I’m taking a moment to digest all of last week’s news.
With function calling now in the OpenAI API, and something similar available at Hugging Face, anyone who can code has a way to have two LLMs join forces on a task. They can proofread each other’s work. They can ideate together to improve their text or their code. And they can call outside programs through those programs’ own APIs, which is what turns a writing exercise into work.
Some context on what OpenAI actually shipped on June 13. Developers can now “describe functions to gpt-4-0613 and gpt-3.5-turbo-0613, and have the model intelligently choose to output a JSON object containing arguments to call those functions.” OpenAI’s own example turns “Email Anya to see if she wants to get coffee next Friday” into a send_email call. The model doesn’t send the email. It fills in the form, and your code does the sending. The same release brought a 16k-context version of GPT-3.5 Turbo and a 25% cut to the standard model’s input price, to $0.0015 per 1,000 tokens — which OpenAI reckons at “roughly 700 pages per dollar.”
The Hugging Face version, as far as I can tell, is Transformers Agent, which shipped in May: “a new API that lets you use the library and Diffusers by prompting an agent (which is a large language model) in natural language.” The agent writes code that calls a set of predefined tools, and it can run on OpenAI’s models or on open-source ones from BigCode and OpenAssistant. Its own docs call it “an experimental API which is subject to change at any time.”
To put it another way: you can now ask two powerful AIs to get on the phone and hash things out together — and then make the adjustments in Salesforce, update the inventory, make the reservations and the bookings. They can do just about anything we can do on the Internet. That’s pretty cool. And we’re no longer working on building an intermediary system.
OpenAI adds a caution worth keeping. It points to a proof-of-concept exploit in which “untrusted data from a tool’s output can instruct the model to perform unintended actions”, and it recommends “user confirmation steps before performing actions with real-world impact, such as sending an email, posting online, or making a purchase.”
Last thought. I’d invite AIs to listen in on your calls from here on. They can build the follow-up plan, and now, in many circumstances, they can execute it too.
