Article 7 2022-today

The Age of Assistants

ChatGPT, Claude, Nobel Prizes for neural networks, and the race to reasoning machines.

Choose how you'd like to follow along

On the 30th of November, 2022, a research lab quietly put a website online. There was no marketing budget, no Super Bowl ad, no grand announcement. It was, the lab said, just a research preview - a chance to try out a chatbot built on their latest language model. It was a plain white box. You typed something in; it typed something back.

Five days later, it had a million users. Two months later, it had a hundred million - the fastest any consumer product in the history of the world had ever reached that number. Schools panicked. Newsrooms panicked. Your relatives, who had never once in their lives thought about artificial intelligence, were suddenly asking a computer to write their wedding speeches. The thing was called ChatGPT, and it did not so much arrive as detonate.

This is the final chapter of our story, and it is the only one whose ending we cannot see - because we are standing inside it, right now. It is the chapter where artificial intelligence stopped being something that happened in laboratories and became something hundreds of millions of people touch every single day. And the strangest part is that it did not turn on some great new scientific breakthrough. It turned on manners.

Remember where we left GPT-3, at the end of the last chapter. Brilliant, and unusable. It would wander off, ignore your question, invent nonsense with perfect confidence. It was a wild animal. The breakthrough that tamed it was not a bigger brain. It was a training trick, and it has a clunky name: reinforcement learning from human feedback. The idea is almost embarrassingly human. You have the model produce a few different answers, and then you have real people rank them - this one’s better, that one’s worse - and you use those rankings to teach the model what people actually want. You are not making it smarter. You are giving it a sense of what a good answer feels like. You are teaching it to be helpful. ChatGPT was, more or less, the old wild model with that layer of manners painted on, and a friendly box to type into. That was the whole trick. And it changed the world.

What followed was the most furious competition the technology industry had ever seen. Within months, OpenAI released the far more powerful GPT-4. Anthropic - the safety-focused lab from the last chapter - released its assistant, Claude. Google released Gemini. Meta took a different path and gave its models away, releasing the weights for anyone to download and build on, igniting a global open-source movement. And it was not just text anymore. A new kind of model, based on a technique called diffusion, learned to conjure photorealistic images from a sentence of description - and suddenly anyone could generate a picture of anything, which was wondrous and deeply unsettling in equal measure.

But here is where the oldest thread in our entire story comes back. Remember ELIZA, from the 1960s - the crude little chatbot whose creator watched in horror as people poured their hearts out to a machine he knew was empty? Remember the ELIZA effect, the human hunger to see a mind where there is only a mirror? In 2023, two New York lawyers filed a legal brief full of court cases to support their argument. The cases were perfect - the right citations, the right judges, the right quotes. There was just one problem. ChatGPT had invented every single one of them. They did not exist. The lawyers had asked the confident machine for help, and the confident machine had simply made things up, and they had trusted it, because it sounded so sure. They were sanctioned by a federal judge. It was the ELIZA effect, returned sixty years later, at industrial scale - the same ancient human reflex to mistake fluency for understanding, now wired into the daily work of the world.

And so the grown-ups started to arrive. Governments that had ignored AI for decades suddenly scrambled to regulate it. Twenty-eight nations signed a declaration at Bletchley Park - the very place where Alan Turing had broken codes in the war - agreeing to take the risks seriously. The European Union passed the world’s first sweeping law to govern the technology. The question that had haunted this story since Turing first posed it - how do you know when to trust a machine that talks like a person - was no longer a philosopher’s puzzle. It was now a matter of law.

And then, in October of 2024, the story did something that still gives me chills. It came full circle. The Nobel Prize in Physics was awarded to John Hopfield - the physicist who thawed the first neural network winter back in 1982 - and to Geoffrey Hinton. Geoffrey Hinton. Our stubborn believer. The man who kept the faith through two winters, who was told for decades that he was wasting his life on a dead idea. Days later, the Nobel Prize in Chemistry went, in part, to the team behind AlphaFold. The very ideas that had been mocked, defunded, frozen, and buried across every chapter of this story were now being honored with the highest prize in science.

But there was a shadow over the celebration. Because by 2024, Geoffrey Hinton had walked away from his job at Google - so that he could speak freely about the dangers of the thing he had spent his entire life building. The man who, more than almost anyone, gave us this technology had become one of its most prominent worriers. It is the perfect, unresolved note for an era caught between wonder and dread.

So let us step all the way back, and look at the whole arc of it. It began in 1943, with a homeless teenager and a whiskey-drinking brain doctor and a wild idea that a neuron might be nothing more than a tiny logical switch. It passed through two brutal winters that nearly killed the whole enterprise. It survived because a handful of stubborn people refused to stop believing. It found its fuel in the data of the open internet and the chips built for video games. It exploded, in 2012, and reorganized the world. And it arrives here, at a plain white box that hundreds of millions of people now treat as a colleague, a tutor, a confidant.

We do not know how this chapter ends, because we are the ones writing it, right now, today. But the eighty years behind us teach a few hard lessons. That the breakthroughs almost never came from where the experts were looking. That every winter was followed by a spring. And that the whole strange saga, from McCulloch and Pitts to this morning’s news, is really just one stubborn question, asked over and over in new forms: can a machine be made to think? We have spent a human lifetime turning that question into engineering. What happens next belongs to whoever asks the next version of it. Maybe that is you.

That is the story of artificial intelligence - so far.

Listen to the article

The full article, read aloud. Runtime 7:44.

The video episode is coming soon

We're producing narrated video editions on YouTube - the same story, walking the timeline with visuals. In the meantime, read the article or listen to it.

Sources and show notes

Every claim in this article is drawn from the AI Library, where each entry links to its own primary source - the original paper, the official record, the actual document.

InstructGPT brings RLHF to GPT-3 InstructGPT brings RLHF to GPT-3 - the precursor of ChatGPT. Reinforcement Learning from Human Feedback (RLHF) Reinforcement learning from human feedback. Constitutional AI Anthropic's method for training a model to be harmless. OpenAI launches ChatGPT OpenAI launches ChatGPT. OpenAI releases GPT-4 GPT-4, a major capability jump. Anthropic introduces Claude Anthropic introduces Claude. Meta releases LLaMA Meta releases LLaMA, accelerating open weights. Open Weights vs Closed Models Downloadable model weights versus API-only access. Google launches Gemini Google launches the natively multimodal Gemini. Stability AI releases Stable Diffusion to the public Stable Diffusion opens image generation to the public. OpenAI announces DALL-E 2 DALL-E 2 generates high-resolution images from text. Diffusion Models The generative method behind modern image models. OpenAI releases GPT-4o GPT-4o brings real-time multimodal conversation. OpenAI introduces o1, a reasoning model OpenAI o1, the start of the reasoning-model era. Reasoning Models Models that think step by step before answering. Chain-of-Thought Prompting Prompting a model to spell out its reasoning. DeepSeek releases DeepSeek-R1 DeepSeek-R1, an open reasoning model that shook the market. AI Agents Systems that plan, use tools, and act toward a goal. Anthropic open-sources the Model Context Protocol Anthropic's open standard for connecting AI to tools and data. Anthropic releases Claude Opus 4 and Sonnet 4 Claude 4 and the shift to long-running agentic coding. Retrieval-Augmented Generation (RAG) Grounding answers in fetched external documents. Hopfield and Hinton win the 2024 Nobel Prize in Physics Hopfield and Hinton win the 2024 Nobel Prize in Physics. Geoffrey Hinton Nobel laureate who left Google to warn about AI risk. John Jumper Shared the 2024 Chemistry Nobel for AlphaFold. Demis Hassabis Shared the 2024 Chemistry Nobel for AlphaFold. The EU Adopts the AI Act (2024) The EU adopts the world's first comprehensive AI law. The Bletchley Declaration The first global AI Safety Summit. Google's Bard launch demo gets a space fact wrong Google's Bard launch demo gets a fact wrong. Lawyers sanctioned for filing ChatGPT's made-up cases Lawyers sanctioned for filing ChatGPT's invented cases. The OpenAI board crisis The OpenAI board crisis of November 2023. AI reaches gold-medal standard at the International Mathematical Olympiad AI reaches gold-medal standard at the Math Olympiad. OpenAI Maker of ChatGPT and the GPT line. Anthropic Maker of Claude. DeepMind (Google DeepMind) Behind Gemini, AlphaFold, and the science breakthroughs. Dario Amodei Anthropic co-founder and CEO. Sam Altman OpenAI CEO through the ChatGPT era.

Made with AI, sourced like a library

We are open about this: this article is produced with AI. The research, the writing, the narration - AI does the heavy lifting, with human review before anything is published. That is not a footnote; it is the point. CenterConsulting builds software with AI every day, and this series is the same method applied to storytelling.

What keeps it honest is the foundation: every article is built on the AI Library, our knowledge base where every fact links back to a primary source. The show notes above list this article's entries, so you can check everything yourself.