ChatGPT has triggered a flood. Suddenly so many contenders have emerged. It almost feels like Game of Thrones. A few noble families are waging war against each other to gain control over the mythical land of AI, while society worries that unleashed forces of AI will threaten the livelihood (or even life) of mankind.
When I wrote an earlier issue of this newsletter (The Rise of ChatGPT: Separating Hype from Reality), I didn’t think I would write another on ChatGPT within a month! But here we are given the pace of rapid developments in this space.
In this issue, you will get to know:
Emerging Large Language Models (LLMs) like ChatGPT, and
Contenders that you should watch out for.
OpenAI is like House Stark right at the frontier. ChatGPT is their Jon Snow seeking adventures “beyond the wall.” But there are times when you feel like yelling, “you know nothing, Jon Snow!”
Google is like House Targaryen. Mighty dragon riders brought other kingdoms to their knees with fire and fury. Now everyone wants to claim their number one spot.
Google invented Transformers, the neural network architecture that powers ChatGPT. They released Bard powered by Language Model for Dialogue Applications (LaMDA) model. Markets punished them harshly even though Bard’s mistakes were no worse than ChatGPT failures. Unlike OpenAI, it has not yet put APIs/services in the hand of people.
Microsoft is like House Lannister. Wealthiest with mines full of gold, shelling out $10B was a no-brainer for forming an alliance with Open AI in its feud with Targaryens, err… Google.
Microsoft gave up its own LLM called Turing-NLG, and has rushed to capitalize on ChatGPT with AI-powered Bing. But its answers can’t be always trusted, it spilled its secrets and is prone to wacky behavior. We will have to wait and watch how this gamble plays out.
Facebook is like House Martell, the most desirable kingdom to live in all of Tech Westeros. It has gained most of its power through marriages with Instagram and WhatsApp rather than conquest.
Facebook has released Galactica late last year. Recently, it released a new model LLaMA (paper) trained only on the publically available dataset. LLaMA-13B runs on a single GPU, and outperforms ChatGPT-like tech despite being 10x smaller.
They are the only ones other than ChatGPT who have put APIs/models in people’s hands. Their model weights got leaked apparently due to a torrent link in a GitHub pull request (alternate weights are available from HuggingFace too).
Amazon is like House Tyrell — wealthy and cunning. Their deal with HuggingFace shows that they know where to invest their silver.
Amazon’s LLM contender, Multimodal-CoT (paper, code), has “under 1 billion parameters outperforms the previous state-of-the-art LLM (GPT-3.5) by 16 percentage points (75.17%->91.68% accuracy) on the ScienceQA benchmark and even surpasses human performance.”
There are sever other LLMs coming up from big corps and startups, including an open-source implementation:
ChatLLaMA: Open-source implementation for LLaMA-based ChatGPT
Baidu: PCL-BAIDU Wenxin
AI21 Labs: Jurassic
At present, you can use only the following 4 APIs/models:
But I expect Google, Amazon, and Microsoft to offer LLM APIs in their cloud offerings soon.
LLMs are a type of Generative AI. And where there’s hype, there’s VC funding!