Meta's New Llama 3.1 AI Model Is Free, Powerful, and Risky

1 month ago 22

Most tech moguls anticipation to merchantability artificial intelligence to the masses. But Mark Zuckerberg is giving distant what Meta considers to beryllium 1 of the world’s champion AI models for free.

Meta released the biggest, astir susceptible mentation of a ample connection exemplary called Llama connected Monday, escaped of charge. Meta has not disclosed the outgo of processing Llama 3.1 but Zuckerberg recently told investors that his institution is spending billions connected AI development.

Through this latest release, Meta is showing that the closed attack favored by astir AI companies is not the lone mode to make AI. But the institution is besides putting itself astatine the halfway of statement astir the dangers posed by releasing AI without controls. Meta trains Llama successful a mode that prevents the exemplary from producing harmful output by default, but the exemplary tin beryllium modified to region specified safeguards.

Meta says that Llama 3.1 is arsenic clever and utile arsenic the champion commercialized offerings from companies similar OpenAI, Google, and Anthropic. In definite benchmarks that measurement advancement successful AI, Meta says the exemplary is the smartest AI connected Earth.

“It’s precise exciting,” says Percy Liang, an subordinate prof astatine Stanford University who tracks unfastened root AI. If developers find the caller exemplary to beryllium conscionable arsenic susceptible arsenic the industry’s starring ones, including OpenAI’s GPT-4o, Liang says, it could spot galore determination implicit to Meta’s offering. “It volition beryllium absorbing to spot however the usage shifts,” helium says.

In an open letter posted with the merchandise of the caller model, Meta’s CEO Zuckerberg compared Llama to the unfastened root Linux operating system. When Linux took disconnected successful the precocious 90s and aboriginal 2000s galore large tech companies were invested successful closed alternatives and criticized unfastened root bundle arsenic risky and unreliable. Today nevertheless Linux is wide utilized successful unreality computing and serves arsenic the halfway of the Android mobile OS.

“I judge that AI volition make successful a akin way,” Zuckerberg writes successful his letter. “Today, respective tech companies are processing starring closed models. But unfastened root is rapidly closing the gap.”

However, Meta’s determination to springiness distant its AI is not devoid of aforesaid interest. Previous Llama releases person helped the institution unafraid an influential presumption among AI researchers, developers, and startups. Liang besides notes that Llama 3.1 is not genuinely unfastened root due to the fact that Meta imposes restrictions connected its usage, for illustration limiting the standard astatine which the exemplary tin beryllium utilized successful commercialized products.

The caller mentation of Llama has 405 cardinal parameters oregon tweakable elements. Meta has already released 2 smaller versions of Llama 3, 1 with 70 cardinal parameters and different with 8 billion. Meta contiguous besides released upgraded versions of these models branded arsenic Llama 3.1.

Llama 3.1 is excessively large to beryllium tally connected a regular machine but Meta says that galore unreality providers, including Databricks, Groq, AWS, and Google Cloud, volition connection hosting options to let developers to tally customized versions of the model. The exemplary tin besides beryllium accessed astatine Meta.ai.

Read Entire Article