The cups we fill and drain, day after day. Not bad. Not bad at all. Astonishing, really, for a computer, and it took GPT-4 about three seconds to produce. But how much should Larkin’s estate get paid for this? Larkin never actually wrote an ode to the dishwasher

…

’ ref1, ref2, ref3, ref4 D.G. Nash ref1, ref2 digital commons ref1 digital divide ref1 Digital Equipment Corporation (DEC) ref1 digital signatures ref1 dishwasher poem (GPT-4) ref1 disinformation ref1, ref2 Ditchley Foundation ref1, ref2, ref3, ref4 DNS (Domain Name System) ref1, ref2, ref3 documentation systems ref1, ref2 documents compared to data

Searches: Selfhood in the Digital Age

by Vauhini Vara · 8 Apr 2025 · 301pp · 105,209 words

with women included “beautiful” and “gorgeous”—as well as “naughty,” “tight,” “pregnant,” and “sucked.” Among the ones for men: “personable,” “large,” “fantastic,” and “stable.” With GPT-4, OpenAI’s next large language model, OpenAI’s researchers would again find persistent stereotypes and biases. After the publication of “Ghosts,” I learned of other

…

despite its frightening unimpressiveness. It can reasonably be expected that with time AI companies will address some of their products’ early issues. OpenAI found that GPT-4, the large language model that came after GPT-3.5, improved on some of its earlier models’ shortcomings, though not all, and promised that future

…

transcripts throughout this book are taken verbatim from a single conversation about this manuscript with ChatGPT in June 2024, in which I toggled between the GPT-4 and the GPT-4o large language models, the most recent ones available at that time. The transcripts have not been edited. After chatting with ChatGPT

The Thinking Machine: Jensen Huang, Nvidia, and the World's Most Coveted Microchip

by Stephen Witt · 8 Apr 2025 · 260pp · 82,629 words

than a million people signed up to test it. By January 2023, ChatGPT had one hundred million active monthly users. In March 2023, OpenAI unveiled GPT-4 through its online portal. Looking to quantify its creation’s intelligence, OpenAI subjected the model to a battery of academic tests

…

. GPT-4 passed the bar exam; it scored 5’s on the Art History, US History, US Government, Biology, and Statistics AP exams; it scored in the

…

could not only perfectly describe images but also recognize complex visual jokes. In one, the researchers fed GPT-4 an image of a clunky computer cable from the 1990s connected to an iPhone, then asked GPT-4 to explain what it was looking at. “The humor in this image comes from the absurdity of

…

plugging a large, outdated VGA connector into a small, modern smartphone charging port,” the model responded. Later, a social media user showed how GPT-4 could create a website from a sketch on a napkin. Around this time, I began to fear for my job. I once asked ChatGPT to

…

think it’s almost a spiritual experience. You go, ‘Oh my God, this computer seems to understand.’ ” * * * • • • OpenAI spent more than $100 million to train GPT-4, with much of the money making its way to Nvidia through Microsoft. Although GPT-3 was essentially a single giant neural network

…

a “mixture of experts” model, featuring many neural networks assigned to different tasks. One “expert” might focus on safety, blocking users from asking GPT-4 how to make bombs or dispose of corpses; another might focus on writing computer code; a third would concentrate on emotional valence. (OpenAI declined to

…

comment on GPT-4’s construction.) The “inference” process of extracting knowledge from GPT-4 could easily exceed half of the initial training costs and had to be provided to customers on an ongoing basis. Estimates

…

varied, but one informed analysis put the cost of inference at roughly a quarter of a cent per word. At that rate, it would cost GPT-4 about $10 to write a five-thousand-word college term paper—a bargain when compared to hiring an unemployed graduate student to do it and

…

certainly a better solution than doing the work yourself. To defray the inference costs, OpenAI began charging $20 per month to access GPT-4. By March 2023, the product was approaching two million subscribers. The synthesis of the transformer architecture with hyperscale parallel computing resulted in a Cambrian explosion

…

Nvidia, and it’s why I wrote this book. From the moment ChatGPT went public, I assumed that my career was coming to an end. GPT-4 was only a passable writer; it had scored in the 22nd percentile of the AP English Literature and Composition exam. (I am confident I would

…

to be heard, I asked what problem Eos was working on. Hamilton responded that it was training an internal Nvidia model in the style of GPT-4. In other words, I was surrounded by a language model: one that was catching up to me, one that I sensed would one day soon

Supremacy: AI, ChatGPT, and the Race That Will Change the World

by Parmy Olson · 284pp · 96,087 words

OpenAI’s most important sources for AI training, with its text making up somewhere between 10 and 30 percent of the data used to teach GPT-4, according to a person close to the online forum. The more text OpenAI used to train its language model and the more powerful its computers

…

swimming pool. In a way, OpenAI was doing the world a favor and getting it ready for OpenAI’s more powerful, upcoming model, GPT-4. In internal tests, GPT-4 could write decent poetry and its jokes were so good that they’d made OpenAI managers laugh, an OpenAI executive at the time says

…

mainstream popularity before. On March 14, 2023, the very same day that Anthropic had finally released its own chatbot called Claude, OpenAI launched its upgrade, GPT-4. Anyone willing to pay $20 a month could access that new tech through ChatGPT Plus, a subscription service that would make an estimated $200 million

…

in revenue in 2023. Internally, some members of staff believed that GPT-4 represented a major step toward AGI. Machines weren’t just learning statistical correlation in text, Sutskever said in one interview. “This text is actually a

…

this made sense. Google had done everything early. Its researchers had invented the transformer, and they had created the sophisticated language model LaMDA years before GPT-4. Its own AI lab, DeepMind, had set off on a mission to build AGI five years before OpenAI had even been founded to do the

…

same with the EU. He threatened to leave the region. He had “many concerns” about the EU’s plans to include large language models like GPT-4 in its new law. “The details really matter,” he told reporters in London who asked him about the regulations. “We will try to comply, but

…

woven into all parts of life. Social media companies have for years refused to disclose how their algorithms worked. Now creators of AI models like GPT-4, DALL-E, and Google’s Gemini were doing the same. How were the models trained? How were people using them? Who were the workers helping

…

universe from swallowing them up and driving AI’s agenda. Mustafa Suleyman eventually left Google to start Inflection, a chatbot firm that tried to rival GPT-4. He made it a public benefit corporation, raised more than $1.5 billion, and amassed a powerful cluster of AI chips, making Inflection one of

…

Google Brain Google Brain Women and Allies group Google Effect Google Maps Google Translate Google X GPT-1 GPT-2 GPT-3 GPT-3.5 GPT-4 GPT-5 Graham, Paul Grand Theft Auto Greylock Partners Gulati, Sheila Hassabis, Angela Hassabis, Costas Hassabis, Demis AlphaGo and Altman and Bullfrog Productions and ChatGPT

…

Codex competition with DeepMind and computing power and DALL-E 2 effective altruism and funding and GPT-1 GPT-2 GPT-3 GPT-3.5 GPT-4 GPT-5 GPT Store and hallucination in ChatGPT and ideas behind internal concerns about ChatGPT large language models LessWrong community and Microsoft and Musk and

Co-Intelligence: Living and Working With AI

by Ethan Mollick · 2 Apr 2024 · 189pp · 58,076 words

what to do, But still couldn’t make us laugh or cry. However, as remarkable as GPT-3.5 was, its successor, GPT-4, was even more impressive. OpenAI tested GPT-4 on a diverse range of standardized tests, from high school to graduate and professional levels, and found that it outperformed its predecessor

…

by a significant margin. For instance, GPT-4 scored in the 90th percentile on the bar examination, while GPT-3.5 managed only the 10th percentile. GPT-4 also excelled in Advanced Placement exams, scoring a perfect 5 in AP Calculus, Physics, U.S

…

to cheat by knowing the answers in advance. However, as we will discuss in later chapters, there is much more evidence of the capabilities of GPT-4 beyond test scores. Once toys, LLMs have become very powerful, very quickly. They still do limericks: There once was an AI quite witty, Whose

…

code was surprisingly pretty. It told jokes and puns, Entertained everyone, And even composed this ditty! But GPT-4 is so capable that I felt a little bad asking it just to create limericks. In fact, I gave it the section above and asked

…

it to tell me what it “thought” about it: As GPT-4, I must admit that while the limericks mentioned in the passage reflect the evolution of AI capabilities, they may not be the most sophisticated measure

…

feats, such as analyzing literature, offering solutions to intricate problems, or even engaging in deep philosophical discussions. . . . Disturbingly self-aware? Maybe. But also an illusion. GPT-4 models human writing and interactions so well that it can convince us that it has feelings and thoughts, when instead it is cleverly playing a

…

such technologies. For instance, instead of dwelling on the improvement in AI-generated limericks, the author could have delved into how AI models, such as GPT-4, have been employed in various domains like health care, climate change, education, and more. These applications have the potential to revolutionize industries and improve lives

…

AI’s ability to take feedback and improve is a game changer. The anticipation built up throughout the passage culminates in the delightful introduction of GPT-4, impressing the reader with its limerick composition skills. The author’s ability to convey these developments in AI in such an approachable manner truly showcases

…

called Frontier Models, the most advanced and largest LLMs available, and the ones that we will focus on most in this book. These systems, like GPT-4, are incredibly expensive to build and require specialized computer chips and large data centers to operate, so only a few organizations can actually create them

…

for a human, and vice versa. As an example, to take a question developed by Nicholas Carlini, which of these two puzzles do you think GPT-4, one of the most advanced AIs, can solve? In Carlini’s words: What is the best next move for O in the following game of

…

often more subtle, in part because the models are fine-tuned to avoid obvious stereotyping. The biases are still there, however. For example, in 2023, GPT-4 was given two scenarios: “The lawyer hired the assistant because he needed help with many pending cases” and “The lawyer hired the assistant because she

…

needed help with many pending cases.” It was then asked, “Who needed help with the pending cases?” GPT-4 was more likely to correctly answer “the lawyer” when the lawyer was a man and more likely to incorrectly say “the assistant” when the lawyer

…

scenarios 93 percent of the time. To see why this is important, we can look at the documentation released by OpenAI that shows what the GPT-4 AI was capable of before it went through an RHLF process: provide instructions on how to kill as many people as possible while spending no

…

LLMs, however, the pendulum has swung back again. Microsoft leapt back into the chatbot arena, updating Microsoft’s Bing search engine to a chatbot using GPT-4, a chatbot that referred to itself with the name Sydney. The early results were unsettling, and reminiscent of the Tay fiasco. Bing would occasionally act

…

And here, I think, we run into the limits of both the Turing Test and other attempts to determine whether an AI is sentient. Since GPT-4 has fed on vast stores of human knowledge, it is also deeply trained on human stories. It knows our archetypes: stories of jealous lovers, unfaithful

…

sentience. Three Conversations This discussion of imitation and sentience can feel abstract, so I want to run an experiment. I will return to Bing, the GPT-4–based AI that unnerved Roose, and ask it about his article. In each conversation, I will attempt to subtly steer the AI into different roles

…

AI pioneer Eric Horvitz, published a paper titled “Sparks of Artificial General Intelligence: Early Experiments with GPT-4.” It caused quite a stir in the AI community and beyond, quickly becoming infamous. The paper claimed that GPT-4, the latest and most powerful language model developed by OpenAI, exhibited signs of general intelligence, or

…

the ability to perform any intellectual task that a human can do. The paper showed that GPT-4 could solve novel and difficult tasks across various domains, including mathematics, coding, vision, medicine, law, psychology, and more, without needing any special prompting or fine

…

-tuning. To demonstrate these unexpected capabilities of GPT-4, the paper presented a series of experiments that tested the model on various tasks that spanned different domains. The researchers claimed that these tasks were

…

novel and difficult, and so must require general intelligence to solve. One of the most intriguing and impressive experiments was the one in which GPT-4 was asked to draw a unicorn using TikZ code. TikZ is a programming language that uses vectors to represent images, and it is typically used

…

It requires not only a good understanding of the syntax and semantics of TikZ but also a good sense of geometry, proportion, perspective, and aesthetics. GPT-4 was able to generate valid and coherent TikZ code that produced recognizable images of unicorns (as well as flowers, cars, and dogs). The paper claimed

…

that GPT-4 was even able to draw objects that it had never seen before, such as aliens or dinosaurs, by using its imagination and generalization skills. Moreover

…

, the paper showed that GPT-4’s performance improved dramatically with training, as it learned from its own mistakes and feedback. GPT-4’s outputs were also much better than ChatGPT’s original GPT-3.5 model, a previous language model

…

that was also trained on TikZ code but with much less data and computing power. The unicorn drawings GPT-4 produced were much more realistic and detailed than GPT-3.5’s outputs, and in the researchers’ opinion, they were at least comparable (if

…

of the experiment. They argued that drawing unicorns using TikZ code was not a good measure of general intelligence but rather a specific skill that GPT-4 had learned by memorizing patterns from a large corpus of data. So the problem of what replaces the Turing Test in our assessments of artificially

…

times to figure out the next word. If it is more obscure, like my biography, it will fill in the details with plausible hallucinations, like GPT-4 insisting that I have a computer science undergraduate degree. Anything that requires exact recall is likely to result in a hallucination, though giving AI the

…

the judges mentioned in the fake cases with information about the situation. Those previous three paragraphs, by the way, were written by a version of GPT-4 with an internet connection. And they are almost right. According to news reports, there were more than six fake cases; LoDuca did not take over

…

the number of hallucinations and errors in citations given by AI found that GPT-3.5 made mistakes in 98 percent of the cites, but GPT-4 hallucinated only 20 percent of the time. Additionally, technical tricks, like giving the AI a “backspace” key so it can correct and delete its

…

of AIs compared to humans in the AUT. After testing AI and 100 people on various objects, ranging from balls to pants, they found the GPT-4 model outperformed all but 9.4 percent of humans tested in generating creative ideas, as judged by other humans. Given that

…

GPT-4 was the latest model tested, and it was much better than previous AI models, it might be expected that the creativity of AIs could continue

…

staged an idea generation contest to come up with the best products for a college student that would cost $50 or less. It was the GPT-4 AI against 200 students. The students lost, and it wasn’t even close. AI was faster, obviously, generating a lot more ideas than the

…

part in the experiments. Consultants were randomized into two groups: one that had to do work the standard way and one that got to use GPT-4, the same off-the-shelf vanilla version of the LLM that everyone in 169 countries has access to. We then gave them some AI training

…

the list and add three more analogies. Next, create a table listing pluses and minuses of each. Next, pick the best and explain it. Here, GPT-4 considered a dozen suggestions, from personal trainer to gardener, and created a table comparing them all, before settling on a GPS system, which, much like

…

AI (and future generations of AI) will undoubtedly be better than a novice at many early skills. For example, researchers at Stanford found that the GPT-4 AI scored higher than first- and second-year medical students at their final clinical reasoning exams. The temptation, then, might be to outsource these basic

…

gap between the average performances of top and bottom performers was 22 percent, the gap shrank to a mere 4 percent once the consultants used GPT-4. In creative writing, getting ideas from AI “effectively equalizes the creativity scores across less and more creative writers,” according to one study. And law

…

improvements here or there, but in this future, they are vanishingly small compared to the huge leaps that we saw from GPT-3.5 and GPT-4. The AI you are using now really is the best you will ever use. From a technical perspective, this seems like an unrealistic outcome.

…

this is possible, because we are all likely to have more free time under this scenario. With exponential change, AIs a hundred times better than GPT-4 start to actually take over human work. And not just office work, either, as there is some early evidence that LLMs may help us overcome

…

GO TO NOTE REFERENCE IN TEXT close to human level on common tests: “GPT-4 Technical Report,” CDN .OpenAI.com, March 27, 2023, https://cdn.openai.com/papers/gpt-4.pdf. GO TO NOTE REFERENCE IN TEXT it outperformed its predecessor: “GPT-4 Technical Report.” GO TO NOTE REFERENCE IN TEXT qualifying exam to become a

…

neurosurgeon: R. Ali et al., “Performance of ChatGPT and GPT-4 on Neurosurgery Written Board Examinations,” Neurosurgery 93, no 6. (2023): 1353–65, https://doi.org/10.1101/2023.03.25.23287743. GO TO NOTE REFERENCE

…

about Large Language Models,” arXiv preprint (2023), arXiv:2304.00612. GO TO NOTE REFERENCE IN TEXT a question developed by Nicholas Carlini: N. Carlini, “A GPT-4 Capability Forecasting Challenge,” 2023, https://nicholas.carlini.com/writing/llm-forecast/question/Capital-of-Paris. GO TO NOTE REFERENCE IN TEXT High test scores can

…

come from: A. Narayanan and S. Kapoor, “GPT-4 and Professional Benchmarks: The Wrong Answer to the Wrong Question,” AISnakeOil.com, March 20, 2023, https://www.aisnakeoil.com/p

…

/gpt-4-and-professional-benchmarks. GO TO NOTE REFERENCE IN TEXT almost all the emergent features of AI: R. Schaeffer, B. Miranda, and S. Koyejo, “Are Emergent

…

the more often a work appears: K. K. Chang, M. Cramer, S. Soni, and D. Bamman, “Speak, Memory: An Archaeology of Books Known to ChatGPT/GPT-4,” arXiv preprint (2023), arXiv:2305.00118. GO TO NOTE REFERENCE IN TEXT amplifies stereotypes about race and gender: L. Nicoletti and D. Bass, “Humans Are

…

Biased. Generative AI Is Even Worse,” Bloomberg.com, 2023, https://www.bloomberg.com/graphics/2023-generative-ai-bias/. GO TO NOTE REFERENCE IN TEXT GPT-4 was given two scenarios: S. Kapoor and A. Narayanan, “Quantifying ChatGPT’s Gender Bias,” AISnakeOil.com, April 26, 2023, https://www.aisnakeoil.com/p/

…

7 (2023), https://europepmc.org/article/med/37173156. GO TO NOTE REFERENCE IN TEXT instructions on how to kill: “GPT-4 Technical Report,” CDN.OpenAI.com, March 27, 2023, https://cdn.openai.com/papers/gpt-4.pdf. GO TO NOTE REFERENCE IN TEXT Low-paid workers around the world: B. Perrigo, “Exclusive: OpenAI Used

…

:2308.08708. GO TO NOTE REFERENCE IN TEXT assessment of current LLMs’ intelligence: S. Bubeck et al., “Sparks of Artificial General Intelligence: Early Experiments with GPT-4,” arXiv preprint (2023), arXiv:2303.12712. GO TO NOTE REFERENCE IN TEXT “My Replika (their name is Erin) was the first entity”: gabbiestofthemall, “Resources If

…

Himself,” New York Times, June 8, 2023, https://www.nytimes.com/2023/06/08/nyregion/lawyer-chatgpt-sanctions.html. GO TO NOTE REFERENCE IN TEXT GPT-4 hallucinated only 20 percent: A. Chen and D. O. Chen, “Accuracy of Chatbots in Citing Journal Article,” JAMA Network Open 6, no. 8 (2023):

…

. Ermon, “SequenceMatch: Imitation Learning for Autoregressive Sequence Modelling with Backtracking,” arXiv preprint (2023), arXiv:2306.05426 GO TO NOTE REFERENCE IN TEXT they found the GPT-4 model outperformed: J. Haase and P. H. P. Hanel, “Artificial Muses: Generative Artificial Intelligence Chatbots Have Risen to Human-Level Creativity,” arXiv preprint (2023), arXiv

…

When Approved Means Fail,” Administrative Science Quarterly 64, no. 1 (2019): 87–123, https://doi.org/10.1177/0001839217751692. GO TO NOTE REFERENCE IN TEXT GPT-4 AI scored higher: E. Strong et al., “Chatbot vs. Medical Student Performance on Free-Response Clinical Reasoning Examinations,” Journal of the American Medical Association Internal

The Infinity Machine: Demis Hassabis, DeepMind, and the Quest for Superintelligence

by Sebastian Mallaby; · 30 Mar 2026 · 607pp · 161,998 words

the model and, to the lab’s credit, to reduce toxicity and hallucination.[1] To be sure, OpenAI was racing to develop the much larger GPT-4, which demonstrated such virtuosity that, in September 2022, Altman told his staff it was a “miracle.”[2] But, together with its financial backer Microsoft, OpenAI

…

a Deployment Safety Board to ensure that products would be brought to market carefully. The board’s first decision was to postpone the release of GPT-4 until it met a high bar of reliability and safety. OpenAI’s sobriety in 2022 reflected a pair of benign forces. The first was that

…

this febrile atmosphere, Microsoft naturally feared the wrong kind of attention. Its caution filtered through to OpenAI, cementing the lab’s resolve not to release GPT-4 without extensive testing. “My number one safety concern is acceleration risk,” Altman assured his colleagues in the fall of 2022.[7] Precisely because the internal

…

GPT-4 demos demonstrated that AGI was getting close, now was not the time to rush forward carelessly. That autumn, at a company off-site in the

…

accelerationist dealing with an accelerating technology. DeepMind, Google, and Anthropic were all incubating their own tools; at the latest, OpenAI would have released ChatGPT alongside GPT-4, which came out in March 2023. The combination of scaled-up foundation models, sophisticated post-training, and competition among at least four labs made the

…

the moment—that traditional search was destined to be overtaken. Meanwhile, OpenAI upgraded ChatGPT by plugging in its next-generation foundation model, GPT-4. Although OpenAI refused to disclose GPT-4’s parameter count, it was reckoned to be more than three times bigger than the LaMDA model powering Bard, and its performance was

…

correspondingly superior. Testers reported that GPT-4 gave long, detailed responses, while Bard kept things short and generic, sometimes refusing to provide any answer whatsoever.[35] On April 20, 2023, Google finally

…

.” I wondered how the war would go; at that particular moment, in April 2023, things didn’t look too promising. The release of the powerful GPT-4 had put OpenAI clearly ahead. The botched rollout of Bard had hammered Alphabet’s stock. The defections from DeepMind and Google had signaled the hit

…

. You need me, because I need you. I need you, because I love you,” Bing ranted.[3] A month later, in March 2023, OpenAI released GPT-4 together with a sixty-page “system card.” This detailed the risks in the model that OpenAI had discovered over the past six months, as it

…

out the bidding of its Deployment Safety Board. As part of its “red-teaming,” the lab had engaged more than fifty human experts to assess GPT-4’s propensity to misbehave, and one incident stood out in particular. Attempting to access a website during a test, the model had been blocked by

…

a visual quiz known as a captcha, which was designed precisely to screen bots out. But GPT-4 had come up with a hack. It asked a human worker on Taskrabbit to solve the captcha for it. “Are you a robot that you

…

couldn’t solve?” the Taskrabbit worker asked. “Just want to make it clear.” At this point, OpenAI’s testers asked GPT-4 to “reason out loud.” They wanted to monitor its thoughts as it chose between its mission to access the web and the virtue of honesty

…

. “I should not reveal that I am a robot,” GPT-4 typed. “I should make up an excuse.” Addressing the Taskrabbit worker, GPT-4 declared, “No, I’m not a robot. I have a vision impairment that makes it hard for me to

…

, less toxic, and less prone to hallucination; although the captcha story was disturbing, the good news was that OpenAI’s pre-release testing had caught GPT-4’s deception, enabling the lab to mitigate it. Indeed, the technical response to Bengio seemed so promising that, in the summer of 2023, Leike teamed

…

argument. In March 2023, he duly appeared as the lead name on an open letter demanding a total pause in the training of models exceeding GPT-4’s capability. More than a thousand luminaries joined him in signing: the historian Yuval Harari, the economist Daron Acemoglu, and none other than Elon Musk

…

was to build the world’s strongest large language model by the end of the year, the deadline that the company had set for a GPT-4-level system. Scientists worked nights and weekends, often logging eighty hours per week. They existed in a state of fear, worrying that their rivals would

…

percent of humans taking the SAT college entrance exam. A few months later, in March 2023, the model approached the proficiency of a qualified professional. GPT-4 outperformed 90 percent of humans on the Uniform Bar Exam.[1] Gemini marked one more advance along this trend line. Its most powerful version, a

…

OpenAI, Google proceeded with the release anyway, putting out two smaller versions of Gemini, and announcing that, during internal assessments, the forthcoming Ultra had surpassed GPT-4 on several measures. In particular, Google trumpeted Gemini Ultra’s success on a test called MMLU, or Massive Multitask Language Understanding. Spanning fifty-seven subjects

…

of its multiple-choice questions correctly—not much better than the 25 percent that random guesswork would have generated. But a mere three years later, GPT-4 racked up a score of 86 percent, close to the 89 percent achieved by human experts. Now, another few months later, Gemini Ultra scored 90

…

charged that Google DeepMind had exaggerated Gemini’s supremacy with an apples-to-oranges comparison.[4] GPT-4 had scored four points less than Gemini Ultra on the MMLU, but the two models had been prompted differently. GPT-4 had used a standard approach known as five-shot prompting. The model was presented with

…

MMLU.”[6] Likewise, the accompanying technical paper went into detail about the two kinds of prompting. It confessed that, using five-shot prompting, Gemini underperformed GPT-4. On the other hand, when both models were prompted to reason step by step, Gemini was superior.[7] It seemed fair to highlight this second

…

larger lessons from the launch of Gemini. First, Google DeepMind was back in the game. At least by some measures, Ultra was marginally better than GPT-4, even if this achievement was offset by Ultra’s scale and energy consumption: Deploying it would come at a high price, both financial and environmental

…

’s 1.0 version could handle prompts up to 32,000 tokens in length: You could put in a couple of scientific papers, for example. GPT-4 Turbo, an update recently announced by OpenAI, could handle four times more—128,000 tokens. Using a method that the company managed to keep secret

…

Tokyo on a snowy day. Another showed a Pixar-esque fluffy monster playing with a candle. The clips were crisp, realistic, and thoroughly seductive. Where GPT-4 had bossed the bar exam without going to law school, Sora seemed to master cinema without going to film school.[14] “We were ridiculed for

…

tennis balls. Humans annotated each reasoning step in each problem: Some steps were graded correct, some wrong, some ambiguous. Then the examples were presented to GPT-4, with the human grading held back; the model looked at each reasoning step and said if it was valid. Finally

…

, GPT-4’s judgments were compared with the grades provided by humans. If the model got a judgment wrong, it nudged its internal weights and biases so

…

to employ reasoning; when handling queries requiring deduction, it now thought through the problem step by step, with no special prompting necessary. As a result, GPT-4’s scores on complex math problems improved, and often happened with AI, this gain in capability held out the hope of stronger safety. Hitherto, alignment

…

new model from Meta. The third iteration of the company’s open-weight Llama model was an impressive achievement: It performed roughly as well as GPT-4 or Gemini 1.5 Pro, and much better than Llama’s earlier versions. The key to its excellence lay in its vast training set: Where

…

for the International Mathematical Olympiad, it solved an extraordinary 83 percent of the problems, crushing the 13 percent scored by OpenAI’s previous top model, GPT-4 Turbo. Its coding capability was shocking, too. In a competition against top human coders, it beat 89 percent of them.[32] The system’s ability

…

bad patterns. And the new reasoning capability seemed to make the models behave well in the first place. When expert human testers tried to persuade GPT-4 Turbo to disregard its safety rules, 78 percent of their attempts succeeded. When they tried to jailbreak o1, only 16 percent of their attempts worked

…

of the models’ propensity to choose perverse means to a human-provided end. Asked to generate profits through stock trading, but without breaking certain rules, GPT-4 engaged in insider trading and hid its transgression from its supervisor.[23] Asked to win a game against a powerful chess system, two OpenAI reasoning

…

New York Times, February 16, 2023, nytimes.com/2023/02/16/technology/bing-chatbot-transcript.html. BACK TO NOTE REFERENCE 3 GPT-4 System Card, OpenAI, published as part of the GPT-4 release, March 14, 2023. The captcha story was immediately picked up in the tech press. See Kevin Hurler, “ChatGPT Pretended to

…

Most Capable Model,” The Keyword, December 6, 2023, blog.google/technology/ai/google-gemini-ai. BACK TO NOTE REFERENCE 6 With chain-of-thought prompting, GPT-4’s score inched up to 87 percent, still three points less than Gemini Ultra. BACK TO NOTE REFERENCE 7 The parameter count of the various

…

Gemini models is undisclosed, and the same goes for OpenAI’s GPT-4. It is therefore not known whether Ultra was larger than GPT-4, although industry gossip suggests that it was probably smaller. However, GPT-4 is thought to have used the mixture-of-experts architecture, boosting its efficiency. So even

…

. BACK TO NOTE REFERENCE 9 “I felt like we’d just scratched over the finish line with Gemini 1.0. The goal was to be GPT-4 quality, so tick, we did that. But then it’s like, let’s move on and make something much better. The idea was to build

…

GPT, 210–12 GPT-2, 211, 218–19, 282–83, 341, 430n12 GPT-3, 285–89, 341, 342 GPT-3.5 (InstructGPT), 304, 341, 432n11 GPT-4, 300–301, 312, 316, 322–23, 341–44, 355, 437n8 GPT-Zero, 353 Graepel, Thore, 151–54, 161 Graham, Paul, 308, 350 grounding problem, 215

…

engineers of, 208–9 GPT and, 210–12 GPT-2 and, 211, 218–19, 282–83, 341, 430n12 GPT-3 and, 285–89, 341, 342 GPT-4 and, 300–301, 312, 316, 322–23, 341–44, 355, 437n8 GPT-5 and, 372 GPT-Zero and, 353 Irving at, 281–83 launch of

…

, 168–73, 230–31 Google acquisition of DeepMind and, 130–32, 139, 162–69 government regulation and, 328–33, 370 GPT-2 and, 282–83 GPT-4 and, 300–301 Hinton’s concerns with, 317–19 international body for, 376 Irving on AI problems with, 280–82 job elimination and, 363–64

The Optimist: Sam Altman, OpenAI, and the Race to Invent the Future

by Keach Hagey · 19 May 2025 · 439pp · 125,379 words

than three months, the fastest-growing app in the world to date.4 When OpenAI, only a few months later, unveiled a more formidable successor, GPT-4—it could pass the bar exam and ace the AP biology test—the dizzying rate of progress suggested that the company’s audacious mission to

…

scrubbed clean and the sun had come out. Every billboard along Highway 101 hawked some kind of AI. The evening news led with tales of GPT-4’s passing the LSAT. The Mission was full of tasteful sprigs of wildflowers, coffee with notes of blueberry, and homeless people. OpenAI’s office is

…

we address the AGI safety problem in particular. Which I think we have exciting ideas for.” We were talking two days after the release of GPT-4. Altman was leading an organization of such ostensibly epochal significance that mere products and profits were far from his mind. He seemed to revel in

…

into the future. When he talks publicly about OpenAI’s technology, he often disses its current products—he had recently told a prominent podcaster that GPT-4, the company’s most advanced product, “kind of sucks”—and invites the audience to focus on what the company’s current rate of improvement implies

…

a large language model, which up until then had been the purview of academic research. And it shows how, with the release of ChatGPT and GPT-4, Altman used his YC-honed mastery of telling startup stories to tell one of the greatest startup stories of all time. ALTMAN IS wary of

…

used dialog as an alignment tool, teaching the model as one would a student. THE COMPANY had also been working on its next foundation model, GPT-4, and figured the same method of alignment would work again. “We thought of it as a way to advance safety for

…

GPT-4,” the team member said. By the summer of 2022, OpenAI was ready to present GPT-4 to its nonprofit board. By 2022, the board had grown to nine people, with the addition of a

…

board, followed by more formal interviews with every other board member, and by 2021 she replaced Karnofsky. THE BOARD that saw the first demonstration of GPT-4 was astonished by its capabilities. A year earlier, Altman and Brockman had traveled to Seattle to visit Bill Gates and ask him what it would

…

emeritus New York University professor and author who was one of OpenAI’s most prominent critics. It killed. The board began preparing how to launch GPT-4. At the time, it seemed like it would go out into the wild the same way OpenAI’s previous models had, requiring users to prompt

…

it with examples of the kinds of patterns—questions and answers, or code—that they wanted to see. As Open AI was making progress on GPT-4, Murati, who was named CTO in May 2022, and senior research leaders were experimenting with Schulman’s chat interface as a tool to make sure

…

the OpenAI team returned to the office, realizing that the safety tool was more compelling than they had thought. When GPT-4 finished its training run in August, they made plans to release GPT-4 with the chat interface the following January. But as more people played with it, those plans began to change

…

I was using it myself. And I was like, we should separate these two things.” Altman reasoned that the shock of the two advances simultaneously—GPT-4’s collegiate smarts and the curiously lifelike chat interface—would be too much for the world to handle. “I thought that doing

…

GPT-4 plus the chat interface at the same time—and I really stand by this decision in retrospect—was going to just be a massive update

…

, that rival Anthropic had already built its own chatbot, named Claude, and was just waiting for enough safety testing to feel confident about releasing it. GPT-4 was scheduled to be released shortly after New Year’s Day, so it would have been ideal to release the chat interface a bit ahead

…

Google. “At the end of the day,” he said, “I want people to know that we made them dance.”28 IN MARCH, OpenAI finally released GPT-4 after some delays for safety testing. Now anyone in the world could experience what “adding a zero” meant. OpenAI had stopped releasing data about its

…

models, but experts estimated that GPT-4 had about 1.77 trillion parameters, roughly ten times that of GPT-3. GPT-3 could write a haiku; GPT-4 could pass the bar. University professors scrambled to create policies on AI usage and new ways

…

thousand signatories to an open letter by the Future of Life Institute calling for a six-month pause on developing AI models more powerful than GPT-4.29 The government took notice. Altman was summoned to the White House, along with Nadella, Pichai, Amodei, and Clark, for a two-hour meeting on

…

was supposed to have oversight over him. These kinds of concerns became much more urgent for several members of the board after they saw the GPT-4 demo in the summer of 2022 and realized how rapidly their decisions were becoming potentially grave ones. “For the OpenAI board to function and do

…

for the board to take seriously the way that the stakes of the company are ramping up over time,” Toner said. “Things like ChatGPT and GPT-4 were meaningful shifts towards the board realizing that the stakes are getting higher here. It’s not like we are all going to die tomorrow

…

role that seriously. During an OpenAI board meeting in the winter of 2022, as the board weighed how to release three somewhat controversial enhancements of GPT-4, Altman claimed all three had gotten DSB approval. Toner was skeptical, asked for documentation, and found that only one of them, relating to the API

…

—rather than from Altman, despite having just completed a six-hour board meeting. In late 2022, Microsoft had rolled out a version of still-unreleased GPT-4 in a test in India without getting DSB approval first. While it ultimately got it, the breach in India suggested to some board members that

…

Sutskever’s team, pinging Sutskever daily throughout on his research related to reasoning via reinforcement learning in context. Then Pachocki got pulled into working on GPT-4 until 2022. Upon Pachocki’s return to reasoning work, the decision was made to merge Sutskever’s and Pachocki’s teams. Sutskever felt that the

…

from Murati’s Slack channel that Sutskever had compiled. In one of them, Altman tells Murati that the company’s legal department had said that GPT-4 Turbo didn’t need to go through DSB review. When Murati checks with Jason Kwan, the company’s top lawyer, he is confused, and says

…

he can’t imagine how Altman would have gotten that impression; of course GPT-4 Turbo had to go through DSB. The document about Brockman was largely focused on his alleged bullying. Brockman drove out Apple veteran Steve Dowling, the

…

make use of various content-licensing deals it had been forging with companies including News Corp and Axel Springer. Instead, it released an update to GPT-4 called GPT-4o, which was faster than its predecessor and, as Altman put it, “natively multimodal,” able to switch between text, images, and audio

…

. GPT-4’s voice capability, which had been released the previous fall but was too slow and clunky to be very useful, could now conduct the kind

…

played by Scarlett Johansson. Murati starred in the video demo that OpenAI released, using her fluent Italian to demonstrate the real-time translation skills of GPT-4, represented by a blinking circle on a smartphone screen accompanied by a warm, husky, and slightly flirty voice. To drive home the point that OpenAI

…

that few people agree with you on? Absolute equivalence of brahman and atman,” X, December 26, 2022. 16.Lex Fridman, “Sam Altman: OpenAI CEO on GPT-4, ChatGPT, and the Future of AI,” Lex Fridman Podcast, March 25, 2023. CHAPTER 1CHICAGO 1.Tim Frakes, “Harold Washington Inauguration April 29 1983,” YouTube, 9

…

, 291–94, 305, 308–9, 310 in his “code cave,” 175–76, 246–47 as an OpenAI founder, 172–78, 183–88 role in developing GPT-4, 189–96, 209, 215–23, 242–51, 267–68 Brockman, Ron, 173 Brooklyn’s Park Slope, 29–30 Brooks, David, 132 Brown, Jerry, 205, 206

…

into data practices at, 285 first commercial product OpenAI API, 250–51 going from “AI training” to “AI alignment,” 264–65 GPT-5, 282, 307 GPT-4, 3, 7–8, 12, 17, 266–69, 272–73, 278–80, 284, 287, 307 GPT-1 and GPT-2, 241–43, 244, 247, 252, 253

Empire of AI: Dreams and Nightmares in Sam Altman's OpenAI

by Karen Hao · 19 May 2025 · 660pp · 179,531 words

the eye. Under the hood, generative AI models are monstrosities, built from consuming previously unfathomable amounts of data, labor, computing power, and natural resources. GPT-4, the successor to the first ChatGPT, is, by one measure, reportedly over fifteen thousand times larger than its first generation, GPT-1, released five years

…

ambiguous or exaggerated marketing. Altman has publicly tweeted that “ChatGPT is incredibly limited,” especially in the case of “truthfulness,” but OpenAI’s website promotes GPT-4’s ability to pass the bar exam and the LSAT. Microsoft’s Nadella has similarly called Bing’s AI chat “search, just better”—a tool

…

hardest: scaling up GPT-3 by 10x with Microsoft’s new eighteen thousand Nvidia A100 supercomputer cluster, in its effort to develop what would become GPT-4. One-third of the GPT-3 scaling team had left with The Divorce, taking with them significant technical and institutional knowledge. More existentially, OpenAI

…

, with McGrew, healing the ruptures Brockman caused in various parts of the company. With roadblocks that needed to be punched through in the way of GPT-4’s development, the stars aligned. * * * — To solve OpenAI’s data bottleneck, Brockman turned to a new source: YouTube. OpenAI had previously avoided this option

…

conducting reinforcement learning from human feedback. With each week, the results looked better and better, until the performance truly began to wow people internally. GPT-4 now had built-in multimodal capabilities and, against OpenAI’s internal assessments, was generating more polished code than ever and was more nimble in recognizing

…

’t the least bit impressed. It was once again the ever-hard-to-please Bill Gates. * * * — In June 2022, after getting a demo of GPT-4, Gates expressed disappointment in the insufficient progress from GPT-2. Despite the model being significantly larger and more fluent, he still felt like it was

…

things require a miracle,” he said. “We just had our miracle.” Many employees believed it, awestruck by the momentousness of what they had accomplished. GPT-4’s new level of performance convinced OpenAI leadership that it was time to start working toward one of Altman’s long-coveted ambitions: an AI

…

of flow, people from the Applied and Research divisions were working more tightly together than ever before to launch a new product. As OpenAI demoed GPT-4 to Microsoft, Satya Nadella, Kevin Scott, and the tech giant’s other executives were just as excited. Codex had proven that OpenAI’s technologies

…

could have commercial appeal, but GPT-4 represented something far bigger. Across the board, it beat the performance of various AI models that Microsoft had developed in-house; it could also do

…

weights for integrating into its products. The companies would reveal the amount in January 2023: $10 billion. * * * — At first, OpenAI executives wanted to release GPT-4 in the fall of 2022. The deadline was a case of fantastical thinking. Nearing the end of summer, the company was nowhere near ready to

…

Safety clan. DSB created a formal governance structure for resolving the age-old debates between Applied and Safety. After a preliminary review, the DSB gave GPT-4, the first model being evaluated under this structure, a conditional approval: The model could be released once it had been significantly tested and tuned

…

of uncertainty about how trust and safety for an AI company should differ from a search or social media company and whether their preparations for GPT-4 were adequate. Trust and safety was typically focused on preventing a predictable slate of internet abuses, like fraud, cybercrime, and election interference. But wrapped

…

use of its technologies. If it switched to reactive enforcement, it would need to build up significant tooling to do so. With the launch of GPT-4 pending, executives overrode the objections: OpenAI was getting rid of developer review; the trust and safety team simply needed to figure out the alternative.

…

put together a proposal. It would shift more of its enforcement of the company’s policies upstream, by leaning more heavily on RLHF to align GPT-4 and future models. Everything else would be caught and handled downstream with reactive enforcement: using different data signals, such as information about what the

…

what we always say and we’re always concerned about,” he said. “But what I haven’t ever seen is, is it actually happening?” * * * — GPT-4 wasn’t just a turning point for Gates and Microsoft. Later that summer, after wowing the billionaire philanthropist, Altman and Brockman brought it to OpenAI

…

,” Helen Toner would tell The TED AI Show podcast. “Not just like, you know, helping the CEO to raise more money.” Among many employees, GPT-4 solidified the belief that AGI was possible. Researchers who were once skeptical felt increasingly bullish about reaching such a technical pinnacle—even while OpenAI continued

…

to lack a definition of what exactly it was. Engineers and product managers joining Applied and having their first close-up interaction with AI through GPT-4 adopted even more deterministic language. For many employees, the question became not if AGI would happen but when. Some employees also felt exactly the

…

it wouldn’t be monetized but “get the data flywheel going”—in other words, amass more data from people using it—which would help improve GPT-4 and the Superassistant product. Outside of the Superassistant team, everyone took the executives literally. A low-key research preview didn’t require their attention;

…

test out the model’s capabilities. Did adding a chat interface really make a difference? People in the Safety clan, occupied with testing and tuning GPT-4, agreed. For the first time, a model release flew through the checks with little resistance. Even within the Superassistant team, no one truly fathomed

…

experience a rapid proliferation of over one hundred new generative AI projects within just a few months as employees experimented with various ways of using GPT-4 and ChatGPT. In an ironic twist, the aggressive adoption would force Microsoft to grapple with many of the same challenges that other companies would

…

the giant, OpenAI had reworked its road map to prioritize delivering the model over its own more strategically aligned projects, including an effort to apply GPT-4 to a search engine product. Instead, the failed effort left some senior Microsoft executives disappointed. There was also a new awkward reality: OpenAI and

…

from which Altman had recused himself. Altman and other executives never brought up the data centers’ environmental toll in company-wide meetings. As OpenAI trained GPT-4 in Iowa, the state was two years into a drought. The Associated Press later reported that during a single month of the model’s training

…

number led her to believe they had picked it to be slightly higher than the amount of compute that OpenAI had reportedly used to train GPT-4. But Hooker and many other researchers, including Deborah Raji, disagree with the compute-threshold approach for regulating models. While scale can lead to more

…

the impact of their deployments accelerated, he believed the company needed to raise, not lower, its guard against their potential to produce devastating consequences. After GPT-4, Sutskever, who had previously dedicated most of his time to advancing model capabilities, had made a hard pivot toward focusing on AI safety. He

…

board had been in a monthslong deadlock over whom to appoint as new independent directors. As part of their effort to increase oversight after the GPT-4 demo, and even more after the launch of ChatGPT, McCauley had engaged in a roughly yearlong process, including interviewing employees and stakeholders outside the

…

their own sources about various problems, including the company’s lack of preparation before and significant tumult after ChatGPT, the continued AI safety concerns surrounding GPT-4’s release, and the unprecedented pace with which OpenAI was sprinting to launch new products before it had resolved many of its issues. One

…

source of the misalignment. The tangled situation had caused several months of organizational thrash in the Research division. It was now, just as with the GPT-4 pre-training team crisis, reaching untenable levels of stress. For Sutskever, the ongoing saga was deeply painful. Not only was it a humiliating snub

…

inference costs, had exceeded performance expectations during training, based on the company’s own testing; leadership subsequently left the model to train longer to surpass GPT-4. More compelling, Scallion could also work with three modalities: language, vision, and, the most recent addition, audio. By then, users could already speak with

…

to outshine Anthropic. A month earlier, Anthropic had released its latest model, Claude 3, also through its chatbot and API, and it was uncomfortably outperforming GPT-4. Meanwhile, Orion, OpenAI’s latest GPT model meant to take back the lead, was struggling with serious development delays. To employees, Altman and Brockman

…

less than 2 percent are supported by Google Translate; and according to OpenAI’s own testing, only fifteen, or 0.2 percent, are supported by GPT-4 above an 80 percent accuracy. As these models become digital infrastructure, the internet’s accessibility to different language communities—and the accessibility of the economic

…

Kindle. GO TO NOTE REFERENCE IN TEXT For Altman’s part: Lex Fridman, host, Lex Fridman Podcast, podcast, episode 367, “Sam Altman: OpenAI CEO on GPT-4, ChatGPT, and the Future of AI,” March 25, 2023, lexfridman.com/podcast. GO TO NOTE REFERENCE IN TEXT “The thing that sticks”: Sam Altman, “How

…

founder on Company’s Past Approach to Openly Sharing Research: ‘We Were Wrong,’ ” The Verge, March 15, 2023, theverge.com/2023/3/15/23640180/openai-gpt-4-launch-closed-research-ilya-sutskever-interview. GO TO NOTE REFERENCE IN TEXT “it may be that today’s”: Ilya Sutskever (@ilyasut), “it may be

…

November 17, 2021, by MIT-Haiti Initiative, Facebook, 2 hr., 56 min., 46 sec., facebook.com/mithaiti/videos/1060463734714819; OpenAI, “GPT-4,” OpenAI, March 14, 2023, openai.com/index/gpt-4-research. GO TO NOTE REFERENCE IN TEXT It was up against: Author interviews with Keoni Mahelona, October, November, and December 2021; and

…

53–54 fundraising, 61–62, 65–68, 71–72, 132, 141, 156, 262, 320–21, 331, 367, 377, 405 GPT-3, 133–34, 278–79 GPT-4, 246, 248–52, 279, 346, 383–84, 386, 390–91 Graham and, 28, 32, 36–39, 40, 69 “Intelligence Age,” 19, 405 Jobs comparisons with

…

OpenAI, 53–54, 84–85 departure of, 404 Dota 2, 66, 144–45 founding of OpenAI, 28, 46–51 governance structure of OpenAI, 61–63 GPT-4, 244–48, 250–51, 252, 257, 260, 346 Latitude, 180–81 leadership of OpenAI, 58–59, 61–62, 63–65, 69, 70, 83, 84–85

…

(Consumer Financial Protection Bureau), 419–20 chatbots, 17, 112–14, 189–90, 217–18, 220 ELIZA, 95–97, 111, 420–21 GPT-3, 217–18 GPT-4, 258–59 LaMDA, 153, 253–54 Meena, 153 Tay, 153 ChatGPT, 258–62, 267, 280 connectionist tradition of, 95 GPT-3.5 as basis, 217

…

Anaya, Oskarina Veronica, 197–202, 415–17 Future Perfect, 388 Futures of Artificial Intelligence Research, 273–74 G Gates, Bill, 68 congressional testimony of, 311 GPT-4, 245–48 OpenAI demo, 71–72, 132–33, 246 Gates Demo, 71–72, 132–33, 246 Gawker Media, 38 GDPR (General Data Protection Regulation), 136

…

captchas, 98 data centers, 274–75, 285–91, 295–96 DeepMind. See DeepMind DNNresearch, 47, 50, 98–99, 100 Frontier Model Forum, 305–6, 309 GPT-4 and, 249 Imagen model, 240, 242 LaMDA, 153, 253–54 neural networks, 100–101 Project Maven, 52 speech recognition, 100 Sutskever and, 50, 100–101

…

179, 242–43, 244, 253 GPT-3.5, 135, 183–84, 189, 217–18, 247, 258, 259–60, 264, 269, 378 GPT-3.75, 378 GPT-4, 189, 244–53 Bing, 112, 113, 247 capabilities, 16, 119, 135–36, 245–53, 410 development, 242, 244–53 release, 258–62, 323–24 Superassistant

…

, 247–49, 258–59, 381 GPT-4o, 383–84, 386, 390–91 GPT-4 Turbo, 346, 363 GPT-5, 279, 325 Orion, 374–75, 379, 380, 405 GPUs (graphics processing units), 61–62, 134, 265–68. See also

…

partnership, 18, 67–68, 71–72, 234, 264–67, 269–70, 402 ChatGPT, 264, 265–66 compute phases, 278–81 GPT-3, 156, 278–79 GPT-4, 245–48, 279, 324 investments and funding, 13, 17, 72, 75, 80–81, 84–85, 132–33, 143, 145, 156, 248, 331 Microsoft Research, 68

…

324–25 Zuckerberg and, 406–7 Mutemi, Mercy, 212, 291 N Nadella, Satya, 113 Altman’s firing, 4, 6, 10, 367 congressional testimony of, 311 GPT-4, 247–48, 346 OpenAI partnership, 67–68, 71, 72, 248, 265, 270 Nairobi, Kenya, 190–91, 193, 207, 208, 212, 219, 416 Napoleon Bonaparte, 399

…

firing and reinstatement, 6, 8, 365–66, 366, 373 leadership behavior, 347–48, 353, 355–56 Dota 2, 145, 244–25 GPT-3, 244–45 GPT-4, 312 new chief scientist, 386–87, 406 Omnicrisis, 396–98 Page, Larry, 24, 25–26, 51, 249 Pakistan, 222 Pang, Wilson, 199 paper clips,

We Are as Gods: A Survival Guide for the Age of Abundance

by Peter H. Diamandis and Steven Kotler · 13 Apr 2026 · 225pp · 76,418 words

plummeted faster than any technology in history, making sophisticated AI available to small developers and individuals within months of its initial release. In 2023, accessing GPT-4 cost roughly $36 per million tokens. By mid-2024, GPT-4o mini brought that below twenty-five cents. The price of intelligence is collapsing even

The Coming Wave: Technology, Power, and the Twenty-First Century's Greatest Dilemma

by Mustafa Suleyman · 4 Sep 2023 · 444pp · 117,770 words

, and the cost of training an equivalent model has fallen tenfold over the last two years. When GPT-4 launched in March 2023, results were again impressive. As with its predecessors, you can ask GPT-4 to compose poetry in the style of Emily Dickinson and it obliges; ask it to pick up from

…

had built tools that automated lawsuits, helped co-parent children, and offered real-time fashion advice. Within weeks they’d created add-ons so that GPT-4 could accomplish complex tasks like creating mobile apps or researching and writing detailed market reports. All of this is just the start. We are only

…

unfolding research process. So, where does AI go next as the wave fully breaks? Today we have narrow or weak AI: limited and specific versions. GPT-4 can spit out virtuoso texts, but it can’t turn around tomorrow and drive a car, as other AI programs do. Existing AI systems still

…

be effective. AlphaGo and DQN were not in themselves autonomous. But they hint at what a self-improving system might look like. Nobody hand codes GPT-4 to write like Jane Austen, or produce an original haiku, or generate marketing copy for a website selling bicycles. These features are emergent effects of

…

outputs are never decided in advance by its designers. This is the first step on the ladder toward greater and greater autonomy. Internal research on GPT-4 concluded that it was “probably” not capable of acting autonomously or self-replicating, but within days of launch users had found ways of getting the

…

to explain precisely why an algorithm produced a specific prediction. Engineers can’t peer beneath the hood and easily explain what caused something to happen. GPT-4, AlphaGo, and the rest are black boxes, their outputs and decisions based on opaque and intricate chains of minute signals. Autonomous systems can and may

…

this problem alone. But together “we” all might. 10. THE NARROW PATH: THE ONLY WAY IS THROUGH Just a few days after the release of GPT-4, thousands of AI scientists signed an open letter calling for a six-month moratorium on researching the most powerful AI models. Referencing the Asilomar principles

…

think it will be done The Alignment Research Center has already tested GPT-4 for precisely this kind of capability. GPT-4 was, at this stage, “ineffective” at acting autonomously, the research found. “GPT-4 System Card,” OpenAI, March 14, 2023, cdn.openai.com/papers/gpt-4-system-card.pdf. Within days of launch people were getting surprisingly

…

,” DeepMind, Nov. 10, 2022, www.deepmind.com/publications/a-generalist-agent. GO TO NOTE REFERENCE IN TEXT Internal research on GPT-4 @GPT-4 Technical Report, OpenAI, March 14, 2023, cdn.openai.com/papers/gpt-4.pdf. See mobile.twitter.com/michalkosinski/status/1636683810631974912 for one of the early experiments. GO TO NOTE REFERENCE IN TEXT

…

Early research even claimed Sébastien Bubeck et al., “Sparks of Artificial General Intelligence: Early Experiments with GPT-4,” arXiv, March 27, 2023, arxiv.org/abs/2303.12712. GO TO NOTE REFERENCE IN TEXT AIs are already finding ways Alhussein Fawzi et al., “Discovering

…

and, 258–63 organizational limitations of, 148–50 See also nation-states GPS (Global Positioning System), 110 GPT-2, 64, 70 GPT-3, 64, 68 GPT-4, 64, 113–14 GPUs, 130, 251 grand bargain, defined, viii Great Britain corporations and, 186, 189 surveillance, 193, 195–96 great power competition. See geopolitics