AlphaGo: the next generation

TL;DR

Alphago Zero, a revolutionary AI, learns strategies from scratch, surpassing human knowledge in just three days.

Transcript

hello we're at new scientists new office in Covent Garden in London I'm Sam Wang and I've got Jacob Aaron here and Sally ad to talk about alphago and the new the new incarnation of deep minds go playing AI so Jacob what's different about this new version last year I think could beat the best human player but they keep beating like whatever the best... Read More

Key Insights

• 🚂 Alphago Zero surpassed human knowledge by training from scratch, without relying on human strategies.
• 🤗 The departure from training data in AI development opens up possibilities for solving complex problems like protein folding.
• 😨 The application of the Alphago Zero approach to real-world scenarios like driverless cars faces challenges due to the infinite possibilities in such environments.
• 👾 DeepMind's next challenge, Starcraft, poses a greater problem due to the vast decision space compared to go.

Questions & Answers

Q: How did Alphago Zero differ from the previous version?

Alphago Zero started from scratch and learned go strategies without any human training data, surpassing its predecessor.

Q: Why is the departure from training data significant in AI development?

Departing from training data allows AIs like Alphago Zero to learn from scratch and discover new knowledge without relying on existing human strategies.

Q: Can the approach used in Alphago Zero be applied to other fields like protein folding?

Yes, the method of training AIs without massive data could potentially be used in complex fields like protein folding where training data is limited.

Q: What are the challenges in applying the Alphago Zero approach to real-world scenarios like driverless cars?

The Alphago Zero approach is limited to bounded environments like games; adapting it to scenarios with infinite possibilities, like driving, poses significant challenges.

Summary & Key Takeaways

• Alphago Zero, a new iteration of the AI, surpassed the previous version by winning 100 out of 100 games.

• Unlike its predecessor, Alphago Zero didn't rely on human strategies but learned from scratch, discovering new tactics.

• This new approach of training AIs without massive data could potentially be applied to other fields like protein folding.