Forward Future News Episode 7

Leaked Google AI Memo: “We Have No Moat”

In one of the most exciting reads in AI news recently, an internal document from a researcher at Google was leaked and illustrated some fear they have over falling behind in AI. But who are they afraid of? It’s not OpenAI, it’s open-source models! In fact, they mention OpenAI as having the same existential risk as they do.

Now, of course, this is a leaked document from just one or a small group of people within Google and might not come to fruition, but many of the points made in this leaked memo are accurate.

The main points of the memo revolve around large models vs. smaller models. Specifically, the rate of innovation is minimal when you can only train a model over months and for millions of dollars.

Open-source models are iterating incredibly quickly, which the highly with this timeline of releases:

I’m a big fan of open-source models, but I would be naive to say that companies like OpenAI and especially Google have no moat. Although he didn’t provide any reasoning, Logan, the head of developer relations at OpenAI, said this about the leaked memo:

I truly believe open-sourced LLMs pose a serious risk for Google and OpenAI. Right now, GPT-4 is by far the best LLM out there. However, many open-source models are making enormous progress daily and weekly toward quality parity. I made a video reviewing the leaked memo in full, which you can watch here:

GPT4All Progress

Speaking of open-source progress, GPT4All has been making incredible progress seemingly every day. Here are some new features that are now baked into GPT4All UI:

• Easier installation: now the installation is easier than ever. They also were approved as a developer by Apple so you no longer get the warning before installing GPT4All.

• Data-sharing: you can now (optionally) share your data with NomicAI into the GPT4All data lake, which helps them improve future models.

• Automatic updates: one-click check and install new updates from the GPT4All team, which seem to be coming fast and furiously.

• Tons of models: GPT4All keeps adding new models (up to 9 now), including ones created by other companies such as Vicuña and WizardLM. Just like everything else with GPT4All, the installation process is super easy.

If you want the best local LLM interface and some of the best models, I highly recommend downloading GPT4All.

tl;dr AI News:

• HuggingChat is released - an LLM made by HuggingFace, based on OpenAssistant. (video review)

• AI can now read your mind - in an exciting paper out of Tokyo University, researchers were able to greatly increase the quality of fMRI scan outputs by using Stable Diffusion. (video explanation)

• AI music is going viral again and again - Drake, Rihanna, Bad Bunny, and more have had deepfakes of their voice used to create viral hits. The music industry has been taking its typical legal approach, which is the worst way to handle new tech.

• BabyAGI adds a UI - now you can easily use BabyAGI with a simple interface that looks a lot like ChatGPT.

Looking for input!

I’m putting together a consistent rubric for testing new LLMs. So far, here are the prompts I will be testing:

What else should I add to this rubric? Your input is appreciated!