Llama 2 - More like Llama1 went to the gym, trained hard and got lean muscle

𝗪𝗵𝘆 𝗶𝘀 𝗟𝗹𝗮𝗺𝗮𝟮 𝗮 𝗯𝗶𝗴 𝗱𝗲𝗮𝗹?

Meta just released a bigger and commerical (conditionally) open-source Large Language Model - Llama 2. 4 Months after it released Llama 1

Netflix takes a whole year to release a new season, while AI research is on a rip in the last year

❓ 𝗪𝗮𝗶𝘁, 𝗪𝗵𝗮𝘁’𝘀 𝗟𝗹𝗮𝗺𝗮?

Llama is a large language model like OpenAI’s GPT, except much smaller and open-source.

GPT3 has 175B Parameters, while Llama comes in Starbucks sizes like Tall(7B) , Grande (13B), Venti (65B).

Why do parameters matter? More the parameters, better it can learn from data and produce more nuanced results. Visualize a color palette for a painter; more colors at hand, more vibrant the picture. However, “more” always comes with cost.

Denser models need more computational resources, so more parameters isn’t always good. Universally facepalmed response to any question applies here as well “it depends”. Remember the times when we used Dial-Up internet based on number of minutes used like calling cards? We're in that age of LLMs now. Lean models with reasonably good performance reduces cost thats needed to move to the broadband, 3G equivalent.

Llama’s source code is available for anyone to peek into. Open Source invokes transparency and trust, that’s foundational for something like AI that scares the life out of most people.

Open-Source model available for download, also means anyone can download the model and run it on any machine.GPT is synonymous to a Cloud Software - always runs on someone else’s machine (for now)

🤷 𝗢𝗸, 𝗛𝗼𝘄’𝘀 𝗟𝗹𝗮𝗺𝗮𝟮 𝗯𝗲𝘁𝘁𝗲𝗿?

1️⃣ Fine-tuned Llama2 used for chat - Llama-2-Chat - outperforms open-source chat models and on par with some popular closed-source models like ChatGPT and PaLM

2️⃣ Llama2 is available for use commercially (special license for monthly active exceeding 700M users)

3️⃣ Llama2 is also available as a service offered through Microsoft Azure’s compute.LLaMA will also be available through AWS, Hugging Face, and other providers

4️⃣ Llama2 can run locally on devices without cloud dependency.

Qualcomm announced it is working with Meta to bring LLaMa to laptops, phones, and headsets starting from 2024 onward for AI-powered apps that work without relying on cloud services.

Google also announced PaLM2 to have many sizes, smallest capable for running on phones. Judging by the pace of innovation thus far, don’t think we need to hold our breath for too long

🏃 𝗠𝗲𝘁𝗮 𝗿𝗮𝗻 𝘁𝗵𝗲 𝗿𝗮𝗰𝗲, 𝗯𝘂𝘁 𝗠𝗶𝗰𝗿𝗼𝘀𝗼𝗳𝘁 𝗽𝗶𝗰𝗸𝘀 𝘂𝗽 𝘁𝗵𝗲 𝗺𝗲𝗱𝗮𝗹 🏆

In all this AI innovation, there one silent winner - Microsoft. Infrastructure providers laid the roads and are happily collecting tolls

AWS will soon join the party for sure, but Microsoft is clearly on full-volume with rolled down windows playing “I Ain’t Worried” for everyone to dance