OpenAI just released the latest version of its primary large language model, the GPT-4. This new one is even smarter and has been trained on more data, according to the company, which can help deliver better results.
The startup said Microsoft Azure is used to training the new model (the Redmond giant has invested billions in the new technology) – unfortunately no details were given on how GPT-4 was trained or what hardware was used, information , which are extensively kept behind closed doors due to increasing competition.
Microsoft announced this Tuesday (14) that Bing’s AI chatbot is already using GTP-4, and it looks like this latest version will likely be adopted by entertainment product chatbots in the coming weeks.
Human-level performance, fewer mistakes and more
For the past six months, OpenAI’s GPT language has powered many of the AI demos that have hypnotized many of us, but what was already pretty good is now even better; The startup claims the new model can respond with fewer factually incorrect answers, as well as fewer conversations about forbidden topics.
However, one of the most interesting features has to do with what is called human-level performance, which the company says results in GPT-4 scoring even better than a normal human on SAT exams. On simulated tests, the new language scored 93% on an SAT reading test, 89% on an SAT math test, and 90% on a bar exam.
It’s not all rainbows and butterflies – but will it get there?
For all its power, GPT-4 still doesn’t fare particularly well when it comes to inventing things — in a recent blog post, the company said its new product has limitations when it comes to social bias, hallucinations, and adversary prompting but as it seems, improvements are already being made.
OpenAI also said in a blog post that the differences between GPT-4 and GPT 3.5 are subtle in casual conversation; The difference becomes apparent when the complexity of the task reaches a sufficient threshold, as the newer version is both more creative, more reliable, and also better prepared to deal with different nuances of instructions.
In terms of availability, GPT-4 will be released first for ChatGPT subscribers (and will also be available as part of an API allowing its integration into the apps). When it comes to pricing, the startup charges:
- 3 cents for approximately 750 word prompts.
- 6 cents for ~750 word answers.
Filed in . Read more about AI (Artificial Intelligence) and Microsoft.
This article was previously published on Source link