But hold on to your hats: the true story here isn't the glamorous conference but a low-key computer science article casually stowed away on the not-so-exclusive arXiv.Org. Meta researchers appear to have a talent for making ground-breaking announcements with little publicity. The paper introduces Llama 2 Long, an enhanced version of Meta's open-source Llama 2. This AI has apparently been going to the gym and doing some intensive pretraining, resulting in superhuman AI powers.
So, how did Llama 2 Long come to be? Well, the researchers took the original Llama 2, which comes in various training parameter sizes like a box of chocolates. They decided to feed it a hearty diet of more extended text data sources. To be precise, they added a whopping 400 billion tokens worth of data. It's like giving your AI model a literary feast to feast upon.
But here's the twist: they didn't touch with the architecture of Llama 2. Nope, they left that alone. Only a "necessary modification" to the positional encoding was done. Positional encoding, in other words, is the AI equivalent of GPS. It assists kids in determining where items are in relation to one another. In this example, they used a technique known as Rotary Positional Embedding (RoPE). It's like going from a paper map to a high-tech GPS device. They just changed the rotation angle of the RoPE from Llama 2 to Llama 2 Long, and presto! Suddenly, the AI could identify even the most remote tokens, those loners who don't get along with others.
Meta introduces LLAMA 2 Long
— AK (@_akhaliq) September 29, 2023
- context windows of up to 32,768 tokens
- the 70B variant can already surpass gpt-3.5-turbo-16k’s overall performance on a suite of long-context tasks https://t.co/uzsVslLUkX pic.twitter.com/aXyPmeLXMo
To make Llama 2 Long even more innovative, the researchers threw in some reinforcement learning from human feedback (RLHF). It's like teaching a dog new tricks; only these tricks involve coding, math, language understanding, common sense reasoning, and answering questions from humans. The AI gets a treat (in the form of rewards) when it gets things right, and humans double-check its homework. It's a win-win for both the AI and its human overlords.
The primary question now is, why is everyone talking about Llama 2 Long? Llama 2 Long is the bee's knees when it comes to everyday tasks done by large language models (LLMs). It outperforms not only its younger sibling, Llama 2, but even the powerful GPT-3.5 Turbo and Claude 2. It's like the new kid on the block showing up as the incumbent champions. Let the applause begin!
The open-source AI community, not one to hold back on their feelings, has been showering Llama 2 Long with admiration and excitement. Reddit, Twitter, and Hacker News have turned into virtual cheerleading squads for Meta's "open source" approach to AI. It's a bit like the classic tale of David versus Goliath, where open source takes on the closed source, "pay-to-play" models championed by deep-pocketed startups.
In conclusion, Meta has dropped a game-changer in the AI realm with Llama 2 Long. This AI powerhouse is the digital equivalent of a superhero, surpassing the competitors and stealing the show. Llama 2 Long has captivated the hearts of the open-source community and is giving the big boys a run for their money with its remarkable capabilities. So, GPT-3.5 Turbo and Claude 2, take note: there's a new AI sheriff in town, and it's armed with llama-like wisdom and a whole lot of style. Yeehaw!
Read next: How Much Time Do Teens Spend On Social Media in 2023?
by Rubah Usman via Digital Information World
No comments:
Post a Comment