Greg Brockman reveals the inside story of ChatGPT at TED

Imagine planning dinner without opening three different apps. You tell an AI what you want to eat, and it creates the recipe, renders a picture of the finished dish, drafts a social post, and drops every single ingredient straight into your delivery cart.

On April 18 at TED2023 in Vancouver, OpenAI co-founder Greg Brockman plugged his laptop into the big theater screen to demo exactly that, showing off unreleased plugins running on top of GPT-4.

It is the clearest sign yet that OpenAI does not want ChatGPT to stay a text sandbox. It wants it running real-world errands.

The chat box becomes an operating system

During his live demo, Brockman showed how ChatGPT can fluidly chain disparate services together without the user ever leaving the prompt box.

After sketching out a meal, the bot generated an image of the plate, drafted a tweet about it, and populated a full grocery list in Instacart ready for checkout.

Brockman also showed off the bot's emerging analytical and self-policing chops. ChatGPT demonstrated an ability to fact-check its own output with human-verifiable citations, as well as make sense of data-heavy spreadsheets based on relatively vague prompts.

Once the demo wrapped, TED head Chris Anderson joined Brockman on stage to press him on ChatGPT's rollout timeline and the swelling wave of safety concerns.

The broader session underscored that unease. Cognitive scientist Gary Marcus called for global regulatory bodies, Yejin Choi warned about resource centralization and the lack of common sense, and Eliezer Yudkowsky cautioned that opaque models are hurtling toward existential risks.

Why it matters

We think this presentation marks a critical pivot from passive question-answering to active execution.

For everyday users and knowledge workers, it shows a future where software interfaces collapse into a single text prompt. You provide the goal, and the model handles cross-platform coordination, data crunching, and commerce behind the scenes. If this works reliably, standalone single-utility apps will face a massive existential threat.

However, letting a probabilistic chatbot take real actions in the outside world brings obvious friction. Hallucinations are annoying in text, but they can be costly when connected to credit cards and third-party services.

OpenAI is trying to address this by having the model cite sources and audit its own claims, which looks like a clever patch. Still, whether automated fact-checking is robust enough to protect users from high-stakes mistakes remains an open question.