I really like this channel and am grateful that someone has the ambition, patience, and intelligence to read all papers and reports about AI development. I'm super interested and always want to go to the source of information, so I might have the ambition but lack the patience and intelligence to read it myself:-)
But I'm happy I have a community with you with whom I can share interesting findings. And Thanks to the ChatGPT, I can deliver the link to this video with a summary that will help you decide if this is worth your time to watch.
- GPT-5 will be released incrementally, unlike the historical approach of infrequent major model upgrades.
- The AI will be incrementally improved through subsequent checkpoints of a given training run.
- OpenAI may still have an order of magnitude more data to use and could buy more high-quality data.
- 50% of AI researchers believe there's a 10% or greater chance of humans going extinct from AI.
- GPT-4 performs better on safety metrics than GPT-3.5, but there is a concern about whether a smarter model can ever really be safer. The release of GPT-5 is estimated to be in mid-2024. Reliability may be a weakness of the model.