The news: On Tuesday, OpenAI released GPT-4, a large multimodal AI model that surpasses its predecessor, GPT-3.5, in capabilities.
- The model can accept text and image inputs and spit out generative text with “human-level performance,” OpenAI says, per Ars Technica.
- The startup also claims that the model passed a simulated bar exam, scoring around the top 10% of test takers, compared with GPT-3.5, which scored in the bottom 10%.
- In addition to the new functionality of describing images in detail, GPT-4 is more precise, has higher accuracy, can accurately summarize an article, and can exhibit humor, per The New York Times.
- The upgraded model is available via ChatGPT Plus, and its API has a waitlist.
We’ve met GPT-4 before: Alongside the release, Microsoft disclosed that its Bing AI chatbot is based on GPT-4.
- Reviewing the success and pitfalls of Bing AI’s debut gives some indication of GPT-4’s capabilities.
- We can expect GPT-4 to at times present fiction as fact, struggle with current events, and be unpredictable.
- The revelation shows that increasing model complexity alone won’t solve all of generative AI’s problems and could make some issues worse.
What this means for Google: The upgrade coming mere months after ChatGPT emerged could be a sign that we’ll see steady commercial releases of increasingly sophisticated generative AI models.
- Google hasn’t released its Bard, yet it’s now talking about its more powerful Big Bard, likely as a result of pressure to show that it’s working on something impressive.
- GPT-4’s functionality means Google has to play catch-up again and give Big Bard multimodal features that surpass GPT-4 capabilities.
- The challenge for Google is to hit a bullseye in its release of a model that will offer more than the competition and that doesn’t embarrass the company.