• Everyday AI
  • Posts
  • AI's Exponential Explosion: AGI Timelines Accelerate

AI's Exponential Explosion: AGI Timelines Accelerate

With every passing week, Artificial General Intelligence moves from science fiction to imminent reality.

March 25, 2023

Everyday AI Logo

AI's Exponential Explosion: AGI Timelines Accelerate

With every passing week, Artificial General Intelligence moves from science fiction to imminent reality.

The Big Stuff

Last week in our newsletter, "AI News Hurricane", we argued that it was possibly the most innovative week in history. We couldn't imagine another week rivaling it for a while. Well, this week topped it. Apparently, it was only the first wave in the set. First, a quick chronological summary of the week, and then we'll dive in, covering some of the biggest news first.Monday: Gen2 text + image to video by Runway ResearchTuesday: Google Bard, NVIDIA AI Foundations, Bing Image Creator, Adobe Firefly, Agility RoboticsWednesday: Github Copilot X, Opera Contextual Prompts, and Microsoft Research AGI PaperThursday: Canva AI, ChatGPT PluginsFriday: GPT4 - One Trillion Parameters, CONFIRMED

ChatGPT Plugins

Until this week, ChatGPT was limited to responses directly from the model, and limited to the training data set with knowledge cutoff of August 2021. This made queries like, "Who won the Oscar for best actor in 2023" impossible. Plugins changes this. Plugins allows for third parties to provide relevant responses to ChatGPT queries. OpenAI released three core plugins: browsing, code interpreter, and retrieval, as well as a small handful of 3rd party plugins.

The browsing plugin allows ChatGPT to retrieve data on the web, so it would be able to give a detailed answer for "who won best actor in the Oscars in 2023".The code interpreter plugin allows you to upload files and execute Python code. If this sounds abstract, let us give you a concrete example. The demo OpenAI showed was uploading a video file, and asking ChatGPT to extract the first five seconds of the video. ChatGPT then writes a Python script to extract the first five seconds, and allows the user to download it. If you want to see some amazing examples of Code Interpreter, check out this post by Andrew Mayne.The retrieval plugin allows you to upload your own data sets, and use ChatGPT to retrieve information from this data. Imagine uploading your emails and notes, and being able to have ChatGPT summarize things.

Finally, there were many 3rd party plugins announced as well from Expedia, Kayak, OpenTable, Wolfram, Zapier, Shopify, Klarna, Fiscal Note, Instacart, Milo Family AI, and Speak.

Everyday AI's take: This is a big breakthrough that is likely a bigger development than the App Store. It could quickly move us to a world where "One app rules them all". Why download 50 apps on your iPhone, when you could download one app that performs the functions of every app in the app store? This announcement makes OpenAI a direct threat to Apple, and the threat to Google search also intensifies. The plugins interface is unlike anything we've ever seen, or others have seen and it's blowing them away. Whereas plugins have traditionally required code, these do not. They are entirely based on natural language, and not too much of it. In fact, here is Klarna's API integration. That is the entirety, there is nothing more to it. Excuse my language, but WTF?!

Microsoft: Sparks of Artificial General Intelligence

Artificial General Intelligence (AGI) has been the holy grail of AI research for decades. AGI exhibits the human traits of intelligence: learning, understanding, thinking, adapting, problem-solving, creativity, and generalization. About six months ago, the consensus on achieving AGI was 2053. One week ago, it was 2043, and now it's 2033. Timelines are rapidly accelerating.This week, Microsoft released a paper where they evaluated OpenAI GPT-4's capabilities, and concluded that GPT-4 is starting to show signs of meeting some of the criteria of AGI. To test this, they came up with novel and interesting questions that could not be the result of memorization. The results are nothing short of astonishing, or as twitter user @8teAPIi put it, "An amazing human achievement and a milestone in history. About the equivalent of the invention of fire, writing, antibiotics." Here are some of the findings. GPT-4 can:

  • Draw in 2D and 3D - it is able to draw and output to TikZ

  • Pass an Amazon technical interview on Leet code with a perfect score

  • Compose music (not harmony yet)

  • Compile LaTeX in natural language (a notoriously complicated language that is used to print math equations)

  • Reverse engineer code compiled to assembly language - this will have huge ramifications in security

  • Execute Python code without a computer

  • Manage calendar and email - it was able to coordinate a meeting over email with other humans by using calendar and email APIs only.

  • Navigate MUD-like interfaces and then draw a map of it using Pyplot

The 156 page paper is a long, yet fascinating read. The gravity of the discoveries makes me think of how it must have felt reading Ben Franklin's "Experiments and Observations on Electricity" right after it was published, and knowing the world was about to change.

Google Bard

Google released Bard, it's 137B parameter LLM. Released in the context of GPT-4, there aren't many things to say about it. It's nowhere near as good as GPT-3.5, let alone GPT-4. It does not code well, nor is it useful for creative writing. Google released way too early, with a product they hoped would be perceived as Google's answer to ChatGPT. It did not deliver on that promise, and they no doubt lost people. Despite having a world class AI team, Google is having difficulty getting products to market. This year, they fumbled their AI presentation, and their response to ChatGPT was lackluster. Google is visibly struggling in the AI competition. Under different circumstances, this setback might be manageable. However, as AI advances rapidly, catching up becomes increasingly challenging each month, in a non-linear manner.

Runway Research Gen-2: Text to Video

Diffusion models like Stable Diffusion and Midjourney gave us text-to-image capabilities. Gen-2 takes this one step further with text-to-video. The results are pretty astounding. Click to watch the demo video:

NVIDIA Foundations

Foundations is a suite of AI services, and APIs for the enterprise, including text generation, summarization, image generation, video generation, 3D content generation, protein prediction, biomolecular generation, and molecular docking. These products are currently in early access, so there is no feedback yet on the product.

Github Copilot X

Copilot has already increased developer productivity by 55.8%. Copilot X introduces new capabilities that will no doubt increase that even more. Software engineers can ask in-context questions, help debug code, generate unit tests, and facilitate learning (release video, release blog)

Over the last two weeks the number of releases have been difficult to keep up. However, as hard as it may to believe, it appears that Microsoft isn't done yet. We can't wait to see they ship next.

Smaller But Still Cool Things:

Going Deeper

Tweets of the Week

Eye Candy

2024 election is going to be lit!

Do you have 30 seconds for a quick survey to help us improve Everday AI?

We'd love your feedback! Click here.

Do you like what you're reading? Share it with a friend.