On July 18, 2023, Meta surprised the world by announcing the release of the next iteration of its language learning model: Llama 2. This release, made in partnership with Microsoft, followed hot on the heels of its predecessor, Llama, in February of the same year.
But what was perhaps most surprising about the release of Llama 2 was its price tag, or lack of one. Yes, the language learning model is free to download and use under a community license, be it for research or commercial purposes. This has massive implications, both for Meta and the AI industry as a whole.
But what makes Llama 2 tick, and what does its free release mean for the industry and you? This article will discuss just that.
The Llama That Was
Meta released the first Large Language Model Meta AI (or Llama for short) on February 24, 2023. Like other foundational generative AI models, it was proficient in reasoning, accessing information, and coding.
It wasn’t as large or comprehensive as competitors like ChatGPT or PaLM, offering users and developers anywhere from 7-65 billion parameters. This may sound like a lot, but consider ChatGPT’s 175B or PaLM’s 540B capabilities and the difference quickly becomes apparent.
This smaller size, though, was also Llama’s advantage in the generative AI space. Fewer parameters meant greater flexibility on different systems or devices. Researchers who lacked the massive infrastructure of companies like OpenAI could still make use of advanced AI functions. This flexibility helped Llama fill a very important niche in the AI landscape.
The first Llama was released as a product for commercial use and required licensing and Meta’s permission to use. That said, the model was leaked and many developers and users were quick to adapt it for various ends and on all manner of devices.
It’s arguable, but perhaps this ready embrace of Meta’s foundational language model was one aspect that led to its decision on how to release the next iteration.
The New Llama in the Pen
On the face of it, Llama 2 doesn’t seem to offer much more than its predecessor. The parameter range has been increased up to 70 billion, but this is still a far cry from the competition. Peek under the hood though, and the difference quickly becomes apparent.
Llama 2 is even better in its various roles thanks to its more extensive training. Meta reports that this iteration was trained on two trillion tokens, an overall data increase of 40% from the original Llama. What makes this data special is its origin.
Llama 2 was trained solely on publicly available online data. This effectively negates the thorny issue of intellectual property and copyright abuses that have plagued AI since its introduction into the mainstream. The model has scraped no data and those using it can do so with a clear conscience.
Like its predecessor, it excels at single and multi-turn prompts while falling behind the competition in areas like coding. In power, it functions at the same capacity as ChatGPT 3.5. It uses up-sampling to check its generation against factual sources in order to reduce hallucination, another major AI issue.
Furthermore, it makes use of reinforcement learning human feedback (RLHF) to further refine its behavior and responses. This allows for users and developers to train the model even further for any purpose in mind.
How Does Meta Benefit?
A company doesn’t just spend a reported $20 million on a foundational language learning model and simply release it to the public for free. At least, that’s what many of us would believe. Meta’s generosity in making the coding and licensing of Llama 2 free for all must have some form of benefit to the company, after all.
While they haven’t given an official statement on the matter, here are a number of factors that could’ve motivated their decision.
Community Innovation
As mentioned before, the first leak to the public (despite Llama being only commercially available at the time) led to widespread adoption and tinkering by various users, eager to put the new model through its paces. Users were able to get the model working on different devices and to different ends.
Decentralizing the development process, from an in-house team to users worldwide, leads to innovation and responses to unique problems that the original designers would never consider. What’s more, it leads to faster development and evolution than one team would be capable of.
By placing Llama 2 in the hands of the community, Meta has taken steps to democratize the AI landscape.
Standardization
Llama 2 is a foundational language model, and is meant to be trained, tweaked, and developed further in a variety of scenarios. No matter how much it changes though, it’ll have always started from the same base.
By making Llama 2 free, it may be Meta’s hope that more users will adopt this technology for their own purposes, creating a standardized format for AI tools that the entire industry can benefit from in turn. It’s easier to develop with a baseline in mind, after all.
What’s more, the company that provides this standard, like Microsoft did with Windows, can obtain a stronger footing in the industry, as the most familiar with the baseline.
Brand Image
Giving tools away for free is a surefire way to win good will, and Meta has a lot to gain from distributing a language model trained off publicly available data, especially considering its past with dubious data handling practices.
By distributing Llama 2, Meta ensures not only that its brand image and presence remain strong in the industry, but is also training a future generation of AI specialists on its own tools. This’ll make future recruitment and expertise a cinch for the tech giant.
What Does Llama 2 Matter to You?
You could be a developer hunting for an AI tool that doesn’t need a supercomputer to run, or a user that wants a chatbot with a memory longer than a goldfish. Either way, Llama 2’s release has some exciting and game-changing implications for its userbase as well as Meta.
Free
Easily the biggest benefit of Llama 2 is its lack of a price tag. Meta has made the tool free for all under a community license. Only sites with over 700 million monthly visitors are required to contact Meta directly to obtain a license, likely to prevent misuse of the tool by competition.
This has huge implications for the industry. Not only does it set an important precedent and democratizes access, but it frees up company and business resources to focus on other areas. For example, users can afford to spend more on more high-powered machines to handle the tool, rather than the tool itself.
Versatility
Llama 2 continues its predecessors important work of being a middle ground for developers, providing a powerful language model without needing the infrastructure of competitors. This allows Meta to meet developers where they are and in turn allows developers to design with fewer limits.
Safety
Llama 2 is a sturdy, reliable base to work from. Issues that plague AI, like toxicity and inherent bias, have been extensively tested and removed. The RLHF testing process was instrumental in removing harmful activities, unqualified advice, and any illegal suggestions.
The language model’s training background on publicly available data also means that users aren’t complicit in ethically dubious data gathering techniques.
It’s Open
There’s been a lot of anxiety surrounding AI and the processing of inputted data by third parties. While Meta distributes the software and the community license for Llama 2, it has decentralized the tool. This means that unlike sites like OpenAI, Llama2 is free form third-party interference or data processing. The data you input or train the model on remains your own.
Relatable Conversation
AI has often been parodied for sounding stiff and repetitive, something which developers have been trying to overcome in more recent iterations. A more casual and conversational AI is more relatable and creates less unease.
Meta has placed a greater emphasis on dialogue with Llama 2, from its aptitude for single and multi-turn prompts to its more relaxed tone of voice. This makes it ideal for constructing and training chatbots, something most users associate with AI.
Toolkit
Though Llama 2 is still relatively new to the scene, it’s already seen widespread adoption and use by the industry. There a plenty of toolkits being developed and libraries to accommodate them, such as the Llama 2-Accessory.
Ready access to such toolkits greatly speeds up development and software evolution as a whole. Meta’s Llama Recipes make development and benchmarking a cinch with pre-made scripts for various hardware configurations and devices.
Besides these development tools, the Llama 2 ecosystem is rapidly growing. Chatbot services like Perplexity Labs, for example, are already using Llama 2 as part of their foundation. More adoption means more resources for developers and a healthier community to build from.
Language Models for the Future
Llama 2 may not be the most powerful language model on the block, but its free release, versatility, and rapid adoption means that its impact on the AI landscape will be felt for years to come. It places development power squarely back in the hands of developers, giving them the sandbox to create models and tools that Meta would never dream possible.
Make no mistake. Llama 2 is an important first step towards a new AI landscape.
