OpenAI didn’t pack it in after they unleashed ChapGPT unto the world; they tried to upgrade their next AI language model, which seems like it could make its cousin feel pretty stupid. It’s called GPT-4, and it comes equipped to help people skate by on projects and tasks without doing the work even more.
OpenAI made a bold claim as they introduced GPT-4 — saying it is not as capable in real-world scenarios as humans. Still, GPT -4 makes up for it in theoretical and academic applications. There might be no use for humans when they get to their sixth AI upgrade.
During a livestream with developers, OpenAI reps showed off GPT-4s skills as it went through problem-solving image recognition, describing images, and creating a functional website. The last one is pretty crazy; it simulated doing taxes.
According to a press release from OpenAI, GPT-4 is pretty spectacular at taking tests. It scored 298/400 on the Unified Bar Exam. That’s in the 90th percentile. It could be a lawyer tomorrow if it wanted.
It could get into college if it wanted, scoring a 710 on reading and writing on the SAT. That’s in the 93rd percentile.
They put the AI through the following AP classes; Art History, biology, Marco-and Microeconomics, Psychology, Statistics, US Government, and US History. It scored a 5 on all of them.
Let’s get back to GPT-4 doing simulated taxes — it used a system prompt called “TaxGPT” and a promo that included large parts of the federal tax code. The user saved a call to H&R Block by telling it to estimate 2018 taxes for a married couple with one child. It reasoned out of the answers using the tax code, and for the big finish, it came up with the family’s standard deduction and tax liability estimate for the year.
Add comment