Meet the AIs that can write — and code
Two new AI models out this week show the power of artificial intelligence to read text, write it — and even convert it into computer code.
Why it matters: Natural language processing (NLP) is one of the most exciting areas in AI research, with major implications for how we'll communicate and work in the years ahead.
- But it also opens the door to a future in which you'll be able to "talk to your computer and get it to do what you're asking in a capable, reliable way," notes Greg Brockman, a co-founder and chief technology officer of OpenAI.
Driving the news: On Wednesday morning, the Israeli startup AI21 Labs is releasing a line of language-generating AI models called Jurassic.
- With 178 billion parameters — the values that a neural network tries to optimize during training — Jurassic-1 Jumbo is the largest such model in the world, slightly bigger than OpenAI's GPT-3, which was released last summer.
- The models will power AI21 Studios, which aims to provide NLP as a service for businesses. Developers will be able to sign up for free to try the service, and can then apply for access to custom models for commercial aims "to use our technology to quickly build any text-based applications they can dream up," says Dan Padnos, VP of platform at AI21.
How it works: Like other large language models, including GPT-3, the AI21 models have been trained on massive amounts of text, which enables the system to learn the statistical relationships between words and use that to read text prompts from users and write text in response.
- The AI21 models, however, aim to provide a more customizable and user-friendly interface, highlighting specific functions, including headline writing, summarization, even an app to dejargonize language.
- During an online demo, Padnos fed a number of stories from the Axios What's Next newsletter into the model, which was then able to produce a one-line summary — or what we here call "the bottom line."
- It wasn't quite good enough to put me out of a job, but it was good enough to make me worry about that future possibility.
Situational awareness: On Tuesday, OpenAI itself released Codex, an updated descendant of its GPT-3 model, for a private developer beta test.
- Codex was trained on both huge amounts of text and billions of lines of publicly available computer code.
- Users can issue commands in written English, and Codex will produce computer code capable of carrying out those instructions, essentially making it an English to computer code translator.
What they're saying: Codex — which in action can feel like you're talking to a computer, "Star Trek"-style — can "remove the barriers of entry" for ordinary people to code and move us toward a future "where the entire world is programmable," says Brockman.
By the numbers: According to one estimate, the global market for NLP is projected to grow from roughly $21 billion in 2021 to over $127 billion in 2028.
- When you consider just how much of business revolves around the analysis and production of language, that might be conservative.
- "We're just at the beginning of where adoption will end up with these tools," says Henrik Roth, a co-founder and CMO of the marketing company neuroflash, which has experimented with GPT-3.
The catch: The models still make errors, have problems with bias and struggle with longer strings of text or code.
- They could also be employed by malign actors — for example, to soup up a disinformation campaign — although both OpenAI and AI23 say they'll control access to prevent misuse.
The bottom line: Let's give AI21's Jurassic-1 Jumbo a chance by inputting this story into the model:
- "Wednesday's release of Jurassic-1 Jumbo and Tuesday's release of Codex are the latest examples of AI's rapid progress, but they're still a long way from being perfect," the model wrote.
- Not bad — and maybe on the part of the AI, a bit of a humblebrag.