Where the previous models were roaming the purposeless or stuck in the loops, the cloud 3.7 is planned ahead, remembers its goals, and when the initial strategy fails, it collapses.
Critical skills to fight Pixlated Jim leaders. And, we also offer to solve real -world problems. pic.twitter.com/scvisp14xg
– Anthropic (@anthropicai) February 25, 2025
Harsha said, “The biggest thing to prevent the current version of the cloud is to improve is” when it receives a good strategy, I do not think it is necessary to know a strategy to know itself. [it] Is better than the other. And this is not a minor problem to solve.
Nevertheless, Harshi said he looks at the “low hanging fruits” to improve the Pokémon game of the cloud while improving the model understanding of the game boy screenshots. Harshi said, “I think it can defeat the game on this occasion if it has a perfect feeling of what is on the screen,” Harshi said, “Harsha said that such a model would probably” do a little less than a human “.
Harshi said expanding the context window for future cloud models would probably allow these models to “make more reasoning than long -time frames and handle things in a longer time.” He added that after remembering future models, it would be better to look at an integrated set of an integrated set needed to try to develop. “
Tweech Chat responded with a flooding emojis when Claude ended the 78+ hour episode from Pokemon’s Mount Moon.
Although you think about the improvement in AI models, the current performance of the cloud in Pokemon does not do so that it is ready to launch a human level, completely artificial intelligence explosion. And Harshi allows Claude 3.7 Swant to be trapped on Mount Moon for 80 hours or it “looks like a model who doesn’t know what it is doing.”
But Harshi is still impressed that the new reasoning model of the cloud will occasionally show some glow of awareness and “it doesn’t know what he is doing and knows what he needs to do something different. And ‘can’t do it’ and ‘can do like this’ for me a great thing for these AI things.” “You know, when something can do something in some way, it usually means that we are really close to being able to do something really well.”