Some observers professed whiplash at OpenAI’s one-two punch: creating a for-profit component and reaching an exclusive deal with Microsoft. How did a company that promised to remain patent-free, open source, and totally transparent wind up giving an exclusive license of its tech to the world’s biggest software company? Elon Musk’s remarks were particularly lacerating. “This does seem like the opposite of open—OpenAI is essentially captured by Microsoft,” he posted on Twitter. On CNBC, he elaborated with an analogy: “Let’s say you founded an organization to save the Amazon rainforest, and instead you became a lumber company, chopped down the forest, and sold it.”
Musk’s jibes might be dismissed as bitterness from a rejected suitor, but he wasn’t alone. “The whole vision of it morphing the way it did feels kind of gross,” says John Carmack. (He does specify that he’s still excited about the company’s work.) Another prominent industry insider, who prefers to speak without attribution, says, “OpenAI has turned from a small, somewhat open research outfit into a secretive product-development house with an unwarranted superiority complex.”
Even some employees had been turned off by OpenAI’s venture into the for-profit world. In 2019, several key executives, including head of research Dario Amodei, left to start a rival AI company called Anthropic. They recently told The New York Times that OpenAI had gotten too commercial and had fallen victim to mission drift.
Another OpenAI defector was Rewon Child, a main technical contributor to the GPT-2 and GPT-3 projects. He left in late 2021 and is now at Inflection AI, a company led by former DeepMind cofounder Mustafa Suleyman.
Altman professes not to be bothered by defections, dismissing them as simply the way Silicon Valley works. “Some people will want to do great work somewhere else, and that pushes society forward,” he says. “That absolutely fits our mission.”
Until November of last year, awareness of OpenAI was largely confined to people following technology and software development. But as the whole world now knows, OpenAI took the dramatic step of releasing a consumer product late that month, built on what was then the most recent iteration of GPT, version 3.5. For months, the company had been internally using a version of GPT with a conversational interface. It was especially important for what the company called “truth-seeking.” That means that via dialog, the user could coax the model to provide responses that would be more trustworthy and complete. ChatGPT, optimized for the masses, could allow anyone to instantly tap into what seemed to be an endless source of knowledge simply by typing in a prompt—and then continue the conversation as if hanging out with a fellow human who just happened to know everything, albeit one with a penchant for fabrication.
Within OpenAI, there was a lot of debate about the wisdom of releasing a tool with such unprecedented power. But Altman was all for it. The release, he explains, was part of a strategy designed to acclimate the public to the reality that artificial intelligence is destined to change their everyday lives, presumably for the better. Internally, this is known as the “iterative deployment hypothesis.” Sure, ChatGPT would create a stir, the thinking went. After all, here was something anyone could use that was smart enough to get college-level scores on the SATs, write a B-minus essay, and summarize a book within seconds. You could ask it to write your funding proposal or summarize a meeting and then request it to do a rewrite in Lithuanian or as a Shakespeare sonnet or in the voice of someone obsessed with toy trains. In a few seconds, pow, the LLM would comply. Bonkers. But OpenAI saw it as a table-setter for its newer, more coherent, more capable, and scarier successor, GPT-4, trained with a reported 1.7 trillion parameters. (OpenAI won’t confirm the number, nor will it reveal the data sets.)
