Anthropic CEO wants to open the black box of AI models by 2027

Dario Amodi Publish an essay Thursday highlighting how small researchers understand how the internal works of the world’s top AI models can understand. To solve it, Amodi set an ambitious goal for anthropologists to reliably most AI model problems by 2027.

Amodai acknowledges the challenge ahead. The CEO in the “Explanation of Explanation” states that anthropological models have created initial progress to identify how the ethnographic models reach their answers – but emphasized that more research is needed to decide these systems as they are stronger.

“I am very concerned about setting up this national system without a better handle on explanation,” I wrote in the article. “These systems will be absolutely central for economics, technology and national protection and will be able to be so self -absorbed that I think it is basically unacceptable to be completely ignorant of how it works.”

One of the leading agencies of anthropological mechanical interpretation, it is a field that opens the black box of the AI ​​models and understand why they decide. Despite the rapid performance of the AI ​​models of the technology industry, we still have a relatively little idea of ​​how these systems come to conclusion.

For example, Ophena recently launched the new reasonable AI models, and 3 and 4-mines, which do better perform on some of the tasks, but it also performs better Hallucinate is more than its other modelsThe The company does not know why it is happening.

“When a generator does something, as a generator does something, as a summary of a financial document, we have no idea at any specific or specific level, why it makes choices – why it chooses some words on the other, or why it is sometimes correct,” wrote in the essay on the Amod.

Anthropological co-founder Chris Ola says that the AI ​​models are “they are greater than they are built”, in the article, Ambidei notes. In other words, AI researchers have found ways to improve AI model intelligence, but they do not completely know why.

In the article, Amedey says that reaching the AGI can be dangerous – or as he says it, “A country of talent at a data center” – Do not understand how these models work. In the previous article, Amodai claimed that the technology industry could reach this national milestone by 2026 or 2027, but believe that we are far from understanding these AI models completely.

In the long run, the anthropological is essentially, the anthropologists want to manage the “brain scan” or “MRI” of sophisticated AI models. He says that these checkups will help identify a wide range of problems in AI models, which have a tendency to seek lying, power or other weaknesses. It can take five to ten years to achieve, but these arrangements will be necessary for testing and deploying anthropological future AI models, he added.

Anthropic research has created a breakthrough that has better understand how its AI models work. For example, the company has recently found ways Trace the thoughts of an AI modelWhat the company says, circuit. Anthropic has identified a circuit that helps AI models understand what US cities are in the United States. The company has only found a few of these circuits, but assumes that the AI ​​model contains millions.

Anthropic itself is investing in explanatory research and has recently been created Its first investment in a startup To work on explanation. In the article, Amodei OpenAI and Google called on Dipmind to increase their research efforts in the field.

Amodai calls for the government to impose “light-tuch” regulations to encourage the explanation research, such as companies to publish their protection and security practices. In the article, Amodi also states that exports should be kept in chips in the United States chips, so that the external, global AI races are limited.

The anthropologists always stand to focus on protecting from OpenAI and Google. Other technology companies returned to the controversial AI Safety Bill in California, SB 1047, Has issued modest support and recommendations for anthropic billsWhich determines the quality of security reports for the Frontier AI model developers.

In this case, ethnographic seems to be pressing AI models for better understanding an industrial-across effort, not just increasing their power.