Inside Google’s big AI shuffle — and how it plans to stay competitive, with Google DeepMind CEO Demis Hassabis
Interesting. interview Demisa Hassabis, CEO of Google DeepMind, with Nilay Patel as part of Verge's Decoder podcast, which was given 3 Weeks ago in early July.
During the interview, issues were discussed: Deepmind and Google Brain will merge into Google Deepmind, which will do more product stories inside Google than abstract things like Alpha Go. (go) Alpha Fold, which Deepmind was famous for (here An interesting documentary about Alpha Go) The interviewer’s hypothesis was that OpenAI had made a breakthrough with LLM and that Google now needed to focus and accelerate as a catch-up player.
- about the cause of hype near LLM - the hypothesis is that LLM solve problems that are understandable to most ordinary people and solve them well, and the previous things from Deepmind are too difficult to understand ordinary people and understand only specialists in narrow areas Cultural conflicts between the two divisions (Brain and Deepmind)They've become one.
- about the approach to solving problems that Demis himself uses - here Demis talks about chess, which he has been engaged in for a long time, about visualizing the final result and backcasting from the target solution back to the current one. (ala Amazon working backwards) A mix of deep new research (This will replace the LLM.) and scale current decisions (More options, more training examples, more power) About Google products in Bard and SGE format (Search Generative Experience) This story is about current grocery stuff using Google’s LLM and experience using it.
- about the famous note "Google 'We Have No Moat, And Neither Does OpenAI'" Demis said he thought the note was real, but he disagreed with the conclusions in the note.
- about the timing of the onset of AGI - Demis' estimate ~ 10 years Learn more about models with the help of people who rate LLM answers as correct or not, and so on.
- about risks and AI regulation, as Demis signed A letter from the Center for AI Safety
- about the combination of models of ala LLM + specific models for solving problems from the subject area - about the same thing told Stephen Wolfram in the book "What Is ChatGPT Doing ... and Why Does It Work?" (detail here)
- about the study.Stochastic Parrots"inside Google, which led to a slowdown in Google's LLM and a further dismissal of authors after the ChatGPT boost from OpenAI"
- about how not to get into the learning cycle of neural networks on the material generated by neural networks, which can lead to deterioration of models - it is proposed to create special watermarks that are built into the generated materials, which will allow them to be recognized when training new models
#AI #ML #Engineering #Software #DataScience #SoftwareDevelopment