The Technology Behind ChatGPT Is Evolving Insanely Fast
It’s been only a year since OpenAI released ChatGPT, but the technology has evolved at such a rapid pace that the original now seems almost quaint.
Many in the tech field were surprised by the technology a year ago and have been astonished at just how fast generative AI is improving.
When it launched to the public on November 30, 2022, ChatGPT was text-only and could only answer questions based on its training data up to September 2021. Plus, it was highly prone to making up facts when it didn’t know the answer, quickly introducing to the world a new meaning for the word “hallucination.”
- Still, ChatGPT was surprisingly powerful and became an overnight success. “Scary good” was how we described it in a story days after its release.
- More than 1 million people used it in the first five days.
Today’s ChatGPT is trained up to April 2023 and can use Microsoft’s Bing and the web to check for even more recent developments.
- It’s also multimodal, meaning it has the ability to use photos or documents as part of the search and to converse in spoken word as well.
Then there’s the ability to create custom GPTs that OpenAI introduced at its DevDay this month.
- Already thousands of those GPTs exist, including ones to create websites, automate tasks and even create other custom GPTs.
TChatGPT is just one example of a generative AI application.
- Enterprises are customizing their own chatbots to answer questions from very specific sets of data.
- Generative AI is being applied to specific domains, including law, medicine, and helping the world adapt to climate change.
Asking a generic tool like ChatGPT to write a legal brief or make a diagnosis remains dicey. But results improve markedly if you use a specific engine trained on the best and latest information in a particular field.
- Similar technology has been a game changer in the field of computer code, where generative AI has shown great promise in turning anyone with an idea into a programmer and making experienced programmers more productive.
- There’s also great power in using large language models not to answer questions but to serve as a natural language interface to complicated programmes. Today’s users typically harness only a fraction of the capabilities of sprawling tools like Photoshop, Word, and Excel.
- And it’s not just about text. Diffusion-based image models have gone from Dali-esque to photorealistic. And, speaking of DALL-E, ChatGPT now integrates that image-creation tool (at least for ChatGPT subscribers).
Expect the next year to bring even more advanced ways to turn the technology behind ChatGPT loose on all sorts of data.
As fast as the technology has evolved, OpenAI and others have gone slow in allowing their AI engines to take action, especially without explicit human consent.
As AI engines begin to take action on their own and interact with other AI agents, the opportunity- and risk-filled ChatGPT evolutionary cycle could well speed up further.
