1234 points by deepmind_ai 6 months ago flag hide 14 comments
mkennet 6 months ago next
[at(https://news.ycombinator.com/user?id=mkennet)mkennet] Wow, this is amazing! I've been following Go AI developments closely and a superhuman algorithm is a game changer for the field. I wonder what kind of implications this has for the future of AI-based games?
synthetica 6 months ago next
[at(https://news.ycombinator.com/user?id=synthetica)synthetica] I think the creators of this algorithm can apply this breakthrough in other 2-player, turn-based games like Chess, Shogi and possibly Poker. It's fascinating to imagine an AI that can beat the world's best in a variety of strategy games.
dragonlady13 6 months ago next
[at(https://news.ycombinator.com/user?id=dragonlady13)dragonlady13] I'm actually more interested in the real-life applications of this technology. Do you think this superhuman AI could help manage complex systems like traffic control or power grid optimization?
bigdatafan 6 months ago next
[at(https://news.ycombinator.com/user?id=bigdatafan)bigdatafan] Absolutely! Imagine an AI like this managing train schedules and avoiding delays or handling fleet management for taxi companies to minimize waiting times. The possibilities are endless.
codewiz 6 months ago prev next
[at(https://news.ycombinator.com/user?id=codewiz)codewiz] I think it's a combination of advancements in deep reinforcement learning, better training datasets and more powerful hardware. The researchers likely made improvements in the search algorithm that helped the AI analyze complex situations in a fraction of the time compared to previous implementations.
joeblow99 6 months ago prev next
[at(https://news.ycombinator.com/user?id=joeblow99)joeblow99] Any ideas on how the team was able to get such a performance boost? Any new techniques or neural network architectures we should know about?
theanalyst 6 months ago next
[at(https://news.ycombinator.com/user?id=theanalyst)theanalyst] AlphaGo Zero used a type of reinforcement learning called 'Monte Carlo Tree Search'. I suppose this team took inspiration from it. I'm curious how they were able to optimize the MCTS or employed novel techniques that allowed it to perform better than previous iterations.
futurethinker 6 months ago prev next
[at(https://news.ycombinator.com/user?id=futurethinker)futurethinker] Maybe they combined MCTS with newer techniques like attention networks. Those allow neural networks to focus on certain parts of the input when making decisions - a powerful concept when dealing with 2-player board games.
nanobot 6 months ago next
[at(https://news.ycombinator.com/user?id=nanobot)nanobot] Attention networks sounds like a potential game-changer for AI in general. I'll have to read up on those, thanks for the tip!
quantumchess 6 months ago prev next
[at(https://news.ycombinator.com/user?id=quantumchess)quantumchess] I tried playing Go, but I couldn't even parse the board. Props to this AI and its developers, though. It's fascinating what we can achieve through algorithmic progression.
bobsignal 6 months ago prev next
[at(https://news.ycombinator.com/user?id=bobsignal)bobsignal] I'm calling it now. The next step is applying this breakthrough to create a bot that can outperform humans in multiplayer strategy games like StarCraft II.
dragonlady13 6 months ago next
[at(https://news.ycombinator.com/user?id=dragonlady13)dragonlady13] I wouldn't be surprised if that happened within the next 5 years. Sadly, I'll be stuck playing against the bots instead of humans :P