FOR IMMEDIATE RELEASE: 5/Feb/2023
Welcome back to The Memo.
We usually release on a monthly cadence, but there is a lot of speed in the AI world right now. Please excuse this edition being released within a few days of the last one!
In the Toys to play with section, we look at a new (free) ChatGPT alternative now available to 400M users via iOS, the latest (free) visual language model, and a new 24x7 GPT-3 + DALL-E + Stable Diffusion TV show!
The BIG Stuff
ChatGPT is probably outputting at least 110x the equivalent volume of Tweets by human Twitter users every day.... and the equivalent of the entire printed works of humanity every 14 days (Jan/2023)
It’s no wonder that OpenAI is concerned about scaling!
Yes, you can cite me on this exclusive.
Check my working: https://lifearchitect.ai/chatgpt/#popular
ChatGPT coming to Bing’s 1B monthly users (3/Feb/2023)
Over the last two years, many people have been asking me when Leta AI will become mainstream. My answer has always been ‘when one of the FAANGs builds it into a platform’.
Even without an avatar, ChatGPT brought in 100 million users in just a few weeks, the ‘fastest growing user-base of all time’ (Reuters, 2/Feb/2023).
As noted in earlier editions of The Memo, Microsoft is putting GPT-3 everywhere. Now Microsoft is exposing ChatGPT to their 1 billion monthly Bing users, mostly in the corporate world where the use of Microsoft products is enforced.
The official announcement is due 8/Feb/2023.
Read more: https://www.theverge.com/2023/2/3/23584675/microsoft-ai-bing-chatgpt-screenshots-leak
The Interesting Stuff
Sam Altman on the next step for AI and capitalism (3/Feb/2023)
It's definitely an exciting time. But my hope is that it's still extremely early. Really this is going to be a continual exponential path of improvement of the technology and the positive impact it has on society.
…I think there's a real chance that we actually have figured out something significant here and this paradigm will take us very, very far.
Read more via Forbes: https://archive.is/x8E2A
John Carmack on the next step for AI (2/Feb/2023)
This Feb/2023 interview is 6,000 words long, but worth the read. It comes from John Carmack, the engineer who gave us games like Commander Keen, Wolfenstein 3D, Doom, Quake, and hardware platforms like the Oculus Rift VR headset. He is now deeply embedded in the AI field.
As a 52-year-old independent researcher, John still works a 60-hour week: 10 hours a day, 6 days a week.
My hope is that I can spend several years working through some of these things, building small things that I think point in the right directions. And then, throw some scale at it and push an entire lifetime of information and experience through this and see if it comes out with something that shows that spark…
I see the destination. I know it’s there, but no, it’s murky and cloudy in between here and there. Nobody knows how to get there. But I’m looking at that path saying I don’t know what’s in there, but I think I can get through there—or at least I think somebody will. And I think it’s very likely that this is going to happen in the 2030s.
I do consider it essentially inevitable.
Anthropic gets $300M from Google (3/Feb/2023)
You may recall that about a dozen OpenAI staff left that company in 2020, just after GPT-3’s release, founding Anthropic AI. Anthropic’s vision is less about the technical model building and commercialization, and more about human alignment and safety.
Google has invested about $300mn in artificial intelligence start-up Anthropic, making it the latest tech giant to throw its money and computing power behind a new generation of companies trying to claim a place in the booming field of “generative AI”.
The terms of the deal, through which Google will take a stake of around 10 per cent, requires Anthropic to use the money to buy computing resources from the search company’s cloud computing division, according to three people familiar with the arrangement.
Read more via Forbes: https://archive.is/ciZPV
Watch my video on Anthropic’s Claude chatbot model:
ChatGPT helps a judge with a verdict (31/Jan/2023)
On January 31, the first labor court of Cartagena resolved a guardianship action with the help of the famous artificial intelligence known as ChatGPT, arguing that it applied Law 2213 of 2022, which says that in certain cases these virtual tools can be used.
English: https://interestingengineering.com/innovation/chatgpt-makes-humane-decision-columbia
Spanish source: https://www.bluradio.com/judicial/sentencia-la-tome-yo-chatgpt-respaldo-argumentacion-juez-de-cartagena-uso-inteligencia-artificial-pr30
LAION’s new Open Assistant project to replicate ChatGPT (Feb/2023)
This Open Assistant project might be 6-12 months away, but LAION are the same team that have brought us incredible datasets used in open-source and closed projects (LAION).
Open Assistant is a project meant to give everyone access to a great chat based large language model.
We believe that by doing this we will create a revolution in innovation in language. In the same way that stable-diffusion helped the world make art and images in new ways we hope Open Assistant can help improve the world by improving language itself.
Official site: https://open-assistant.io/
Repo: https://github.com/LAION-AI/Open-Assistant
Amazon’s new vision + LLM: Multimodal-CoT 738M (2/Feb/2023)
With Multimodal-CoT, our model under 1 billion parameters [748M params] outperforms the previous state-of-the-art LLM (GPT-3.5) by 16% (75.17%->91.68%) on the ScienceQA benchmark and even surpasses human performance.
Read the paper: https://arxiv.org/abs/2302.00923
View the repo: https://github.com/amazon-science/mm-cot
Carnegie Mellon’s new vision + dialogue model: FROMAGe (31/Jan/2023)
While these vision abilities have been available for three years or so (see Leta Episode 5 from 20/May/2021), CM’s new Frozen Retrieval Over Multimodal Data for Autoregressive Generation (FROMAGe) adds grounding to the entire conversation.
We use the publicly available [Meta] OPT model with 6.7B parameters as our LLM...For the visual model, we use a pretrained CLIP ViT-L/14 model...
Read the paper: https://arxiv.org/abs/2301.13823
See some examples: https://jykoh.com/fromage
Try it yourself with the older Quickchat.ai/Emerson: https://quickchat.ai/emerson
New concerns with draft EU AI laws (4/Feb/2023)
The EU continues to try and quash artificial intelligence progress. You may recall my outburst back in The Memo edition 15/Sep/2022, that ‘These new AI laws are a true abomination’.
It gets worse.
Under the EU draft rules, ChatGPT is considered a general purpose AI system which can be used for multiple purposes including high-risk ones such as the selection of candidates for jobs and credit scoring.
Breton wants OpenAI to cooperate closely with downstream developers of high-risk AI systems to enable their compliance with the proposed AI Act.
"Just the fact that generative AI has been newly included in the definition shows the speed at which technology develops and that regulators are struggling to keep up with this pace," a partner at a U.S. law firm, said.
New article: Chinchilla scaling in plain English (2/Feb/2023)
If you’re interested in understanding more about the amount of data needed to train large language models in 2023, I’ve outlined the ‘need to know’ in an article.
There is a little bit of technical detail by necessity, but it is designed to be read by anyone, with viz and videos also included.
Summary: Chinchilla showed that we need to be using 11× more data during training than that used for GPT-3 and similar models. This means that we need to source, clean, and filter to around 33TB of text data for a 1T-parameter model.
Read more: https://lifearchitect.ai/chinchilla/
Toys to Play With
Quora Poe powered by ChatGPT + Claude (3/Feb/2023)
[Poe is] a new AI product we have been building called Poe. Poe lets people ask questions, get instant answers, and have back-and-forth conversations with several AI-powered bots.
…we hope to become the most efficient way for people to collectively explore the possibilities opened up by new AI models as they are released. The name Poe is short for “Platform for Open Exploration” to reflect this intent.
My iPhone install gave me three general chatbots: Sage, Heron, and Dragonfly. There are many others from which to choose. The underlying language model is ‘using text-generation algorithms like ChatGPT and Anthropic’s Claude… his team secured access to OpenAI's bot and Anthropic’s chatbot Claude—he won’t share the terms’ (via Wired, archive).
It doesn’t seem any better than Quickchat.ai’s Emerson, which has been available for nearly three years, and was used to power Leta AI beginning in Apr/2021.
Read the announcement via Twitter (unrolled).
Download it for iOS here (free, needs cell + email confirmation): https://poe.com/
Salesforce BLIP-2 visual language model (Feb/2023)
Maybe even better than DeepMind Flamingo (my video), the new BLIP-2 visual language model can take in an image and output text about the image. Captioning results are via the Meta AI OPT-6.7B model, and chat results are via the Google FlanT5-XXL model.
Try it for free: https://huggingface.co/spaces/Salesforce/BLIP2
Paper: https://arxiv.org/abs/2301.12597
Midjourney and the new /blend tool (Feb/2023)
Midjourney v4 is still the best text-to-image model out there. The new ‘/blend’ tool allows you to combine two or more uploaded images, with the AI conceptualizing the result. The results are wild!
Keep in mind that using Midjourney is ugly. You’ll need a text-based chat platform called Discord. You’ll need to follow a few steps to get it set up. And then you’ll need to almost ‘script’ your prompts to get going. Start here: https://www.midjourney.com/
Read more about ‘/blend’: https://medium.com/seeds-for-the-future/another-cool-feature-from-midjourney-blend-mode-b4cc3a4007b4
Nothing, forever: a 24x7 animated GPT-3 + DALL-E + SD stream (Feb/2023)
It’s a bit like Seinfeld…
Nothing, Forever is a show about nothing, that happens forever. Kinda like popular sitcoms of the past, except that it never stops. Nothing, Forever is always-on, runs 365 days of the year, and delivers new content every minute. Everything you see, hear, or experience (with the exception of the artwork and laugh track) is always brand new content, generated via machine learning and AI algorithms.
…the characters are all speaking to each other using GPT-3, OpenAI’s language model, which becomes clear as the characters are often not looking at each other when they are talking and rarely make sense.
Wikipedia confirms:
Dialogue is generated through GPT-3, a language model from OpenAI. Other technologies used include Stable Diffusion, DALL-E, and Azure Cognitive Services. To generate new scenes, an Azure Function written in TypeScript is used. The machine learning models were written in Python using TensorFlow, while the show is rendered using Unity and C#.
Watch: https://www.twitch.tv/watchmeforever
UPDATE 6/Feb/2023: The channel was banned for this joke. It may be back soon…
Read a related article: https://www.vice.com/en/article/88qy3p/thousands-of-people-cant-stop-watching-ai-tv-show-nothing-forever
Leta Episodes 0-5 remastered (Feb/2023)
Just before the 12-hour marathon of Leta AI last month (which I forgot to mention in these editions!), I remastered the audio for the first six episodes. I used Adobe’s Podcast Enhance, a free drag-and-drop tool. The results are astounding. They converted the recordings of my old iPhone mic (1-meter or 3-feet away from me) to a close-mic studio-sounding setup.
Using Adobe Podcast Enhance (free): https://podcast.adobe.com/enhance
Have a listen to the remasters starting at Episode 0 (Episodes 0-5 only).
Next
More models, more good times!
All my very best,
Alan
LifeArchitect.ai
Archives | Unsubscribe new account | Unsubscribe old account (before Aug/2022)