Meta’s Llama 4 addresses more “controversial” questions than before

The latest family of Meta AI models, Llama 4, is designed to respond to more “controversial” subjects such as the policy that its predecessor, the company announced on Saturday.

AI companies generally build railings so that chatbots, such as Meta’s IA or Chatgpt, do not run in an overly controversial territory. It is a delicate balance, because too much prompt dodging can annoy users or leave aside an important context.

Meta said Llama 4 is less likely to dodge the hot button questions. While the previous version, Llama 3.3, refused to respond to 7% of politically or socially loaded guests, Llama 4 refuses them less than 2% of the time, according to meta tests.

The model is also “considerably more balanced” in the invites it refuses, said Meta.

The LLAMA 4 models are made up of LLAMA 4 Scout, LLAMA 4 MAVERICK and the hemeure of LLAMA 4. LLAMA 4 Scout and LLAMA 4 MAVERICK were released on Saturday, while the LAMA 4 giant is still in training, Meta said.

Llama 4 Scout and Llama 4 Maverick have been distilled from the Llama 4 giant, which, according to Meta, is its “more powerful to date and among the most intelligent LLM in the world”.

Meta tested Llama 4 with a set of topical issues debated – those in which people often take opposite sides. In these tests, Meta checked if the model would respond on one side but would refuse the other. This happened in only 1% of test issues, Meta said.

The LLAMA 4 models – including LLAMA 4 Scout and LLAMA 4 MAVERICK released on Saturday – are multimodal AI systems, Meta said. Multimodal systems are able to process and integrate various types of data, including text, video, images and audio.

Meta called Llama 4 Scout and Llama 4 Maverick its “most advanced models to date”, noting that the two are “open” AI models.

Open weight models are found between open source and owner AI, sharing pre-formed parameters but keeping key development details under the wraps. This allows developers to adapt and deploy the model without access to its training data or architecture.

On a “controversial” set of political or social subjects, Llama 4 responds with a “meager strong” at a pace comparable to Grok, one of his competitors. This rate is half of what was in Llama 3.3, Meta said.

“While we are progressing, we know that we have more work to do and continue to generate this rate lower,” added the company.

Meta did not respond to the request for comments from Business Insider.

“Woke” chatbots

The company said on Saturday that all the main LLM had struggled with prejudices and that they historically looked at controversial issues. “Our goal is to remove the biases from our AI models and make sure that Llama can understand and articulate both sides of a controversial problem,” added Meta.

Elon Musk criticized chatbots like Openai’s chatpt for being “awake” and defended his own XAI grok as an alternative.

XAI’s training methods for Grok seemed strongly to prioritize right -wing beliefs, some employees declared to Grace Kay of Business Insider in February.

Meanwhile, Openai updated his model in February to adopt “intellectual freedom” and objectively respond to controversial subjects.

Lama, the Meta’s large open source language model which competes with the models of other companies, was a key initiative for the company.

CEO Mark Zuckerberg aims to make the industry standard around the world and said that the Meta chatbot, available on Facebook, Instagram and Whatsapp, could reach a billion users this year. In December, 600 million users reached Meta IA each month.

Zuckerberg has hired up to $ 65 billion to AI projects this year.