ChatGPT can now access the Internet and run the code it wrote
Linguistic model AIs teach themselves the art of communication and problem solving based on limited training data. In the case of GPT-4, that data is quite out of date, as the deadline is the end of 2021. That’s where all the ChatGPT “knowledge” up to this point came from, and its only output, at least in the service. public can use – was text. Now, after today’s launch of the plugin ecosystem, GPT is back on the rise with some impressive new abilities.
First of all, it’s now accessible to the Internet, which means it can scour the web for answers if it decides you need up-to-date information that isn’t in its knowledge base. To do this, it formulates relevant search strings, sends them to Bing, looks at the results, then goes and reads the links it deems worthy until it decides it’s got a good answer for you. You can watch exactly what it’s talking about while it does it, and when your answer comes in, it’s neatly labeled with links you can click on to go and research the relevant sources yourself.
For now, its web browser activity is read-only, rather than sending “fetch” requests to Bing. It can’t fill out forms or do anything else online, so it can’t just go ahead and put its unfettered copies on a hidden server somewhere and start engaging in the kind of “power-seeking behavior” that has already been caught. showing.
However, OpenAI keeps everything that happens within its search API separate from the rest of its infrastructure, just to be sure. It can’t visit sites that aren’t accessible through Bing’s “safe mode,” and it won’t visit sites that ask not to dive into their robots.txt files.
Second, it can now run the code it wrote. OpenAI has given it a working Python interpreter sitting in a “sandboxed, firewalled execution environment,” along with some disk space that remains available for the duration of your chat session or until it ends. It can also now upload and download files.
So if you ask it a question that requires some serious number crunching, it’s now able to code task-specific software and run that code to complete your task. You can feed it data in certain file formats and it will perform operations on that data and give you something back, possibly in a different format if that’s what you ask for.
This is some pretty embarrassing stuff. It will take a spreadsheet and create annotated graphs for you. It will accept JPGs, tell them what they look like, and write and run code to resize those images or convert them to grayscale.
And it gets access to a bunch of premium third-party plugins, with tons more to come. For example, the Expedia, OpenTable, and Kayak apps can search for and manage flight, restaurant, lodging, and rental car reservations. Instacart, Klarna and Shop plugins can find and compare products and order. The Wolfram|Alpha application gives GPT access to mathematical and computational power, as well as real-time data streams.
At this stage, it seems that its capabilities are mostly limited to setting things up rather than actually doing transactions with your money; you will have to click through and adjust the amount yourself.
Finally, the Zapier app acts as a gateway through which GPT can now access nearly 5,000 other apps, including Gmail, Google Sheets, Trello, HubSpot, and Salesforce. This begins to position GPT as the ultimate personal assistant, with access to vast amounts of your personal and company information and possibly permissions to log in and perform a variety of tasks for you. Unusual things.
These plugins are gradually becoming available to paid users and developers via a waiting list. And new plugins are going to spread at an extraordinary rate because no one even needs to code them. “You write an OpenAPI manifest for your API, use human language descriptions for everything, and that’s it,” wrote developer Mitchell Hashimoto. “You let the model figure out how to validate, chain calls, process the data, format it for viewing, and so on. Glue code is absolutely zero.”
I’ve developed a lot of plugin systems, and the OpenAI ChatGPT plugin interface might be the craziest and most impressive approach I’ve ever seen in my entire life in computing.
— Mitchell Hashimoto (@mitchellh) March 23, 2023
The pace of progress in OpenAI has been absolutely dizzying over the past few months. This insanely advanced AI seems to be getting a massive overhaul, with extraordinary new abilities every time we blink. These new plugins represent ChatGPT starting to reach outside of the box it’s housed in and out into the real world.
For now, its capabilities will be extremely limited, as OpenAI knows more about the potential dangers of this unique technology than anyone else. But assuming these guys are good guys and they’ve taken the time to make sure it’s done safely, the massively disruptive appearance of GPT will certainly prompt other, less principled and less capable actors to rush to develop competing AIs and make them compete. powers.
The possibilities here are absolutely incredible and the risks are unprecedented with every step of this technology. At this stage we are in uncharted territory, with very limited forward vision and the gas pedal stuck to the floor. What a time to live!