Despite recent viral analysis StuartJRitchie says AIs are showing no decline in their cognitive abilities
The reason this is even possible is that OpenAI constantly fine-tunes the model behind the scenes. The changes they make – sometimes part of attempts to improve efficiency, sometimes to ensure the model doesn’t produce harmful information – can inadvertently affect the answers it produces.
In part, the analysis looked at prime numbers. Apparently, back in March, GPT-4 could tell you whether a number was prime with 97.6 per cent accuracy. Now, its accuracy is 2.4 per cent. The analysis went viral, with many GPT users saying it fit their impression of a drop-off in performance in recent months., they argued that the analysis was mistaken: it only tested GPT-4 on prime numbers, and didn’t throw in any non-prime ones. So it missed the fact that GPT-4 pretty much always said a number is prime back in March, and pretty much always says it isn’t prime now. In other words, GPT-4 was never good at assessing prime numbers – it’s just become bad in the opposite direction.
The Princeton researchers argued that there’s a difference between the underlying capability of the AI, and the way it behaves. ChatGPT never had much of a capability for prime numbers . But it now behaves differently in response to questions about them. That’s cold comfort to people who rely on ChatGPT for work and now have to check if the behaviour shifts mean they need to change their set-up. But if it’s true, it means that, perhaps unlike humans, AIs are showing no decline in their cognitive abilities.
Belgique Dernières Nouvelles, Belgique Actualités
Similar News:Vous pouvez également lire des articles d'actualité similaires à celui-ci que nous avons collectés auprès d'autres sources d'information.
GPT-4 and ChatGPT study shows LLMs are getting dumberBehavior of OpenAI models about as consistent as Office 365's uptime
Lire la suite »
Apple is reportedly working on Apple GPT, a ChatGPT rival you won't see anytime soonApple is reportedly working on a ChatGPT rival, AppleGPT
Lire la suite »
In an effort to surpass ChatGPT, Meta is open-sourcing its large language AI modelThe possibility for more AI tools, for free from Meta.
Lire la suite »
What to Know About Claude 2, Anthropic's Rival to ChatGPTClaude 2, built by Anthropic, is the latest large language model AI to challenge OpenAI's ChatGPT.
Lire la suite »
We asked ChatGPT to write a poem about Northamptonshire and its pretty good'Northamptonshire, where the meadows dance and history weaves its timeless trance'
Lire la suite »
AI: Facebook owner Meta announces its rival to ChatGPT will be free to useThis version of the model is trained on 40% more data compared with Llama 1, according to Meta.
Lire la suite »