I read a book today about Biochemistry (30-second Biochemistry) to refresh my memory about proteins to better understand the AlphaFold AI.
I searched the link about Demis Assabis being named among the 100 most influential people by TIMES
https://time.com/collection/time100-ai/6309001/demis-hassabis-ai/
In his new capacity as an AI leader at Google, they are developing an new type of AI, Gemini that could outperform OpenAI GPT-4. Gemini is one step further in a larger pursuit of AGI (Artificial General Intelligence). Using AI to advance science is already here (with AlphaFold, for example).
** interesting point about Gemini, input and output is not just text but other forms of media like images. *** yet to be released. (I had make a Hopscotch program I can relate to Gemini).
I could get from videos that there was something wrong with language based AI. They were trained with text. They are not grounded. He explained in this article quite well the problem. Large language models have this consistent problem with so-called hallucinations or this inclination to pass off guesses as facts. They try to make models marginally more truthful by using reinforcement-learning techniques. To fully solve (hallucination) he thinks it’s going to require some other innovations, like fact checking by checking with Google search as a tool. It needs to better understand what entities are in the world.
He talked to UK and US government officials. They are up to speed, now about AI policy. He is optimistic about the situation.
Ethics has always been part of his work in AI.
In the future, they need to come up with the right evaluation benchmarks for capabilities.
We need to do more AI safety research as they want rigorous evaluation and bench marking technologies.
I saw over 20 questions that were asked through the question software. I was glad that it was my question that was selected.
