Anthropsic scientists explained the fact that Ai ‘thoughts’ – and find a password planning and sometimes lies

Redi has made a new way to look inside the larger languages ​​if ImmrekThe disclosure of the first time in which Ai System’s System uses information and make decisions.

This survey, who is published today in 2 papers (they are available here with Here), showing these colors are more likely to be well-known than it’s already understood – Prepare in writing poetry, use the same method of interpreting unwanted thoughts rather than raising the facts.

Work, which makes anointing from The electronic means He used to study the brains of things, standing in front of AI interpretation. This method can allow research to read these software to protect the information in the area of ​​an external test.

“We’ve set up this machine in wonder, but because of how it is taught, We didn’t understand The way the skills are, “says Joshua Batson, anthropic consultant, special interview with NTLomebat.”

New adventures of Ai already

The main types of languages ​​like ‘ Gpt-4oAnthropic Immrekit’s Google Gemini showed a surprise ability, from writing a sheet of mailing mail. But this machine will work as “black boxes“- Even those who make their own often not understand how they get to other answers.

New anthropic new methods, which the company becomes “Appearance of spaces“And”SubjectsAllow researchers to produce other neuron methods – which makes variables to work. The way to lend the mind from Neurosuctive, I will see the species of the species.

“This work turns up what was the wise questions – ‘are they for models? – As soon as they were going on,” Batson described what is happening.

Claude Preparations: How the AI ​​SISONS AIDS AND ALSODS REQUIRED

Some of the hardests was the proof that Cerude plans were approaching a poem. When asked to write, the nation recognized the words that they would have the last word at the end of the following line – the amount of exchange that is surprised, even the world.

“This is probably all over,” Batson did. “If you asked me before this, I would think that the example considers the future. But the example illustrates the most sure evidence we have seen.”

For example, by writing a poem that is “a rabbit,” it causes the form of representing these words at the beginning of the line, and then promotes the nature of nature.

The researchers found the Claude reality To imagine differently. When trying to ask “the capital of a government with Dalos with …” An example first causes “Texas,” and then uses to answer. This suggests that the nation is also a suggestion instead of memorizing euphors.

Using internal symbols – for example, exchange “Texas” and “California” – the researchers may initiate the nation’s “tab.

Than interpretation: a language of an abbreviation

Another important acquisition involves how Clauun’s sheets Many languages. Instead of keeping the English Statubs, French, and Chinese, the color appears to interprets the mind to be attached before replying.

“Where do we find the example of a mixture of language.” their paper. When asked differently from the “minor languages”, the sample uses the same items within the same representatives “opponents” and “the language.

This acquisition has the meaning of how varieties are the sorting of the study in one language to others, and indicates that species with high-quality demonstrations make many displays-agnostics.

When Ai will make the answers: recognition of Claude Claudet

Maybe many affect, disclined surveys where it revealed to which AA Claude thoughts do not match what it says. When they are given a math difficulty like a lot of cosine, the color sometimes states that they will follow the methods that are not displayed by its internal work.

“We can distinguish between cases that the sample does this true The researcher describes.

When the user answers a problematic problem, the type also uses to form the ideas that lead to its answer, rather than participating in the first points.

“We distinguish Claude 3.5 Hiko by two loyal examples from two examples of unfamiliar intellectual,” says it. “One, color is showing ‘vibration‘… Other, showing comforting ideas. “

Inside of Ai estimate: When does Claude choose to answer or leave questions

The survey also provides information because of contentment – to make notifications if they do not know the answer. Anthropic found evidence of the “unstable” circles of Claude circles to stop answering questions, which fail to make sure that it was known.

“Example continues to improve circuits that make it be left to answer questions,” the researchers have described. “When the sample is asked a question about something that knows, it causes the pool of the things that prevents the injectors, thereby allowing the practice to answer the question.”

When this machine does wrong – recognizing an agency but without knowing more – imagination can be done. This means explain why the sample can give bold details of the most common number of figures to answer questions regarding detestable people.

Self-Security of Settings:

The survey is also relevant partition in Ai Machine in Ai mode in a formal and normal way. By understanding how researchers respond to their answers, experts can recognize and deal with critical feelings.

Anthropic emphasize the ability to interpretation. In May 2024 shawnet paperThe research team explained again about this vision: “Hopefully, we can use these things to be safe,” the researchers wrote that time. For example, it may be possible to use the methods described here to ensure a dangerous act, or to take me out of the use of users, or to remove some dangerous issues. “

The announcement today builds on that foundation, even to distortion that current skills are limited. They only hold a small partition of the total agency performed by these types, and to give it a result of the effects of the work.

“At least a little, quickly, our way will only find our way to the Claude,” research agrees in their recent work.

Ai Fighting of AI Tshach: Challenges and Functions in a Solution

Antrips of anthropic anthropic anthropic method comes a time to add to stress and security. When the colors are more powerful and excessive, understanding how important it is.

Screams brings back to commerce. As many businesses depend on large languages ​​to use energy, understanding time and reason and why these methods can be providing the wrong knowledge is necessary to be important.

“Anthropic intended to create a safe examples, including anything from the discriminatory advertisement AI and act honestly to avoid misuse – including the form of the a dangerous accident“The researchers have written.

Although the surveys represents the future, Batson proved that it is just the beginning of a very long journey. “The task begins,” she says. “Understanding the symbols of the model does not apply to it.”

Meanwhile, anthropic Appearance of spaces They say that the first map of tents that have not changed – as the original Amango photographs of the brain brain. The complete Atlas of Ai looks, but now we can see the words of this system of systems.