New blog posts
3 November, 2023 by Reza Moezzi
We are delighted to share some remarkable news...
Universe composed of water? Large exoplanet's environment contains methane and atmospheric carbon dioxide.
22 September, 2023 by Reza Moezzi
Hubble Space Telescope observations provided...
16 September, 2023 by Reza Moezzi
In the world of academia and research, certain...
Calendar of Events
Psychologists have shown that GPT-3 has the same level of reasoning ability as a college student
Posted on 4 August, 2023 by Reza Moezzi
Standardised test-like logic problems were no match for the artificial intelligence language model GPT-3, which performed as well as college students. The experiment's authors argue that their findings raise the question of whether the technology is emulating human thinking or employing a novel cognitive mechanism. To get an answer, you'd need to go inside the code that powers GPT-3 and other AI programmes.
UCLA psychologists have shown that the AI language model GPT-3 does as well as college freshmen when presented with the types of reasoning difficulties normally seen in IQ testing and standardised exams like the SAT. Nature Human Behaviour has published the study.
However, the authors of the publication state that the research prompts the following question: Is GPT-3 employing a fundamentally different form of cognitive process, or is it just a result of its large language training dataset that makes it behave like a human brain?
Since OpenAI, the business that developed GPT-3, is protecting its secretive inner workings, the scientists at UCLA cannot definitively comment on the nature of GPT-3's reasoning skills. They also note that despite GPT-3's impressive performance in some areas of reasoning, the widely used AI tool still falls short in others.
"It's important to emphasise that this system has major limitations," said Taylor Webb, the study's first author and a postdoctoral researcher in psychology at UCLA.
Forty first-year students at UCLA were given the identical issues to tackle by the researchers.
According to the study's principal author and UCLA psychology professor Hongjing Lu: "Surprisingly, not only did GPT-3 do about as well as humans but it made similar mistakes as well."
GPT-3 was successful at solving 80% of the questions, which is above the average score of slightly around 60% for human participants and within the range of the top human scores.
The only way to find out is to gain access to the programme and the data used to train the software, and then to give the software tests that it hasn't already been given, which is a daunting task. They claimed it would be the next stage in determining the proper direction for AI.
Webb said that "having the backend to GPT models would be very useful for AI and cognitive researchers." To paraphrase, "We're just doing inputs and getting outputs, and it's not as decisive as we'd like it to be."
Today In History
Here are some interesting facts ih history happened on 4 December.
- America's 1st Thanksgiving Day (Va)
- A hen in Rome lays an egg imprinted with figure of great comet 1st seen on December 16 1680
- 1st Gen Assembly in Penn (Chester)
- Washington bids officers farewell at Fraunce's Tavern NYC
- 1st Grey Cup game (University of Toronto 26 Toronto Parkdale 6)
- Roald Amundsen reaches South pole
- Panama Pacific International Exposition opens
- 1st edition of Chase's Annual Events published
- Gemini 7 launched with 2 astronauts
- Islanders end 15 game undefeated streak (13-0-2) (Col Rockies)
- South Africa grants Transkei independence