GPT-4 Has Arrived! Everything You Need To Know Now
OpenAI Announced The Rollout Of Their Latest Language Model GPT4
In its usual casual manner, OpenAI announced today that the most powerful language model available will be rolled out to developers and people with OpenAI API access. 🎉
This is the news everyone has been waiting for since a regional Microsoft CTO got the rumor mill churning last week.
In their latest blog post, OpenAI mentioned that GPT-4 is actually already in use in apps by Duolingo, Be My Eyes, Stripe, Morgan Stanley, Khan Academy, and — you guessed it — the Government of Iceland.
Now, here’s the good news: as a ChatGPT Plus subscriber, you can already use GPT-4 (limited to 100 messages/hour). If you’re not using ChatGPT Plus you’ll have to wait a bit or get on the API waitlist here.
So, what’s it all about? Is GPT-4 living up to the expectations?
Well, new benchmarks have been achieved, and new capabilities have been unlocked, but old issues remain as well.
I’ll try to sum up everything you need to know about today’s release of GPT-4.
New capabilities unlocked
Here’s OpenAI's initial statement in their GPT-4 release notes:
In a casual conversation, the distinction between GPT-3.5 and GPT-4 can be subtle. The difference comes out when the complexity of the task reaches a sufficient threshold — GPT-4 is more reliable, creative, and able to handle much more nuanced instructions than GPT-3.5. — OpenAI GPT4 release notes
I tried GPT-4 via the ChatGPT Plus interface briefly and noticed it seems to give better results in some more complicated storytelling tasks such as multi-perspective storylines, building and tweaking story arcs etc.
However, the new reasoning capabilities are impressively illustrated with a diagram that shows the improvement of GPT-4 in various tests compared to its predecessors:
Most notably, GPT-4 shines on the USABO (USA BioOlympics) semi-final exam and the GRE Verbal test (the world’s most widely used college and graduate school admissions test). And in the UBE (lawyers’ Uniform Bar Exam), GPT-4 improves dramatically, leaving GPT3.5’s abilities far behind.
In some areas, the new “vision capability” (more on that later) boosts GPT-4’s reasoning power even further. Here’s an overview of some of the simulated tests:
Language competence
GPT-4 outperformed GPT-3.5 and other language models for multiple-choice problems spanning 57 subjects in 24 languages including low-resource languages like Latvian, Welsh, and Swahili.
Multimodality: Visual input
GPT-4 can accept prompts consisting of both text and images. This allows us to specify any visual or language task combining these input modalities. However, the image inputs are still in the research stage and are not yet publicly available.
In the meantime, you can play around with Microsoft’s VisualGPT.
However, it is impressive to see how far the understanding of images has already advanced with GPT-4! The new model reads and interprets documents, solves visual puzzles, and “gets” humor:
Steerability
With GPT-4, it will be possible to alter the so-called “system” message in order to change the AI’s verbosity, tone, and conversation style. A feature that has been already available to developers working with GPT3.5turbo will soon be available to all ChatGPT users:
Limitations, risks & mitigations
Of course, there are still limitations as well. The problem of hallucinating facts, for example, or reasoning errors. However, GPT-4 has improved in this regard as well and there is also progress with erroneous behavior and sensitive content. Although, OpenAI says that there is still “much to do”:
Our mitigations have significantly improved many of GPT-4’s safety properties compared to GPT-3.5. We’ve decreased the model’s tendency to respond to requests for disallowed content by 82% compared to GPT-3.5, and GPT-4 responds to sensitive requests (e.g., medical advice and self-harm) in accordance with our policies 29% more often. — OpenAI GPT4 release notes
Further Information
For more stories and insights regarding ChatGPT, GPT-3/GPT-4, and prompt engineering, have a look at this list:
Sources for this article:
OpenAI release notes: https://openai.com/research/gpt-4
OpenAI blog post about GPT-4: https://openai.com/product/gpt-4
➡️ If you like my content, why not leave a “clap” at the end of this article, so more people can see it?
