tl;dr:
New GPTs: HuggingGPT & BloombergGPT
Local LLMs Accelerate
Autonomous AI Experiments
More AI News
HuggingGPT
HuggingGPT is described in a new paper by HuggingFace, the company behind huggingface.com, which hosts an incredible amount of artificial intelligence models, machine learning models, and tooling to help with AI/ML.
This paper describes a new architecture where a generalized LLM (ChatGPT) can control a suite of specialized (fine-tuned) LLMs for specific tasks. For example, if you want to summarize a multi-media presentation, AI must be able to read and digest text, understand and caption images, read charts and graphs, and much more.
This architecture is akin to how the human brain works and could be the architecture necessary to enable AGI (artificial general intelligence). The human brain has many specialized areas for different tasks, called functional specialization, and it seems as though AI is converging more and more on the architecture that powers the brain.
HuggingGPT was co-authored by Microsoft and Microsoft has already put out their own version called Jarvin, which is already a Github Repo you can download and play with!
BloombergGPT
Bloomberg, the financial behemoth with software products, news products, and so much more, has also released their own research paper about BloombergGPT. Bloomberg married their incredibly dense and proprietary financial data set with large public datasets for a hybrid approach to training their model. They used public data from sources such as Wikipedia, Github, C4, YouTube Subtitles, HackerNews, and even Enron Emails 😂.
They discovered that this hybrid approach was able to achieve great results with financial prompts, as shown in this graph:
As more companies release their own LLMs, it’s becoming clear that the true differentiator between models is the data. Not only is a proprietary dataset important, but having clean and well-labeled data is just as important. Here’s the video I made about BloombergGPT:
Local LLMs Everywhere!
The number of local LLMs is accelerating quickly. From GPT4All to LLaMA and Alpaca, installing a fully functional LLM on any device, even without an internet connection is a reality. Check out the tweet below which shows Brian Roemmele having installed GPT4All on a TI-84 Plus graphing calculator! Math tests will never be the same.
Although these local models still aren’t up to the quality of GPT-4, they are accelerating quickly. As more people adopt edge models (LLMs installed on devices vs. used through the cloud like OpenAI), end users will be able to participate in making them better by contributing their data. This feels very much like the early days of the internet when AOL was trying to control everything and the open and free internet we know today ended up breaking out and being the dominant form we know today.
As new open-source LLMs get released and iterate and add features, I’ll continue to test them and create tutorial videos.
Autonomous AI
The newest hot topic in the world of AI right now is fully autonomous AI. Projects like AutoGPT and BabyAGI are blowing up on GitHub, receiving tons of stars and forks. These AI projects take ChatGPT to the next level by allowing a user to start with a simple described goal and then be able to create a plan, a task list, prioritization, and autonomously execute the given goal.
For example, with AutoGPT, you can tell it your goal is to increase your Twitter follower count. It can then research on the internet for how to do that, come up with a list of action items, present them to you, and then actually execute those tasks continuously.
I plan on making tutorial videos for both of these projects, so stay tuned!
Other AI News
I want this newsletter to give you everything you need to know about the world of AI, so here are a few hand-picked pieces of news from this week:
