Sparks of Artificial General Intelligence: Early experiments with GPT-4
22 In March, an interesting article was published. article from Microsoft Research 154 The results of experiments with GPT-4which they had been doing for almost six months. But read 154 It is a rather time-consuming job, so I doubt I would get to read this article. Fortunately, for lazy people like me, literally a day after the article was published on the AI Explained channel, it appeared. video. In this discussion, the author notes the following points:
- Tool use - GPT-4 Unlimited availability for researchers since October 2022 year or earlier
- Image understanding - about the generation of images
- Coding - about completing tasks on LeetCode on average better people
- 3D games - about creating 3D toys
- Mathlete - about solving problems from mathematical Olympiads
- Fermi Qs - about answers to complex questions with an estimate of the number of golf balls fit in the Boeing and the like
- Actual PA - about a real personal assistant
- AI handyman - about help in repair
- Mapping - about making maps
- ToM (Theory of Mind) - Theory of thought.
- Joke punchline problem - about the problem of creating jokes, when you have to go from the end of the punchline and then create the beginning of the joke, and GPT -4 He can't.)
- Misinformation problem – generation of fake stories, for example, for propaganda purposes or conspiracy theories
- The data admission problem is that people are likely to want to control how much they contribute to the information that is used to train the neural network. Conventionally, content creators like artists, writers, etc.
- Intrinsic Motives: Empowering models with intrinsic motivation can be an important area of further work for AGI development. (artificial general intelligence)
- Thought on urgency: The urgency of further research to understand how and why GPT is needed.4 It has all the above properties.
#AI #Software #SoftwareDevelopment