Anthropic’s New Model Excels at Reasoning and Planning—and Has the Pokémon Skills to Prove It

Claud 3.7 Sonnet played the game when it was in some challenges: It spent on it “A few dozen hours“There was a problem to be trapped in a city and identifying nonplayer characters, which severely stunned its progress in the game. With Claud 4 Opus, Harshie Claud noticed an improvement of long -term memories and planning capacity when he saw a complex Pokemon navigate a certain energy after spending a certain energy, before ai, ayi, before it, before it, before it, before it, ayi, before AII, before AII. Before AII, instant reactions, showing a new level of solidarity, which means the model has better power to track.

Harshi says “This is one of my favorite ways to know a model” This new model we are about to find out and how to work with it my simply my way to come “”

Everyone wants an agent

Anthropic’s Pokémon Research is a fancy view of dealing with a prosecisting problem – how do we understand what AI is deciding when approaching complex tasks and preventing it in the right direction?

The answer to this question is inseparable-II for the advancement of the high-hypeide AI agents in the industry that can deal with complex tasks with relative freedom. In Pokémon, it is important that the model is not the context in hand or “forgetting”. It is also said to automato a workflow that applies to AI agents – even a few hours.

“When a job goes out of a five -minute job 30 minutes of work, you can see the ability to keep the model consistent, you can see all the things you need to perform it [the task] Successfully get worse over time, ”said Harshi.

Ethnic, Like many other AI labsHope to create strong agents to sell as products for consumers. Criger says that this year anthropological “top objective” is Claud Ho “Working for you a few hours.”

“This model is now supplying it-we have seen that one of our primary access customers closed the model for seven hours,” Criger said, “often refer to the process of reconstruction of a large amount of code for more efficient and organized.”

Companies like Google and Open are working towards this future. Google has released Mariner’s release earlier this week, An AI agent built on Chrome It can do the jobs like buying grocery (at $ 249.99 per month). OpenAI recently Revealing a coding agentAnd a few months ago It has launched the operatorAn agent that can browse the web for a user.

Compared to its competitors, the ethnographic is often seen as a more vigilant mover, but slowly moving forward in research but slow on deployment. And with strong AI, this is probably a positive: an agent can be wrong with an agent that has access to sensitive information like the user inbox or bank logins. “We have significantly reduced the behavior where models use shortcuts or lufles to complete the tasks,” said an anthropologist at a blog post on Thursday. The company also says that both Claud 4 Oppus and Clock Sonnet 4 are less likely to be involved in this behavior 65 percent, which is known as reward hacking than the previous models – at least in specific coding functions.